ci: the AI provider workflow runs only the provider tests and passes every provider's key

docs(ai): the image reference and quotas page cover Replicate's safety reruns
style: restore the repo formatting in files the AI stack reflowed
2026-11-13 20:12:57 -10:00 · 2026-11-13 20:31:22 -07:00 · 2026-11-13 20:31:22 -07:00 · 2026-11-13 20:31:22 -07:00 · 2026-11-13 20:31:22 -07:00 · 2026-11-13 20:31:21 -07:00
429 changed files with 8051 additions and 6133 deletions
@@ -2,7 +2,7 @@

 # The last matching rule takes precedence.

- @Salazareo
+* @Salazareo

 # GUI

@@ -18,7 +18,10 @@
 # Backend AI

 /src/backend/controllers/puterai/ @404oops @ProgrammerIn-wonderland @Salazareo
-/src/backend/drivers/ai-\*/ @404oops @ProgrammerIn-wonderland @Salazareo
+/src/backend/drivers/ai/ @404oops @ProgrammerIn-wonderland @Salazareo
+/src/backend/util/aiCostFactor* @404oops @ProgrammerIn-wonderland @Salazareo
+/src/backend/util/aiPrices* @404oops @ProgrammerIn-wonderland @Salazareo
+/src/backend/util/aiProviderKey* @404oops @ProgrammerIn-wonderland @Salazareo

 # Backend: AI, hosting, workers, networking, and WebDAV

@@ -29,7 +32,7 @@
 /src/backend/drivers/workers/ @ProgrammerIn-wonderland @Salazareo
 /src/backend/services/localworker/ @ProgrammerIn-wonderland @Salazareo
 /src/backend/stores/subdomain/ @ProgrammerIn-wonderland @Salazareo
-/extensions/workerSandbox\* @ProgrammerIn-wonderland @Salazareo
+/extensions/workerSandbox* @ProgrammerIn-wonderland @Salazareo
 /src/worker/ @ProgrammerIn-wonderland @Salazareo
 /src/worker-types/ @ProgrammerIn-wonderland @Salazareo
 /src/backend/controllers/peer/ @ProgrammerIn-wonderland @Salazareo
@@ -42,7 +45,7 @@
 /src/backend/services/permission/ @jfcastro92 @Salazareo
 /src/backend/services/acl/ @jfcastro92 @Salazareo
 /src/backend/services/team/ @jfcastro92 @Salazareo
-/src/backend/services/apps/AppPermissionService\* @jfcastro92 @Salazareo
+/src/backend/services/apps/AppPermissionService* @jfcastro92 @Salazareo
 /src/backend/stores/share/ @jfcastro92 @Salazareo
 /src/backend/stores/permission/ @jfcastro92 @Salazareo
 /src/backend/stores/group/ @jfcastro92 @Salazareo
@@ -59,7 +62,7 @@
 /src/puter-js/tests/api/suites/hosting.suite.ts @ProgrammerIn-wonderland @Salazareo
 /src/puter-js/tests/api/suites/workers.suite.ts @ProgrammerIn-wonderland @Salazareo
 /src/puter-js/tests/api/suites/net.suite.ts @ProgrammerIn-wonderland @Salazareo
-/src/puter-js/src/modules/Peer\* @ProgrammerIn-wonderland @Salazareo
+/src/puter-js/src/modules/Peer* @ProgrammerIn-wonderland @Salazareo

 # SDK: sharing and permissions

@@ -31,10 +31,9 @@ jobs:
              run: npm ci

            - name: Run AI provider integration tests
-              # Vitest treats positional args as filename substring filters.
-              # The pattern matches every `*.integration.test.ts` file and
-              # nothing else — the regular backend test suite stays out.
-              run: npm run test:backend -- integration.test
+              # Only the provider tests: the other `*.integration.test.ts`
+              # files need no credentials and run with the backend suite.
+              run: npm run test:backend -- $(git ls-files 'src/backend/drivers/ai/**/*.integration.test.ts')
              env:
                  CI: 'true'
                  PUTER_TEST_AI_CLAUDE_API_KEY: ${{ secrets.PUTER_TEST_AI_CLAUDE_API_KEY }}
@@ -49,6 +48,15 @@ jobs:
                  PUTER_TEST_AI_MOONSHOT_API_KEY: ${{ secrets.PUTER_TEST_AI_MOONSHOT_API_KEY }}
                  PUTER_TEST_AI_ZAI_API_KEY: ${{ secrets.PUTER_TEST_AI_ZAI_API_KEY }}
                  PUTER_TEST_AI_ELEVENLABS_API_KEY: ${{ secrets.PUTER_TEST_AI_ELEVENLABS_API_KEY }}
+                  PUTER_TEST_AI_ALIBABA_API_KEY: ${{ secrets.PUTER_TEST_AI_ALIBABA_API_KEY }}
+                  PUTER_TEST_AI_AZURE_OPENAI_API_KEY: ${{ secrets.PUTER_TEST_AI_AZURE_OPENAI_API_KEY }}
+                  PUTER_TEST_AI_AZURE_OPENAI_API_URL: ${{ secrets.PUTER_TEST_AI_AZURE_OPENAI_API_URL }}
+                  PUTER_TEST_AI_BYTEPLUS_API_KEY: ${{ secrets.PUTER_TEST_AI_BYTEPLUS_API_KEY }}
+                  PUTER_TEST_AI_HOONIFY_API_KEY: ${{ secrets.PUTER_TEST_AI_HOONIFY_API_KEY }}
+                  PUTER_TEST_AI_INFRON_API_KEY: ${{ secrets.PUTER_TEST_AI_INFRON_API_KEY }}
+                  PUTER_TEST_AI_KLING_API_KEY: ${{ secrets.PUTER_TEST_AI_KLING_API_KEY }}
+                  PUTER_TEST_AI_META_API_KEY: ${{ secrets.PUTER_TEST_AI_META_API_KEY }}
+                  PUTER_TEST_AI_MINIMAX_API_KEY: ${{ secrets.PUTER_TEST_AI_MINIMAX_API_KEY }}

            # Realtime conversations through the SDK against live OpenAI. Builds
            # the SDK bundle these tests load itself, so it runs even when the
@@ -9,6 +9,7 @@ Use these as the source of truth before exploring further:
 - [README.md](README.md) — project overview and quickstart.
 - [doc/architecture.md](doc/architecture.md) — backend layered stack (controllers → drivers → services → stores → clients), `PuterServer` wiring, `Context` (ALS), actors, and extensions.
 - [doc/contributing-apis.md](doc/contributing-apis.md) — adding and maintaining public APIs end to end (backend surface → puter.js → types → docs → tests). Follow it for any API work.
+- [doc/ai-drivers.md](doc/ai-drivers.md) — where AI driver code lives (interfaces, providers, protocols, shared), the import rules between those folders, and how to add a provider or an AI interface.
 - [doc/pagination.md](doc/pagination.md) — the one pagination convention for list APIs (limit/cursor/offset/includeTotal, envelope shape, cursor semantics).
 - [doc/alarms.md](doc/alarms.md) — raising alarms and picking a severity (what pages, what only gets recorded), plus the config that routes them.
 - [doc/self-hosting.md](doc/self-hosting.md) — running Puter outside hosted infra.
@@ -70,7 +71,7 @@ Commit messages and PR descriptions never credit anyone or mention a bug report.
 - **Test new behavior.** Every new function, endpoint, driver method, or logic branch gets a test; every bug fix gets a regression test that fails before the fix. If something is genuinely hard to test, skip it but say so in the PR.
 - **Boy Scout Rule, proportional to the change.** In files you're already touching, fix the typo, the dead import, the missing test, the bit you had to read twice. Don't ride a refactor along with a bug fix.
 - **Understand what you commit.** AI assistance is fine; shipping code you couldn't have written, debugged, or defended in review is not.
- **A limit change is not done until the docs change with it.** Every rate limit, concurrency cap, quota, and allowance in the code is published in [src/docs/src/rate-limits-and-quotas.md](src/docs/src/rate-limits-and-quotas.md) — an undisclosed limit is one developers discover as a service failure. If a PR moves any of these numbers, the same PR updates that page. The numbers live in `src/backend/controllers/fs/limits.ts` (filesystem), `src/backend/drivers/util/aiLimits.ts` (all AI drivers), per-driver `rateLimit`/`concurrent` configs, and `src/backend/data/subPolicies/` (free-tier allowances).
+- **A limit change is not done until the docs change with it.** Every rate limit, concurrency cap, quota, and allowance in the code is published in [src/docs/src/rate-limits-and-quotas.md](src/docs/src/rate-limits-and-quotas.md) — an undisclosed limit is one developers discover as a service failure. If a PR moves any of these numbers, the same PR updates that page. The numbers live in `src/backend/controllers/fs/limits.ts` (filesystem), `src/backend/drivers/ai/shared/aiLimits.ts` (all AI drivers), per-driver `rateLimit`/`concurrent` configs, and `src/backend/data/subPolicies/` (free-tier allowances).

 ---

@@ -97,6 +98,10 @@ Layer rules:
 - **Drivers and controllers stay thin.** Domain rules a second entry point needs (signup, app CRUD, worker deploy, AI routing/billing) live in a service, not in the first driver or controller that needed them.
 - **Prefer explicit arguments over [`Context`](src/backend/core/context.ts) (ALS).** Use it only for request-scoped values that would otherwise thread through many layers; today that's mostly `actor` and `req`.

+### AI drivers
+
+AI drivers live in [src/backend/drivers/ai/](src/backend/drivers/ai/): `interfaces/` holds the Puter-facing drivers, `providers/<vendor>/<interface>/` holds vendor code, `protocols/` holds wire protocols several vendors speak, and `shared/` holds helpers more than one interface uses. Follow [doc/ai-drivers.md](doc/ai-drivers.md) when adding or moving AI code.
+
 ### Extensions

 [extensions/](extensions/) is for **non-crucial parts of the system**: Puter must still work with the extension removed. If core needs to call it, it belongs in core ([whoami](extensions/whoami.ts) is the cautionary example). Inside an extension, follow the same layered structure, unless it only needs a few route handlers, in which case the `extension.get/post/...` helpers are enough on their own. The `extension` API is in [doc/architecture.md](doc/architecture.md#extension-api).
@@ -62,10 +62,6 @@
    // direct socket peer. NEVER set to `true` in prod — it trusts every hop
    // and makes X-Forwarded-For forgeable.
    "trust_proxy": false,
-    // Request header a TLS-terminating proxy in front of the server sets to
-    // the client's TLS fingerprint (JA3/JA4). Only set it when that proxy
-    // overwrites any value the client sends. Unset disables it.
-    // "tlsFingerprintHeader": "x-tls-fingerprint",
    "no_browser_launch": false,

    // ── Dev watcher (devWatcher extension) ──────────────────────────────
@@ -0,0 +1,153 @@
+# AI Drivers
+
+Where the AI driver code lives under [src/backend/drivers/ai/](../src/backend/drivers/ai/), what each folder owns, and how to add a provider or a new AI interface. The general layer rules are in [architecture.md](architecture.md) and [AGENTS.md](../AGENTS.md#backend). This page covers what is specific to AI.
+
+## Layout
+
+```text
+src/backend/drivers/ai/
+├── interfaces/    Puter-facing drivers, one folder per AI interface
+├── providers/     Vendor code, as providers/<vendor>/<interface>/
+├── protocols/     Wire protocols that several vendors speak
+└── shared/        Helpers that more than one interface uses
+```
+
+| Folder | Owns | Example |
+| --- | --- | --- |
+| `interfaces/<interface>/` | The driver class, the interface's contracts in `types.ts`, its provider base class where one exists, and the helpers and tests that apply to every provider of that interface. | [interfaces/chat/ChatCompletionDriver.ts](../src/backend/drivers/ai/interfaces/chat/ChatCompletionDriver.ts) |
+| `providers/<vendor>/<interface>/` | One vendor's implementation of one interface: the adapter, its model catalog and prices, its fixtures, and its tests. | [providers/gemini/video/](../src/backend/drivers/ai/providers/gemini/video/) |
+| `providers/<vendor>/` | Helpers that several of that vendor's interfaces use. | [providers/gemini/geminiRest.ts](../src/backend/drivers/ai/providers/gemini/geminiRest.ts) |
+| `protocols/openaiCompatible/` | The OpenAI wire protocol: the base provider class, message and usage conversion, and parameter mapping for Chat Completions and Responses. | [protocols/openaiCompatible/OpenAICompatProvider.ts](../src/backend/drivers/ai/protocols/openaiCompatible/OpenAICompatProvider.ts) |
+| `shared/` | Helpers that more than one interface uses: limits, the provider registry, the remote catalog cache, job polling, streamed charging, and media input. | [shared/aiLimits.ts](../src/backend/drivers/ai/shared/aiLimits.ts) |
+
+Helpers that non-AI code also uses stay in [src/backend/drivers/util/](../src/backend/drivers/util/): file input, data URIs, upstream error mapping, and stream piping.
+
+AI code that a service also needs lives in [src/backend/util/](../src/backend/util/), because a service can't import from the driver layer. The realtime relay in `services/realtime/` bills through the same cost factor as the drivers, so these three files are there:
+
+| File | Holds |
+| --- | --- |
+| [aiCostFactor.ts](../src/backend/util/aiCostFactor.ts) | `withAiCostFactor`, the metering view every AI driver and the realtime relay record usage through, and `withAiCreditHold`. |
+| [aiProviderKey.ts](../src/backend/util/aiProviderKey.ts) | `readProviderKey`, which reads an API key from `config.providers` blocks. |
+| [aiPrices.ts](../src/backend/util/aiPrices.ts) | Vendor prices that a driver and a service both bill. |
+
+### Interfaces
+
+| Folder | Driver | `driverInterface` | `driverName` | Key in `puterDrivers` |
+| --- | --- | --- | --- | --- |
+| `interfaces/chat/` | `ChatCompletionDriver` | `puter-chat-completion` | `ai-chat` | `aiChat` |
+| `interfaces/image/` | `ImageGenerationDriver` | `puter-image-generation` | `ai-image` | `aiImage` |
+| `interfaces/video/` | `VideoGenerationDriver` | `puter-video-generation` | `ai-video` | `aiVideo` |
+| `interfaces/textToSpeech/` | `TTSDriver` | `puter-tts` | `ai-tts` | `aiTts` |
+| `interfaces/speechToText/` | `SpeechToTextDriver` | `puter-speech2txt` | `ai-speech2txt` | `aiSpeech2Txt` |
+| `interfaces/speechToSpeech/` | `VoiceChangerDriver` | `puter-speech2speech` | `ai-speech2speech` | `aiSpeech2Speech` |
+| `interfaces/ocr/` | `OCRDriver` | `puter-ocr` | `ai-ocr` | `aiOcr` |
+
+Folder names are not runtime identifiers. `driverInterface`, `driverName`, driver aliases, the keys under `providers` in config, provider ids, model ids, and usage keys are public contracts. Moving or renaming a file never changes them.
+
+## Request flow
+
+A call goes from the interface to a provider, and from there to a shared protocol when the vendor speaks one:
+
+```text
+interfaces/chat/ChatCompletionDriver.ts
+    validates the call, picks the model and provider, holds credits
+  → providers/deepseek/chat/DeepSeekProvider.ts
+        DeepSeek's catalog, endpoint, and request differences
+    → protocols/openaiCompatible/OpenAICompatProvider.ts
+          the OpenAI-compatible request, response, and stream handling
+```
+
+A provider whose vendor has its own API, such as `ClaudeProvider` or `ElevenLabsTTSProvider`, implements the interface's contract directly and uses no protocol.
+
+## Import rules
+
+- `shared/` imports nothing from `interfaces/`, `providers/`, or `protocols/`.
+- Nothing in `services/`, `stores/`, or `clients/` imports from `drivers/ai/`. Code that a service and a driver both need goes in `src/backend/util/`.
+- `providers/` and `protocols/` import an interface's `types.ts` and helpers. They never import a driver class. Their tests can, to drive a provider through the real driver.
+- Only an interface's driver file imports providers, to register them. The interface's helpers don't.
+- An interface doesn't import another interface. When a second interface needs a helper, move the helper to `shared/`, or to `providers/<vendor>/` when it belongs to one vendor.
+- A vendor folder imports another vendor's folder only when it serves that vendor's models. `providers/azure/` reads the OpenAI and xAI catalogs for that reason.
+
+## Where a file goes
+
+| The code is | It goes in |
+| --- | --- |
+| Input validation, gates, routing, credit holds, or anything else every provider of the interface gets | `interfaces/<interface>/` |
+| A request or response contract that providers implement | `interfaces/<interface>/types.ts` |
+| A vendor's endpoint, request shape, catalog, prices, or quirks | `providers/<vendor>/<interface>/` |
+| A vendor's auth, REST plumbing, or constants that two of its interfaces use | `providers/<vendor>/` |
+| Behavior of a wire protocol that several vendors speak | `protocols/<protocol>/` |
+| A helper that two or more interfaces use and that belongs to no vendor | `shared/` |
+| A helper that non-AI drivers or controllers also use | `src/backend/drivers/util/` |
+| AI code that a service also uses | `src/backend/util/`, in a file named `ai<Something>.ts` |
+
+Keep a helper next to its only consumer. Move it up when a second consumer appears, not before.
+
+When one Puter option becomes a different request shape per vendor, the interface reads the option and each dialect maps it. The chat `compaction` option is read in `interfaces/chat/utils/compaction.js`, mapped for OpenAI by `toOpenAiContextManagement` in `protocols/openaiCompatible/openaiParams.ts`, and mapped for Anthropic by `toAnthropicContextManagement` in `providers/claude/chat/contextManagement.ts`. Tools work the same way: `normalize_tools_object` in the interface, then `toolsToWire` for OpenAI and `claudeToolPolicy` for Anthropic.
+
+Tests sit next to the code they test. Fixtures go in a `fixtures/` folder next to the test that reads them.
+
+## Adding a provider
+
+1. Create `providers/<vendor>/<interface>/`. Use the vendor's existing folder if it has one, so all of a vendor's interfaces stay together.
+2. Implement the interface's contract from `interfaces/<interface>/types.ts`:
+
+    | Interface | Contract | Base class |
+    | --- | --- | --- |
+    | chat | `IChatProvider` | `OpenAICompatProvider` for an OpenAI-compatible vendor |
+    | image | `IImageProvider` | none |
+    | video | `IVideoProvider` | `VideoProvider` |
+    | textToSpeech | `ITTSProvider` | none |
+    | speechToText | `ISpeechToTextProvider` | `SpeechToTextProvider` |
+    | ocr | `IOcrProvider` | none |
+
+    Speech to speech has one provider, so `VoiceChangerDriver` uses `ElevenLabsSpeechToSpeechProvider` directly. Add a contract to its `types.ts` when a second provider needs one.
+
+3. Put the model catalog and prices in the same folder, as `models.ts` or `costs.ts`. A provider that loads its catalog from the vendor wraps the fetch in `cachedRemoteCatalog` from `shared/`.
+4. Register the provider in the interface's driver: in `#registerProviders()`, or in `onServerStart()` for OCR and speech to speech. A chat provider built from an API key and an optional base URL is one row in `KEYED_PROVIDERS` instead. Text to speech and speech to text also list the provider id and its aliases in `providerAliases.ts`.
+5. Read credentials from `config.providers[<key>]`. The key is part of the self-hosting config contract, so reuse the vendor's existing key when it has one.
+6. Add tests next to the provider. Mock the vendor SDK or `fetch` at the network boundary and drive the provider through the real driver where you can. A live test is a `*.integration.test.ts` file that skips itself, with `skipUnlessEnv` from [testEnv.ts](../src/backend/testEnv.ts), unless its `PUTER_TEST_AI_*` variable is set. Helpers for provider tests are in [shared/integrationTestUtil.ts](../src/backend/drivers/ai/shared/integrationTestUtil.ts).
+7. For a chat provider, add its static catalog to `CATALOGS` in [modelCatalogs.test.ts](../src/backend/drivers/ai/interfaces/chat/modelCatalogs.test.ts) and the provider to [providerConsistency.test.ts](../src/backend/drivers/ai/interfaces/chat/providerConsistency.test.ts). The catalog test fails when a `providers/*/chat/` file with "model" in its name exports a catalog that isn't listed.
+8. Add the provider's config key to the `providers` config of the interface's [fuzz suite](#fuzz-tests), and its id to the list of providers the suite expects.
+9. Update the published docs for the models and limits the provider adds, following [contributing-apis.md](contributing-apis.md).
+
+### OpenAI-compatible vendors
+
+A vendor that speaks the OpenAI Chat Completions protocol extends `OpenAICompatProvider` and overrides only what differs. It still gets its own `providers/<vendor>/chat/` folder.
+
+Vendors share one constants-only folder only when at least six of them differ in nothing but base URL and model catalog. Each then becomes one data file under `providers/openaiCompatible/providers/`, read by a single implementation. No vendor qualifies today, because every subclass overrides request, catalog, or billing behavior.
+
+## Fuzz tests
+
+Two seeded suites send malformed input through the real drivers and providers. They run with the rest of the backend tests.
+
+| Suite | Covers |
+| --- | --- |
+| [interfaces/chat/ChatCompletionDriver.fuzz.test.ts](../src/backend/drivers/ai/interfaces/chat/ChatCompletionDriver.fuzz.test.ts) | Chat requests with a few values changed, and valid requests against wrong statuses, bodies of the wrong shape, and broken event streams. |
+| [interfaces/mediaDrivers.fuzz.test.ts](../src/backend/drivers/ai/interfaces/mediaDrivers.fuzz.test.ts) | Image, video, speech and OCR requests with a few values changed. Every vendor refuses the request. |
+
+What the suites assert:
+
+- A malformed request never fails with an error that has no status, which `/drivers/call` would report as an unclassified 500.
+- For chat, with the vendors answering, a malformed request also never fails with a 5xx and never marks a route unhealthy.
+- When the upstream misbehaves, chat ends with an answer or an `HttpError`. A stream only has to end: the error line a stream carries isn't checked, so a malformed event can still end one with a 500.
+- Requests reach the vendors, and with the vendors answering some chat requests are answered, so a suite that stops exercising the providers fails.
+
+The vendors are played at the HTTP boundary by `upstream` in [shared/fuzzTestUtil.ts](../src/backend/drivers/ai/shared/fuzzTestUtil.ts), which stands in for global `fetch`, undici `fetch` and axios, and makes `https.request` throw so an SDK with its own client fails the suite. The vendors' SDKs run, except Groq's, which is swapped for the OpenAI SDK, and the AWS clients, whose `send` is stubbed. The media suite calls the drivers as a user with unlimited metering, so a request gets past file input and the credit hold to the vendor. A provider that loads its catalog from the vendor registers no models there, so the suites skip it.
+
+A failure lists each distinct finding with its case number and input. Each case is generated from the seed and its own number, so the same case fails again on the next run. Routes are picked by rendezvous hashing: a model added to a catalog takes only the cases it wins, and every other case keeps its route. To search further, raise the case count:
+
+```sh
+PUTER_FUZZ_CASES=100000 npx vitest run --config src/backend/vitest.config.ts src/backend/drivers/ai/interfaces/chat/ChatCompletionDriver.fuzz.test.ts
+```
+
+When a suite finds a crash, validate the value where the request enters, in the driver or in `normalize_single_message`, so every provider is covered. Reject only what crashed: a value a provider ignored or refused before must still reach it. An option a driver reads as text, a number or a flag goes in the driver's `SCALAR_OPTIONS`, and `assertScalarOptions` from `shared/scalarOptions.ts` answers 400 for a value there that can't be converted to text. Add a unit test for the case next to the fix.
+
+## Adding an interface
+
+1. Create `interfaces/<interface>/` with the driver class and a `types.ts` for the contracts providers implement.
+2. Register the driver in [src/backend/drivers/index.ts](../src/backend/drivers/index.ts).
+3. Use `AI_RATE_LIMIT` and `AI_CONCURRENT` from `shared/aiLimits.ts`, and publish any new limit in [rate-limits-and-quotas.md](../src/docs/src/rate-limits-and-quotas.md).
+4. Meter through `withAiCostFactor` from `src/backend/util/aiCostFactor.ts`, and route between providers with `ProviderRegistry` from `shared/providerRegistry.ts`.
+5. List the options the providers read as text, a number or a flag in a `SCALAR_OPTIONS` array, and call `assertScalarOptions` from `shared/scalarOptions.ts` first in each method. Add the driver to [mediaDrivers.fuzz.test.ts](../src/backend/drivers/ai/interfaces/mediaDrivers.fuzz.test.ts).
+6. Follow [contributing-apis.md](contributing-apis.md) for the puter.js method, types, docs, and tests.
@@ -42,6 +42,19 @@ block-beta

 Extensions sit alongside this stack and can register into any layer; see [Extensions](#extensions).

+### AI drivers
+
+The AI drivers have their own layout under [src/backend/drivers/ai/](../src/backend/drivers/ai/):
+
+| Folder | Holds |
+| --- | --- |
+| `interfaces/` | The Puter-facing drivers, one folder per AI interface (chat, image, video, text to speech, speech to text, speech to speech, OCR). |
+| `providers/` | Vendor code, as `providers/<vendor>/<interface>/`. |
+| `protocols/` | Wire protocols that several vendors speak, such as the OpenAI-compatible one. |
+| `shared/` | Helpers that more than one interface uses. |
+
+A call goes from an interface's driver to a provider, and from there to a shared protocol when the vendor speaks one. [ai-drivers.md](ai-drivers.md) has the ownership rules, the import rules, and the steps for adding a provider.
+
 ## Entry point: `PuterServer`

 `PuterServer` is the bootstrap. It:
@@ -27,7 +27,7 @@ import { DatabaseClientFactory } from './index.js';
 import { SqliteDatabaseClient } from './SqliteDatabaseClient.js';

 /** Highest schema version the migration table can reach. */
-const CURRENT_SCHEMA_VERSION = 101;
+const CURRENT_SCHEMA_VERSION = 99;

 /**
 * These suites migrate real files on disk. Idle they finish in well under a
@@ -735,12 +735,7 @@ describe('SqliteDatabaseClient — boot and migrations', { timeout: DISK_MIGRATI
                    '`recipient_email`) VALUES (?, ?, ?)',
                ['share-invite', user.id, '[email protected]'],
            );
-            // The full schema as it stood before 0103, so 0103 reruns; 0104
-            // is undone so it can apply again after it.
-            await before.write('DROP INDEX idx_user_signup_tls_fingerprint');
-            await before.write(
-                'ALTER TABLE `user` DROP COLUMN `signup_tls_fingerprint`',
-            );
+            // The full schema as it stood before 0103, so only 0103 reruns.
            await before.write('PRAGMA user_version = 98');
            before.onServerShutdown();

@@ -138,8 +138,6 @@ const AVAILABLE_MIGRATIONS: [number, string[]][] = [
    [96, ['0101_user-scoped-handlers.sql']],
    [97, ['0102_notification-scope-id-index.sql']],
    [98, ['0103_share-entry-owner-backfill.sql']],
-    [99, ['0104_signup-tls-fingerprint.sql']],
-    [100, ['0105_sessions_kind_debug_app.sql']],
 ];

 export class SqliteDatabaseClient extends AbstractDatabaseClient {
@@ -1,51 +0,0 @@
-- Copyright (C) 2024-present Puter Technologies Inc.
--
-- This file is part of Puter.
--
-- Puter is free software: you can redistribute it and/or modify
-- it under the terms of the GNU Affero General Public License as published
-- by the Free Software Foundation, either version 3 of the License, or
-- (at your option) any later version.
--
-- This program is distributed in the hope that it will be useful,
-- but WITHOUT ANY WARRANTY; without even the implied warranty of
-- MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-- GNU Affero General Public License for more details.
--
-- You should have received a copy of the GNU Affero General Public License
-- along with this program.  If not, see <https://www.gnu.org/licenses/>.
-
-- `user`.`signup_tls_fingerprint`. See
-- sqlite/0104_signup-tls-fingerprint.sql for the rationale. No per-file
-- applied-state tracking, so the column and index are added only when missing
-- and a replay does nothing.
-
-DROP PROCEDURE IF EXISTS _puter_add_user_signup_tls_fingerprint;
-DELIMITER //
-CREATE PROCEDURE _puter_add_user_signup_tls_fingerprint()
-BEGIN
-  IF NOT EXISTS (
-    SELECT 1 FROM INFORMATION_SCHEMA.COLUMNS
-    WHERE TABLE_SCHEMA = DATABASE()
-      AND TABLE_NAME = 'user'
-      AND COLUMN_NAME = 'signup_tls_fingerprint'
-  ) THEN
-    ALTER TABLE `user`
-      ADD COLUMN `signup_tls_fingerprint` varchar(64) DEFAULT NULL,
-      ALGORITHM=INSTANT;
-  END IF;
-  IF NOT EXISTS (
-    SELECT 1 FROM INFORMATION_SCHEMA.STATISTICS
-    WHERE TABLE_SCHEMA = DATABASE()
-      AND TABLE_NAME = 'user'
-      AND INDEX_NAME = 'idx_user_signup_tls_fingerprint'
-  ) THEN
-    ALTER TABLE `user` ADD INDEX `idx_user_signup_tls_fingerprint`
-      (`signup_tls_fingerprint`), ALGORITHM=INPLACE, LOCK=NONE;
-  END IF;
-END//
-DELIMITER ;
-
-CALL _puter_add_user_signup_tls_fingerprint();
-
-DROP PROCEDURE IF EXISTS _puter_add_user_signup_tls_fingerprint;
@@ -1,46 +0,0 @@
-- Copyright (C) 2024-present Puter Technologies Inc.
--
-- This file is part of Puter.
--
-- Puter is free software: you can redistribute it and/or modify
-- it under the terms of the GNU Affero General Public License as published
-- by the Free Software Foundation, either version 3 of the License, or
-- (at your option) any later version.
--
-- This program is distributed in the hope that it will be useful,
-- but WITHOUT ANY WARRANTY; without even the implied warranty of
-- MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-- GNU Affero General Public License for more details.
--
-- You should have received a copy of the GNU Affero General Public License
-- along with this program.  If not, see <https://www.gnu.org/licenses/>.
-
-- Add 'debug_app' to the `sessions.kind` ENUM: a short-lived app session an
-- operator creates on a user's behalf. Mirrors SQLite migration 0105.
--
-- Idempotent: guarded against COLUMN_TYPE so re-running the directory is a
-- no-op once 'debug_app' is in the ENUM.
-
-DROP PROCEDURE IF EXISTS _puter_sessions_kind_debug_app;
-DELIMITER //
-CREATE PROCEDURE _puter_sessions_kind_debug_app()
-BEGIN
-  IF NOT EXISTS (
-    SELECT 1 FROM INFORMATION_SCHEMA.COLUMNS
-    WHERE TABLE_SCHEMA = DATABASE()
-      AND TABLE_NAME = 'sessions'
-      AND COLUMN_NAME = 'kind'
-      AND FIND_IN_SET('debug_app', REPLACE(REPLACE(REPLACE(COLUMN_TYPE, 'enum(', ''), ')', ''), '''', '')) > 0
-  ) THEN
-    ALTER TABLE `sessions`
-      MODIFY COLUMN `kind`
-        ENUM('web', 'app', 'access_token', 'asset', 'worker', 'debug_app')
-        NOT NULL DEFAULT 'web',
-      ALGORITHM=INSTANT;
-  END IF;
-END//
-DELIMITER ;
-
-CALL _puter_sessions_kind_debug_app();
-
-DROP PROCEDURE IF EXISTS _puter_sessions_kind_debug_app;
@@ -1,22 +0,0 @@
-- Copyright (C) 2024-present Puter Technologies Inc.
--
-- This file is part of Puter.
--
-- Puter is free software: you can redistribute it and/or modify
-- it under the terms of the GNU Affero General Public License as published
-- by the Free Software Foundation, either version 3 of the License, or
-- (at your option) any later version.
--
-- This program is distributed in the hope that it will be useful,
-- but WITHOUT ANY WARRANTY; without even the implied warranty of
-- MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-- GNU Affero General Public License for more details.
--
-- You should have received a copy of the GNU Affero General Public License
-- along with this program.  If not, see <https://www.gnu.org/licenses/>.
-
-- "user".signup_tls_fingerprint. Mirrors SQLite migration 0104. Idempotent
-- via IF NOT EXISTS.
-
-ALTER TABLE "user" ADD COLUMN IF NOT EXISTS signup_tls_fingerprint varchar(64);
-CREATE INDEX IF NOT EXISTS idx_user_signup_tls_fingerprint ON "user" (signup_tls_fingerprint);
@@ -1,24 +0,0 @@
-- Copyright (C) 2024-present Puter Technologies Inc.
--
-- This file is part of Puter.
--
-- Puter is free software: you can redistribute it and/or modify
-- it under the terms of the GNU Affero General Public License as published
-- by the Free Software Foundation, either version 3 of the License, or
-- (at your option) any later version.
--
-- This program is distributed in the hope that it will be useful,
-- but WITHOUT ANY WARRANTY; without even the implied warranty of
-- MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-- GNU Affero General Public License for more details.
--
-- You should have received a copy of the GNU Affero General Public License
-- along with this program.  If not, see <https://www.gnu.org/licenses/>.
-
-- Add 'debug_app' to the `sessions.kind` CHECK constraint: a short-lived app
-- session an operator creates on a user's behalf. Mirrors SQLite migration
-- 0105. Drop-then-add keeps it idempotent across directory replays.
-
-ALTER TABLE sessions DROP CONSTRAINT IF EXISTS sessions_kind_check;
-ALTER TABLE sessions ADD CONSTRAINT sessions_kind_check
-    CHECK (kind IN ('web', 'app', 'access_token', 'asset', 'worker', 'debug_app'));
@@ -1,22 +0,0 @@
-- Copyright (C) 2024-present Puter Technologies Inc.
--
-- This file is part of Puter.
--
-- Puter is free software: you can redistribute it and/or modify
-- it under the terms of the GNU Affero General Public License as published
-- by the Free Software Foundation, either version 3 of the License, or
-- (at your option) any later version.
--
-- This program is distributed in the hope that it will be useful,
-- but WITHOUT ANY WARRANTY; without even the implied warranty of
-- MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-- GNU Affero General Public License for more details.
--
-- You should have received a copy of the GNU Affero General Public License
-- along with this program.  If not, see <https://www.gnu.org/licenses/>.
-
-- `user`.`signup_tls_fingerprint`: the TLS fingerprint of the request that
-- created the account, when a fronting proxy supplies one. Indexed so the
-- accounts created from one TLS stack can be listed.
-ALTER TABLE `user` ADD COLUMN `signup_tls_fingerprint` varchar(64) DEFAULT NULL;
-CREATE INDEX IF NOT EXISTS idx_user_signup_tls_fingerprint ON `user` (`signup_tls_fingerprint`);
@@ -1,89 +0,0 @@
-- Copyright (C) 2024-present Puter Technologies Inc.
--
-- This file is part of Puter.
--
-- Puter is free software: you can redistribute it and/or modify
-- it under the terms of the GNU Affero General Public License as published
-- by the Free Software Foundation, either version 3 of the License, or
-- (at your option) any later version.
--
-- This program is distributed in the hope that it will be useful,
-- but WITHOUT ANY WARRANTY; without even the implied warranty of
-- MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-- GNU Affero General Public License for more details.
--
-- You should have received a copy of the GNU Affero General Public License
-- along with this program.  If not, see <https://www.gnu.org/licenses/>.
-
-- Add 'debug_app' to the `kind` CHECK constraint on `sessions`. Mirrors the
-- MySQL ENUM extension in mysql_mig_59.
--
-- SQLite cannot ALTER a CHECK constraint in place, so we follow the same
-- table-rebuild as migration 0056: create `sessions_new` with the corrected
-- constraint, copy rows, drop the old table, rename, then recreate every
-- index that lived on the original.
-
-CREATE TABLE `sessions_new` (
-    "id" INTEGER PRIMARY KEY AUTOINCREMENT,
-    "user_id" INTEGER NOT NULL,
-    "uuid" TEXT NOT NULL,
-    "meta" JSON DEFAULT NULL,
-    "created_at" INTEGER DEFAULT 0,
-    "last_activity" INTEGER DEFAULT 0,
-    "kind" TEXT NOT NULL DEFAULT 'web'
-        CHECK (`kind` IN ('web', 'app', 'access_token', 'asset', 'worker', 'debug_app')),
-    "label" TEXT,
-    "parent_session_id" TEXT,
-    "last_ip" TEXT,
-    "last_user_agent" TEXT,
-    "revoked_at" INTEGER,
-    "expires_at" INTEGER,
-    "app_uid" TEXT,
-    "legacy_token_uid" TEXT,
-    "created_via" TEXT,
-    "auth_id" TEXT,
-    "access_token_uid" TEXT,
-    FOREIGN KEY("user_id") REFERENCES "user" ("id") ON DELETE CASCADE ON UPDATE CASCADE
-);
-
-INSERT INTO `sessions_new` (
-    `id`, `user_id`, `uuid`, `meta`, `created_at`, `last_activity`,
-    `kind`, `label`, `parent_session_id`, `last_ip`, `last_user_agent`,
-    `revoked_at`, `expires_at`, `app_uid`, `legacy_token_uid`,
-    `created_via`, `auth_id`, `access_token_uid`
-)
-SELECT
-    `id`, `user_id`, `uuid`, `meta`, `created_at`, `last_activity`,
-    `kind`, `label`, `parent_session_id`, `last_ip`, `last_user_agent`,
-    `revoked_at`, `expires_at`, `app_uid`, `legacy_token_uid`,
-    `created_via`, `auth_id`, `access_token_uid`
-FROM `sessions`;
-
-DROP TABLE `sessions`;
-
-ALTER TABLE `sessions_new` RENAME TO `sessions`;
-
-CREATE INDEX IF NOT EXISTS `idx_sessions_user_revoked`
-    ON `sessions` (`user_id`, `revoked_at`);
-
-CREATE INDEX IF NOT EXISTS `idx_sessions_parent`
-    ON `sessions` (`parent_session_id`);
-
-CREATE UNIQUE INDEX IF NOT EXISTS `idx_sessions_user_app_active`
-    ON `sessions` (`user_id`, `app_uid`)
-    WHERE `kind` = 'app' AND `revoked_at` IS NULL;
-
-CREATE UNIQUE INDEX IF NOT EXISTS `idx_sessions_legacy_token_active`
-    ON `sessions` (`legacy_token_uid`)
-    WHERE `legacy_token_uid` IS NOT NULL AND `revoked_at` IS NULL;
-
-CREATE INDEX IF NOT EXISTS `idx_sessions_kind_user`
-    ON `sessions` (`kind`, `user_id`);
-
-CREATE INDEX IF NOT EXISTS `idx_sessions_access_token_uid`
-    ON `sessions` (`access_token_uid`)
-    WHERE `access_token_uid` IS NOT NULL;
-
-CREATE UNIQUE INDEX IF NOT EXISTS `idx_sessions_user_worker_active`
-    ON `sessions` (`user_id`, IFNULL(`app_uid`, ''), json_extract(`meta`, '$.worker_name'))
-    WHERE `kind` = 'worker' AND `revoked_at` IS NULL;
@@ -241,8 +241,6 @@ export type EventMap = {
        trail_id?: string;
        /** Device signal forwarded verbatim from the signup request body. */
        fingerprint?: string | null;
-        /** TLS fingerprint of the signup request, when a proxy supplies one. */
-        tls_fingerprint?: string | null;
        /**
         * Set by the abuse harness — require SMS phone verification
         * post-signup.
@@ -132,28 +132,6 @@ describe('PreludeClient', () => {
        });
    });

-    it('forwards a JA4 TLS fingerprint and drops any other shape', async () => {
-        fetchMock.mockResolvedValue(okJson({ id: 'v', status: 'success' }));
-        const client = makeClient('sk_test');
-        const ja4 = 't13d1516h2_8daaf6152771_02713d6af862';
-
-        await client.createVerification('+14155550123', {
-            tls_fingerprint: ja4,
-        });
-        await client.createVerification('+14155550123', {
-            ip: '203.0.113.7',
-            // A JA3 hash.
-            tls_fingerprint: 'e7d705a3286e19ea42f587b344ee6865',
-        });
-
-        expect(JSON.parse(fetchMock.mock.calls[0][1].body).signals).toEqual({
-            ja4_fingerprint: ja4,
-        });
-        expect(JSON.parse(fetchMock.mock.calls[1][1].body).signals).toEqual({
-            ip: '203.0.113.7',
-        });
-    });
-
    it('omits the signals object entirely when none are supplied', async () => {
        fetchMock.mockResolvedValue(okJson({ id: 'v', status: 'success' }));
        const client = makeClient('sk_test');
@@ -25,8 +25,6 @@ const PRELUDE_API_BASE = 'https://api.prelude.dev/v2';
 const REQUEST_TIMEOUT_MS = 8000;
 /** OTP length — matches the 6-box code UI. Prelude allows 4–8. */
 const PRELUDE_CODE_SIZE = 6;
-const JA4_SHAPE =
-    /^[tqd][0-9a-z]{2}[di]\d{4}[0-9a-z]{2}_[0-9a-f]{12}_[0-9a-f]{12}$/;
 /**
 * Channel Prelude prioritizes for delivery. RCS is far cheaper than SMS, so we
 * prefer it by default; Prelude routes to the next reachable channel (SMS) when
@@ -136,8 +134,6 @@ export class PreludeClient extends PuterClient {
            ip?: string;
            device_id?: string;
            user_agent?: string;
-            /** Forwarded only when it is a JA4 value. */
-            tls_fingerprint?: string;
            dispatch_id?: string;
        } = {},
    ): Promise<{
@@ -172,8 +168,6 @@ export class PreludeClient extends PuterClient {
        if (signals.ip) sig.ip = signals.ip;
        if (signals.device_id) sig.device_id = signals.device_id;
        if (signals.user_agent) sig.user_agent = signals.user_agent;
-        if (signals.tls_fingerprint && JA4_SHAPE.test(signals.tls_fingerprint))
-            sig.ja4_fingerprint = signals.tls_fingerprint;
        if (Object.keys(sig).length > 0) body.signals = sig;
        // dispatch_id is a top-level field, not a member of `signals`.
        if (signals.dispatch_id) body.dispatch_id = signals.dispatch_id;
@@ -722,36 +722,6 @@ describe('AuthController.handleSignup', () => {
        expect(audit.ip_fwd).toBe(forged);
    });

-    it('records the request TLS fingerprint on the validate hook and the user row', async () => {
-        const username = `s_${uniq()}`;
-        const tlsFingerprint = 't13d1516h2_8daaf6152771_02713d6af862';
-        let seenTlsFingerprint: unknown;
-
-        await withSignupValidateOverride(
-            (event) => {
-                seenTlsFingerprint = (event as { tls_fingerprint?: unknown })
-                    .tls_fingerprint;
-            },
-            async () => {
-                await controller.handleSignup(
-                    {
-                        ...makeReq({
-                            username,
-                            email: `${username}@test.local`,
-                            password: 'correct-horse-battery',
-                        }),
-                        tlsFingerprint,
-                    },
-                    makeRes(),
-                );
-            },
-        );
-
-        expect(seenTlsFingerprint).toBe(tlsFingerprint);
-        const persisted = await server.stores.user.getByUsername(username);
-        expect(persisted!.signup_tls_fingerprint).toBe(tlsFingerprint);
-    });
-
    it('forces phone verification on every signup when always_require_phone_verification is set', async () => {
        const cfg = (controller as { config: Record<string, unknown> }).config;
        const prev = cfg.always_require_phone_verification;
@@ -1510,7 +1510,6 @@ export class AuthController extends PuterController {
                    ip,
                    device_id: req.deviceFingerprint ?? undefined,
                    user_agent: userAgent,
-                    tls_fingerprint: req.tlsFingerprint,
                    dispatch_id: dispatchId,
                },
            );
@@ -4487,34 +4486,6 @@ export class AuthController extends PuterController {
        res.json({ token: guiToken });
    }

-    @Post('/auth/first-party-session', {
-        subdomain: 'api',
-        rateLimit: SESSION_LIMIT,
-        // Hands out a session, so only the GUI's own page script may ask, and
-        // only as the user's sign-in itself.
-        guiOriginOnly: true,
-        requireUserActor: true,
-    })
-    async handleFirstPartySession(req: Request, res: Response): Promise<void> {
-        const { origin } = req.body ?? {};
-        if (typeof origin !== 'string' || !origin) {
-            throw new HttpError(400, 'Missing `origin`', {
-                legacyCode: 'bad_request',
-            });
-        }
-        const token = await this.services.auth.createFirstPartySessionToken(
-            req.actor!,
-            origin,
-            {
-                ip: req.ip || req.socket?.remoteAddress,
-                user_agent: req.headers?.['user-agent'],
-                origin: req.headers?.origin,
-                host: req.headers?.host,
-            },
-        );
-        res.json({ token });
-    }
-
    @Get('/session/sync-cookie', {
        rateLimit: SESSION_LIMIT,
        // Installs the session cookie. Only page script on our own origin
@@ -4539,14 +4510,6 @@ export class AuthController extends PuterController {
            res.status(400).end();
            return;
        }
-        // A session handed to a first-party page never gets a cookie.
-        if (req.actor.session.parentUid) {
-            throw new HttpError(
-                403,
-                'This session cannot install a session cookie',
-                { legacyCode: 'forbidden' },
-            );
-        }
        // As in `/get-gui-token`: the actor is the user's row as auth read it.
        const sessionToken = this.services.auth.createSessionTokenForSession(
            req.actor.user as UserRow,
@@ -1,282 +0,0 @@
-/*
- * Copyright (C) 2024-present Puter Technologies Inc.
- *
- * This file is part of Puter.
- *
- * Puter is free software: you can redistribute it and/or modify
- * it under the terms of the GNU Affero General Public License as published
- * by the Free Software Foundation, either version 3 of the License, or
- * (at your option) any later version.
- *
- * This program is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
- * GNU Affero General Public License for more details.
- *
- * You should have received a copy of the GNU Affero General Public License
- * along with this program.  If not, see <https://www.gnu.org/licenses/>.
- */
-
-import { afterAll, beforeAll, describe, expect, it } from 'vitest';
-import type { IConfig } from '../../types.js';
-import { setupPuterTestEnv, type PuterTestEnv } from '../../testUtil.js';
-
-const PAGE = 'https://desk.first-party.test';
-// Seeded with godmode on.
-const DEV_CENTER = 'app-0b37f054-07d4-4627-8765-11bd23e889d4';
-
-/**
- * One of the deployment's own pages, served apart from the GUI, signs in with a
- * `web` session of its own under the user's sign-in.
- */
-describe('first-party page sessions', () => {
-    let env: PuterTestEnv;
-    let userId = 0;
-
-    beforeAll(async () => {
-        env = await setupPuterTestEnv({
-            first_party_session_origins: [`${PAGE}/`],
-        } as IConfig);
-        const row = await env.server.stores.user.getByUsername(
-            env.users.user.username,
-        );
-        userId = row!.id;
-    }, 120_000);
-
-    afterAll(async () => {
-        await env?.shutdown();
-    });
-
-    const call = (
-        method: string,
-        path: string,
-        token: string,
-        body?: unknown,
-        headers: Record<string, string> = {},
-    ) =>
-        fetch(new URL(path, env.apiOrigin), {
-            method,
-            headers: {
-                'content-type': 'application/json',
-                authorization: `Bearer ${token}`,
-                ...headers,
-            },
-            ...(body === undefined ? {} : { body: JSON.stringify(body) }),
-        });
-
-    /** A sign-in of its own, so revoking it leaves the shared one up. */
-    const signIn = async () => {
-        const user = await env.server.stores.user.getById(userId);
-        const { gui_token, session } =
-            await env.server.services.auth.createSessionToken(user!);
-        return { token: gui_token, uuid: session.uuid as string };
-    };
-
-    const mint = async (token: string, origin = PAGE) => {
-        const res = await call('POST', '/auth/first-party-session', token, {
-            origin,
-        });
-        expect(res.status).toBe(200);
-        return ((await res.json()) as { token: string }).token;
-    };
-
-    const actorOf = async (token: string) =>
-        (await env.server.services.auth.authenticate(token)).actor;
-
-    const sessions = async (token: string) => {
-        const res = await call('GET', '/auth/list-sessions', token);
-        expect(res.status).toBe(200);
-        return (await res.json()) as Array<Record<string, unknown>>;
-    };
-
-    it('hands the page a session of its own under the sign-in', async () => {
-        const root = await signIn();
-        const token = await mint(root.token);
-
-        const actor = await actorOf(token);
-        expect(actor?.app ?? null).toBeNull();
-        expect(actor?.accessToken ?? null).toBeNull();
-        expect(actor?.session).toMatchObject({
-            kind: 'web',
-            parentUid: root.uuid,
-        });
-        expect(actor?.session?.uid).not.toBe(root.uuid);
-
-        const row = (await sessions(root.token)).find(
-            (s) => s.uuid === actor!.session!.uid,
-        );
-        expect(row).toMatchObject({
-            kind: 'web',
-            parent_session_id: root.uuid,
-            first_party_origin: PAGE,
-            current: false,
-        });
-    });
-
-    it('reuses the session on a second ask', async () => {
-        const root = await signIn();
-        const first = await actorOf(await mint(root.token));
-        const second = await actorOf(await mint(root.token));
-        expect(second?.session?.uid).toBe(first?.session?.uid);
-    });
-
-    it('acts as the account: launches apps, godmode ones included', async () => {
-        const root = await signIn();
-        const token = await mint(root.token);
-
-        const res = await call('POST', '/auth/get-user-app-token', token, {
-            app_uid: DEV_CENTER,
-        });
-        expect(res.status).toBe(200);
-        expect(await res.json()).toMatchObject({ godmode: true });
-    });
-
-    it('refuses an origin that is not listed', async () => {
-        const root = await signIn();
-        const res = await call(
-            'POST',
-            '/auth/first-party-session',
-            root.token,
-            {
-                origin: 'https://elsewhere.test',
-            },
-        );
-        expect(res.status).toBe(403);
-    });
-
-    it('refuses a page on another origin, the listed page included', async () => {
-        const root = await signIn();
-        for (const origin of [PAGE, 'https://elsewhere.test']) {
-            const res = await call(
-                'POST',
-                '/auth/first-party-session',
-                root.token,
-                { origin: PAGE },
-                { origin },
-            );
-            expect(res.status).toBe(403);
-        }
-    });
-
-    it('refuses app tokens, access tokens and a first-party session', async () => {
-        const root = await signIn();
-        const page = await mint(root.token);
-        const app = (await (
-            await call('POST', '/auth/get-user-app-token', root.token, {
-                origin: 'https://some-app.test',
-            })
-        ).json()) as { token: string };
-        const godmode = (await (
-            await call('POST', '/auth/get-user-app-token', root.token, {
-                app_uid: DEV_CENTER,
-            })
-        ).json()) as { token: string };
-
-        for (const token of [page, app.token, godmode.token]) {
-            const res = await call('POST', '/auth/first-party-session', token, {
-                origin: PAGE,
-            });
-            expect(res.status).toBe(403);
-        }
-    });
-
-    it('never gets a session cookie', async () => {
-        const root = await signIn();
-        const page = await mint(root.token);
-
-        const syncCookie = (token: string) =>
-            fetch(new URL('/session/sync-cookie', env.origin), {
-                headers: { authorization: `Bearer ${token}` },
-            });
-        expect((await syncCookie(root.token)).status).toBe(204);
-
-        const res = await syncCookie(page);
-        expect(res.status).toBe(403);
-        expect(res.headers.get('set-cookie')).toBeNull();
-    });
-
-    it('ends with the sign-in it was handed out under', async () => {
-        const root = await signIn();
-        const page = await mint(root.token);
-        expect(await actorOf(page)).toBeDefined();
-
-        await env.server.services.auth.revokeSession(root.uuid);
-        const result = await env.server.services.auth.authenticate(page);
-        expect(result.actor).toBeUndefined();
-        expect(result.reauth?.reason).toBe('session_revoked');
-    });
-
-    it('takes the tokens minted under it along when its sign-in ends', async () => {
-        const root = await signIn();
-        const page = await mint(root.token);
-        const godmode = (await (
-            await call('POST', '/auth/get-user-app-token', page, {
-                app_uid: DEV_CENTER,
-            })
-        ).json()) as { token: string };
-        const scoped = (await (
-            await call('POST', '/auth/create-access-token', godmode.token, {
-                permissions: ['service:foo:ii:read'],
-                expiresIn: '7d',
-            })
-        ).json()) as { token: string };
-        expect(await actorOf(scoped.token)).toBeDefined();
-
-        await env.server.services.auth.revokeSession(root.uuid);
-        for (const token of [godmode.token, scoped.token]) {
-            expect(
-                (await env.server.services.auth.authenticate(token)).actor,
-            ).toBeUndefined();
-        }
-    });
-
-    it('is revoked on its own without ending the sign-in', async () => {
-        const root = await signIn();
-        const page = await mint(root.token);
-        const pageUuid = (await actorOf(page))!.session!.uid;
-
-        await env.server.services.auth.revokeSession(pageUuid);
-        expect(
-            (await env.server.services.auth.authenticate(page)).actor,
-        ).toBeUndefined();
-        expect(await actorOf(root.token)).toBeDefined();
-    });
-
-    it('keeps the whole current sign-in when revoking all others', async () => {
-        const root = await signIn();
-        const page = await mint(root.token);
-        const other = await signIn();
-        const otherPage = await mint(other.token);
-
-        // Asked from the page: its sign-in stays, and so does the page.
-        await env.server.services.auth.revokeAllSessions(
-            (await actorOf(page))!,
-        );
-        expect(await actorOf(page)).toBeDefined();
-        expect(await actorOf(root.token)).toBeDefined();
-        expect(
-            (await env.server.services.auth.authenticate(other.token)).actor,
-        ).toBeUndefined();
-        expect(
-            (await env.server.services.auth.authenticate(otherPage)).actor,
-        ).toBeUndefined();
-
-        // Asked from the sign-in: the page under it stays too.
-        await env.server.services.auth.revokeAllSessions(
-            (await actorOf(root.token))!,
-        );
-        expect(await actorOf(page)).toBeDefined();
-    });
-
-    it('is not counted as a sign-in', async () => {
-        const root = await signIn();
-        const before = await env.server.stores.session.listSignIns(userId, {
-            limit: 100,
-        });
-        await mint(root.token);
-        const after = await env.server.stores.session.listSignIns(userId, {
-            limit: 100,
-        });
-        expect(after.length).toBe(before.length);
-    });
-});
@@ -324,7 +324,7 @@ describe('DriverController — concurrent acquire/release', () => {
        const driver = {
            ...makeSyntheticDriver(),
            ping: async () => {
-                Context.get('keepDriverSlot')?.(
+                Context.get('keepConcurrencySlot')?.(
                    new Promise<void>((resolve) => {
                        settleWork = resolve;
                    }),
@@ -51,7 +51,7 @@ import { PermissionUtil } from '../../services/permission/permissionUtil.js';
 import type { WithLifecycle } from '../../types';
 import { withSpan } from '../../util/span.js';
 import { PuterController } from '../types.js';
-import { CallSlot } from './callSlot.js';
+import { CallSlot } from '../../core/http/callSlot.js';

 type DriverInstance = WithLifecycle & Record<string, unknown>;

@@ -446,7 +446,7 @@ export class DriverController extends PuterController {
                });
            }
            slot.attach(handle, res);
-            Context.set('keepDriverSlot', (work) => slot.keep(work));
+            Context.set('keepConcurrencySlot', (work) => slot.keep(work));
        }

        // Stash the requested driver name in Context so multi-provider
@@ -44,7 +44,7 @@ import {
 } from 'vitest';

 import type { Actor } from '../../core/actor.js';
-import type { ChatCompletionDriver } from '../../drivers/ai-chat/ChatCompletionDriver.js';
+import type { ChatCompletionDriver } from '../../drivers/ai/interfaces/chat/ChatCompletionDriver.js';
 import { PuterServer } from '../../server.js';
 import { inRequestScope, setupTestServer } from '../../testUtil.js';
 import { PuterAIController } from './PuterAIController.js';
@@ -49,8 +49,8 @@ import type { RouteOptions } from '../../core/http/index.js';
 import {
    COUNT_TOKENS,
    type ChatCompletionDriver,
-} from '../../drivers/ai-chat/ChatCompletionDriver.js';
-import { AI_CONCURRENT, AI_RATE_LIMIT } from '../../drivers/util/aiLimits.js';
+} from '../../drivers/ai/interfaces/chat/ChatCompletionDriver.js';
+import { AI_CONCURRENT, AI_RATE_LIMIT } from '../../drivers/ai/shared/aiLimits.js';
 import { PuterServer } from '../../server.js';
 import { inRequestScope, setupTestServer } from '../../testUtil.js';
 import { PuterAIController } from './PuterAIController.js';
@@ -30,25 +30,28 @@ import type { PuterRouter } from '../../core/http/PuterRouter.js';
 import {
    COUNT_TOKENS,
    type ChatCompletionDriver,
-} from '../../drivers/ai-chat/ChatCompletionDriver.js';
+} from '../../drivers/ai/interfaces/chat/ChatCompletionDriver.js';
 import type {
    IChatCompleteResult,
    IChatMessageResult,
    IChatModel,
    ICompleteArguments,
    UsageDetails,
-} from '../../drivers/ai-chat/types.js';
+} from '../../drivers/ai/interfaces/chat/types.js';
 import {
    outputFormatFromResponseFormat,
    rejectStatefulResponsesFields,
    toolChoiceFromWire,
-} from '../../drivers/ai-chat/utils/openaiParams.js';
+} from '../../drivers/ai/protocols/openaiCompatible/openaiParams.js';
 import {
    promoteStopForToolCalls,
    toFinishReason,
-} from '../../drivers/ai-chat/utils/stopReason.js';
+} from '../../drivers/ai/interfaces/chat/utils/stopReason.js';
 import { isDriverStreamResult } from '../../drivers/meta.js';
-import { AI_CONCURRENT, AI_RATE_LIMIT } from '../../drivers/util/aiLimits.js';
+import {
+    AI_CONCURRENT,
+    AI_RATE_LIMIT,
+} from '../../drivers/ai/shared/aiLimits.js';
 import { PuterController } from '../types.js';
 import { parseAnthropicRequest, toAnthropicMessage } from './anthropicWire.js';
 import { pipeNdjsonStream } from './ndjsonStream.js';
@@ -1301,13 +1304,18 @@ export class PuterAIController extends PuterController {
        return driver;
    }

-    /** A completion that stops, and stops billing, when the caller hangs up. */
+    /**
+     * A completion whose caller may hang up. It runs on to its usage report
+     * either way, so it holds the route's concurrency slot until it settles.
+     */
    #complete(
        res: Response,
        args: ICompleteArguments,
    ): Promise<IChatCompleteResult> {
        Context.set('abortSignal', abortOnDisconnect(res));
-        return this.#driver().complete(args);
+        const completion = this.#driver().complete(args);
+        Context.get('keepConcurrencySlot')?.(completion);
+        return completion;
    }
 }

@@ -18,7 +18,7 @@
 */

 import { describe, expect, it } from 'vitest';
-import type { IChatMessageResult } from '../../drivers/ai-chat/types.js';
+import type { IChatMessageResult } from '../../drivers/ai/interfaces/chat/types.js';
 import {
    anthropicUsage,
    parseAnthropicRequest,
@@ -32,7 +32,7 @@

 import type { IncomingHttpHeaders } from 'node:http';
 import { HttpError } from '../../core/http/HttpError.js';
-import { sanitizeCacheControl } from '../../drivers/ai-chat/providers/claude/anthropicPolicy.js';
+import { sanitizeCacheControl } from '../../drivers/ai/providers/claude/chat/anthropicPolicy.js';
 import type {
    ICompleteArguments,
    IChatMessageResult,
@@ -42,9 +42,9 @@ import type {
    ThinkingConfig,
    ToolChoice,
    UsageDetails,
-} from '../../drivers/ai-chat/types.js';
-import { needsOpenAICoercion } from '../../drivers/ai-chat/utils/normalizeToOpenAI.js';
-import { fromFinishReason } from '../../drivers/ai-chat/utils/stopReason.js';
+} from '../../drivers/ai/interfaces/chat/types.js';
+import { needsOpenAICoercion } from '../../drivers/ai/interfaces/chat/utils/normalizeToOpenAI.js';
+import { fromFinishReason } from '../../drivers/ai/interfaces/chat/utils/stopReason.js';

 const toStringOrEmpty = (v: unknown): string =>
    typeof v === 'string' ? v : '';
@@ -20,15 +20,14 @@
 /**
 * SSE transport (`startSse`) plus the Anthropic `/v1/messages` streaming writer
 * (`AnthropicSseWriter`), which turns the driver's NDJSON chunk protocol
- * (`drivers/ai-chat/utils/Streaming.js`) into the exact Anthropic event
- * sequence: `message_start` → `content_block_start/delta/stop`* →
+ * (`drivers/ai/interfaces/chat/utils/Streaming.js`) into the exact Anthropic
+ * event sequence: `message_start` → `content_block_start/delta/stop`* →
 * `message_delta` → `message_stop`.
 */

 import type { Response } from 'express';
-import { onResponseClose } from '../../core/http/onResponseClose.js';
-import type { UsageDetails } from '../../drivers/ai-chat/types.js';
-import { promoteStopForToolCalls } from '../../drivers/ai-chat/utils/stopReason.js';
+import type { UsageDetails } from '../../drivers/ai/interfaces/chat/types.js';
+import { promoteStopForToolCalls } from '../../drivers/ai/interfaces/chat/utils/stopReason.js';
 import { anthropicUsage } from './anthropicWire.js';

 // -- SSE transport ------------------------------------------------------------
@@ -80,7 +79,7 @@ export const startSse = (res: Response, opts: { ping?: boolean } = {}): Sse => {
        timer.unref?.();
    }

-    onResponseClose(res, () => {
+    res.once('close', () => {
        ended = true;
        if (timer) clearInterval(timer);
    });
@@ -1,210 +0,0 @@
-/*
- * Copyright (C) 2024-present Puter Technologies Inc.
- *
- * This file is part of Puter.
- *
- * Puter is free software: you can redistribute it and/or modify
- * it under the terms of the GNU Affero General Public License as published
- * by the Free Software Foundation, either version 3 of the License, or
- * (at your option) any later version.
- *
- * This program is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
- * GNU Affero General Public License for more details.
- *
- * You should have received a copy of the GNU Affero General Public License
- * along with this program.  If not, see <https://www.gnu.org/licenses/>.
- */
-
-import { GetObjectCommand } from '@aws-sdk/client-s3';
-import { randomBytes } from 'node:crypto';
-import { Readable } from 'node:stream';
-import { afterAll, beforeAll, describe, expect, it, vi } from 'vitest';
-import {
-    RESPONSE_CLOSE_LISTENER_BUDGET,
-    setupPuterTestEnv,
-    TEST_USER_CREDENTIALS,
-    trackResponseCloseListeners,
-    type PuterTestEnv,
-} from '../../testUtil.js';
-
-/** File downloads over WebDAV: the bytes, ranges, and either end failing. */
-describe('WebDAV GET over HTTP', () => {
-    let env: PuterTestEnv;
-
-    beforeAll(async () => {
-        env = await setupPuterTestEnv();
-    }, 60_000);
-
-    afterAll(async () => {
-        await env?.shutdown();
-    });
-
-    const dav = (path: string, init: RequestInit = {}) => {
-        const port = new URL(env.apiOrigin).port;
-        const auth = Buffer.from(
-            `${TEST_USER_CREDENTIALS.username}:${TEST_USER_CREDENTIALS.password}`,
-        ).toString('base64');
-        return fetch(`http://dav.puter.localhost:${port}${encodeURI(path)}`, {
-            ...init,
-            headers: {
-                Authorization: `Basic ${auth}`,
-                // Uncompressed, so the bytes on the wire are the stored ones.
-                'Accept-Encoding': 'identity',
-                ...(init.headers as Record<string, string> | undefined),
-            },
-        });
-    };
-
-    const writeFile = async (name: string, content: Buffer) => {
-        const { username } = env.users.user;
-        const user = await env.server.stores.user.getByUsername(username);
-        const path = `/${username}/Documents/${name}`;
-        const { fsEntry } = await env.server.services.fs.write(user!.id, {
-            fileMetadata: {
-                path,
-                size: content.byteLength,
-                contentType: 'application/octet-stream',
-            },
-            fileContent: content,
-        });
-        return { path, uuid: fsEntry.uuid };
-    };
-
-    it('returns the stored bytes, whole or as a range', async () => {
-        const content = randomBytes(128 * 1024 + 3);
-        const file = await writeFile('dav-get.bin', content);
-
-        const whole = await dav(file.path);
-        expect(whole.status).toBe(200);
-        expect(whole.headers.get('content-length')).toBe(
-            String(content.byteLength),
-        );
-        expect(Buffer.from(await whole.arrayBuffer())).toEqual(content);
-
-        const range = await dav(file.path, {
-            headers: { Range: 'bytes=10-19' },
-        });
-        expect(range.status).toBe(206);
-        expect(range.headers.get('content-range')).toBe(
-            `bytes 10-19/${content.byteLength}`,
-        );
-        expect(Buffer.from(await range.arrayBuffer())).toEqual(
-            content.subarray(10, 20),
-        );
-    });
-
-    /**
-     * Storage hands over one chunk of `file` and then goes quiet, as a large
-     * object does while the client is still reading.
-     */
-    const stallStorage = (file: { uuid: string }) => {
-        const s3 = env.server.stores.s3Object;
-        const client = env.server.clients.s3.get(s3.resolveRegion(null));
-        const send = client.send.bind(client);
-        const bodies: Readable[] = [];
-        const spy = vi.spyOn(client, 'send').mockImplementation(((
-            command: unknown,
-        ) => {
-            if (
-                !(command instanceof GetObjectCommand) ||
-                command.input.Key !== file.uuid
-            ) {
-                return send(command as never);
-            }
-            let pushed = false;
-            const body = new Readable({
-                read() {
-                    if (!pushed) {
-                        pushed = true;
-                        this.push(Buffer.alloc(16 * 1024, 1));
-                    }
-                },
-            });
-            bodies.push(body);
-            return Promise.resolve({
-                Body: body,
-                ContentLength: 1024 * 1024,
-                ContentType: 'application/octet-stream',
-            });
-        }) as never);
-        return { bodies, restore: () => spy.mockRestore() };
-    };
-
-    const expectFullDownload = async (path: string) => {
-        const response = await dav(path);
-        expect(response.status).toBe(200);
-        expect((await response.arrayBuffer()).byteLength).toBe(1024 * 1024);
-    };
-
-    it('releases the storage stream when the client aborts', async () => {
-        const file = await writeFile(
-            'dav-abort.bin',
-            Buffer.alloc(1024 * 1024, 1),
-        );
-        const storage = stallStorage(file);
-        try {
-            const abort = new AbortController();
-            const response = await dav(file.path, { signal: abort.signal });
-            expect(response.status).toBe(200);
-            await response.body!.getReader().read();
-            abort.abort();
-
-            expect(storage.bodies).toHaveLength(1);
-            await vi.waitFor(() =>
-                expect(storage.bodies[0]!.destroyed).toBe(true),
-            );
-        } finally {
-            storage.restore();
-        }
-
-        await expectFullDownload(file.path);
-    });
-
-    it('cuts the download off when storage fails mid-stream', async () => {
-        const file = await writeFile(
-            'dav-reset.bin',
-            Buffer.alloc(1024 * 1024, 1),
-        );
-        const storage = stallStorage(file);
-        try {
-            const response = await dav(file.path);
-            expect(response.status).toBe(200);
-            const reader = response.body!.getReader();
-            await reader.read();
-            storage.bodies[0]!.destroy(new Error('upstream reset'));
-
-            const outcome = await (async () => {
-                try {
-                    while (!(await reader.read()).done);
-                    return 'ended';
-                } catch {
-                    return 'cut off';
-                }
-            })();
-            expect(outcome).toBe('cut off');
-        } finally {
-            storage.restore();
-        }
-
-        await expectFullDownload(file.path);
-    });
-
-    it('stays within the close listener budget per response', async () => {
-        const file = await writeFile('dav-listeners.bin', randomBytes(4096));
-        const tracker = trackResponseCloseListeners();
-        try {
-            const response = await dav(file.path);
-            expect(response.status).toBe(200);
-            await response.arrayBuffer();
-        } finally {
-            tracker.restore();
-        }
-        const peak = tracker.peak(
-            (req) => req.method === 'GET' && req.url === file.path,
-        );
-        expect(peak).toBeGreaterThan(0);
-        expect(peak).toBeLessThanOrEqual(RESPONSE_CLOSE_LISTENER_BUDGET);
-    });
-});
@@ -1365,7 +1365,6 @@ describe('WebDAVController verbs', () => {
                callback();
            },
        });
-        sink.on('error', () => {});
        sink.on('finish', () => {
            captured.ended = true;
        });
@@ -21,6 +21,7 @@ import type { Team2faLookup } from '../../core/http/middleware/gates.js';
 import { compare as bcryptCompare } from 'bcrypt';
 import type { Request, Response } from 'express';
 import { posix as pathPosix } from 'node:path';
+import { pipeline } from 'node:stream/promises';
 import type { Actor } from '../../core/actor.js';
 import { HttpError } from '../../core/http/HttpError.js';
 import { pipeToResponse } from '../../core/http/pipeToResponse.js';
@@ -550,7 +551,12 @@ export class WebDAVController extends PuterController {
                'Content-Length': String(result.contentLength ?? 0),
            });
        }
-        await pipeToResponse(result.body, res);
+        try {
+            await pipeline(result.body, res);
+        } catch {
+            // Client disconnect or upstream stream error: pipeline already
+            // tore down both ends, so nothing is left dangling to crash on.
+        }
    }

    // -- PROPFIND ----------------------------------------------------
@@ -76,14 +76,8 @@ export interface Actor {
    /**
     * Set when authenticated by a session token or an app-under-user token
     * carrying one. `kind` mirrors the session row (`web`, `app`, `worker`).
-     * `parentUid` is the sign-in a `web` session was handed out under, set only
-     * on a first-party page's session.
     */
-    session?: {
-        uid: string;
-        kind?: string | null;
-        parentUid?: string | null;
-    } | null;
+    session?: { uid: string; kind?: string | null } | null;
    /**
     * Events handler runs behind this actor's writes; set only from an events
     * handler token or a token one minted. Provenance, not authority, so
@@ -69,11 +69,11 @@ export interface KnownContextFields {
     */
    abortSignal: AbortSignal;
    /**
-     * Keeps the call's concurrency slot until `work` settles (set by
-     * DriverController when the method is capped), for work that outlives the
-     * response.
+     * Keeps the request's concurrency slot until `work` settles, for work that
+     * outlives the response. Set by DriverController for a capped method and by
+     * the route concurrency gate.
     */
-    keepDriverSlot: (work: Promise<unknown>) => void;
+    keepConcurrencySlot: (work: Promise<unknown>) => void;
    /**
     * Aborts the chat driver's upstream request. A client disconnect doesn't
     * fire it: the upstream bills a request it has accepted either way, so the
@@ -102,7 +102,7 @@ const KNOWN_KEY_FLAGS: Record<keyof KnownContextFields, true> = {
    requestId: true,
    driverName: true,
    abortSignal: true,
-    keepDriverSlot: true,
+    keepConcurrencySlot: true,
    upstreamSignal: true,
    strictUpstreamErrors: true,
    platformWrite: true,
@@ -19,7 +19,6 @@

 import type { Response } from 'express';
 import { HttpError } from './HttpError.js';
-import { onResponseClose } from './onResponseClose.js';

 /**
 * A signal that fires when the caller hangs up before its response is done, so
@@ -39,6 +38,6 @@ export const abortOnDisconnect = (res: Response): AbortSignal => {
        }
    };
    if (res.destroyed) onClose();
-    else onResponseClose(res, onClose);
+    else res.once('close', onClose);
    return abort.signal;
 };
@@ -18,7 +18,6 @@
 */

 import type { Response } from 'express';
-import { onResponseClose } from '../../core/http/onResponseClose.js';

 // Under the concurrency store's one-hour orphan TTL.
 const SLOT_RENEW_MS = 10 * 60 * 1000;
@@ -29,9 +28,9 @@ interface SlotHandle {
 }

 /**
- * A driver call's concurrency slot, held until the response is done and the
- * call's work has settled, including work a driver carries on after the caller
- * hung up or was answered.
+ * A request's concurrency slot, held until the response is done and the work
+ * registered with `keep` has settled, including work that carries on after the
+ * caller hung up or was answered.
 */
 export class CallSlot {
    #handle?: SlotHandle;
@@ -47,7 +46,7 @@ export class CallSlot {
            this.#maybeRelease();
        };
        res.once('finish', responded);
-        onResponseClose(res, responded);
+        res.once('close', responded);
        if (res.destroyed || res.writableFinished) responded();
        if (!this.#released && this.#work.size > 0) this.#startRenewing();
    }
@@ -104,13 +104,6 @@ declare global {
             */
            deviceFingerprint?: string;

-            /**
-             * The client's TLS fingerprint, read from the header named by
-             * `tlsFingerprintHeader`. Shared by every client on the same TLS
-             * stack, so it identifies software, not a person.
-             */
-            tlsFingerprint?: string;
-
            /**
             * Parsed cookies, populated by the global `cookie-parser`
             * middleware.
@@ -28,7 +28,6 @@ import {
 } from '../../../services/metering/costs.js';
 import type { UsageInput } from '../../../services/metering/types';
 import '../expressAugmentation';
-import { onResponseClose } from '../onResponseClose';

 /**
 * Subset of the metering service this middleware needs. Metering is optional
@@ -196,7 +195,7 @@ export const createEgressMeteringMiddleware = (
        // 'close' rather than 'finish': it fires for a response that completed
        // AND for one whose connection died mid-stream, and an aborted download
        // still sent whatever it sent.
-        onResponseClose(res, () => {
+        res.once('close', () => {
            try {
                if (!isMeteredHost(req)) return;
                const actor = resolveEgressActor(req);
@@ -23,11 +23,8 @@ import { createFingerprintMiddleware } from './fingerprint';

 const VALID_FP = 'dev-fingerprint-abc123';

-const run = (
-    reqPartial: Partial<Request> & Record<string, unknown>,
-    opts?: Parameters<typeof createFingerprintMiddleware>[0],
-) => {
-    const middleware = createFingerprintMiddleware(opts);
+const run = (reqPartial: Partial<Request> & Record<string, unknown>) => {
+    const middleware = createFingerprintMiddleware();
    const req = {
        headers: {},
        ...reqPartial,
@@ -93,21 +90,4 @@ describe('createFingerprintMiddleware', () => {
        const { req } = run({ ip: '203.0.113.1' });
        expect(req.deviceFingerprint).toBeUndefined();
    });
-
-    it('reads the TLS fingerprint from the configured header only', () => {
-        const ja4 = 't13d1516h2_8daaf6152771_02713d6af862';
-        const headers = { 'x-tls-fp': ja4 };
-
-        expect(run({ headers }).req.tlsFingerprint).toBeUndefined();
-        expect(
-            run({ headers }, { tlsFingerprintHeader: 'X-TLS-FP' }).req
-                .tlsFingerprint,
-        ).toBe(ja4);
-        expect(
-            run(
-                { headers: { 'x-tls-fp': 'not a fingerprint!' } },
-                { tlsFingerprintHeader: 'x-tls-fp' },
-            ).req.tlsFingerprint,
-        ).toBeUndefined();
-    });
 });
@@ -33,20 +33,6 @@ const DEVICE_FINGERPRINT_SHAPE = /^[A-Za-z0-9._-]{8,128}$/;
 /** Header the GUI may send the device fingerprint on for non-signup requests. */
 const DEVICE_FINGERPRINT_HEADER = 'x-puter-device-fingerprint';

-/** JA3 / JA4-style values; anything else is dropped. */
-const TLS_FINGERPRINT_SHAPE = /^[A-Za-z0-9_]{10,64}$/;
-
-function readTlsFingerprint(
-    req: { headers?: Record<string, unknown> },
-    header: string | undefined,
-): string | undefined {
-    if (!header) return undefined;
-    const value = req.headers?.[header];
-    return typeof value === 'string' && TLS_FINGERPRINT_SHAPE.test(value)
-        ? value
-        : undefined;
-}
-
 function readDeviceFingerprint(req: {
    body?: unknown;
    headers?: Record<string, unknown>;
@@ -79,22 +65,15 @@ function readDeviceFingerprint(req: {
 *   and spoofable, but stable per real device across IP rotation. The rate
 *   limiter's 'fingerprint' strategy appends it to the network hash so each
 *   device behind a shared network gets its own bucket.
- * - `req.tlsFingerprint` — set only when `tlsFingerprintHeader` is configured and
- *   the request carries a well-shaped value in that header. It names the
- *   client's TLS stack, so every client on the same build shares it.
 *
 * Install AFTER the body parsers (so the body fingerprint is readable) and
 * before `requestContext` (so the snapshot into ALS already carries them).
 * Never rejects — a missing/invalid device fingerprint is simply absent.
 */
-export const createFingerprintMiddleware = (
-    opts: { tlsFingerprintHeader?: string } = {},
-): RequestHandler => {
-    const tlsHeader = opts.tlsFingerprintHeader?.toLowerCase();
+export const createFingerprintMiddleware = (): RequestHandler => {
    return (req, _res, next) => {
        req.networkFingerprint = computeNetworkFingerprint(req);
        req.deviceFingerprint = readDeviceFingerprint(req);
-        req.tlsFingerprint = readTlsFingerprint(req, tlsHeader);
        next();
    };
 };
@@ -20,7 +20,6 @@
 import type { Request, RequestHandler, Response } from 'express';
 import { actorUid } from '../../actor';
 import { HttpError } from '../HttpError';
-import { onResponseClose } from '../onResponseClose';
 import { createJsonBodyParser } from './jsonBody';

 /** The global JSON parser's limit, which applies before auth. */
@@ -153,7 +152,7 @@ export const createLargeJsonBodyParser = ({
        }
        inFlightBytes += declared;
        inFlightByUser.set(user, userBytes + declared);
-        onResponseClose(res, () => {
+        res.once('close', () => {
            inFlightBytes -= declared;
            const left = (inFlightByUser.get(user) ?? 0) - declared;
            if (left > 0) inFlightByUser.set(user, left);
@@ -26,7 +26,7 @@ import type { IConfig } from '../../../types';
 * Browsers send none of those variations, but config values do — a hand-edited
 * `origin` with a trailing slash would otherwise reject every real request.
 */
-export const normalizeOrigin = (raw: string | undefined): string =>
+const normalizeOrigin = (raw: string | undefined): string =>
    (raw ?? '').trim().replace(/\/+$/, '').toLowerCase();

 /**
@@ -20,8 +20,9 @@

 import crypto from 'node:crypto';
 import { withSpan } from '../../../util/span.js';
+import { Context } from '../../context.js';
+import { CallSlot } from '../callSlot.js';
 import { HttpError } from '../HttpError.js';
-import { onResponseClose } from '../onResponseClose.js';
 import { subscriptionOverride } from '../../../services/metering/consts.js';

 /**
@@ -413,8 +414,8 @@ async function peekKv(kv, key, limit) {
 //
 // Lifecycle is the wedge between rate and concurrent limits: rate just
 // records a tick, concurrent has to track "still in flight" → "done"
-// across a request boundary. The route middleware hooks `res.finish` /
-// `res.close`; the driver helper wraps the invocation in `try/finally`.
+// across a request boundary. The route middleware and the driver controller
+// both hold the slot in a `CallSlot`.

 // Orphan TTL safety nets — used when the process dies between acquire
 // and release. Memory backend doesn't need one (the process is also
@@ -1138,8 +1139,9 @@ async function resolveSubscriptionLimit(req, opts) {
 * bySubscription: { user_free: 2, unlimited: 50 } } } { concurrent: { limit:
 * 10, backend: 'redis', scope: 'expensive-op' } }
 *
- * On accept, schedules release on `res.finish` / `res.close` so even aborted
- * requests give their slot back. On reject, 429 with the same
+ * On accept, holds the slot until the response is done and any work the handler
+ * registered through `keepConcurrencySlot` has settled, so work that outlives a
+ * caller who hung up still counts. On reject, 429 with the same
 * `too_many_requests` legacyCode as the rate gate (clients already handle that
 * branch). Fails open on backend error.
 */
@@ -1165,13 +1167,11 @@ export function concurrencyGate(opts) {
            return;
        }

-        // `finish` (response sent) and `close` (connection closed, possibly
-        // aborted before finish) can both fire; the guard releases once.
-        const { release } = guardSlot(slot, '[concurrent]');
-        const onDone = () => void release();
-        res.once('finish', onDone);
-        onResponseClose(res, onDone);
-        if (res.destroyed || res.writableFinished) onDone();
+        const held = new CallSlot();
+        held.attach(guardSlot(slot, '[concurrent]'), res);
+        if (Context.current()) {
+            Context.set('keepConcurrencySlot', (work) => held.keep(work));
+        }
        next();
    };
 }
@@ -29,6 +29,7 @@ import {
    vi,
 } from 'vitest';
 import { EventEmitter } from 'node:events';
+import { Context, runWithContext } from '../../context.js';
 import { isHttpError } from '../HttpError.js';
 import { setupTestServer } from '../../../testUtil.ts';
 import {
@@ -977,6 +978,27 @@ describe('concurrencyGate — memory backend', () => {
        expect(isHttpError(third.err)).toBe(true);
    });

+    it('holds the slot past the response for work the handler keeps', async () => {
+        const opts = { limit: 1, key: 'ip', scope: 'cg-kept-work' };
+        let finishWork;
+        const work = new Promise((resolve) => {
+            finishWork = resolve;
+        });
+        const res = makeRes();
+        await runWithContext({}, async () => {
+            await concurrencyGate(opts)(baseReq(), res, () => {
+                Context.get('keepConcurrencySlot')(work);
+            });
+        });
+        res.emit('close');
+        await new Promise((r) => setImmediate(r));
+        expect((await runGate(opts, baseReq())).err?.statusCode).toBe(429);
+
+        finishWork();
+        await new Promise((r) => setImmediate(r));
+        expect((await runGate(opts, baseReq())).err).toBeUndefined();
+    });
+
    it('different keys do not share slots', async () => {
        const opts = { limit: 1, key: 'ip', scope: 'cg-keyed' };
        const a = await runGate(opts, baseReq({ ip: '1.1.1.1' }));
@@ -1,105 +0,0 @@
-/*
- * Copyright (C) 2024-present Puter Technologies Inc.
- *
- * This file is part of Puter.
- *
- * Puter is free software: you can redistribute it and/or modify
- * it under the terms of the GNU Affero General Public License as published
- * by the Free Software Foundation, either version 3 of the License, or
- * (at your option) any later version.
- *
- * This program is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
- * GNU Affero General Public License for more details.
- *
- * You should have received a copy of the GNU Affero General Public License
- * along with this program.  If not, see <https://www.gnu.org/licenses/>.
- */
-
-import { EventEmitter } from 'node:events';
-import { createServer, get, type Server } from 'node:http';
-import type { AddressInfo } from 'node:net';
-import { afterEach, describe, expect, it, vi } from 'vitest';
-import { onResponseClose } from './onResponseClose.js';
-
-let server: Server | undefined;
-afterEach(async () => {
-    server?.closeAllConnections();
-    await new Promise((resolve) => server?.close(resolve) ?? resolve(null));
-    server = undefined;
-});
-
-describe('onResponseClose', () => {
-    it('shares one close listener and runs every callback once, in order', () => {
-        const res = new EventEmitter();
-        const calls: string[] = [];
-        onResponseClose(res, () => calls.push('a'));
-        onResponseClose(res, () => calls.push('b'));
-        onResponseClose(res, () => calls.push('c'));
-        expect(res.listenerCount('close')).toBe(1);
-
-        res.emit('close');
-        res.emit('close');
-        expect(calls).toEqual(['a', 'b', 'c']);
-        expect(res.listenerCount('close')).toBe(0);
-    });
-
-    it('keeps responses apart', () => {
-        const first = new EventEmitter();
-        const second = new EventEmitter();
-        const onFirst = vi.fn();
-        const onSecond = vi.fn();
-        onResponseClose(first, onFirst);
-        onResponseClose(second, onSecond);
-
-        first.emit('close');
-        expect(onFirst).toHaveBeenCalledOnce();
-        expect(onSecond).not.toHaveBeenCalled();
-    });
-
-    it('runs the rest when one callback throws, and still surfaces the error', async () => {
-        const res = new EventEmitter();
-        const after = vi.fn();
-        const failure = new Error('callback broke');
-        const uncaught = vi.fn();
-        const listeners = process.listeners('uncaughtException');
-        process.removeAllListeners('uncaughtException');
-        process.on('uncaughtException', uncaught);
-        try {
-            onResponseClose(res, () => {
-                throw failure;
-            });
-            onResponseClose(res, after);
-            res.emit('close');
-            expect(after).toHaveBeenCalledOnce();
-            await vi.waitFor(() =>
-                expect(uncaught).toHaveBeenCalledWith(
-                    failure,
-                    expect.anything(),
-                ),
-            );
-        } finally {
-            process.off('uncaughtException', uncaught);
-            for (const listener of listeners)
-                process.on('uncaughtException', listener);
-        }
-    });
-
-    it('fires when a client hangs up on a real response', async () => {
-        const closed = new Promise<boolean>((resolve) => {
-            server = createServer((_req, res) => {
-                onResponseClose(res, () => resolve(res.writableFinished));
-                res.write('partial');
-            });
-        });
-        await new Promise<void>((resolve) => server!.listen(0, resolve));
-        const { port } = server!.address() as AddressInfo;
-        const req = get(`http://127.0.0.1:${port}/`, (res) => {
-            res.once('data', () => req.destroy());
-        });
-        req.on('error', () => {});
-
-        expect(await closed).toBe(false);
-    });
-});
@@ -1,57 +0,0 @@
-/*
- * Copyright (C) 2024-present Puter Technologies Inc.
- *
- * This file is part of Puter.
- *
- * Puter is free software: you can redistribute it and/or modify
- * it under the terms of the GNU Affero General Public License as published
- * by the Free Software Foundation, either version 3 of the License, or
- * (at your option) any later version.
- *
- * This program is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
- * GNU Affero General Public License for more details.
- *
- * You should have received a copy of the GNU Affero General Public License
- * along with this program.  If not, see <https://www.gnu.org/licenses/>.
- */
-
-/** Only what's used here, so test doubles and express's Response both fit. */
-interface ClosingResponse {
-    once(event: 'close', listener: () => void): unknown;
-}
-
-const callbacksByResponse = new WeakMap<ClosingResponse, Array<() => void>>();
-
-/**
- * Runs `callback` once when the response closes. Every callback on a response
- * shares one `close` listener, since Node warns past ten and request tracing
- * adds its own. Callbacks run in the order added; one that throws doesn't stop
- * the rest.
- */
-export const onResponseClose = (
-    res: ClosingResponse,
-    callback: () => void,
-): void => {
-    const callbacks = callbacksByResponse.get(res);
-    if (callbacks) {
-        callbacks.push(callback);
-        return;
-    }
-    const pending = [callback];
-    callbacksByResponse.set(res, pending);
-    res.once('close', () => {
-        for (const run of pending.splice(0)) {
-            try {
-                run();
-            } catch (e) {
-                // Surfaces as an uncaught error, as it would from a listener
-                // of its own.
-                queueMicrotask(() => {
-                    throw e;
-                });
-            }
-        }
-    });
-};
@@ -1,147 +0,0 @@
-/*
- * Copyright (C) 2024-present Puter Technologies Inc.
- *
- * This file is part of Puter.
- *
- * Puter is free software: you can redistribute it and/or modify
- * it under the terms of the GNU Affero General Public License as published
- * by the Free Software Foundation, either version 3 of the License, or
- * (at your option) any later version.
- *
- * This program is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
- * GNU Affero General Public License for more details.
- *
- * You should have received a copy of the GNU Affero General Public License
- * along with this program.  If not, see <https://www.gnu.org/licenses/>.
- */
-
-import { afterAll, beforeAll, describe, expect, it } from 'vitest';
-import {
-    RESPONSE_CLOSE_LISTENER_BUDGET,
-    setupPuterTestEnv,
-    trackResponseCloseListeners,
-    type PuterTestEnv,
-} from '../../testUtil.js';
-
-/**
- * The routes that hold a response open the longest, where middleware, gates and
- * the handler each want to hear when it closes.
- */
-describe('close listeners per response over HTTP', () => {
-    let env: PuterTestEnv;
-    let filePath = '';
-
-    beforeAll(async () => {
-        env = await setupPuterTestEnv({
-            meteringEnforcement: { subscriptions: false },
-        } as never);
-        const { username } = env.users.user;
-        const user = await env.server.stores.user.getByUsername(username);
-        filePath = `/${username}/Documents/close-listeners.txt`;
-        await env.server.services.fs.write(user!.id, {
-            fileMetadata: {
-                path: filePath,
-                size: 4096,
-                contentType: 'text/plain',
-            },
-            fileContent: Buffer.alloc(4096, 97),
-        });
-    }, 120_000);
-
-    afterAll(async () => {
-        await env?.shutdown();
-    });
-
-    const streamedChat = {
-        model: 'fake',
-        stream: true,
-        messages: [{ role: 'user', content: 'hi' }],
-        max_tokens: 16,
-    };
-    const post = (path: string, token: string, body: unknown) =>
-        fetch(new URL(path, env.apiOrigin), {
-            method: 'POST',
-            headers: {
-                Authorization: `Bearer ${token}`,
-                'Content-Type': 'application/json',
-                Origin: env.apiOrigin,
-            },
-            body: JSON.stringify(body),
-        });
-
-    const routes: Array<[string, string, () => Promise<Response>]> = [
-        [
-            'a streamed driver call',
-            '/drivers/call',
-            () =>
-                post('/drivers/call', env.users.user.token, {
-                    interface: 'puter-chat-completion',
-                    service: 'ai-chat',
-                    method: 'complete',
-                    args: streamedChat,
-                }),
-        ],
-        [
-            'an Anthropic-style streamed completion',
-            '/puterai/anthropic/v1/messages',
-            () =>
-                post(
-                    '/puterai/anthropic/v1/messages',
-                    env.users.user.apiToken,
-                    streamedChat,
-                ),
-        ],
-        [
-            'an OpenAI-style streamed completion',
-            '/puterai/openai/v1/chat/completions',
-            () =>
-                post(
-                    '/puterai/openai/v1/chat/completions',
-                    env.users.user.apiToken,
-                    streamedChat,
-                ),
-        ],
-        [
-            'a file read',
-            '/read',
-            () => {
-                const url = new URL('/read', env.apiOrigin);
-                url.searchParams.set('file', filePath);
-                return fetch(url, {
-                    headers: {
-                        Authorization: `Bearer ${env.users.user.token}`,
-                    },
-                });
-            },
-        ],
-        [
-            'a v2 file read',
-            '/fs/read',
-            () => {
-                const url = new URL('/fs/read', env.apiOrigin);
-                url.searchParams.set('path', filePath);
-                return fetch(url, {
-                    headers: {
-                        Authorization: `Bearer ${env.users.user.token}`,
-                    },
-                });
-            },
-        ],
-    ];
-
-    it.each(routes)('%s stays within the budget', async (_name, path, call) => {
-        const tracker = trackResponseCloseListeners();
-        try {
-            const response = await call();
-            expect(response.status).toBe(200);
-            await response.arrayBuffer();
-        } finally {
-            tracker.restore();
-        }
-        const peak = tracker.peak((req) => req.url?.split('?')[0] === path);
-        expect(peak).toBeGreaterThan(0);
-        expect(peak).toBeLessThanOrEqual(RESPONSE_CLOSE_LISTENER_BUDGET);
-    });
-});
@@ -21,7 +21,6 @@ import type { Request, RequestHandler } from 'express';
 import { actorUid } from '../actor';
 import type { EventClient } from '../../clients/event/EventClient';
 import type { LifecycleRejectStatus } from '../../clients/event/types';
-import { onResponseClose } from './onResponseClose';
 import type { RouteMethod } from './types';

 type RoutePathValue = string | RegExp | Array<string | RegExp> | undefined;
@@ -180,7 +179,7 @@ export const createRouteLifecycleMiddleware = (
            );
        };
        res.once('finish', () => settle(false));
-        onResponseClose(res, () => settle(!res.writableFinished));
+        res.once('close', () => settle(!res.writableFinished));
        next();
    };
 };
@@ -263,10 +263,10 @@ export interface RouteOptions {
    rateLimit?: RouteRateLimit | RouteRateLimit[];

    /**
-     * Cap simultaneous in-flight requests per key; slots are released on
-     * response finish or close. `bySubscription` overrides `limit` per
-     * `SubscriptionPolicy.id` when metering is wired in. `key`, `scope`,
-     * `backend` as in `rateLimit`.
+     * Cap simultaneous in-flight requests per key; a slot is released once the
+     * response is done and any work the handler kept it for has settled.
+     * `bySubscription` overrides `limit` per `SubscriptionPolicy.id` when
+     * metering is wired in. `key`, `scope`, `backend` as in `rateLimit`.
     */
    concurrent?: {
        limit: number;
@@ -1,132 +0,0 @@
-/*
- * Copyright (C) 2024-present Puter Technologies Inc.
- *
- * This file is part of Puter.
- *
- * Puter is free software: you can redistribute it and/or modify
- * it under the terms of the GNU Affero General Public License as published
- * by the Free Software Foundation, either version 3 of the License, or
- * (at your option) any later version.
- *
- * This program is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
- * GNU Affero General Public License for more details.
- *
- * You should have received a copy of the GNU Affero General Public License
- * along with this program.  If not, see <https://www.gnu.org/licenses/>.
- */
-
-import { describe, expect, it } from 'vitest';
-import {
-    messagesHaveCompaction,
-    toAnthropicContextManagement,
-    toOpenAiContextManagement,
-    wantsCompaction,
-} from './compaction.js';
-
-describe('toOpenAiContextManagement', () => {
-    it('returns undefined when compaction is off', () => {
-        expect(toOpenAiContextManagement({})).toBeUndefined();
-        expect(
-            toOpenAiContextManagement({ compaction: false }),
-        ).toBeUndefined();
-    });
-
-    it('builds a compaction entry from `true`', () => {
-        expect(toOpenAiContextManagement({ compaction: true })).toEqual([
-            { type: 'compaction' },
-        ]);
-    });
-
-    it('maps trigger_tokens to compact_threshold', () => {
-        expect(
-            toOpenAiContextManagement({ compaction: { trigger_tokens: 1000 } }),
-        ).toEqual([{ type: 'compaction', compact_threshold: 1000 }]);
-    });
-
-    it('passes a raw context_management payload through verbatim', () => {
-        const raw = [{ type: 'compaction', compact_threshold: 5 }];
-        expect(
-            toOpenAiContextManagement({
-                compaction: true,
-                context_management: raw,
-            }),
-        ).toBe(raw);
-    });
-});
-
-describe('toAnthropicContextManagement', () => {
-    it('returns undefined when compaction is off', () => {
-        expect(toAnthropicContextManagement({})).toBeUndefined();
-    });
-
-    it('builds a compact_20260112 edit from `true`', () => {
-        expect(toAnthropicContextManagement({ compaction: true })).toEqual({
-            edits: [{ type: 'compact_20260112' }],
-        });
-    });
-
-    it('maps trigger_tokens to an input_tokens trigger', () => {
-        expect(
-            toAnthropicContextManagement({
-                compaction: { trigger_tokens: 2000 },
-            }),
-        ).toEqual({
-            edits: [
-                {
-                    type: 'compact_20260112',
-                    trigger: { type: 'input_tokens', value: 2000 },
-                },
-            ],
-        });
-    });
-
-    it('passes a raw context_management payload through verbatim', () => {
-        const raw = { edits: [{ type: 'compact_20260112' }] };
-        expect(
-            toAnthropicContextManagement({ context_management: raw }),
-        ).toBe(raw);
-    });
-});
-
-describe('wantsCompaction', () => {
-    it('is false without opt-in', () => {
-        expect(wantsCompaction({})).toBe(false);
-        expect(wantsCompaction({ compaction: false })).toBe(false);
-    });
-
-    it('is true for the neutral opt-in or a raw passthrough', () => {
-        expect(wantsCompaction({ compaction: true })).toBe(true);
-        expect(wantsCompaction({ compaction: { trigger_tokens: 1 } })).toBe(
-            true,
-        );
-        expect(wantsCompaction({ context_management: [] })).toBe(true);
-    });
-});
-
-describe('messagesHaveCompaction', () => {
-    it('detects a round-tripped compaction content block', () => {
-        expect(
-            messagesHaveCompaction([
-                { role: 'user', content: [{ type: 'text', text: 'hi' }] },
-                {
-                    role: 'assistant',
-                    content: [
-                        { type: 'compaction', encrypted_content: 'ENC' },
-                    ],
-                },
-            ]),
-        ).toBe(true);
-    });
-
-    it('is false for ordinary messages or non-arrays', () => {
-        expect(
-            messagesHaveCompaction([
-                { role: 'user', content: [{ type: 'text', text: 'hi' }] },
-            ]),
-        ).toBe(false);
-        expect(messagesHaveCompaction(undefined)).toBe(false);
-        expect(messagesHaveCompaction('nope')).toBe(false);
-    });
-});
@@ -1,718 +0,0 @@
-/*
- * Copyright (C) 2024-present Puter Technologies Inc.
- *
- * This file is part of Puter.
- *
- * Puter is free software: you can redistribute it and/or modify
- * it under the terms of the GNU Affero General Public License as published
- * by the Free Software Foundation, either version 3 of the License, or
- * (at your option) any later version.
- *
- * This program is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
- * GNU Affero General Public License for more details.
- *
- * You should have received a copy of the GNU Affero General Public License
- * along with this program.  If not, see <https://www.gnu.org/licenses/>.
- */
-
-import {
-    DetectDocumentTextCommand,
-    InvalidS3ObjectException,
-    TextractClient,
-    UnsupportedDocumentException,
-} from '@aws-sdk/client-textract';
-import { Mistral } from '@mistralai/mistralai';
-import { Actor } from '../../core/actor.js';
-import { Context } from '../../core/context.js';
-import { HttpError } from '../../core/http/HttpError.js';
-import {
-    awsCredentialsFrom,
-    type AwsCredentials,
-} from '../../util/awsCredentials.js';
-import { mimeFromName } from '../../util/fileSigning.js';
-import {
-    assertActorHasCredits,
-    insufficientCreditsError,
-} from '../../services/metering/enforcement.js';
-import type { CreditHold } from '../../services/metering/types.js';
-import { PuterDriver } from '../types.js';
-import {
-    type AiMeteringService,
-    withAiCostFactor,
-} from '../util/aiCostFactor.js';
-import { AI_CONCURRENT, AI_RATE_LIMIT } from '../util/aiLimits.js';
-import {
-    type LoadedFile,
-    type ResolvedFileInput,
-    resolveFileInput,
-} from '../util/fileInput.js';
-import {
-    type ProviderCatalog,
-    ProviderRegistry,
-    readProviderKey,
-} from '../util/providerRegistry.js';
-import { OCR_COSTS } from './costs.js';
-import {
-    DEFAULT_OCR_MODEL,
-    findOcrModel,
-    OCR_MAX_INPUT_BYTES,
-    OCR_MAX_PAGES,
-    RETIRED_OCR_MODELS,
-    type OcrModel,
-    type OcrProviderId,
-} from './models.js';
-import { countPdfPages } from './pdfPages.js';
-import { textractRates } from './textractPricing.js';
-
-/**
- * Driver implementing `puter-ocr` — document OCR. Two providers: •
- * `aws-textract` — AWS Textract (region-aware clients; direct S3 source when
- * available) • `mistral` — Mistral OCR (URL/data-URL based)
- */
-interface RecognizeArgs {
-    source?: unknown;
-    file?: unknown;
-    provider?: string;
-    model?: string;
-    // Mistral-specific options — ignored by Textract.
-    pages?: number[];
-    includeImageBase64?: boolean;
-    imageLimit?: number;
-    imageMinSize?: number;
-    bboxAnnotationFormat?: unknown;
-    documentAnnotationFormat?: unknown;
-    documentAnnotationPrompt?: string;
-    tableFormat?: string;
-    extractHeader?: boolean;
-    extractFooter?: boolean;
-    test_mode?: boolean;
-}
-
-/** One recognized line, in the same shape for every provider. */
-interface OcrBlock {
-    type: string;
-    text: string;
-    confidence?: number;
-    page: number;
-}
-
-interface TextractBlock {
-    BlockType?: string;
-    Confidence?: number;
-    Text?: string;
-}
-
-interface MistralOcrResponse {
-    model?: string;
-    pages?: Array<{
-        index?: number;
-        markdown?: string;
-        images?: unknown[];
-        dimensions?: unknown;
-    }>;
-    documentAnnotation?: string | null;
-    usageInfo?: { pagesProcessed?: number };
-}
-
-interface MistralOcrClient {
-    ocr: {
-        process: (
-            payload: Record<string, unknown>,
-        ) => Promise<MistralOcrResponse>;
-    };
-}
-
-// Aliases callers may use in place of a canonical provider id. Resolved here
-// rather than in the SDK so a new alias reaches every caller at once.
-const OCR_CATALOG: ProviderCatalog = {
-    label: 'OCR',
-    ids: ['aws-textract', 'mistral'],
-    defaultId: 'aws-textract',
-    aliases: {
-        aws: 'aws-textract',
-        'aws-textract': 'aws-textract',
-        textract: 'aws-textract',
-        mistral: 'mistral',
-        'mistral-ocr': 'mistral',
-    },
-};
-
-/** Recognizes one input with one provider's model. */
-type OcrRecognizer = (
-    input: ResolvedFileInput,
-    args: RecognizeArgs,
-    model: OcrModel,
-    actor: Actor,
-) => Promise<unknown>;
-
-const NOT_CONFIGURED: Record<OcrProviderId, string> = {
-    'aws-textract': 'AWS credentials not configured',
-    mistral: 'Mistral OCR not configured',
-};
-
-const TABLE_FORMATS = new Set(['markdown', 'html']);
-
-const DEFAULT_TEXTRACT_REGION = 'us-west-2';
-const TEXTRACT_PAGE_USAGE_TYPE = 'aws-textract:detect-document-text:page';
-
-const badRequest = (message: string) =>
-    new HttpError(400, message, { legacyCode: 'bad_request' });
-
-/**
- * Accept a Mistral annotation format in the REST spelling
- * (`json_schema.schema`) or the SDK spelling (`jsonSchema.schemaDefinition`)
- * and return the SDK shape.
- */
-const toMistralResponseFormat = (
-    name: string,
-    value: unknown,
-): Record<string, unknown> => {
-    const format = value as Record<string, unknown> | null;
-    const jsonSchema = (format?.jsonSchema ?? format?.json_schema) as
-        Record<string, unknown> | undefined;
-    const schema = jsonSchema?.schemaDefinition ?? jsonSchema?.schema;
-    if (
-        !format ||
-        typeof format !== 'object' ||
-        (format.type !== undefined && format.type !== 'json_schema') ||
-        !schema ||
-        typeof schema !== 'object'
-    ) {
-        throw badRequest(
-            `\`${name}\` must be { type: 'json_schema', json_schema: { name, schema } }`,
-        );
-    }
-    return {
-        type: 'json_schema',
-        jsonSchema: {
-            name:
-                typeof jsonSchema?.name === 'string'
-                    ? jsonSchema.name
-                    : 'annotation',
-            schemaDefinition: schema,
-            ...(typeof jsonSchema?.description === 'string' && {
-                description: jsonSchema.description,
-            }),
-            ...(typeof jsonSchema?.strict === 'boolean' && {
-                strict: jsonSchema.strict,
-            }),
-        },
-    };
-};
-
-const inputMimeType = (loaded: LoadedFile): string =>
-    loaded.mimeType ??
-    mimeFromName(loaded.filename) ??
-    'application/octet-stream';
-
-/**
- * Declared documents (PDF, DOCX, PPTX, ...) and PDF bytes are documents; images
- * and untyped bytes are read as a single image.
- */
-const isDocumentInput = (loaded: LoadedFile): boolean => {
-    const mime = inputMimeType(loaded);
-    return (
-        loaded.buffer.subarray(0, 4).toString('latin1') === '%PDF' ||
-        loaded.filename.toLowerCase().endsWith('.pdf') ||
-        (!mime.startsWith('image/') && mime !== 'application/octet-stream')
-    );
-};
-
-/**
- * The most pages a call can bill, known before the provider runs: a PDF's page
- * count, or a size-based guess for other documents, capped by a `pages`
- * selection and by what the provider reads in one call. Reads the input only
- * for a provider that takes more than one page.
- */
-const estimateOcrPages = async (
-    input: ResolvedFileInput,
-    provider: OcrProviderId,
-    selection?: unknown,
-): Promise<number> => {
-    const maxPages = OCR_MAX_PAGES[provider];
-    if (maxPages <= 1) return 1;
-    const loaded = await input.load();
-    if (!isDocumentInput(loaded)) return 1;
-    // A document we can't count is billed in proportion to its size, reaching
-    // the provider's page limit at its size limit (20 pages per MB for
-    // Mistral). Small files stay cheap; a large one can't pass as one page.
-    const documentPages =
-        (await countPdfPages(loaded.buffer, maxPages)) ??
-        Math.ceil(
-            (loaded.buffer.length * maxPages) / OCR_MAX_INPUT_BYTES[provider],
-        );
-    const selected = Array.isArray(selection)
-        ? new Set(
-              selection.filter(
-                  (page) => Number.isInteger(page) && (page as number) >= 0,
-              ),
-          ).size
-        : 0;
-    return Math.max(1, Math.min(documentPages, maxPages, selected || Infinity));
-};
-
-export class OCRDriver extends PuterDriver {
-    readonly driverInterface = 'puter-ocr';
-    readonly driverName = 'ai-ocr';
-
-    // Shared AI policy — see `drivers/util/aiLimits.ts` for the tier table.
-    readonly rateLimit = AI_RATE_LIMIT;
-    readonly concurrent = AI_CONCURRENT;
-
-    /** Metering scoped to this driver. Lazy: services wire up after drivers. */
-    get #aiMetering(): AiMeteringService {
-        return withAiCostFactor(
-            this.services.metering,
-            this.clients.event,
-            this.driverName,
-        );
-    }
-
-    /**
-     * Textract is reported at the configured region's first-tier rate. A call
-     * that reads its file from a bucket in another region is billed at that
-     * region's rate instead, which this per-usage-type report can't show.
-     */
-    override getReportedCosts() {
-        const textractRegion =
-            this.#awsConfig?.region ?? DEFAULT_TEXTRACT_REGION;
-        return Object.entries(OCR_COSTS).map(([usageType, ucentsPerUnit]) => ({
-            usageType,
-            ucentsPerUnit:
-                usageType === TEXTRACT_PAGE_USAGE_TYPE
-                    ? this.#textractRates.peek(textractRegion)
-                    : ucentsPerUnit,
-            unit: 'page',
-            source: 'driver:aiOcr',
-        }));
-    }
-
-    // Older SDK bundles name the provider in the driver slot instead of
-    // passing `{ provider }`; `#resolveModel` reads the requested alias
-    // back off the Context.
-    readonly driverAliases = [...OCR_CATALOG.ids];
-    readonly isDefault = true;
-
-    // Textract state — one client per region.
-    #textractClients: Record<string, TextractClient> = {};
-    #awsConfig: { credentials: AwsCredentials; region: string } | null = null;
-    /** DetectDocumentText rates per region, from AWS's public price list. */
-    #textractRates = textractRates(OCR_COSTS[TEXTRACT_PAGE_USAGE_TYPE]);
-
-    // Mistral state.
-    #mistral: MistralOcrClient | null = null;
-
-    /** Configured providers, each a recognizer bound to its state above. */
-    #providers = new ProviderRegistry<OcrRecognizer>(OCR_CATALOG);
-
-    override onServerStart() {
-        const providers = this.config.providers ?? {};
-
-        const textract = providers['aws-textract'] as
-            Record<string, unknown> | undefined;
-        const textractAws = (textract?.aws ?? textract) as
-            Record<string, unknown> | undefined;
-        const textractCredentials = awsCredentialsFrom(textractAws);
-        const textractRegion =
-            (textractAws?.region as string | undefined) ??
-            (textract?.region as string | undefined) ??
-            DEFAULT_TEXTRACT_REGION;
-        if (textractCredentials) {
-            this.#awsConfig = {
-                credentials: textractCredentials,
-                region: textractRegion,
-            };
-            // Warm the configured region so the cost report has its rate.
-            void this.#textractRates.rate(textractRegion);
-            this.#providers.register(
-                'aws-textract',
-                (input, _args, model, actor) =>
-                    this.#textractRecognize(input, model, actor),
-            );
-        }
-
-        const mistralKey = readProviderKey(providers['mistral-ocr']);
-        if (mistralKey) {
-            try {
-                this.#mistral = new Mistral({
-                    apiKey: mistralKey,
-                }) as unknown as MistralOcrClient;
-                this.#providers.register(
-                    'mistral',
-                    (input, args, model, actor) =>
-                        this.#mistralRecognize(input, args, model, actor),
-                );
-            } catch (e) {
-                console.warn(
-                    '[OCRDriver] Failed to init Mistral:',
-                    (e as Error).message,
-                );
-            }
-        }
-    }
-
-    async recognize(args: RecognizeArgs) {
-        if (args.test_mode) return sampleResponse();
-
-        const model = this.#resolveModel(args);
-
-        const actor = Context.get('actor');
-        if (!actor)
-            throw new HttpError(401, 'Authentication required', {
-                legacyCode: 'unauthorized',
-            });
-
-        const input = args.source ?? args.file;
-        if (!input) throw badRequest('`source` is required');
-
-        const recognize = this.#providers.get(model.provider);
-        if (!recognize)
-            throw new HttpError(500, NOT_CONFIGURED[model.provider], {
-                legacyCode: 'internal_error',
-            });
-
-        // Cheap, cached check: reject a zero-balance actor before paying for
-        // the file load and page estimate, which a real credit check would
-        // reject anyway once it runs.
-        await assertActorHasCredits(this.services.metering, actor, this.config);
-
-        // A Puter file is resolved and access-checked here but read only if
-        // the provider needs its bytes; Textract reads it straight from S3.
-        const resolved = await resolveFileInput(
-            this.stores,
-            this.services.fs,
-            actor,
-            input,
-            {
-                acceptWebInput: true,
-                maxBytes: OCR_MAX_INPUT_BYTES[model.provider],
-            },
-        );
-
-        return recognize(resolved, args, model, actor);
-    }
-
-    /**
-     * Decide which model handles a call. A `model` picks its own provider; an
-     * explicit `provider`, then the legacy driver alias the caller dispatched
-     * through, then whichever provider is configured pick that provider's
-     * default model.
-     */
-    #resolveModel(args: RecognizeArgs): OcrModel {
-        const provider = args.provider
-            ? (this.#providers.resolve(args.provider) as OcrProviderId)
-            : undefined;
-
-        if (args.model !== undefined) {
-            if (typeof args.model !== 'string' || !args.model.trim())
-                throw badRequest('`model` must be a non-empty string');
-            const retired = RETIRED_OCR_MODELS[args.model.trim().toLowerCase()];
-            if (retired)
-                throw badRequest(
-                    `${args.model} is no longer available: ${retired}`,
-                );
-            const model = findOcrModel(args.model);
-            if (!model) throw badRequest(`Unknown OCR model: ${args.model}`);
-            if (provider && model.provider !== provider)
-                throw badRequest(
-                    `OCR model ${args.model} is not served by provider ${args.provider}`,
-                );
-            return model;
-        }
-
-        const hinted =
-            provider ?? this.#providers.normalize(Context.get('driverName'));
-        if (!hinted && this.#providers.names().length === 0)
-            throw new HttpError(500, 'No OCR provider configured', {
-                legacyCode: 'internal_error',
-            });
-        const chosen = (hinted ?? this.#providers.defaultId()) as OcrProviderId;
-        return findOcrModel(DEFAULT_OCR_MODEL[chosen])!;
-    }
-
-    /**
-     * Refuse a call the balance can't cover, and hold its cost while the
-     * provider runs so the account's concurrent calls see it. Usage is still
-     * metered from the pages the provider reports.
-     */
-    async #holdCredits(
-        actor: Actor,
-        usageType: string,
-        cost: number,
-    ): Promise<CreditHold> {
-        // Priced at the cost factor usage is recorded at.
-        const hold = await this.#aiMetering.reserveAiCredits(
-            actor,
-            usageType,
-            cost,
-        );
-        if (!hold) throw insufficientCreditsError();
-        return hold;
-    }
-
-    // -- AWS Textract -------------------------------------------------
-
-    #textractClientFor(region: string): TextractClient {
-        const cached = this.#textractClients[region];
-        if (cached) return cached;
-        const client = new TextractClient({
-            credentials: this.#awsConfig!.credentials,
-            region,
-        });
-        this.#textractClients[region] = client;
-        return client;
-    }
-
-    async #textractRecognize(
-        input: ResolvedFileInput,
-        model: OcrModel,
-        actor: Actor,
-    ) {
-        // Prefer S3 direct source if the file is FS-backed; fall back to raw bytes.
-        const { fsEntry } = input;
-        const s3Info =
-            fsEntry && fsEntry.bucket && fsEntry.bucketRegion
-                ? {
-                      bucket: fsEntry.bucket,
-                      bucketRegion: fsEntry.bucketRegion,
-                      key: fsEntry.uuid,
-                  }
-                : null;
-
-        const configuredRegion =
-            this.#awsConfig!.region ?? DEFAULT_TEXTRACT_REGION;
-
-        // Both regions the call may run in are priced up front, so the hold
-        // covers either and a refresh mid-call can't change the charge.
-        const [configuredRate, s3Rate] = await Promise.all([
-            this.#textractRates.rate(configuredRegion),
-            s3Info ? this.#textractRates.rate(s3Info.bucketRegion) : undefined,
-        ]);
-
-        const tryRun = async (useS3: boolean) => {
-            const region =
-                s3Info && useS3 ? s3Info.bucketRegion : configuredRegion;
-            const client = this.#textractClientFor(region);
-            const document =
-                s3Info && useS3
-                    ? { S3Object: { Bucket: s3Info.bucket, Name: s3Info.key } }
-                    : { Bytes: (await input.load()).buffer };
-            return client.send(
-                new DetectDocumentTextCommand({ Document: document }),
-            );
-        };
-
-        const hold = await this.#holdCredits(
-            actor,
-            model.pageUsageType,
-            Math.max(configuredRate, s3Rate ?? 0) *
-                (await estimateOcrPages(input, model.provider)),
-        );
-        try {
-            let response;
-            let costPerPage: number;
-            try {
-                try {
-                    response = await tryRun(Boolean(s3Info));
-                    costPerPage = s3Rate ?? configuredRate;
-                } catch (err) {
-                    if (!(s3Info && err instanceof InvalidS3ObjectException))
-                        throw err;
-                    response = await tryRun(false);
-                    costPerPage = configuredRate;
-                }
-            } catch (err) {
-                if (err instanceof UnsupportedDocumentException)
-                    throw badRequest(
-                        'AWS Textract reads JPEG, PNG, TIFF and single-page PDF documents; use a Mistral OCR model for multi-page PDFs and other formats',
-                    );
-                throw err;
-            }
-
-            const blocks: OcrBlock[] = [];
-            let pageCount = 0;
-            for (const block of (response.Blocks ?? []) as TextractBlock[]) {
-                if (block.BlockType === 'PAGE') {
-                    pageCount += 1;
-                    continue;
-                }
-                if (block.BlockType !== 'LINE' || !block.Text) continue;
-                blocks.push({
-                    type: 'text/textract:LINE',
-                    text: block.Text,
-                    confidence: Number(block.Confidence ?? 0),
-                    page: Math.max(pageCount - 1, 0),
-                });
-            }
-
-            const pages = pageCount || 1;
-            this.#aiMetering.incrementUsage(
-                actor,
-                model.pageUsageType,
-                pages,
-                costPerPage * pages,
-            );
-            return {
-                model: model.id,
-                blocks,
-                text: blocks.map((b) => b.text).join('\n'),
-            };
-        } finally {
-            await hold.release();
-        }
-    }
-
-    // -- Mistral OCR --------------------------------------------------
-
-    async #mistralRecognize(
-        input: ResolvedFileInput,
-        args: RecognizeArgs,
-        model: OcrModel,
-        actor: Actor,
-    ) {
-        const loaded = await input.load();
-        const payload: Record<string, unknown> = {
-            model: model.id,
-            document: this.#mistralBuildChunk(loaded),
-        };
-        if (args.pages) payload.pages = args.pages;
-        if (args.includeImageBase64 !== undefined)
-            payload.includeImageBase64 = args.includeImageBase64;
-        if (typeof args.imageLimit === 'number')
-            payload.imageLimit = args.imageLimit;
-        if (typeof args.imageMinSize === 'number')
-            payload.imageMinSize = args.imageMinSize;
-        if (args.bboxAnnotationFormat !== undefined)
-            payload.bboxAnnotationFormat = toMistralResponseFormat(
-                'bboxAnnotationFormat',
-                args.bboxAnnotationFormat,
-            );
-        if (args.documentAnnotationFormat !== undefined)
-            payload.documentAnnotationFormat = toMistralResponseFormat(
-                'documentAnnotationFormat',
-                args.documentAnnotationFormat,
-            );
-        if (typeof args.documentAnnotationPrompt === 'string')
-            payload.documentAnnotationPrompt = args.documentAnnotationPrompt;
-        if (args.tableFormat !== undefined) {
-            if (!TABLE_FORMATS.has(args.tableFormat))
-                throw badRequest("`tableFormat` must be 'markdown' or 'html'");
-            payload.tableFormat = args.tableFormat;
-        }
-        if (typeof args.extractHeader === 'boolean')
-            payload.extractHeader = args.extractHeader;
-        if (typeof args.extractFooter === 'boolean')
-            payload.extractFooter = args.extractFooter;
-
-        const annotations =
-            payload.documentAnnotationFormat !== undefined ||
-            payload.bboxAnnotationFormat !== undefined;
-        const costPerPage =
-            OCR_COSTS[model.pageUsageType] +
-            (annotations && model.annotationUsageType
-                ? OCR_COSTS[model.annotationUsageType]
-                : 0);
-        const hold = await this.#holdCredits(
-            actor,
-            model.pageUsageType,
-            costPerPage *
-                (await estimateOcrPages(input, model.provider, args.pages)),
-        );
-
-        try {
-            const response = await this.#mistral!.ocr.process(payload);
-            this.#recordMistralUsage(response, model, actor, annotations);
-            return this.#normalizeMistralResponse(response, model);
-        } finally {
-            await hold.release();
-        }
-    }
-
-    #mistralBuildChunk(loaded: LoadedFile): Record<string, unknown> {
-        const mime = inputMimeType(loaded);
-        const dataUrl = `data:${mime};base64,${loaded.buffer.toString('base64')}`;
-        if (isDocumentInput(loaded)) {
-            return {
-                type: 'document_url',
-                documentUrl: dataUrl,
-                documentName: loaded.filename,
-            };
-        }
-        return { type: 'image_url', imageUrl: { url: dataUrl } };
-    }
-
-    #normalizeMistralResponse(response: MistralOcrResponse, model: OcrModel) {
-        const pages = response?.pages ?? [];
-        const blocks: OcrBlock[] = [];
-        for (const [position, page] of pages.entries()) {
-            if (typeof page?.markdown !== 'string') continue;
-            const lines = page.markdown
-                .split('\n')
-                .map((l) => l.trim())
-                .filter(Boolean);
-            for (const line of lines) {
-                blocks.push({
-                    type: 'text/mistral:LINE',
-                    text: line,
-                    page: page.index ?? position,
-                });
-            }
-        }
-        return {
-            model: response?.model ?? model.id,
-            pages,
-            usage_info: response?.usageInfo,
-            ...(typeof response?.documentAnnotation === 'string' && {
-                document_annotation: response.documentAnnotation,
-            }),
-            blocks,
-            text: blocks.map((b) => b.text).join('\n'),
-        };
-    }
-
-    #recordMistralUsage(
-        response: MistralOcrResponse,
-        model: OcrModel,
-        actor: Actor,
-        annotations: boolean,
-    ) {
-        try {
-            const pagesProcessed =
-                response?.usageInfo?.pagesProcessed ??
-                (Array.isArray(response?.pages) ? response.pages.length : 1);
-            this.#aiMetering.incrementUsage(
-                actor,
-                model.pageUsageType,
-                pagesProcessed,
-                OCR_COSTS[model.pageUsageType] * pagesProcessed,
-            );
-            if (annotations && model.annotationUsageType) {
-                this.#aiMetering.incrementUsage(
-                    actor,
-                    model.annotationUsageType,
-                    pagesProcessed,
-                    OCR_COSTS[model.annotationUsageType] * pagesProcessed,
-                );
-            }
-        } catch {
-            // Non-critical.
-        }
-    }
-}
-
-function sampleResponse() {
-    return {
-        model: 'test-mode',
-        blocks: [
-            {
-                type: 'text/puter:sample-output',
-                confidence: 1,
-                text: 'test_mode is enabled; this is a sample OCR response.',
-                page: 0,
-            },
-        ],
-        text: 'test_mode is enabled; this is a sample OCR response.',
-    };
-}
@@ -1,511 +0,0 @@
-/*
- * Copyright (C) 2024-present Puter Technologies Inc.
- *
- * This file is part of Puter.
- *
- * Puter is free software: you can redistribute it and/or modify
- * it under the terms of the GNU Affero General Public License as published
- * by the Free Software Foundation, either version 3 of the License, or
- * (at your option) any later version.
- *
- * This program is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
- * GNU Affero General Public License for more details.
- *
- * You should have received a copy of the GNU Affero General Public License
- * along with this program.  If not, see <https://www.gnu.org/licenses/>.
- */
-
-import { pipeline, Readable, Transform } from 'node:stream';
-import { Context } from '../../core/context.js';
-import { HttpError } from '../../core/http/HttpError.js';
-import type { DriverStreamResult } from '../meta.js';
-import { OUTPUT_FORMATS } from '../ai-tts/providers/elevenlabs/ElevenLabsTTSProvider.js';
-import { PuterDriver } from '../types.js';
-import {
-    type AiMeteringService,
-    withAiCostFactor,
-    withStreamedAiCharge,
-} from '../util/aiCostFactor.js';
-import { AI_CONCURRENT, AI_RATE_LIMIT } from '../util/aiLimits.js';
-import { loadFileInput } from '../util/fileInput.js';
-import {
-    type ProviderCatalog,
-    ProviderRegistry,
-    readProviderKey,
-} from '../util/providerRegistry.js';
-import { SAMPLE_AUDIO_URL } from '../util/testMode.js';
-import { upstreamBodyStream, upstreamFetch } from '../util/upstreamErrors.js';
-import { VOICE_CHANGER_COSTS } from './costs.js';
-
-/**
- * Driver implementing `puter-speech2speech` — voice changer. Currently a single
- * provider (ElevenLabs).
- */
-
-const DEFAULT_MODEL = 'eleven_multilingual_sts_v2';
-const DEFAULT_VOICE_ID = '21m00Tcm4TlvDq8ikWAM';
-const DEFAULT_OUTPUT_FORMAT = 'mp3_44100_128';
-const MAX_AUDIO_FILE_SIZE = 25 * 1024 * 1024;
-// Covers the upload as well as the conversion.
-const CONVERT_TIMEOUT_MS = 10 * 60 * 1000;
-
-const SUPPORTED_MODELS = Object.keys(VOICE_CHANGER_COSTS).map(
-    (key) => key.split(':')[1]!,
-);
-
-const VOICE_CHANGER_CATALOG: ProviderCatalog = {
-    label: 'Speech-to-speech',
-    ids: ['elevenlabs'],
-    defaultId: 'elevenlabs',
-    aliases: { elevenlabs: 'elevenlabs' },
-};
-
-interface ElevenLabsConfig {
-    apiKey: string;
-    baseUrl: string;
-    defaultVoiceId: string;
-    defaultModelId: string;
-}
-
-interface ConvertArgs {
-    audio: unknown;
-    provider?: string;
-    voice?: string;
-    voice_id?: string;
-    voiceId?: string;
-    model?: string;
-    model_id?: string;
-    voice_settings?: unknown;
-    voiceSettings?: unknown;
-    seed?: number;
-    remove_background_noise?: boolean;
-    output_format?: string;
-    file_format?: string;
-    optimize_streaming_latency?: number;
-    enable_logging?: boolean;
-    test_mode?: boolean;
-}
-
-export class VoiceChangerDriver extends PuterDriver {
-    readonly driverInterface = 'puter-speech2speech';
-    readonly driverName = 'ai-speech2speech';
-    // Older SDK bundles name the provider in the driver slot.
-    readonly driverAliases = ['elevenlabs-voice-changer'];
-    readonly isDefault = true;
-
-    // Shared AI policy — see `drivers/util/aiLimits.ts` for the tier table.
-    readonly rateLimit = AI_RATE_LIMIT;
-    readonly concurrent = AI_CONCURRENT;
-
-    /** Metering scoped to this driver. Lazy: services wire up after drivers. */
-    get #aiMetering(): AiMeteringService {
-        return withAiCostFactor(
-            this.services.metering,
-            this.clients.event,
-            this.driverName,
-        );
-    }
-
-    override getReportedCosts(): Record<string, unknown>[] {
-        return Object.entries(VOICE_CHANGER_COSTS).map(
-            ([usageType, ucentsPerUnit]) => ({
-                usageType,
-                ucentsPerUnit,
-                unit: 'second',
-                source: 'driver:aiSpeech2Speech',
-            }),
-        );
-    }
-
-    #providers = new ProviderRegistry<ElevenLabsConfig>(VOICE_CHANGER_CATALOG);
-
-    override onServerStart() {
-        const elevenlabs = this.config.providers?.elevenlabs;
-        const apiKey = readProviderKey(elevenlabs);
-        if (!apiKey) return;
-        this.#providers.register('elevenlabs', {
-            apiKey,
-            baseUrl:
-                (elevenlabs?.apiBaseUrl as string | undefined) ??
-                'https://api.elevenlabs.io',
-            defaultVoiceId:
-                (elevenlabs?.defaultVoiceId as string | undefined) ??
-                DEFAULT_VOICE_ID,
-            defaultModelId:
-                (elevenlabs?.speechToSpeechModelId as string | undefined) ??
-                DEFAULT_MODEL,
-        });
-    }
-
-    async convert(
-        args: ConvertArgs,
-    ): Promise<DriverStreamResult | { url: string; content_type: string }> {
-        // Only one provider exists today, but naming a different one should
-        // fail loudly rather than quietly convert with this one.
-        const providerName = this.#providers.resolve(args.provider);
-
-        if (args.test_mode) {
-            return { url: SAMPLE_AUDIO_URL, content_type: 'audio/mpeg' };
-        }
-
-        const elevenlabs = this.#providers.get(providerName);
-        if (!elevenlabs) {
-            throw new HttpError(500, 'ElevenLabs API key not configured', {
-                legacyCode: 'internal_error',
-            });
-        }
-
-        const actor = Context.get('actor');
-        if (!actor)
-            throw new HttpError(401, 'Authentication required', {
-                legacyCode: 'unauthorized',
-            });
-
-        if (!args.audio) {
-            throw new HttpError(400, '`audio` is required', {
-                legacyCode: 'bad_request',
-            });
-        }
-
-        const modelId =
-            args.model_id || args.model || elevenlabs.defaultModelId;
-        const voiceId =
-            args.voice_id ||
-            args.voiceId ||
-            args.voice ||
-            elevenlabs.defaultVoiceId;
-        if (!voiceId)
-            throw new HttpError(400, '`voice` is required', {
-                legacyCode: 'bad_request',
-            });
-        // `voiceId` lands in the request URL path, which carries our API key,
-        // so anything outside ElevenLabs' `[A-Za-z0-9_-]+` id shape could
-        // steer the request at a different endpoint.
-        if (!/^[A-Za-z0-9_-]+$/.test(voiceId))
-            throw new HttpError(400, '`voice` must be alphanumeric', {
-                legacyCode: 'bad_request',
-            });
-
-        // Gate on the cost table: an id we can't price is one we can't bill
-        // for, and ElevenLabs bills us for it either way. This also keeps
-        // anything but a known id out of the multipart `model_id` field.
-        const usageKey = `elevenlabs:${modelId}:second`;
-        if (!Object.hasOwn(VOICE_CHANGER_COSTS, usageKey)) {
-            throw new HttpError(
-                400,
-                `Invalid model: ${modelId}. Expected: ${SUPPORTED_MODELS.join(', ')}`,
-                {
-                    legacyCode: 'field_invalid',
-                    fields: {
-                        key: 'model',
-                        expected: SUPPORTED_MODELS,
-                        got: modelId,
-                    },
-                },
-            );
-        }
-        const ucentsPerSecond = VOICE_CHANGER_COSTS[usageKey]!;
-
-        const loaded = await loadFileInput(
-            this.stores,
-            this.services.fs,
-            actor,
-            args.audio,
-            { maxBytes: MAX_AUDIO_FILE_SIZE, acceptWebInput: true },
-        );
-
-        // The hold estimates seconds from the input at 128 kbps: compact audio
-        // is under-held rather than refused. Usage is metered from the length
-        // of the audio that comes back.
-        const estimatedSeconds = Math.max(
-            1,
-            Math.ceil(loaded.buffer.byteLength / 16000),
-        );
-        const estimatedCost = ucentsPerSecond * estimatedSeconds;
-
-        const formData = new FormData();
-        const blob = new Blob([loaded.buffer as BlobPart], {
-            type: loaded.mimeType ?? 'application/octet-stream',
-        });
-        formData.append('audio', blob, loaded.filename);
-        formData.append('model_id', modelId);
-
-        const settings = args.voice_settings ?? args.voiceSettings;
-        if (settings !== undefined && settings !== null) {
-            formData.append(
-                'voice_settings',
-                typeof settings === 'string'
-                    ? settings
-                    : JSON.stringify(settings),
-            );
-        }
-        if (args.seed !== undefined && args.seed !== null) {
-            formData.append('seed', String(args.seed));
-        }
-        if (typeof args.remove_background_noise === 'boolean') {
-            formData.append(
-                'remove_background_noise',
-                String(args.remove_background_noise),
-            );
-        }
-        if (args.file_format) {
-            formData.append('file_format', args.file_format);
-        }
-
-        const searchParams = new URLSearchParams();
-        const outputFormat = args.output_format || DEFAULT_OUTPUT_FORMAT;
-        if (outputFormat) searchParams.set('output_format', outputFormat);
-        if (
-            args.optimize_streaming_latency !== undefined &&
-            args.optimize_streaming_latency !== null
-        ) {
-            searchParams.set(
-                'optimize_streaming_latency',
-                String(args.optimize_streaming_latency),
-            );
-        }
-        if (args.enable_logging !== undefined && args.enable_logging !== null) {
-            searchParams.set('enable_logging', String(args.enable_logging));
-        }
-
-        const url = new URL(
-            `/v1/speech-to-speech/${voiceId}`,
-            elevenlabs.baseUrl,
-        );
-        const search = searchParams.toString();
-        if (search) url.search = search;
-
-        const metering = this.#aiMetering;
-        const output = measuredOutput(
-            outputBytesPerSecond(outputFormat),
-            outputFormat,
-        );
-        return withStreamedAiCharge(
-            metering,
-            actor,
-            usageKey,
-            estimatedCost,
-            async (): Promise<DriverStreamResult> => {
-                const response = await upstreamFetch(
-                    'elevenlabs',
-                    url,
-                    {
-                        method: 'POST',
-                        headers: { 'xi-api-key': elevenlabs.apiKey },
-                        body: formData,
-                    },
-                    {
-                        providerLabel: 'ElevenLabs',
-                        timeoutMs: CONVERT_TIMEOUT_MS,
-                        stream: true,
-                    },
-                );
-                return {
-                    dataType: 'stream',
-                    content_type:
-                        response.headers.get('content-type') ?? 'audio/mpeg',
-                    stream: output.measure(upstreamBodyStream(response)),
-                };
-            },
-            async () => {
-                const measured = await output.seconds;
-                // A stream cut short was still converted in full upstream.
-                const seconds =
-                    measured.complete && measured.seconds
-                        ? measured.seconds
-                        : Math.max(measured.seconds ?? 0, estimatedSeconds);
-                return metering.incrementUsage(
-                    actor,
-                    usageKey,
-                    seconds,
-                    ucentsPerSecond * seconds,
-                );
-            },
-        );
-    }
-}
-
-/**
- * Bytes per second of an ElevenLabs `output_format`. A format outside the
- * documented set is sized as the default, so a made-up bitrate can't shrink the
- * bill.
- */
-export function outputBytesPerSecond(format: string): number {
-    const known = OUTPUT_FORMATS.has(format) ? format : DEFAULT_OUTPUT_FORMAT;
-    const [codec, rate, kbps] = known.split('_');
-    if (codec === 'mp3' || codec === 'opus') return Number(kbps) * 125;
-    if (codec === 'ulaw' || codec === 'alaw') return Number(rate);
-    // pcm and wav: 16-bit mono.
-    return Number(rate) * 2;
-}
-
-// Past this, a WAV whose data chunk hasn't started is billed by raw length.
-const WAV_HEADER_MAX_BYTES = 64 * 1024;
-
-/**
- * Where the samples of a RIFF/WAVE stream start, once `head` reaches its `data`
- * chunk; undefined until then, or if `head` isn't a WAV.
- */
-export function wavDataOffset(head: Buffer): number | undefined {
-    if (head.length < 12) return undefined;
-    if (
-        head.toString('ascii', 0, 4) !== 'RIFF' ||
-        head.toString('ascii', 8, 12) !== 'WAVE'
-    )
-        return undefined;
-    let offset = 12;
-    while (offset + 8 <= head.length) {
-        if (head.toString('ascii', offset, offset + 4) === 'data')
-            return offset + 8;
-        const size = head.readUInt32LE(offset + 4);
-        offset += 8 + size + (size & 1);
-    }
-    return undefined;
-}
-
-const OGG_HEADER_BYTES = 27;
-const NO_GRANULE = 2n ** 64n - 1n;
-
-/**
- * Walks an Ogg Opus stream page by page as it arrives. Opus is variable
- * bitrate, so its length comes from the last page's granule position (48 kHz
- * samples, less the stream's pre-skip) rather than its byte count.
- */
-export class OggOpusClock {
-    #header = Buffer.alloc(0);
-    #skip = 0;
-    #firstBody: Buffer | undefined = Buffer.alloc(0);
-    #firstBodyBytes = 0;
-    #granule: bigint | undefined;
-    #preSkip = 0;
-    #broken = false;
-
-    push(chunk: Buffer): void {
-        let offset = 0;
-        while (!this.#broken && offset < chunk.length) {
-            if (this.#skip > 0) {
-                const n = Math.min(this.#skip, chunk.length - offset);
-                if (this.#firstBody && this.#firstBody.length < 19) {
-                    this.#firstBody = Buffer.concat([
-                        this.#firstBody,
-                        chunk.subarray(offset, offset + n),
-                    ]);
-                }
-                this.#skip -= n;
-                offset += n;
-                if (this.#skip === 0 && this.#firstBody) this.#readOpusHead();
-                continue;
-            }
-            const want =
-                this.#header.length < OGG_HEADER_BYTES
-                    ? OGG_HEADER_BYTES
-                    : OGG_HEADER_BYTES + this.#header[26]!;
-            const n = Math.min(
-                want - this.#header.length,
-                chunk.length - offset,
-            );
-            this.#header = Buffer.concat([
-                this.#header,
-                chunk.subarray(offset, offset + n),
-            ]);
-            offset += n;
-            if (this.#header.length < want) continue;
-            if (this.#header.toString('ascii', 0, 4) !== 'OggS') {
-                this.#broken = true;
-                return;
-            }
-            if (want === OGG_HEADER_BYTES && this.#header[26]! > 0) continue;
-            const granule = this.#header.readBigUInt64LE(6);
-            if (granule !== NO_GRANULE) this.#granule = granule;
-            let body = 0;
-            for (const lacing of this.#header.subarray(OGG_HEADER_BYTES)) {
-                body += lacing;
-            }
-            this.#header = Buffer.alloc(0);
-            this.#skip = body;
-            if (this.#firstBody) this.#firstBodyBytes = body;
-        }
-    }
-
-    /** Whole seconds so far, once the stream has proven to be Ogg Opus. */
-    get seconds(): number | undefined {
-        if (this.#broken || this.#granule === undefined) return undefined;
-        const samples = Number(this.#granule) - this.#preSkip;
-        return samples > 0
-            ? Math.max(1, Math.ceil(samples / 48_000))
-            : undefined;
-    }
-
-    #readOpusHead(): void {
-        const head = this.#firstBody!;
-        this.#firstBody = undefined;
-        if (
-            this.#firstBodyBytes < 19 ||
-            head.toString('ascii', 0, 8) !== 'OpusHead'
-        ) {
-            this.#broken = true;
-            return;
-        }
-        this.#preSkip = head.readUInt16LE(10);
-    }
-}
-
-/**
- * Counts the bytes of a converted stream. `seconds` settles when the stream
- * ends or closes: its length in whole seconds (undefined when nothing came
- * back), and whether the upstream sent all of it. A WAV's header isn't audio,
- * so it isn't counted; Opus is timed by its pages.
- */
-function measuredOutput(bytesPerSecond: number, format: string) {
-    let bytes = 0;
-    let head: Buffer | undefined = format.startsWith('wav_')
-        ? Buffer.alloc(0)
-        : undefined;
-    const opus = format.startsWith('opus_') ? new OggOpusClock() : undefined;
-    let headerBytes = 0;
-    let settle!: (measured: { seconds?: number; complete: boolean }) => void;
-    const seconds = new Promise<{ seconds?: number; complete: boolean }>(
-        (resolve) => {
-            settle = resolve;
-        },
-    );
-    const lengthOf = () => {
-        const timed = opus?.seconds;
-        if (timed !== undefined) return timed;
-        const audio = bytes - headerBytes;
-        return audio > 0
-            ? Math.max(1, Math.ceil(audio / bytesPerSecond))
-            : undefined;
-    };
-    const measure = (source: Readable): Readable => {
-        const counter = new Transform({
-            transform(chunk: Buffer, _encoding, callback) {
-                bytes += chunk.length;
-                opus?.push(chunk);
-                if (head) {
-                    head = Buffer.concat([head, chunk]);
-                    const offset = wavDataOffset(head);
-                    if (offset !== undefined) headerBytes = offset;
-                    if (
-                        offset !== undefined ||
-                        head.length >= WAV_HEADER_MAX_BYTES
-                    )
-                        head = undefined;
-                }
-                callback(null, chunk);
-            },
-            flush(callback) {
-                settle({ seconds: lengthOf(), complete: true });
-                callback();
-            },
-        });
-        counter.once('close', () =>
-            settle({ seconds: lengthOf(), complete: false }),
-        );
-        pipeline(source, counter, () => {});
-        return counter;
-    };
-    return { measure, seconds };
-}
@@ -19,9 +19,9 @@

 import { createServer, type Server } from 'node:http';
 import { afterAll, beforeAll, beforeEach, describe, expect, it } from 'vitest';
-import { setupPuterTestEnv, type PuterTestEnv } from '../../testUtil.js';
+import { setupPuterTestEnv, type PuterTestEnv } from '../../../../testUtil.js';
 import { clearUnhealthyRoutes } from './utils/providerHealth.js';
-import { makeActor } from '../../core/actor.js';
+import { makeActor } from '../../../../core/actor.js';

 const policyMessage =
    'Policy Violation: this user has been blocked for a previous policy violation. Learn more: https://vendor.test/private';
@@ -129,6 +129,17 @@ describe('model availability through real HTTP', () => {
            meteringEnforcement: { subscriptions: false },
        } as never);
    }, 120000);
+    it('serves qualified chat model names through HTTP', async () => {
+        const response = await fetch(
+            new URL('/puterai/chat/models', env.apiOrigin),
+        );
+        expect(response.status).toBe(200);
+        const { models } = (await response.json()) as { models: string[] };
+        expect(models).toContain('openrouter:vendor/availability-test');
+        expect(models.every((id) => /^[^:/]+:[^/]+\/.+/.test(id))).toBe(true);
+        expect(models).not.toContain('vendor/availability-test');
+    });
+
    afterAll(async () => {
        await env?.shutdown();
        upstream?.closeAllConnections();
@@ -19,8 +19,8 @@

 import { createServer, type Server, type ServerResponse } from 'node:http';
 import { afterAll, beforeAll, describe, expect, it, vi } from 'vitest';
-import { makeActor, type Actor } from '../../core/actor.js';
-import { setupPuterTestEnv, type PuterTestEnv } from '../../testUtil.js';
+import { makeActor, type Actor } from '../../../../core/actor.js';
+import { setupPuterTestEnv, type PuterTestEnv } from '../../../../testUtil.js';

 const routes = [
    '/drivers/call',
@@ -0,0 +1,225 @@
+/*
+ * Copyright (C) 2024-present Puter Technologies Inc.
+ *
+ * This file is part of Puter.
+ *
+ * Puter is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU Affero General Public License as published
+ * by the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU Affero General Public License for more details.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program.  If not, see <https://www.gnu.org/licenses/>.
+ */
+
+import { createServer, type Server, type ServerResponse } from 'node:http';
+import { afterAll, beforeAll, describe, expect, it, vi } from 'vitest';
+import { setupPuterTestEnv, type PuterTestEnv } from '../../../../testUtil.js';
+
+/**
+ * A completion runs on to its usage report after the caller hangs up, so it
+ * keeps counting against the caller's concurrency cap until it settles.
+ */
+describe('chat concurrency over HTTP after the caller hangs up', () => {
+    let upstream: Server;
+    let env: PuterTestEnv;
+    let onRequest: (res: ServerResponse, body: string) => void;
+
+    beforeAll(async () => {
+        upstream = createServer((req, res) => {
+            if (req.url === '/api/tags') {
+                res.setHeader('Content-Type', 'application/json');
+                res.end(JSON.stringify({ models: [{ name: 'slot-test' }] }));
+                return;
+            }
+            let body = '';
+            req.setEncoding('utf8');
+            req.on('data', (chunk: string) => {
+                body += chunk;
+            });
+            req.once('end', () => onRequest(res, body));
+        });
+        await new Promise<void>((resolve) =>
+            upstream.listen(0, '127.0.0.1', resolve),
+        );
+        const port = (upstream.address() as { port: number }).port;
+        env = await setupPuterTestEnv({
+            providers: {
+                ollama: { apiBaseUrl: `http://127.0.0.1:${port}` },
+            },
+            meteringEnforcement: { subscriptions: false },
+        } as never);
+        await vi.waitFor(async () => {
+            expect(
+                (await env.server.drivers.aiChat.models()).some(
+                    (model) => model.id === 'ollama:ollama/slot-test',
+                ),
+            ).toBe(true);
+        });
+    }, 120_000);
+
+    afterAll(async () => {
+        await env?.shutdown();
+        upstream?.closeAllConnections();
+        if (upstream)
+            await new Promise<void>((resolve) =>
+                upstream.close(() => resolve()),
+            );
+    });
+
+    const call = (route: string, stream: boolean, signal?: AbortSignal) => {
+        const args = {
+            model: 'ollama:ollama/slot-test',
+            stream,
+            messages: [{ role: 'user', content: 'hi' }],
+            max_tokens: 16,
+        };
+        return fetch(new URL(route, env.apiOrigin), {
+            method: 'POST',
+            headers: {
+                'Content-Type': 'application/json',
+                Authorization: `Bearer ${env.users.user.apiToken}`,
+                Origin: env.apiOrigin,
+            },
+            body: JSON.stringify(
+                route === '/drivers/call'
+                    ? {
+                          interface: 'puter-chat-completion',
+                          method: 'complete',
+                          args,
+                      }
+                    : args,
+            ),
+            signal,
+        });
+    };
+
+    const usage = { prompt_tokens: 3, completion_tokens: 5, total_tokens: 8 };
+    const chunk = (delta: object, finish: string | null, extra = {}) =>
+        `data: ${JSON.stringify({
+            id: 'chatcmpl_1',
+            object: 'chat.completion.chunk',
+            choices: [{ index: 0, delta, finish_reason: finish }],
+            ...extra,
+        })}\n\n`;
+
+    const finishAnswer = (res: ServerResponse, body: string) => {
+        if (!(JSON.parse(body) as { stream?: boolean }).stream) {
+            res.setHeader('Content-Type', 'application/json');
+            res.end(
+                JSON.stringify({
+                    id: 'chatcmpl_1',
+                    object: 'chat.completion',
+                    created: 0,
+                    model: 'slot-test',
+                    choices: [
+                        {
+                            index: 0,
+                            message: { role: 'assistant', content: 'done' },
+                            finish_reason: 'stop',
+                        },
+                    ],
+                    usage,
+                }),
+            );
+            return;
+        }
+        if (!res.headersSent)
+            res.setHeader('Content-Type', 'text/event-stream');
+        res.end(
+            `${chunk({ content: 'done' }, 'stop')}data: ${JSON.stringify({
+                id: 'chatcmpl_1',
+                object: 'chat.completion.chunk',
+                choices: [],
+                usage,
+            })}\n\ndata: [DONE]\n\n`,
+        );
+    };
+
+    /** Starts a completion, waits until it is generating, then hangs up. */
+    const abandon = async (route: string, stream: boolean) => {
+        let received!: () => void;
+        const reached = new Promise<void>((resolve) => {
+            received = resolve;
+        });
+        const pending: { res: ServerResponse; body: string }[] = [];
+        onRequest = (res, body) => {
+            pending.push({ res, body });
+            if (stream) {
+                res.setHeader('Content-Type', 'text/event-stream');
+                res.write(chunk({ role: 'assistant', content: 'partial' }, null));
+            }
+            received();
+        };
+        const abort = new AbortController();
+        const request = call(route, stream, abort.signal);
+        const outcome = await Promise.race([
+            reached.then(() => 'generating' as const),
+            request.then((response) => response),
+        ]);
+        if (outcome !== 'generating') {
+            return { refused: outcome, pending };
+        }
+        if (stream) {
+            const response = await request;
+            const reader = response.body!.getReader();
+            let text = '';
+            while (!text.includes('partial')) {
+                const { value, done } = await reader.read();
+                if (done) break;
+                text += new TextDecoder().decode(value);
+            }
+        }
+        abort.abort();
+        await request.catch(() => undefined);
+        return { refused: undefined, pending };
+    };
+
+    it.each([
+        { route: '/drivers/call', stream: true },
+        { route: '/drivers/call', stream: false },
+        { route: '/puterai/openai/v1/chat/completions', stream: true },
+        { route: '/puterai/openai/v1/chat/completions', stream: false },
+        { route: '/puterai/anthropic/v1/messages', stream: true },
+    ])(
+        'counts an abandoned completion until it settles: $route (stream=$stream)',
+        async ({ route, stream }) => {
+            const running: { res: ServerResponse; body: string }[] = [];
+            let refused: Response | undefined;
+            try {
+                for (let i = 0; i < 25 && !refused; i++) {
+                    const attempt = await abandon(route, stream);
+                    running.push(...attempt.pending);
+                    refused = attempt.refused;
+                    // Let the server see the hang-up before the next call.
+                    await new Promise((resolve) => setTimeout(resolve, 50));
+                }
+                expect(refused?.status).toBe(429);
+                expect(await refused!.text()).toContain(
+                    'Too many concurrent requests',
+                );
+                expect(running.length).toBeGreaterThan(0);
+
+                for (const { res, body } of running.splice(0))
+                    finishAnswer(res, body);
+                onRequest = finishAnswer;
+                await expect
+                    .poll(async () => (await call(route, false)).status, {
+                        timeout: 15_000,
+                    })
+                    .toBe(200);
+            } finally {
+                for (const { res, body } of running) finishAnswer(res, body);
+                upstream.closeAllConnections();
+                // The next case starts with a fresh rate-limit window.
+                await new Promise((resolve) => setTimeout(resolve, 10_100));
+            }
+        },
+        60_000,
+    );
+});
@@ -42,22 +42,22 @@ import {
    vi,
 } from 'vitest';

-import { Context } from '../../core/context.js';
-import { HttpError } from '../../core/http/HttpError.js';
-import { PuterServer } from '../../server.js';
-import { setupTestServer } from '../../testUtil.js';
-import { withTestActor } from '../integrationTestUtil.js';
+import { Context } from '../../../../core/context.js';
+import { HttpError } from '../../../../core/http/HttpError.js';
+import { PuterServer } from '../../../../server.js';
+import { setupTestServer } from '../../../../testUtil.js';
+import { withTestActor } from '../../shared/integrationTestUtil.js';
 import { ChatCompletionDriver } from './ChatCompletionDriver.js';
-import { AzureChatProvider } from './providers/azure/AzureChatProvider.js';
-import { FakeChatProvider } from './providers/FakeChatProvider.js';
-import { InfronProvider } from './providers/infron/InfronProvider.js';
-import { NeuralwattProvider } from './providers/neuralwatt/NeuralwattProvider.js';
-import { OpenAiChatProvider } from './providers/openai/OpenAiChatCompletionsProvider.js';
-import { OpenRouterProvider } from './providers/openrouter/OpenRouterProvider.js';
-import { TogetherAIProvider } from './providers/together/TogetherAIProvider.js';
-import { XAI_CHAT_MODELS } from './providers/xai/catalog.js';
-import { XAIProvider } from './providers/xai/XAIProvider.js';
-import { handle_completion_output } from './utils/OpenAIUtil.js';
+import { AzureChatProvider } from '../../providers/azure/chat/AzureChatProvider.js';
+import { FakeChatProvider } from '../../providers/fake/chat/FakeChatProvider.js';
+import { InfronProvider } from '../../providers/infron/chat/InfronProvider.js';
+import { NeuralwattProvider } from '../../providers/neuralwatt/chat/NeuralwattProvider.js';
+import { OpenAiChatProvider } from '../../providers/openai/chat/OpenAiChatCompletionsProvider.js';
+import { OpenRouterProvider } from '../../providers/openrouter/chat/OpenRouterProvider.js';
+import { TogetherAIProvider } from '../../providers/together/chat/TogetherAIProvider.js';
+import { XAI_CHAT_MODELS } from '../../providers/xai/chat/catalog.js';
+import { XAIProvider } from '../../providers/xai/chat/XAIProvider.js';
+import { handle_completion_output } from '../../protocols/openaiCompatible/OpenAIUtil.js';
 import { clearUnhealthyRoutes, isRouteUnhealthy } from './utils/providerHealth.js';

 let server: PuterServer;
@@ -1231,4 +1231,41 @@ describe('ChatCompletionDriver streaming failure handling', () => {
        expect(events).toContainEqual({ type: 'text', text: 'hello' });
        expect(cleanup).toHaveBeenCalledTimes(1);
    });
+
+    it('frees the concurrency slot before the provider cleanup hook finishes', async () => {
+        let finishCleanup!: () => void;
+        const cleanup = vi.fn(
+            () =>
+                new Promise<void>((resolve) => {
+                    finishCleanup = resolve;
+                }),
+        );
+        vi.spyOn(FakeChatProvider.prototype, 'complete').mockResolvedValue({
+            stream: true,
+            init_chat_stream: async ({
+                chatStream,
+            }: {
+                chatStream: { end: (usage?: Record<string, number>) => void };
+            }) => {
+                chatStream.end({ input_tokens: 1, output_tokens: 1 });
+            },
+            finally_fn: cleanup,
+        } as never);
+
+        const kept: Promise<unknown>[] = [];
+        const result = (await withTestActor(() => {
+            Context.set('keepConcurrencySlot', (work) => kept.push(work));
+            return fakeOnlyDriver.complete({
+                model: 'fake',
+                messages: [{ role: 'user', content: 'hi' }],
+                stream: true,
+            } as never);
+        })) as unknown as { stream: Readable };
+        await collect(result.stream);
+
+        expect(kept).toHaveLength(1);
+        await kept[0];
+        await vi.waitFor(() => expect(cleanup).toHaveBeenCalledTimes(1));
+        finishCleanup();
+    });
 });
@@ -0,0 +1,726 @@
+/*
+ * Copyright (C) 2024-present Puter Technologies Inc.
+ *
+ * This file is part of Puter.
+ *
+ * Puter is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU Affero General Public License as published
+ * by the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * Puter is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ * GNU Affero General Public License for more details.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with Puter. If not, see <https://www.gnu.org/licenses/>.
+ */
+
+/**
+ * Seeded fuzzing of `ChatCompletionDriver.complete` through every registered
+ * provider. The vendors are played at the HTTP boundary, so the real SDKs and
+ * the providers' own request and response code run.
+ */
+
+import { afterAll, beforeAll, describe, expect, it, vi } from 'vitest';
+
+import { isHttpError } from '../../../../core/http/HttpError.js';
+import type { PuterServer } from '../../../../server.js';
+import { setupTestServer } from '../../../../testUtil.js';
+import { kv } from '../../../../util/kvSingleton.js';
+import { isDriverStreamResult } from '../../../meta.js';
+import {
+    caseRandom,
+    corruptReply,
+    Findings,
+    fuzzCases,
+    hostileValue,
+    mutate,
+    pick,
+    pickStable,
+    type Random,
+    unclassifiedFailure,
+    upstream,
+    type UpstreamReply,
+    type UpstreamRequest,
+} from '../../shared/fuzzTestUtil.js';
+import {
+    readStreamResult,
+    withTestActor,
+} from '../../shared/integrationTestUtil.js';
+import { ChatCompletionDriver } from './ChatCompletionDriver.js';
+import type { ICompleteArguments } from './types.js';
+import { clearUnhealthyRoutes } from './utils/providerHealth.js';
+
+vi.mock('undici', async (importOriginal) => ({
+    ...(await importOriginal<typeof import('undici')>()),
+    fetch: (input: unknown, init?: RequestInit) => upstream.fetch(input, init),
+}));
+
+// groq-sdk brings its own HTTP client. Groq's API is OpenAI-compatible, so the
+// OpenAI SDK stands in for it and the requests reach the stub like the rest.
+vi.mock('groq-sdk', async () => ({ default: (await import('openai')).OpenAI }));
+
+vi.mock('axios', () => {
+    const request = (config: Parameters<typeof upstream.axios>[0]) =>
+        upstream.axios(config);
+    const get = (url: string, config?: object) =>
+        upstream.axios({ ...config, url, method: 'GET' });
+    return { default: { request, get }, request, get };
+});
+
+// -- The vendors, played at the HTTP boundary --
+
+const TEXT = 'pong';
+const chatUsage = { prompt_tokens: 3, completion_tokens: 5, total_tokens: 8 };
+const chatChunk = (delta: object, finish: string | null) => ({
+    id: 'chatcmpl-1',
+    object: 'chat.completion.chunk',
+    created: 1,
+    model: 'm',
+    choices: [{ index: 0, delta, finish_reason: finish }],
+});
+const responsesMessage = {
+    type: 'message',
+    id: 'msg_1',
+    role: 'assistant',
+    status: 'completed',
+    content: [{ type: 'output_text', text: TEXT, annotations: [] }],
+};
+const responsesBody = {
+    id: 'resp_1',
+    object: 'response',
+    status: 'completed',
+    model: 'm',
+    output: [responsesMessage],
+    usage: {
+        input_tokens: 3,
+        output_tokens: 5,
+        total_tokens: 8,
+        input_tokens_details: { cached_tokens: 0 },
+        output_tokens_details: { reasoning_tokens: 0 },
+    },
+};
+const anthropicUsage = { input_tokens: 3, output_tokens: 5 };
+const geminiBody = {
+    candidates: [
+        {
+            index: 0,
+            content: { role: 'model', parts: [{ text: TEXT }] },
+            finishReason: 'STOP',
+        },
+    ],
+    usageMetadata: {
+        promptTokenCount: 3,
+        candidatesTokenCount: 5,
+        totalTokenCount: 8,
+    },
+};
+
+const anthropicEvents = [
+    {
+        type: 'message_start',
+        message: {
+            id: 'msg_1',
+            type: 'message',
+            role: 'assistant',
+            model: 'm',
+            content: [],
+            stop_reason: null,
+            stop_sequence: null,
+            usage: { input_tokens: 3, output_tokens: 1 },
+        },
+    },
+    {
+        type: 'content_block_start',
+        index: 0,
+        content_block: { type: 'text', text: '' },
+    },
+    {
+        type: 'content_block_delta',
+        index: 0,
+        delta: { type: 'text_delta', text: TEXT },
+    },
+    { type: 'content_block_stop', index: 0 },
+    {
+        type: 'message_delta',
+        delta: { stop_reason: 'end_turn', stop_sequence: null },
+        usage: { output_tokens: 5 },
+    },
+    { type: 'message_stop' },
+];
+
+const responsesEvents = [
+    {
+        type: 'response.created',
+        response: { ...responsesBody, status: 'in_progress', output: [] },
+    },
+    {
+        type: 'response.output_item.added',
+        output_index: 0,
+        item: { ...responsesMessage, status: 'in_progress', content: [] },
+    },
+    {
+        type: 'response.output_text.delta',
+        item_id: 'msg_1',
+        output_index: 0,
+        content_index: 0,
+        delta: TEXT,
+    },
+    {
+        type: 'response.output_item.done',
+        output_index: 0,
+        item: responsesMessage,
+    },
+    { type: 'response.completed', response: responsesBody },
+];
+
+/** A well-formed answer from whichever vendor API the request went to. */
+const healthyUpstream = ({
+    url,
+    json,
+}: UpstreamRequest): UpstreamReply | undefined => {
+    const path = url.pathname;
+    // A caller-supplied media link that leads nowhere.
+    if (url.hostname === 'upstream.test') return { status: 404, body: '' };
+    const streaming = (json as { stream?: unknown } | undefined)?.stream;
+    if (path.endsWith('/chat/completions')) {
+        if (!streaming) {
+            return {
+                body: {
+                    id: 'chatcmpl-1',
+                    object: 'chat.completion',
+                    created: 1,
+                    model: 'm',
+                    choices: [
+                        {
+                            index: 0,
+                            message: { role: 'assistant', content: TEXT },
+                            finish_reason: 'stop',
+                        },
+                    ],
+                    usage: chatUsage,
+                },
+            };
+        }
+        return {
+            sse: [
+                { data: chatChunk({ role: 'assistant', content: TEXT }, null) },
+                { data: chatChunk({}, 'stop') },
+                {
+                    data: {
+                        ...chatChunk({}, null),
+                        choices: [],
+                        usage: chatUsage,
+                    },
+                },
+                { data: '[DONE]' },
+            ],
+        };
+    }
+    if (path.endsWith('/responses')) {
+        return streaming
+            ? {
+                  sse: responsesEvents.map((data) => ({
+                      event: data.type,
+                      data,
+                  })),
+              }
+            : { body: responsesBody };
+    }
+    if (path.endsWith('/messages/count_tokens')) {
+        return { body: { input_tokens: 3 } };
+    }
+    if (path.endsWith('/messages')) {
+        return streaming
+            ? {
+                  sse: anthropicEvents.map((data) => ({
+                      event: data.type,
+                      data,
+                  })),
+              }
+            : {
+                  body: {
+                      id: 'msg_1',
+                      type: 'message',
+                      role: 'assistant',
+                      model: 'm',
+                      content: [{ type: 'text', text: TEXT }],
+                      stop_reason: 'end_turn',
+                      stop_sequence: null,
+                      usage: anthropicUsage,
+                  },
+              };
+    }
+    if (path.endsWith(':generateContent')) return { body: geminiBody };
+    if (path.endsWith(':streamGenerateContent')) {
+        return { sse: [{ data: geminiBody }] };
+    }
+    if (path.endsWith(':countTokens')) return { body: { totalTokens: 3 } };
+    if (path.endsWith('/moderations')) {
+        return {
+            body: {
+                id: 'modr_1',
+                model: 'm',
+                results: [
+                    { flagged: false, categories: {}, category_scores: {} },
+                ],
+            },
+        };
+    }
+    return undefined;
+};
+
+// -- Requests to mutate --
+
+const PNG =
+    'data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mP8z8BQDwAEhQGAhKmMIQAAAABJRU5ErkJggg==';
+
+const weatherTool = {
+    type: 'function',
+    function: {
+        name: 'get_weather',
+        description: 'Get the weather for a city',
+        parameters: {
+            type: 'object',
+            properties: { city: { type: 'string' } },
+            required: ['city'],
+        },
+    },
+};
+
+const CORPUS: Record<string, unknown>[] = [
+    { messages: [{ role: 'user', content: 'hi' }] },
+    {
+        messages: [
+            { role: 'system', content: 'Be brief.' },
+            {
+                role: 'user',
+                content: [
+                    { type: 'text', text: 'What is this?' },
+                    { type: 'image_url', image_url: { url: PNG } },
+                ],
+            },
+        ],
+        max_tokens: 64,
+        temperature: 0.5,
+        top_p: 0.9,
+        stream: true,
+    },
+    {
+        messages: [
+            { role: 'user', content: 'Weather in Paris?' },
+            {
+                role: 'assistant',
+                content: null,
+                tool_calls: [
+                    {
+                        id: 'call_1',
+                        type: 'function',
+                        function: {
+                            name: 'get_weather',
+                            arguments: '{"city":"Paris"}',
+                        },
+                    },
+                ],
+            },
+            { role: 'tool', tool_call_id: 'call_1', content: 'Sunny' },
+        ],
+        tools: [weatherTool],
+        tool_choice: 'auto',
+        parallel_tool_calls: false,
+    },
+    {
+        messages: [
+            {
+                role: 'user',
+                content: [
+                    {
+                        type: 'text',
+                        text: 'Weather in Paris?',
+                        cache_control: { type: 'ephemeral' },
+                    },
+                ],
+            },
+            {
+                role: 'assistant',
+                content: [
+                    { type: 'text', text: 'Checking.' },
+                    {
+                        type: 'tool_use',
+                        id: 'toolu_1',
+                        name: 'get_weather',
+                        input: { city: 'Paris' },
+                    },
+                ],
+            },
+            {
+                role: 'user',
+                content: [
+                    {
+                        type: 'tool_result',
+                        tool_use_id: 'toolu_1',
+                        content: 'Sunny',
+                    },
+                ],
+            },
+        ],
+        tools: [weatherTool, { type: 'web_search' }],
+        tool_choice: { type: 'function', function: { name: 'get_weather' } },
+        stream: true,
+        streamToolInput: true,
+    },
+    {
+        messages: [{ role: 'user', content: 'Reply in JSON.' }],
+        outputFormat: {
+            type: 'json_schema',
+            schema: { type: 'object', properties: { a: { type: 'string' } } },
+        },
+        reasoning: { effort: 'low' },
+        reasoning_effort: 'low',
+        thinking: { type: 'enabled', budget_tokens: 1024 },
+        text: { verbosity: 'low' },
+        verbosity: 'low',
+        stopSequences: ['END'],
+        topK: 5,
+        normalize: true,
+    },
+    {
+        messages: [{ role: 'user', content: 'hi' }],
+        instructions: 'Be brief.',
+        compaction: { trigger_tokens: 150000 },
+        context_management: { edits: [{ type: 'clear_tool_uses_20250919' }] },
+        cacheControl: { type: 'ephemeral', ttl: '1h' },
+        prompt_cache_key: 'k',
+        prompt_cache_retention: '24h',
+        serviceSpeed: 'fast',
+        service_tier: 'flex',
+        speed: 'fast',
+        metadata: { trace: 't' },
+        include: ['reasoning.encrypted_content'],
+        store: false,
+        truncation: 'auto',
+        taskBudget: { total: 1000, remaining: 500 },
+        safeguards: [{ type: 'prompt_injection' }],
+        anthropicBetas: ['context-management-2025-06-27'],
+        moderation: true,
+        response: { normalize: true },
+        fallback: false,
+    },
+];
+
+CORPUS.push(
+    {
+        messages: [
+            {
+                role: 'user',
+                content: [
+                    { type: 'text', text: 'Describe these.' },
+                    { type: 'image_url', image_url: PNG },
+                    {
+                        type: 'image',
+                        source: {
+                            type: 'base64',
+                            media_type: 'image/png',
+                            data: PNG.split(',')[1],
+                        },
+                    },
+                    {
+                        type: 'video_url',
+                        video_url: { url: 'data:video/mp4;base64,AAAA' },
+                    },
+                    {
+                        type: 'file',
+                        file: {
+                            filename: 'a.pdf',
+                            file_data: 'data:application/pdf;base64,JVBERi0=',
+                        },
+                    },
+                    {
+                        type: 'document',
+                        source: {
+                            type: 'base64',
+                            media_type: 'application/pdf',
+                            data: 'JVBERi0=',
+                        },
+                    },
+                ],
+            },
+        ],
+    },
+    {
+        messages: [
+            {
+                role: 'user',
+                content: [
+                    { type: 'text', text: 'Read it.' },
+                    { puter_path: '~/Documents/a.png' },
+                ],
+            },
+        ],
+    },
+    // Model output sent back, in each vendor's shape.
+    {
+        messages: [
+            { role: 'user', content: 'Think, then search.' },
+            {
+                role: 'assistant',
+                content: [
+                    { type: 'thinking', thinking: 'Hm.', signature: 'sig' },
+                    { type: 'redacted_thinking', data: 'x' },
+                    {
+                        type: 'reasoning',
+                        id: 'rs_1',
+                        summary: [{ type: 'summary_text', text: 'Hm.' }],
+                        encrypted_content: 'enc',
+                    },
+                    { type: 'text', text: 'Searching.' },
+                    {
+                        type: 'server_tool_use',
+                        id: 'srvtoolu_1',
+                        name: 'web_search',
+                        input: { query: 'x' },
+                    },
+                    {
+                        type: 'web_search_tool_result',
+                        tool_use_id: 'srvtoolu_1',
+                        content: [],
+                    },
+                    {
+                        type: 'compaction',
+                        id: 'cmp_1',
+                        encrypted_content: 'enc',
+                    },
+                ],
+                reasoning: 'Hm.',
+                refusal: null,
+            },
+            { type: 'compaction', id: 'cmp_2', encrypted_content: 'enc' },
+            { role: 'user', content: 'Go on.' },
+        ],
+        tools: [
+            { type: 'web_search_20250305', name: 'web_search', max_uses: 3 },
+            {
+                type: 'function',
+                function: {
+                    name: 'strict_tool',
+                    strict: true,
+                    parameters: {
+                        type: 'object',
+                        properties: {},
+                        additionalProperties: false,
+                    },
+                },
+            },
+            {
+                name: 'anthropic_style',
+                description: 'x',
+                input_schema: { type: 'object', properties: {} },
+            },
+        ],
+        tool_choice: 'required',
+        thinking: { type: 'adaptive' },
+        prompt_cache_options: { ttl: '30m' },
+        instructions: [{ role: 'system', content: 'Be brief.' }],
+    },
+);
+
+const OPTION_NAMES = [
+    ...new Set(CORPUS.flatMap((request) => Object.keys(request))),
+    'previous_response_id',
+    'conversation',
+    'prompt',
+    'prompt_cache_options',
+    'background',
+    'custom',
+    'customLimitMessage',
+    'test_mode',
+];
+
+// -- Harness --
+
+const key = { apiKey: 'fuzz-key' };
+const providers = {
+    claude: key,
+    'azure-openai': { ...key, apiURL: 'https://azure.upstream.test/openai' },
+    'openai-completion': key,
+    gemini: key,
+    meta: key,
+    groq: key,
+    deepseek: key,
+    mistral: key,
+    xai: key,
+    moonshot: key,
+    minimax: key,
+    zai: key,
+    alibaba: key,
+    byteplus: key,
+    hoonify: key,
+    ollama: { enabled: false },
+};
+
+let server: PuterServer;
+let driver: ChatCompletionDriver;
+// Every provider gets the same share of cases, whatever its catalog size.
+let routes: Map<string, { provider: string; id: string }[]>;
+
+beforeAll(async () => {
+    upstream.install();
+    server = await setupTestServer();
+    driver = new ChatCompletionDriver(
+        { providers } as never,
+        server.clients,
+        server.stores,
+        server.services,
+    );
+    await driver.onServerStart();
+    routes = new Map();
+    for (const model of await driver.models()) {
+        if (!model.provider || model.provider === 'fake-chat') continue;
+        const route = { provider: model.provider, id: model.id };
+        routes.set(route.provider, [
+            ...(routes.get(route.provider) ?? []),
+            route,
+        ]);
+    }
+}, 120_000);
+
+afterAll(async () => {
+    await server?.shutdown();
+    upstream.uninstall();
+});
+
+const pickRoute = (random: Random) =>
+    pickStable(
+        random,
+        routes.get(pickStable(random, [...routes.keys()], (name) => name))!,
+        (route) => route.id,
+    );
+
+/** Runs one completion to its end, stream included, and returns its failure. */
+const complete = async (args: Record<string, unknown>): Promise<unknown> => {
+    try {
+        const result = await withTestActor(() =>
+            driver.complete(
+                structuredClone(args) as unknown as ICompleteArguments,
+            ),
+        );
+        if (isDriverStreamResult(result)) await readStreamResult(result);
+        return undefined;
+    } catch (error) {
+        return error;
+    }
+};
+
+describe('ChatCompletionDriver seeded fuzzing', () => {
+    it('covers every configured provider', () => {
+        expect([...routes.keys()].sort()).toEqual([
+            'alibaba',
+            'azure-openai',
+            'azure-openai-responses',
+            'byteplus',
+            'claude',
+            'deepseek',
+            'gemini',
+            'groq',
+            'hoonify',
+            'meta',
+            'minimax',
+            'mistral',
+            'moonshotai',
+            'openai-completion',
+            'openai-responses',
+            'xai',
+            'zai',
+        ]);
+    });
+
+    it('answers or rejects a malformed request without failing a healthy route', async () => {
+        upstream.respond = healthyUpstream;
+        upstream.unanswered.clear();
+        clearUnhealthyRoutes();
+        const findings = new Findings();
+        let answered = 0;
+
+        for (let index = 0; index < fuzzCases(3000); index++) {
+            const random = caseRandom(0xc4a7, index);
+            const route = pickRoute(random);
+            const args: Record<string, unknown> = {
+                ...mutate(pick(random, CORPUS), random, OPTION_NAMES),
+                provider: route.provider,
+                model: route.id,
+            };
+            // Now and then the route itself is the malformed part.
+            if (random() < 0.05) {
+                args[pick(random, ['provider', 'model'])] =
+                    hostileValue(random);
+            }
+            const label = `${route.provider}:${route.id}`;
+            const error = await complete(args);
+            if (!error) answered++;
+            const unclassified = error && unclassifiedFailure(error);
+            if (unclassified) findings.add(unclassified, index, label, args);
+            else if (isHttpError(error) && error.statusCode >= 500) {
+                const attempts = (error.fields?.attempts ?? []) as {
+                    status?: number;
+                    error: string;
+                }[];
+                findings.add(
+                    `${error.statusCode} ${error.legacyCode}: ${error.message} [${attempts.map((a) => `${a.status ?? 'no status'} ${a.error}`).join(' | ')}]`,
+                    index,
+                    label,
+                    args,
+                );
+            }
+            // The vendors answered every request, so no route failed.
+            for (const marked of kv.keys('aiChat:unhealthyRoute:*')) {
+                findings.add(`marked unhealthy ${marked}`, index, label, args);
+            }
+            clearUnhealthyRoutes();
+        }
+
+        for (const request of upstream.unanswered) {
+            findings.add(`no upstream fixture for ${request}`, -1, '', {});
+        }
+        findings.assertNone();
+        // A run where nothing is answered has stopped reaching the providers.
+        expect(answered).toBeGreaterThan(0);
+    }, 600_000);
+
+    it('ends every request with an answer or a classified error when the upstream misbehaves', async () => {
+        const findings = new Findings();
+        const askedBefore = upstream.asked;
+
+        for (let index = 0; index < fuzzCases(2000); index++) {
+            const random = caseRandom(0x0b5e, index);
+            const route = pickRoute(random);
+            const replies: unknown[] = [];
+            upstream.respond = (request) => {
+                const healthy = healthyUpstream(request);
+                const reply = healthy && corruptReply(healthy, random);
+                replies.push(reply);
+                return reply;
+            };
+            const error = await complete({
+                messages: [{ role: 'user', content: 'hi' }],
+                provider: route.provider,
+                model: route.id,
+                stream: random() < 0.5,
+                fallback: random() < 0.5,
+            });
+            const unclassified = error && unclassifiedFailure(error);
+            if (unclassified) {
+                findings.add(
+                    unclassified,
+                    index,
+                    `${route.provider}:${route.id}`,
+                    replies,
+                );
+            }
+            clearUnhealthyRoutes();
+        }
+        findings.assertNone();
+        expect(upstream.asked).toBeGreaterThan(askedBefore);
+    }, 600_000);
+});
@@ -38,11 +38,11 @@ import {
    vi,
 } from 'vitest';

-import { PuterServer } from '../../server.js';
-import { setupTestServer } from '../../testUtil.js';
-import { withTestActor } from '../integrationTestUtil.js';
+import { PuterServer } from '../../../../server.js';
+import { setupTestServer } from '../../../../testUtil.js';
+import { withTestActor } from '../../shared/integrationTestUtil.js';
 import { ChatCompletionDriver } from './ChatCompletionDriver.js';
-import { FakeChatProvider } from './providers/FakeChatProvider.js';
+import { FakeChatProvider } from '../../providers/fake/chat/FakeChatProvider.js';
 import type { IChatModel, ICompleteArguments } from './types.js';

 // The puter_path resolver reads the user's filesystem; stub it and assert on
@@ -51,7 +51,7 @@ const { processPuterPathUploadsMock } = vi.hoisted(() => ({
    processPuterPathUploadsMock: vi.fn(async (_messages: unknown) => {}),
 }));

-vi.mock('./providers/openai/fileUpload.js', () => ({
+vi.mock('../../providers/openai/chat/fileUpload.js', () => ({
    processPuterPathUploads: processPuterPathUploadsMock,
    MAX_FILE_SIZE: 5 * 1_000_000,
 }));
@@ -38,14 +38,14 @@ import {
    vi,
 } from 'vitest';

-import { Context } from '../../core/context.js';
-import type { UsageInput } from '../../services/metering/types.js';
-import { PuterServer } from '../../server.js';
-import { setupTestServer } from '../../testUtil.js';
-import { withTestActor } from '../integrationTestUtil.js';
+import { Context } from '../../../../core/context.js';
+import type { UsageInput } from '../../../../services/metering/types.js';
+import { PuterServer } from '../../../../server.js';
+import { setupTestServer } from '../../../../testUtil.js';
+import { withTestActor } from '../../shared/integrationTestUtil.js';
 import { ChatCompletionDriver } from './ChatCompletionDriver.js';
 import { clearUnhealthyRoutes, isRouteUnhealthy } from './utils/providerHealth.js';
-import { FakeChatProvider } from './providers/FakeChatProvider.js';
+import { FakeChatProvider } from '../../providers/fake/chat/FakeChatProvider.js';
 import type { IChatCompleteResult } from './types.js';
 import { SDK_TIMEOUT_MS } from './utils/sdkClient.js';
 import type { AIChatStream } from './utils/Streaming.js';
@@ -44,21 +44,21 @@ import type { Request } from 'express';
 import {
    renderAnthropicError,
    renderOpenAIError,
-} from '../../controllers/puterai/wireErrors.js';
-import { Context } from '../../core/context.js';
-import { HttpError } from '../../core/http/HttpError.js';
-import { PuterServer } from '../../server.js';
-import { setupTestServer } from '../../testUtil.js';
-import { withTestActor } from '../integrationTestUtil.js';
+} from '../../../../controllers/puterai/wireErrors.js';
+import { Context } from '../../../../core/context.js';
+import { HttpError } from '../../../../core/http/HttpError.js';
+import { PuterServer } from '../../../../server.js';
+import { setupTestServer } from '../../../../testUtil.js';
+import { withTestActor } from '../../shared/integrationTestUtil.js';
 import { ChatCompletionDriver } from './ChatCompletionDriver.js';
-import { OPEN_AI_MODELS } from './providers/openai/models.js';
+import { OPEN_AI_MODELS } from '../../providers/openai/chat/models.js';
 import {
    clearUnhealthyRoutes,
    isRouteUnhealthy,
    markRouteUnhealthy,
 } from './utils/providerHealth.js';
-import { InfronProvider } from './providers/infron/InfronProvider.js';
-import { UpstreamDeadlineError } from '../util/upstreamErrors.js';
+import { InfronProvider } from '../../providers/infron/chat/InfronProvider.js';
+import { UpstreamDeadlineError } from '../../../util/upstreamErrors.js';

 // -- OpenAI SDK mock ------------------------------------------------
 // Gemini reaches Google through `new openai.OpenAI()` (default export) and
@@ -44,17 +44,17 @@ import {
    vi,
 } from 'vitest';

-import { Context } from '../../core/context.js';
-import { HttpError } from '../../core/http/HttpError.js';
-import { PuterServer } from '../../server.js';
-import { setupTestServer } from '../../testUtil.js';
-import { withTestActor } from '../integrationTestUtil.js';
+import { Context } from '../../../../core/context.js';
+import { HttpError } from '../../../../core/http/HttpError.js';
+import { PuterServer } from '../../../../server.js';
+import { setupTestServer } from '../../../../testUtil.js';
+import { withTestActor } from '../../shared/integrationTestUtil.js';
 import { COUNT_TOKENS, ChatCompletionDriver } from './ChatCompletionDriver.js';
-import { ClaudeProvider } from './providers/claude/ClaudeProvider.js';
-import { FakeChatProvider } from './providers/FakeChatProvider.js';
-import { OllamaChatProvider } from './providers/ollama/OllamaProvider.js';
-import { XAI_CHAT_MODELS } from './providers/xai/catalog.js';
-import { XAIProvider } from './providers/xai/XAIProvider.js';
+import { ClaudeProvider } from '../../providers/claude/chat/ClaudeProvider.js';
+import { FakeChatProvider } from '../../providers/fake/chat/FakeChatProvider.js';
+import { OllamaChatProvider } from '../../providers/ollama/chat/OllamaProvider.js';
+import { XAI_CHAT_MODELS } from '../../providers/xai/chat/catalog.js';
+import { XAIProvider } from '../../providers/xai/chat/XAIProvider.js';
 import type {
    IChatCompleteResult,
    IChatProvider,
@@ -207,6 +207,50 @@ describe('ChatCompletionDriver model map', () => {
        d.onServerShutdown();
    });

+    it('lists qualified model ids without publishing aliases as models', async () => {
+        const ids = [
+            'kimi/kimi-3',
+            'ollama:plain-model',
+            'ollama:vendor/qualified',
+            'vendor/tagged:free',
+        ];
+        vi.spyOn(OllamaChatProvider.prototype, 'models').mockResolvedValue(
+            ids.map((id) => ({
+                id,
+                aliases: ['legacy-' + id],
+                costs_currency: 'usd-cents',
+                costs: {},
+                max_tokens: 8192,
+            })),
+        );
+        const d = newDriver({ ollama: {} });
+        await d.onServerStart();
+        try {
+            expect(
+                (await d.list()).filter(
+                    (id) => !['abuse', 'costly', 'fake'].includes(id),
+                ),
+            ).toEqual([
+                'ollama:kimi/kimi-3',
+                'ollama:ollama/plain-model',
+                'ollama:vendor/qualified',
+                'ollama:vendor/tagged:free',
+            ]);
+            const models = (await d.models()).filter(
+                (model) => model.provider === 'ollama',
+            );
+            expect(models.map((model) => model.id).sort()).toEqual(
+                [...ids].sort(),
+            );
+            for (const model of models) {
+                expect(model.aliases).toContain(model.puterId);
+                expect(model.aliases).toContain('legacy-' + model.id);
+            }
+        } finally {
+            d.onServerShutdown();
+        }
+    });
+
    it('lists the catalog to a caller that asks while it is still loading', async () => {
        const d = newDriver({ ollama: { enabled: false } });
        d.onServerStart();
@@ -2396,3 +2440,97 @@ describe('ChatCompletionDriver service speed', () => {
        },
    );
 });
+
+describe('ChatCompletionDriver.complete option shapes', () => {
+    const hi = [{ role: 'user', content: 'hi' }];
+
+    it.each(['model', 'provider', 'service_tier'])(
+        'rejects a value that cannot be text as `%s`',
+        async (name) => {
+            const args = { model: 'fake', messages: hi, [name]: { toString: null } };
+            await expect(
+                withTestActor(() => driver.complete(args as never)),
+            ).rejects.toMatchObject({
+                statusCode: 400,
+                legacyCode: 'bad_request',
+                message: `\`${name}\` is not a usable value`,
+            });
+            await expect(
+                withTestActor(() => driver[COUNT_TOKENS](args as never)),
+            ).rejects.toMatchObject({ statusCode: 400 });
+        },
+    );
+
+    it('routes by the model when `provider` is an object, as it did before', async () => {
+        const completeSpy = vi.spyOn(FakeChatProvider.prototype, 'complete');
+        await withTestActor(() =>
+            driver.complete({
+                model: 'fake',
+                messages: hi,
+                provider: { order: ['Anthropic'] },
+            } as never),
+        );
+        expect(completeSpy).toHaveBeenCalledTimes(1);
+    });
+
+    it.each([false, 0, '', null])(
+        'treats `tools: %j` as no tools when counting tokens',
+        async (tools) => {
+            const countTokens = vi.fn(async (_args: object) => 1);
+            const fake = FakeChatProvider.prototype as {
+                countTokens?: typeof countTokens;
+            };
+            fake.countTokens = countTokens;
+            try {
+                await withTestActor(() =>
+                    driver[COUNT_TOKENS]({
+                        model: 'fake',
+                        messages: hi,
+                        tools,
+                    } as never),
+                );
+            } finally {
+                delete fake.countTokens;
+            }
+            expect(countTokens.mock.calls[0]![0]).not.toHaveProperty('tools');
+        },
+    );
+
+    it.each([false, 0, '', null])(
+        'treats `tools: %j` as no tools',
+        async (tools) => {
+            const completeSpy = vi.spyOn(FakeChatProvider.prototype, 'complete');
+            await withTestActor(() =>
+                driver.complete({ model: 'fake', messages: hi, tools } as never),
+            );
+            expect(completeSpy.mock.calls[0]![0]).not.toHaveProperty('tools');
+        },
+    );
+
+    it('reports zero tokens in the cost event for a usage count that is not a number', async () => {
+        vi.spyOn(FakeChatProvider.prototype, 'models').mockResolvedValueOnce([
+            {
+                id: 'priced',
+                aliases: [],
+                costs_currency: 'usd-cents',
+                costs: { input_tokens: 1000, output_tokens: 2000 },
+                max_tokens: 8192,
+            },
+        ]);
+        const d = await makeDriver();
+        const costEvents = captureEvent('ai.prompt.cost-calculated');
+        vi.spyOn(FakeChatProvider.prototype, 'complete').mockResolvedValueOnce({
+            message: { role: 'assistant', content: [{ type: 'text', text: 'ok' }] },
+            usage: { input_tokens: 10, output_tokens: { toString: null } },
+            finish_reason: 'stop',
+        } as never);
+
+        await withTestActor(() => d.complete({ model: 'priced', messages: hi }));
+
+        expect(costEvents[0]![1]).toMatchObject({
+            input_tokens: 10,
+            output_tokens: 0,
+            total_ucents: 10 * 1000,
+        });
+    });
+});
@@ -19,56 +19,56 @@

 import crypto from 'node:crypto';
 import { PassThrough } from 'node:stream';
-import { EventMap } from '../../clients/event/types.js';
-import type { Actor } from '../../core/actor.js';
-import { Context } from '../../core/context.js';
-import { HttpError, isHttpError } from '../../core/http/HttpError.js';
-import { insufficientCreditsError } from '../../services/metering/enforcement.js';
-import { isFreeSubscription } from '../../services/metering/consts.js';
-import type { MeteringService } from '../../services/metering/MeteringService.js';
-import type { CreditHold } from '../../services/metering/types.js';
-import { NO_CREDIT_HOLD } from '../../services/metering/types.js';
-import type { DriverStreamResult } from '../meta.js';
-import { PuterDriver } from '../types.js';
+import { EventMap } from '../../../../clients/event/types.js';
+import type { Actor } from '../../../../core/actor.js';
+import { Context } from '../../../../core/context.js';
+import { HttpError, isHttpError } from '../../../../core/http/HttpError.js';
+import { insufficientCreditsError } from '../../../../services/metering/enforcement.js';
+import { isFreeSubscription } from '../../../../services/metering/consts.js';
+import type { MeteringService } from '../../../../services/metering/MeteringService.js';
+import type { CreditHold } from '../../../../services/metering/types.js';
+import { NO_CREDIT_HOLD } from '../../../../services/metering/types.js';
+import type { DriverStreamResult } from '../../../meta.js';
+import { PuterDriver } from '../../../types.js';
 import {
    type AiMeteringService,
    factoredTotal,
    requestAiCosts,
    withAiCostFactor,
-} from '../util/aiCostFactor.js';
-import { AI_CONCURRENT, AI_RATE_LIMIT } from '../util/aiLimits.js';
+} from '../../../../util/aiCostFactor.js';
+import { AI_CONCURRENT, AI_RATE_LIMIT } from '../../shared/aiLimits.js';
 import {
    isCreditExhaustion as isUpstreamCreditExhaustion,
    isUpstreamConnectionError,
    isUpstreamTimeoutError,
    sanitizeUpstreamMessage,
    UpstreamDeadlineError,
-} from '../util/upstreamErrors.js';
-import { AlibabaProvider } from './providers/alibaba/AlibabaProvider.js';
-import { AzureChatProvider } from './providers/azure/AzureChatProvider.js';
-import { AzureResponsesProvider } from './providers/azure/AzureResponsesProvider.js';
-import { BytePlusProvider } from './providers/byteplus/BytePlusProvider.js';
-import { ClaudeProvider } from './providers/claude/ClaudeProvider.js';
-import { DeepSeekProvider } from './providers/deepseek/DeepSeekProvider.js';
-import { FakeChatProvider } from './providers/FakeChatProvider.js';
-import { GeminiChatProvider } from './providers/gemini/GeminiChatProvider.js';
-import { GroqAIProvider } from './providers/groq/GroqAIProvider.js';
-import { HoonifyProvider } from './providers/hoonify/HoonifyProvider.js';
-import { InfronProvider } from './providers/infron/InfronProvider.js';
-import { MetaProvider } from './providers/meta/MetaProvider.js';
-import { MiniMaxProvider } from './providers/minimax/MiniMaxProvider.js';
-import { MistralAIProvider } from './providers/mistral/MistralAiProvider.js';
-import { MoonshotProvider } from './providers/moonshot/MoonshotProvider.js';
-import { NeuralwattProvider } from './providers/neuralwatt/NeuralwattProvider.js';
-import { OllamaChatProvider } from './providers/ollama/OllamaProvider.js';
-import type { ChatProviderConfig } from './providers/OpenAICompatProvider.js';
-import { processPuterPathUploads } from './providers/openai/fileUpload.js';
-import { OpenAiChatProvider } from './providers/openai/OpenAiChatCompletionsProvider.js';
-import { OpenAiResponsesChatProvider } from './providers/openai/OpenAiChatResponsesProvider.js';
-import { OpenRouterProvider } from './providers/openrouter/OpenRouterProvider.js';
-import { TogetherAIProvider } from './providers/together/TogetherAIProvider.js';
-import { XAIProvider } from './providers/xai/XAIProvider.js';
-import { ZAIProvider } from './providers/zai/ZAIProvider.js';
+} from '../../../util/upstreamErrors.js';
+import { AlibabaProvider } from '../../providers/alibaba/chat/AlibabaProvider.js';
+import { AzureChatProvider } from '../../providers/azure/chat/AzureChatProvider.js';
+import { AzureResponsesProvider } from '../../providers/azure/chat/AzureResponsesProvider.js';
+import { BytePlusProvider } from '../../providers/byteplus/chat/BytePlusProvider.js';
+import { ClaudeProvider } from '../../providers/claude/chat/ClaudeProvider.js';
+import { DeepSeekProvider } from '../../providers/deepseek/chat/DeepSeekProvider.js';
+import { FakeChatProvider } from '../../providers/fake/chat/FakeChatProvider.js';
+import { GeminiChatProvider } from '../../providers/gemini/chat/GeminiChatProvider.js';
+import { GroqAIProvider } from '../../providers/groq/chat/GroqAIProvider.js';
+import { HoonifyProvider } from '../../providers/hoonify/chat/HoonifyProvider.js';
+import { InfronProvider } from '../../providers/infron/chat/InfronProvider.js';
+import { MetaProvider } from '../../providers/meta/chat/MetaProvider.js';
+import { MiniMaxProvider } from '../../providers/minimax/chat/MiniMaxProvider.js';
+import { MistralAIProvider } from '../../providers/mistral/chat/MistralAiProvider.js';
+import { MoonshotProvider } from '../../providers/moonshot/chat/MoonshotProvider.js';
+import { NeuralwattProvider } from '../../providers/neuralwatt/chat/NeuralwattProvider.js';
+import { OllamaChatProvider } from '../../providers/ollama/chat/OllamaProvider.js';
+import type { ChatProviderConfig } from '../../protocols/openaiCompatible/OpenAICompatProvider.js';
+import { processPuterPathUploads } from '../../providers/openai/chat/fileUpload.js';
+import { OpenAiChatProvider } from '../../providers/openai/chat/OpenAiChatCompletionsProvider.js';
+import { OpenAiResponsesChatProvider } from '../../providers/openai/chat/OpenAiChatResponsesProvider.js';
+import { OpenRouterProvider } from '../../providers/openrouter/chat/OpenRouterProvider.js';
+import { TogetherAIProvider } from '../../providers/together/chat/TogetherAIProvider.js';
+import { XAIProvider } from '../../providers/xai/chat/XAIProvider.js';
+import { ZAIProvider } from '../../providers/zai/chat/ZAIProvider.js';
 import type {
    IChatCompleteResult,
    IChatMessageResult,
@@ -103,9 +103,13 @@ import {
    normalizeResultToOpenAI,
    shouldPresentAsOpenAI,
 } from './utils/normalizeToOpenAI.js';
-import { isToolChoice, toolChoiceFromWire } from './utils/openaiParams.js';
 import {
-    catalogPrice,
+    isToolChoice,
+    toolChoiceFromWire,
+} from '../../protocols/openaiCompatible/openaiParams.js';
+import { assertScalarOptions } from '../../shared/scalarOptions.js';
+import { catalogPrice } from '../../shared/catalogPrice.js';
+import {
    costKeys,
    costsForRequest,
    isFreeModel,
@@ -121,7 +125,7 @@ import {
    TRANSIENT_UNHEALTHY_TTL_SEC,
    UNHEALTHY_TTL_SEC,
 } from './utils/providerHealth.js';
-import { CATALOG_TTL_MS } from './utils/cachedRemoteCatalog.js';
+import { CATALOG_TTL_MS } from '../../shared/cachedRemoteCatalog.js';
 import { SDK_TIMEOUT_MS } from './utils/sdkClient.js';
 import { fromFinishReason } from './utils/stopReason.js';
 import { AIChatStream } from './utils/Streaming.js';
@@ -178,6 +182,9 @@ export const COUNT_TOKENS = Symbol('aiChat.countTokens');
 */
 const HOLD_RENEW_INTERVAL_MS = 5 * 60 * 1000;

+// Options that reach error messages and request headers as text.
+const SCALAR_OPTIONS = ['model', 'provider', 'service_tier'];
+
 /**
 * A moderation refusal is a completion that was produced and charged, then
 * withheld — not a route failure. Retrying it on a fallback provider would bill
@@ -585,7 +592,7 @@ export class ChatCompletionDriver extends PuterDriver {
    readonly driverName = 'ai-chat';
    readonly isDefault = true;

-    // Shared AI policy — see `drivers/util/aiLimits.ts` for the tier table.
+    // Shared AI policy — see `drivers/ai/shared/aiLimits.ts` for the tier table.
    readonly rateLimit = AI_RATE_LIMIT;
    readonly concurrent = AI_CONCURRENT;

@@ -690,6 +697,7 @@ export class ChatCompletionDriver extends PuterDriver {
                { legacyCode: 'invalid_service_speed' },
            );
        }
+        assertScalarOptions(args, SCALAR_OPTIONS);
        const abortSignal = Context.get('abortSignal');
        abortSignal?.throwIfAborted();
        const actor = Context.get('actor');
@@ -744,6 +752,9 @@ export class ChatCompletionDriver extends PuterDriver {
        args.messages = normalizeMediaParts(normalize_messages(args.messages));
        if (args.tools) {
            normalize_tools_object(args.tools);
+        } else {
+            // `tools: enabled && list` leaves `false` here; it means no tools.
+            delete args.tools;
        }
        // A raw `/drivers/call` caller still sends the OpenAI wire form
        // ('auto'/'none'/'required', {type:'function', function:{name}}) —
@@ -1297,8 +1308,17 @@ export class ChatCompletionDriver extends PuterDriver {
            }, HOLD_RENEW_INTERVAL_MS);
            renewHold.unref?.();

-            // Fire-and-forget — the stream writes happen async while the
-            // response is being piped to the client.
+            // The upstream runs on after a caller hangs up, so the request
+            // keeps its concurrency slot until the upstream is done, not until
+            // the cleanup after it.
+            let upstreamDone!: () => void;
+            const upstreamRunning = new Promise<void>((resolve) => {
+                upstreamDone = resolve;
+            });
+            Context.get('keepConcurrencySlot')?.(upstreamRunning);
+
+            // The stream writes happen async while the response is piped to
+            // the client.
            (async () => {
                try {
                    await init({ chatStream });
@@ -1323,6 +1343,7 @@ export class ChatCompletionDriver extends PuterDriver {
                    }
                    passthrough.end();
                } finally {
+                    upstreamDone();
                    clearInterval(renewHold);
                    clearTimeout(detachedLimit);
                    abortSignal?.removeEventListener('abort', detachStream);
@@ -1490,6 +1511,7 @@ export class ChatCompletionDriver extends PuterDriver {
    async [COUNT_TOKENS](
        args: ICompleteArguments,
    ): Promise<{ input_tokens: number }> {
+        assertScalarOptions(args, SCALAR_OPTIONS);
        let intendedProvider = args.provider || '';
        if (!args.model && !intendedProvider) {
            intendedProvider = 'azure-openai';
@@ -1515,6 +1537,9 @@ export class ChatCompletionDriver extends PuterDriver {
        args.messages = normalizeMediaParts(normalize_messages(args.messages));
        if (args.tools) {
            normalize_tools_object(args.tools);
+        } else {
+            // `tools: enabled && list` leaves `false` here; it means no tools.
+            delete args.tools;
        }

        const provider = this.#providers[model.provider!];
@@ -1848,20 +1873,20 @@ export class ChatCompletionDriver extends PuterDriver {

        const cost = usageCost(usage, model, usageCosts, pricingTime);
        const { inputKey, outputKey } = costKeys(model);
+        // The vendor reports usage, and pricing skips a count that is not a
+        // number; so does this.
+        const count = (value: unknown) =>
+            typeof value === 'number' && Number.isFinite(value) ? value : 0;
        const inputTokens = cost
-            ? Number(
-                  usage[inputKey] ??
-                      usage.prompt_tokens ??
-                      usage.input_tokens ??
-                      0,
+            ? count(
+                  usage[inputKey] ?? usage.prompt_tokens ?? usage.input_tokens,
              )
            : 0;
        const outputTokens = cost
-            ? Number(
+            ? count(
                  usage[outputKey] ??
                      usage.completion_tokens ??
-                      usage.output_tokens ??
-                      0,
+                      usage.output_tokens,
              )
            : 0;
        const inputMicroCents = cost?.inputMicroCents ?? 0;
@@ -2061,14 +2086,39 @@ export class ChatCompletionDriver extends PuterDriver {
                // instance, so they are read and never written: normalizing
                // the id or appending puterId in place would accumulate across
                // instantiations. The bucket gets its own copy instead.
-                const aliases =
-                    entry.puterId &&
-                    !(entry.aliases ?? []).includes(entry.puterId)
-                        ? [...(entry.aliases ?? []), entry.puterId]
-                        : entry.aliases;
+                // Keep upstream ids for provider calls and publish a qualified
+                // routing name even when a catalog omits its vendor namespace.
+                const publicId = entry.puterId || entry.id;
+                // A colon after the first slash belongs to the model name, as
+                // in `vendor/model:free`, and names no service.
+                const serviceSeparator = publicId
+                    .split('/', 1)[0]!
+                    .indexOf(':');
+                const service =
+                    serviceSeparator < 0
+                        ? providerName
+                        : publicId.slice(0, serviceSeparator);
+                const modelPath =
+                    serviceSeparator < 0
+                        ? publicId
+                        : publicId.slice(serviceSeparator + 1);
+                const puterId =
+                    providerName === 'fake-chat'
+                        ? entry.puterId
+                        : `${service}:${modelPath.includes('/') ? modelPath : `${service}/${modelPath}`}`;
+                const aliases = puterId
+                    ? [
+                          ...new Set([
+                              ...(entry.aliases ?? []),
+                              ...(entry.puterId ? [entry.puterId] : []),
+                              puterId,
+                          ]),
+                      ]
+                    : entry.aliases;
                const model = {
                    ...entry,
                    id: normalizeModelKey(entry.id),
+                    ...(puterId ? { puterId } : {}),
                };
                // Assigned only when the entry has names to carry: models()
                // is serialized to the API, and an entry that declared no
@@ -47,29 +47,29 @@
 * on that test.)
 */

-import { readdirSync } from 'node:fs';
+import { existsSync, readdirSync } from 'node:fs';
 import { dirname, join } from 'node:path';
 import { fileURLToPath, pathToFileURL } from 'node:url';

 import { describe, expect, it } from 'vitest';

-import type { IChatModel } from '../types.js';
-import { ALIBABA_MODELS } from './alibaba/models.js';
-import { AZURE_MODELS } from './azure/models.js';
-import { BYTEPLUS_MODELS } from './byteplus/models.js';
-import { CLAUDE_MODELS } from './claude/models.js';
-import { DEEPSEEK_MODELS } from './deepseek/models.js';
-import { GEMINI_MODELS } from './gemini/models.js';
-import { GROQ_MODELS } from './groq/models.js';
-import { HOONIFY_MODELS } from './hoonify/models.js';
-import { META_MODELS } from './meta/models.js';
-import { MINIMAX_MODELS } from './minimax/models.js';
-import { MISTRAL_MODELS } from './mistral/models.js';
-import { MOONSHOT_MODELS } from './moonshot/models.js';
-import { OPEN_AI_MODELS } from './openai/models.js';
-import { OPEN_ROUTER_MODEL_OVERRIDES } from './openrouter/modelOverrides.js';
-import { XAI_MODELS } from './xai/models.js';
-import { ZAI_MODELS } from './zai/models.js';
+import type { IChatModel } from './types.js';
+import { ALIBABA_MODELS } from '../../providers/alibaba/chat/models.js';
+import { AZURE_MODELS } from '../../providers/azure/chat/models.js';
+import { BYTEPLUS_MODELS } from '../../providers/byteplus/chat/models.js';
+import { CLAUDE_MODELS } from '../../providers/claude/chat/models.js';
+import { DEEPSEEK_MODELS } from '../../providers/deepseek/chat/models.js';
+import { GEMINI_MODELS } from '../../providers/gemini/chat/models.js';
+import { GROQ_MODELS } from '../../providers/groq/chat/models.js';
+import { HOONIFY_MODELS } from '../../providers/hoonify/chat/models.js';
+import { META_MODELS } from '../../providers/meta/chat/models.js';
+import { MINIMAX_MODELS } from '../../providers/minimax/chat/models.js';
+import { MISTRAL_MODELS } from '../../providers/mistral/chat/models.js';
+import { MOONSHOT_MODELS } from '../../providers/moonshot/chat/models.js';
+import { OPEN_AI_MODELS } from '../../providers/openai/chat/models.js';
+import { OPEN_ROUTER_MODEL_OVERRIDES } from '../../providers/openrouter/chat/modelOverrides.js';
+import { XAI_MODELS } from '../../providers/xai/chat/models.js';
+import { ZAI_MODELS } from '../../providers/zai/chat/models.js';

 // Providers whose catalog is fetched at runtime (OpenRouter, Ollama, Together,
 // Infron, Neuralwatt) have nothing static to check and are absent by design;
@@ -201,28 +201,34 @@ describe('CATALOGS', () => {
    // silent gap these tests are about, so the list is checked against what
    // is actually on disk.
    //
-    // The scan covers every non-test source file under providers/*/ with
-    // "model" in its name — models.ts, but also siblings like openrouter's
-    // modelOverrides.ts. Importing every provider file regardless of name
-    // would drag in SDK modules for a filename sweep, so that naming
-    // convention is the one assumption left unenforced here: a catalog in a
-    // file named without "model" would escape this net.
-    it('lists every static catalog under providers/', async () => {
-        const here = dirname(fileURLToPath(import.meta.url));
+    // The scan covers every non-test source file under providers/*/chat/
+    // with "model" in its name — models.ts, but also siblings like
+    // openrouter's modelOverrides.ts. Importing every provider file
+    // regardless of name would drag in SDK modules for a filename sweep, so
+    // that naming convention is the one assumption left unenforced here: a
+    // catalog in a file named without "model" would escape this net.
+    it('lists every static catalog under providers/*/chat/', async () => {
+        const providersDir = join(
+            dirname(fileURLToPath(import.meta.url)),
+            '../../providers',
+        );
        const registered = new Map(CATALOGS);
+        const found = new Set<string>();
        const problems: string[] = [];

-        for (const dir of readdirSync(here, { withFileTypes: true })) {
+        for (const dir of readdirSync(providersDir, { withFileTypes: true })) {
            if (!dir.isDirectory()) continue;
+            const chatDir = join(providersDir, dir.name, 'chat');
+            if (!existsSync(chatDir)) continue;

-            for (const file of readdirSync(join(here, dir.name))) {
+            for (const file of readdirSync(chatDir)) {
                if (!/model/i.test(file)) continue;
                if (!file.endsWith('.ts') || file.endsWith('.test.ts')) {
                    continue;
                }

                const module = await import(
-                    pathToFileURL(join(here, dir.name, file)).href
+                    pathToFileURL(join(chatDir, file)).href
                );
                for (const [name, value] of Object.entries(module)) {
                    // A catalog is a non-empty array of entries carrying an
@@ -233,16 +239,19 @@ describe('CATALOGS', () => {
                        value.length > 0 &&
                        typeof value[0]?.id === 'string';
                    if (!isCatalog) continue;
+                    found.add(name);

                    if (!registered.has(name)) {
-                        problems.push(`${dir.name}/${file} exports ${name}`);
+                        problems.push(
+                            `${dir.name}/chat/${file} exports ${name}`,
+                        );
                    } else if (registered.get(name) !== value) {
                        // The row's label names this export but its value is
                        // a different array — the wrong catalog would be the
                        // one getting checked.
                        problems.push(
                            `the CATALOGS row named ${name} does not hold ` +
-                                `the ${name} that ${dir.name}/${file} exports`,
+                                `the ${name} that ${dir.name}/chat/${file} exports`,
                        );
                    }
                }
@@ -250,5 +259,7 @@ describe('CATALOGS', () => {
        }

        expect(problems, problems.join('\n')).toEqual([]);
+        // A scan pointed at the wrong directory finds nothing and passes.
+        expect([...found].sort()).toEqual([...registered.keys()].sort());
    });
 });
@@ -39,41 +39,41 @@

 import { describe, expect, it, vi } from 'vitest';

-import type { MeteringService } from '../../../services/metering/MeteringService.js';
-import { withTestActor } from '../../integrationTestUtil.js';
+import type { MeteringService } from '../../../../services/metering/MeteringService.js';
+import { withTestActor } from '../../shared/integrationTestUtil.js';
 import type {
    IChatMessageResult,
    IChatModel,
    IChatProvider,
-} from '../types.js';
+} from './types.js';
 import {
    needsOpenAICoercion,
    normalizeResultToOpenAI,
-} from '../utils/normalizeToOpenAI.js';
+} from './utils/normalizeToOpenAI.js';

-import { AlibabaProvider } from './alibaba/AlibabaProvider.js';
-import { AzureChatProvider } from './azure/AzureChatProvider.js';
-import { AzureResponsesProvider } from './azure/AzureResponsesProvider.js';
-import { BytePlusProvider } from './byteplus/BytePlusProvider.js';
-import { ClaudeProvider } from './claude/ClaudeProvider.js';
-import { DeepSeekProvider } from './deepseek/DeepSeekProvider.js';
-import { FakeChatProvider } from './FakeChatProvider.js';
-import { GeminiChatProvider } from './gemini/GeminiChatProvider.js';
-import { GroqAIProvider } from './groq/GroqAIProvider.js';
-import { HoonifyProvider } from './hoonify/HoonifyProvider.js';
-import { InfronProvider } from './infron/InfronProvider.js';
-import { MetaProvider } from './meta/MetaProvider.js';
-import { MiniMaxProvider } from './minimax/MiniMaxProvider.js';
-import { MistralAIProvider } from './mistral/MistralAiProvider.js';
-import { MoonshotProvider } from './moonshot/MoonshotProvider.js';
-import { NeuralwattProvider } from './neuralwatt/NeuralwattProvider.js';
-import { OllamaChatProvider } from './ollama/OllamaProvider.js';
-import { OpenAiChatProvider } from './openai/OpenAiChatCompletionsProvider.js';
-import { OpenAiResponsesChatProvider } from './openai/OpenAiChatResponsesProvider.js';
-import { OpenRouterProvider } from './openrouter/OpenRouterProvider.js';
-import { TogetherAIProvider } from './together/TogetherAIProvider.js';
-import { XAIProvider } from './xai/XAIProvider.js';
-import { ZAIProvider } from './zai/ZAIProvider.js';
+import { AlibabaProvider } from '../../providers/alibaba/chat/AlibabaProvider.js';
+import { AzureChatProvider } from '../../providers/azure/chat/AzureChatProvider.js';
+import { AzureResponsesProvider } from '../../providers/azure/chat/AzureResponsesProvider.js';
+import { BytePlusProvider } from '../../providers/byteplus/chat/BytePlusProvider.js';
+import { ClaudeProvider } from '../../providers/claude/chat/ClaudeProvider.js';
+import { DeepSeekProvider } from '../../providers/deepseek/chat/DeepSeekProvider.js';
+import { FakeChatProvider } from '../../providers/fake/chat/FakeChatProvider.js';
+import { GeminiChatProvider } from '../../providers/gemini/chat/GeminiChatProvider.js';
+import { GroqAIProvider } from '../../providers/groq/chat/GroqAIProvider.js';
+import { HoonifyProvider } from '../../providers/hoonify/chat/HoonifyProvider.js';
+import { InfronProvider } from '../../providers/infron/chat/InfronProvider.js';
+import { MetaProvider } from '../../providers/meta/chat/MetaProvider.js';
+import { MiniMaxProvider } from '../../providers/minimax/chat/MiniMaxProvider.js';
+import { MistralAIProvider } from '../../providers/mistral/chat/MistralAiProvider.js';
+import { MoonshotProvider } from '../../providers/moonshot/chat/MoonshotProvider.js';
+import { NeuralwattProvider } from '../../providers/neuralwatt/chat/NeuralwattProvider.js';
+import { OllamaChatProvider } from '../../providers/ollama/chat/OllamaProvider.js';
+import { OpenAiChatProvider } from '../../providers/openai/chat/OpenAiChatCompletionsProvider.js';
+import { OpenAiResponsesChatProvider } from '../../providers/openai/chat/OpenAiChatResponsesProvider.js';
+import { OpenRouterProvider } from '../../providers/openrouter/chat/OpenRouterProvider.js';
+import { TogetherAIProvider } from '../../providers/together/chat/TogetherAIProvider.js';
+import { XAIProvider } from '../../providers/xai/chat/XAIProvider.js';
+import { ZAIProvider } from '../../providers/zai/chat/ZAIProvider.js';

 // ── Upstream SDK mocks ──────────────────────────────────────────────
 // One create-mock per wire dialect; every provider speaking that dialect
@@ -18,7 +18,6 @@
 */

 import { HttpError } from '@heyputer/backend/src/core/http';
-import { claudeToolPolicy } from '../providers/claude/anthropicPolicy.js';

 // Anthropic server/typed tool ids (`web_search_20250305`,
 // `advisor_20260301`, …) and the OpenAI Responses `web_search` tool pass
@@ -124,7 +123,7 @@ export const normalize_tools_object = (tools) => {
        };

        // Claude-only extras, carried on the `{type:'function', ...}`
-        // wrapper so `make_claude_tools` can forward them; a top-level
+        // wrapper so `claudeToolPolicy` can forward them; a top-level
        // `strict` (Claude's own convention) moves onto `function.strict`.
        const extras = {};
        if (tool.cache_control !== undefined)
@@ -169,76 +168,3 @@ export const normalize_tools_object = (tools) => {
    }
    return tools;
 };
-
-// OpenAI's own tool conventions pass through untouched.
-const OPENAI_NATIVE_WEB_SEARCH = new Set(['web_search', 'web_search_preview']);
-const ANTHROPIC_WEB_SEARCH_TYPE = (type) =>
-    type === 'web_search_20250305' || /^web_search_2026/.test(type ?? '');
-
-/**
- * Converts a normalized tools object to the format expected by an OpenAI-family
- * dialect: function tools drop their Anthropic-only extras (`cache_control`,
- * `defer_loading`, `eager_input_streaming`, `input_examples`); every other
- * Anthropic typed/server tool (`web_search_20250305`, `advisor_20260301`,
- * `bash_*`, …) is dropped, since none of it is a wire shape OpenAI recognizes.
- * The one exception: an Anthropic web-search tool is translated to OpenAI's own
- * `web_search` tool for the Responses dialect, which supports it natively.
- *
- * @param {any} tools
- * @param {{ dialect?: 'chat' | 'responses' }} [opts]
- * @returns {any[] | undefined}
- */
-export const make_openai_tools = (tools, opts = {}) => {
-    if (!tools) return tools;
-    const dialect = opts.dialect ?? 'chat';
-    const out = [];
-    for (const tool of tools) {
-        if (!tool || typeof tool !== 'object') continue;
-        if (tool.type === 'function') {
-            const stripped = { ...tool };
-            delete stripped.cache_control;
-            delete stripped.defer_loading;
-            delete stripped.eager_input_streaming;
-            delete stripped.input_examples;
-            out.push(stripped);
-            continue;
-        }
-        if (OPENAI_NATIVE_WEB_SEARCH.has(tool.type)) {
-            out.push(tool);
-            continue;
-        }
-        if (dialect === 'responses' && ANTHROPIC_WEB_SEARCH_TYPE(tool.type)) {
-            out.push({
-                type: 'web_search',
-                ...(tool.allowed_domains
-                    ? { filters: { allowed_domains: tool.allowed_domains } }
-                    : {}),
-                ...(tool.user_location
-                    ? { user_location: tool.user_location }
-                    : {}),
-            });
-            continue;
-        }
-        // Any other Anthropic typed/server tool (web_search_* on the chat
-        // dialect, advisor, bash, memory, web_fetch, tool_search, …) has no
-        // OpenAI equivalent — drop it rather than forward an unrecognized type.
-    }
-    return out;
-};
-
-/**
- * This function will convert a normalized tools object to the format expected
- * by Claude: function tools become `{name, description, input_schema, ...}`,
- * and typed/server tools (`web_search`, `web_search_20250305`,
- * `advisor_20260301`, …) go through the allowlist/clamp/reject policy in
- * `claudeToolPolicy` instead of being treated as malformed function
- * definitions.
- *
- * @param {any} tools
- * @param {{ models?: import('../types.js').IChatModel[] }} [opts]
- * @returns {any[] | undefined}
- */
-export const make_claude_tools = (tools, opts = {}) => {
-    if (!tools) return undefined;
-    return claudeToolPolicy(tools, opts.models ?? []).tools;
-};
@@ -20,13 +20,11 @@
 import { describe, expect, it } from 'vitest';
 // @ts-expect-error — sibling JS module without an adjacent .d.ts
 import {
-    make_claude_tools,
-    make_openai_tools,
    normalize_json_schema,
    normalize_tools_object,
 } from './FunctionCalling.js';

-// All four exports are pure data transforms — these tests just feed
+// Both exports are pure data transforms — these tests just feed
 // inputs and check the normalized output, no service mocks involved.

 // ── normalize_json_schema ───────────────────────────────────────────
@@ -242,159 +240,3 @@ describe('normalize_tools_object', () => {
        expect(out[0].function.parameters.properties).toBe('x');
    });
 });
-
-// ── make_openai_tools ───────────────────────────────────────────────
-
-describe('make_openai_tools', () => {
-    it('keeps a function tool, dropping its Anthropic-only extras', () => {
-        const tools = [
-            {
-                type: 'function',
-                function: { name: 'lookup', parameters: { type: 'object' } },
-                cache_control: { type: 'ephemeral' },
-                defer_loading: true,
-                eager_input_streaming: true,
-                input_examples: [{}],
-            },
-        ];
-        expect(make_openai_tools(tools)).toEqual([
-            {
-                type: 'function',
-                function: { name: 'lookup', parameters: { type: 'object' } },
-            },
-        ]);
-    });
-
-    it('drops an Anthropic web_search tool on the chat dialect', () => {
-        const tools = [{ type: 'web_search_20250305', name: 'web_search' }];
-        expect(make_openai_tools(tools, { dialect: 'chat' })).toEqual([]);
-    });
-
-    it('converts an Anthropic web_search tool to OpenAI web_search on the responses dialect', () => {
-        const tools = [
-            {
-                type: 'web_search_20250305',
-                name: 'web_search',
-                max_uses: 5,
-                allowed_domains: ['example.com'],
-                user_location: { type: 'approximate' },
-            },
-        ];
-        expect(
-            make_openai_tools(tools, { dialect: 'responses' }),
-        ).toEqual([
-            {
-                type: 'web_search',
-                filters: { allowed_domains: ['example.com'] },
-                user_location: { type: 'approximate' },
-            },
-        ]);
-    });
-
-    it('keeps OpenAI web_search/web_search_preview tools verbatim', () => {
-        const tools = [{ type: 'web_search' }, { type: 'web_search_preview' }];
-        expect(make_openai_tools(tools, { dialect: 'responses' })).toEqual(
-            tools,
-        );
-    });
-
-    it('drops other Anthropic typed/server tools (advisor, bash, …)', () => {
-        const tools = [
-            { type: 'advisor_20260301', model: 'claude-haiku-4-5' },
-            { type: 'bash_20250124', name: 'bash' },
-        ];
-        expect(make_openai_tools(tools, { dialect: 'responses' })).toEqual([]);
-    });
-
-    it('passes a falsy tools value through unchanged', () => {
-        expect(make_openai_tools(undefined)).toBeUndefined();
-        expect(make_openai_tools(null)).toBeNull();
-    });
-});
-
-describe('make_claude_tools with a malformed tool', () => {
-    it.each([
-        ['type:function with no function property', { type: 'function' }],
-        ['a null entry', null],
-        ['a bare string entry', 'lookup'],
-    ])('answers 400 for a tool with %s', (_label, tool) => {
-        // A TypeError here has no status, so the retry loop reads it as a
-        // provider failure and marks the route unhealthy for every caller.
-        expect(() => make_claude_tools([tool] as never)).toThrowError(
-            expect.objectContaining({
-                statusCode: 400,
-                legacyCode: 'bad_request',
-            }),
-        );
-    });
-
-    it('answers 400 for an unresolvable advisor model', () => {
-        expect(() =>
-            make_claude_tools([
-                { type: 'advisor_20260301', model: 'not-a-real-model' },
-            ] as never),
-        ).toThrowError(
-            expect.objectContaining({ statusCode: 400, legacyCode: 'bad_request' }),
-        );
-    });
-
-    it('answers 400 for an mcp_toolset tool', () => {
-        expect(() =>
-            make_claude_tools([{ type: 'mcp_toolset' }] as never),
-        ).toThrowError(
-            expect.objectContaining({ statusCode: 400, legacyCode: 'bad_request' }),
-        );
-    });
-});
-
-// ── make_claude_tools ───────────────────────────────────────────────
-
-describe('make_claude_tools', () => {
-    it('returns undefined when tools is undefined', () => {
-        expect(make_claude_tools(undefined)).toBeUndefined();
-    });
-
-    it('returns [] when tools is an empty array', () => {
-        expect(make_claude_tools([])).toEqual([]);
-    });
-
-    it('flattens the OpenAI {function:{...}} wrapper into Claude shape', () => {
-        const out = make_claude_tools([
-            {
-                type: 'function',
-                function: {
-                    name: 'lookup',
-                    description: 'Search the docs',
-                    parameters: {
-                        type: 'object',
-                        properties: { q: { type: 'string' } },
-                    },
-                },
-            },
-        ]);
-        expect(out).toEqual([
-            {
-                name: 'lookup',
-                description: 'Search the docs',
-                input_schema: {
-                    type: 'object',
-                    properties: { q: { type: 'string' } },
-                },
-            },
-        ]);
-    });
-
-    it('maps the OpenAI {type:"web_search"} form to web_search_20250305', () => {
-        const out = make_claude_tools([{ type: 'web_search' }] as never);
-        expect(out).toEqual([
-            { type: 'web_search_20250305', name: 'web_search', max_uses: 10 },
-        ]);
-    });
-
-    it('passes a typed server tool through by its own regexp-matched type', () => {
-        const out = make_claude_tools([
-            { type: 'bash_20250124', name: 'bash' },
-        ] as never);
-        expect(out).toEqual([{ type: 'bash_20250124', name: 'bash' }]);
-    });
-});
@@ -18,6 +18,7 @@
 */

 import { HttpError } from '@heyputer/backend/src/core/http';
+import { convertsToText } from '../../../shared/scalarOptions.js';

 /**
 * Normalizes a single message into a standardized format with role and content
@@ -154,6 +155,39 @@ export const normalize_single_message = (message, params = {}) => {
        }
    }

+    // Providers write a block's type and a tool call's id into text, which
+    // throws on a value that can't be converted.
+    for (const block of message.content) {
+        if (!convertsToText(block.type)) {
+            throw new HttpError(
+                400,
+                'each message content item type must be a string',
+                { legacyCode: 'bad_request' },
+            );
+        }
+        const callId =
+            block.type === 'tool_use'
+                ? block.id
+                : block.type === 'tool_result'
+                  ? block.tool_use_id
+                  : undefined;
+        if (!convertsToText(callId)) {
+            throw new HttpError(400, 'a tool call id must be a string', {
+                legacyCode: 'bad_request',
+            });
+        }
+        // A tool's result is text or content blocks; any other object is sent
+        // as its JSON, like the content of a `tool` message above.
+        if (
+            block.type === 'tool_result' &&
+            block.content &&
+            typeof block.content === 'object' &&
+            !Array.isArray(block.content)
+        ) {
+            block.content = JSON.stringify(block.content);
+        }
+    }
+
    // Remove "text" properties from content blocks with type=tool_use
    for (let i = 0; i < message.content.length; i++) {
        if (message.content[i].type !== 'tool_use') {
@@ -449,3 +449,106 @@ describe('extract_text', () => {
        ).toThrow(expect.objectContaining({ statusCode: 400 }));
    });
 });
+
+describe('normalize_single_message block validation', () => {
+    it.each([
+        [
+            'an assistant tool call',
+            {
+                role: 'assistant',
+                tool_calls: [
+                    { id: { toString: null }, function: { name: 'f', arguments: '{}' } },
+                ],
+            },
+        ],
+        [
+            'a tool_use block',
+            {
+                role: 'assistant',
+                content: [
+                    {
+                        type: 'tool_use',
+                        id: { toString: null },
+                        name: 'f',
+                        input: {},
+                    },
+                ],
+            },
+        ],
+        [
+            'a tool message',
+            { role: 'tool', tool_call_id: [{ toString: null }], content: 'ok' },
+        ],
+        [
+            'a tool_result block',
+            {
+                role: 'user',
+                content: [
+                    {
+                        type: 'tool_result',
+                        tool_use_id: { toString: null },
+                        content: 'ok',
+                    },
+                ],
+            },
+        ],
+    ])('rejects a tool call id that cannot be text on %s', (_label, message) => {
+        expect(() => normalize_single_message(message)).toThrow(
+            expect.objectContaining({
+                statusCode: 400,
+                message: 'a tool call id must be a string',
+            }),
+        );
+    });
+
+    it('keeps a numeric tool call id as the caller sent it', () => {
+        const message = normalize_single_message({
+            role: 'assistant',
+            content: [{ type: 'tool_use', id: 7, name: 'f', input: {} }],
+        });
+        expect(message.content[0].id).toBe(7);
+    });
+
+    it('keeps a tool call that has no id', () => {
+        const message = normalize_single_message({
+            role: 'assistant',
+            tool_calls: [{ function: { name: 'f', arguments: '{}' } }],
+        });
+        expect(message.content).toEqual([
+            { type: 'tool_use', id: undefined, name: 'f', input: '{}' },
+        ]);
+    });
+
+    it('rejects a content item whose type cannot be text', () => {
+        expect(() =>
+            normalize_single_message({
+                role: 'user',
+                content: [{ type: { toString: null }, text: 'hi' }],
+            }),
+        ).toThrow(
+            expect.objectContaining({
+                statusCode: 400,
+                message: 'each message content item type must be a string',
+            }),
+        );
+    });
+
+    it('sends an object tool result as its JSON', () => {
+        const message = normalize_single_message({
+            role: 'user',
+            content: [
+                { type: 'tool_result', tool_use_id: 't1', content: { temp: 21 } },
+            ],
+        });
+        expect(message.content[0].content).toBe('{"temp":21}');
+    });
+
+    it('keeps a tool result made of content blocks', () => {
+        const blocks = [{ type: 'text', text: 'Sunny' }];
+        const message = normalize_single_message({
+            role: 'user',
+            content: [{ type: 'tool_result', tool_use_id: 't1', content: blocks }],
+        });
+        expect(message.content[0].content).toEqual(blocks);
+    });
+});
@@ -18,20 +18,21 @@
 */

 /**
- * Inline-compaction translation helpers.
+ * Inline-compaction opt-in helpers.
 *
 * The driver-facing surface exposes a single provider-neutral opt-in
 * (`compaction: boolean | { trigger_tokens }` on `ICompleteArguments`), plus a
 * raw `context_management` escape hatch for callers hitting `/responses` with
- * the OpenAI-native shape. These helpers map that neutral opt-in to each
- * provider's SDK shape so the providers stay free of opt-in-parsing logic.
+ * the OpenAI-native shape. This module reads the opt-in; each dialect maps it
+ * to its own SDK shape (`toOpenAiContextManagement` in the OpenAI protocol,
+ * `toAnthropicContextManagement` in the Claude provider).
 */

 /**
 * @param {boolean | { trigger_tokens?: number } | undefined} compaction
 * @returns {{ enabled: boolean; trigger_tokens?: number }}
 */
-const readCompaction = (compaction) => {
+export const readCompaction = (compaction) => {
    if (compaction === true) return { enabled: true };
    if (compaction && typeof compaction === 'object') {
        return {
@@ -44,65 +45,6 @@ const readCompaction = (compaction) => {
    return { enabled: false };
 };

-/**
- * Build OpenAI Responses `context_management` from the neutral opt-in. A raw
- * `context_management` passthrough (already in OpenAI shape) wins.
- *
- * @param {{
- *     compaction?: boolean | { trigger_tokens?: number };
- *     context_management?: unknown;
- * }} args
- * @returns {{ type: 'compaction'; compact_threshold?: number }[] | undefined}
- */
-export const toOpenAiContextManagement = (args) => {
-    if (args.context_management !== undefined) {
-        return /** @type {any} */ (args.context_management);
-    }
-    const { enabled, trigger_tokens } = readCompaction(args.compaction);
-    if (!enabled) return undefined;
-    return [
-        {
-            type: 'compaction',
-            ...(trigger_tokens !== undefined
-                ? { compact_threshold: trigger_tokens }
-                : {}),
-        },
-    ];
-};
-
-/**
- * Build Anthropic `context_management` (beta `compact-2026-01-12`) from the
- * neutral opt-in. A raw `context_management` passthrough wins.
- *
- * @param {{
- *     compaction?: boolean | { trigger_tokens?: number };
- *     context_management?: unknown;
- * }} args
- * @returns {{ edits: Record<string, unknown>[] } | undefined}
- */
-export const toAnthropicContextManagement = (args) => {
-    if (args.context_management !== undefined) {
-        return /** @type {any} */ (args.context_management);
-    }
-    const { enabled, trigger_tokens } = readCompaction(args.compaction);
-    if (!enabled) return undefined;
-    return {
-        edits: [
-            {
-                type: 'compact_20260112',
-                ...(trigger_tokens !== undefined
-                    ? {
-                          trigger: {
-                              type: 'input_tokens',
-                              value: trigger_tokens,
-                          },
-                      }
-                    : {}),
-            },
-        ],
-    };
-};
-
 /**
 * Whether the request opted into inline compaction by any route.
 *
@@ -0,0 +1,62 @@
+/*
+ * Copyright (C) 2024-present Puter Technologies Inc.
+ *
+ * This file is part of Puter.
+ *
+ * Puter is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU Affero General Public License as published
+ * by the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU Affero General Public License for more details.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program.  If not, see <https://www.gnu.org/licenses/>.
+ */
+
+import { describe, expect, it } from 'vitest';
+import { messagesHaveCompaction, wantsCompaction } from './compaction.js';
+
+describe('wantsCompaction', () => {
+    it('is false without opt-in', () => {
+        expect(wantsCompaction({})).toBe(false);
+        expect(wantsCompaction({ compaction: false })).toBe(false);
+    });
+
+    it('is true for the neutral opt-in or a raw passthrough', () => {
+        expect(wantsCompaction({ compaction: true })).toBe(true);
+        expect(wantsCompaction({ compaction: { trigger_tokens: 1 } })).toBe(
+            true,
+        );
+        expect(wantsCompaction({ context_management: [] })).toBe(true);
+    });
+});
+
+describe('messagesHaveCompaction', () => {
+    it('detects a round-tripped compaction content block', () => {
+        expect(
+            messagesHaveCompaction([
+                { role: 'user', content: [{ type: 'text', text: 'hi' }] },
+                {
+                    role: 'assistant',
+                    content: [
+                        { type: 'compaction', encrypted_content: 'ENC' },
+                    ],
+                },
+            ]),
+        ).toBe(true);
+    });
+
+    it('is false for ordinary messages or non-arrays', () => {
+        expect(
+            messagesHaveCompaction([
+                { role: 'user', content: [{ type: 'text', text: 'hi' }] },
+            ]),
+        ).toBe(false);
+        expect(messagesHaveCompaction(undefined)).toBe(false);
+        expect(messagesHaveCompaction('nope')).toBe(false);
+    });
+});
@@ -40,7 +40,7 @@ type Part = Record<string, unknown>;
 const isObject = (value: unknown): value is Part =>
    !!value && typeof value === 'object' && !Array.isArray(value);

-export { parseDataUri } from '../../util/dataUri.js';
+export { parseDataUri } from '../../../../util/dataUri.js';

 /** The URL out of a bare string or an OpenAI-style `{ url }` object. */
 export const mediaUrlOf = (value: unknown): string | undefined => {
@@ -18,19 +18,19 @@
 */

 import { describe, expect, it } from 'vitest';
-import type { Actor } from '../../../core/actor.js';
-import { ALIBABA_MODELS } from '../providers/alibaba/models.js';
-import { BYTEPLUS_MODELS } from '../providers/byteplus/models.js';
-import { DEEPSEEK_MODELS } from '../providers/deepseek/models.js';
-import { GROQ_MODELS } from '../providers/groq/models.js';
-import { HOONIFY_MODELS } from '../providers/hoonify/models.js';
-import { MINIMAX_MODELS } from '../providers/minimax/models.js';
-import { MISTRAL_MODELS } from '../providers/mistral/models.js';
-import { MOONSHOT_MODELS } from '../providers/moonshot/models.js';
-import { ZAI_MODELS } from '../providers/zai/models.js';
+import type { Actor } from '../../../../../core/actor.js';
+import { ALIBABA_MODELS } from '../../../providers/alibaba/chat/models.js';
+import { BYTEPLUS_MODELS } from '../../../providers/byteplus/chat/models.js';
+import { DEEPSEEK_MODELS } from '../../../providers/deepseek/chat/models.js';
+import { GROQ_MODELS } from '../../../providers/groq/chat/models.js';
+import { HOONIFY_MODELS } from '../../../providers/hoonify/chat/models.js';
+import { MINIMAX_MODELS } from '../../../providers/minimax/chat/models.js';
+import { MISTRAL_MODELS } from '../../../providers/mistral/chat/models.js';
+import { MOONSHOT_MODELS } from '../../../providers/moonshot/chat/models.js';
+import { ZAI_MODELS } from '../../../providers/zai/chat/models.js';
 import type { IChatModel } from '../types.js';
 import { meterChatUsage } from './meterChatUsage.js';
-import { splitCachedPrompt } from './OpenAIUtil.js';
+import { splitCachedPrompt } from '../../../protocols/openaiCompatible/OpenAIUtil.js';
 import { usageCost } from './pricing.js';

 type Recorded = [Record<string, number>, string, Record<string, number>];
@@ -17,8 +17,8 @@
 * along with this program.  If not, see <https://www.gnu.org/licenses/>.
 */

-import type { Actor } from '../../../core/actor.js';
-import type { MeteringService } from '../../../services/metering/MeteringService.js';
+import type { Actor } from '../../../../../core/actor.js';
+import type { MeteringService } from '../../../../../services/metering/MeteringService.js';
 import type { IChatModel, ServiceSpeed } from '../types.js';
 import { serviceSpeedBillingModel } from './serviceSpeed.js';
 import { buildCostsOverride } from './pricing.js';
@@ -19,7 +19,7 @@
 */

 import { describe, expect, it } from 'vitest';
-import { GEMINI_MODELS } from '../providers/gemini/models.js';
+import { GEMINI_MODELS } from '../../../providers/gemini/chat/models.js';
 import type { IChatModel } from '../types.js';
 import {
    compareModelPreference,
@@ -21,7 +21,6 @@ import { describe, expect, it } from 'vitest';
 import type { IChatModel } from '../types.js';
 import {
    buildCostsOverride,
-    catalogPrice,
    costsForRequest,
    isFreeModel,
    isOutputCostKey,
@@ -480,12 +479,6 @@ it('preserves free classification for legacy hyphenated token keys', () => {
    expect(isFreeModel(model({ 'input-tokens': 0, 'output-tokens': 0 }))).toBe(true);
 });

-it('leaves overflowing catalog unit conversions unknown', () => {
-    expect(catalogPrice(1e308, 100)).toBeUndefined();
-    expect(catalogPrice(0, 100)).toBe(0);
-    expect(catalogPrice('0.0000000014', 100_000_000)).toBeCloseTo(0.14);
-});
-
 describe('costsForRequest prompt-size overrides', () => {
    const model = (minPromptTokensInclusive?: boolean): IChatModel => ({
        id: 'tiered',
@@ -23,29 +23,7 @@ import type {
    ModelPricingOverride,
    UsageDetails,
 } from '../types.js';
-
-/**
- * Absent, negative and placeholder prices remain unpriced, including empty
- * strings.
- */
-export const catalogPrice = (
-    value: unknown,
-    multiplier = 1,
-): number | undefined => {
-    if (typeof value !== 'number' && typeof value !== 'string')
-        return undefined;
-    if (
-        typeof value === 'string' &&
-        !/^(?:\d+(?:\.\d*)?|\.\d+)(?:e[+-]?\d+)?$/i.test(value.trim())
-    )
-        return undefined;
-    const price = Number(value);
-    if (!Number.isFinite(price) || price < 0) return undefined;
-    const converted = price * multiplier;
-    return Number.isFinite(converted) && !(price > 0 && converted === 0)
-        ? converted
-        : undefined;
-};
+import { catalogPrice } from '../../../shared/catalogPrice.js';

 const CENTS_PER_USD = 100;
 const MTOK = 1_000_000;
@@ -360,7 +338,8 @@ export const buildCostsOverride = (
 * What a usage object costs in µ¢, split into its input and output sides.
 * `recordedCosts` are the per-key costs a provider already metered with and win
 * over the model's rates, so the figure matches the ledger. Null when nothing
- * in the usage could be priced. `now` is the instant the request was priced at.
+ * in the usage could be priced. `now` is the instant the request was priced
+ * at.
 */
 export const usageCost = (
    usage: Record<string, unknown>,
@@ -19,7 +19,7 @@
 */

 import { afterEach, describe, expect, it } from 'vitest';
-import { kv } from '../../../util/kvSingleton.js';
+import { kv } from '../../../../../util/kvSingleton.js';
 import {
    clearUnhealthyRoutes,
    isRouteUnhealthy,
@@ -30,7 +30,7 @@
 * blip doesn't hand the model's traffic to a reseller for long.
 */

-import { kv } from '../../../util/kvSingleton.js';
+import { kv } from '../../../../../util/kvSingleton.js';

 /**
 * How long a route stays marked after a failure that won't heal on its own,
@@ -19,21 +19,21 @@

 import { describe, expect, it } from 'vitest';
 import type { IChatModel, ServiceSpeed } from '../types.js';
-import { ALIBABA_MODELS } from '../providers/alibaba/models.js';
-import { AZURE_MODELS } from '../providers/azure/models.js';
-import { BYTEPLUS_MODELS } from '../providers/byteplus/models.js';
-import { CLAUDE_MODELS } from '../providers/claude/models.js';
-import { DEEPSEEK_MODELS } from '../providers/deepseek/models.js';
-import { GEMINI_MODELS } from '../providers/gemini/models.js';
-import { GROQ_MODELS } from '../providers/groq/models.js';
-import { HOONIFY_MODELS } from '../providers/hoonify/models.js';
-import { META_MODELS } from '../providers/meta/models.js';
-import { MINIMAX_MODELS } from '../providers/minimax/models.js';
-import { MISTRAL_MODELS } from '../providers/mistral/models.js';
-import { MOONSHOT_MODELS } from '../providers/moonshot/models.js';
-import { OPEN_AI_MODELS } from '../providers/openai/models.js';
-import { XAI_MODELS } from '../providers/xai/models.js';
-import { ZAI_MODELS } from '../providers/zai/models.js';
+import { ALIBABA_MODELS } from '../../../providers/alibaba/chat/models.js';
+import { AZURE_MODELS } from '../../../providers/azure/chat/models.js';
+import { BYTEPLUS_MODELS } from '../../../providers/byteplus/chat/models.js';
+import { CLAUDE_MODELS } from '../../../providers/claude/chat/models.js';
+import { DEEPSEEK_MODELS } from '../../../providers/deepseek/chat/models.js';
+import { GEMINI_MODELS } from '../../../providers/gemini/chat/models.js';
+import { GROQ_MODELS } from '../../../providers/groq/chat/models.js';
+import { HOONIFY_MODELS } from '../../../providers/hoonify/chat/models.js';
+import { META_MODELS } from '../../../providers/meta/chat/models.js';
+import { MINIMAX_MODELS } from '../../../providers/minimax/chat/models.js';
+import { MISTRAL_MODELS } from '../../../providers/mistral/chat/models.js';
+import { MOONSHOT_MODELS } from '../../../providers/moonshot/chat/models.js';
+import { OPEN_AI_MODELS } from '../../../providers/openai/chat/models.js';
+import { XAI_MODELS } from '../../../providers/xai/chat/models.js';
+import { ZAI_MODELS } from '../../../providers/zai/chat/models.js';
 import { shouldPresentAsOpenAI } from './normalizeToOpenAI.js';
 import {
    isServiceSpeedUnavailable,
@@ -20,19 +20,19 @@
 import { describe, expect, it } from 'vitest';
 import {
    makeMeteringStub,
-    optionalEnv,
    withTestActor,
-} from '../../integrationTestUtil.js';
+} from '../../../shared/integrationTestUtil.js';
+import { optionalEnv } from '../../../../../testEnv.js';
 import type {
    IChatMessageResult,
    IChatProvider,
    ICompleteArguments,
 } from '../types.js';
-import { OpenAiResponsesChatProvider } from '../providers/openai/OpenAiChatResponsesProvider.js';
-import { GeminiChatProvider } from '../providers/gemini/GeminiChatProvider.js';
-import { MiniMaxProvider } from '../providers/minimax/MiniMaxProvider.js';
-import { MoonshotProvider } from '../providers/moonshot/MoonshotProvider.js';
-import { MistralAIProvider } from '../providers/mistral/MistralAiProvider.js';
+import { OpenAiResponsesChatProvider } from '../../../providers/openai/chat/OpenAiChatResponsesProvider.js';
+import { GeminiChatProvider } from '../../../providers/gemini/chat/GeminiChatProvider.js';
+import { MiniMaxProvider } from '../../../providers/minimax/chat/MiniMaxProvider.js';
+import { MoonshotProvider } from '../../../providers/moonshot/chat/MoonshotProvider.js';
+import { MistralAIProvider } from '../../../providers/mistral/chat/MistralAiProvider.js';
 import { selectServiceSpeed } from './serviceSpeed.js';

 type Case = {
@@ -23,7 +23,7 @@ import { AIChatStream } from './Streaming.js';
 import {
    handle_completion_output,
    handle_completion_output_responses_api,
-} from './OpenAIUtil.js';
+} from '../../../protocols/openaiCompatible/OpenAIUtil.js';
 import type { IChatModel, ServiceSpeed } from '../types.js';
 import {
    requestedServiceSpeed,
@@ -33,10 +33,10 @@ import {
    isServiceSpeedUnavailable,
 } from './serviceSpeed.js';
 import { meterChatUsage } from './meterChatUsage.js';
-import { OPEN_AI_MODELS } from '../providers/openai/models.js';
-import { XAI_MODELS } from '../providers/xai/models.js';
-import { GEMINI_MODELS } from '../providers/gemini/models.js';
-import { CLAUDE_MODELS } from '../providers/claude/models.js';
+import { OPEN_AI_MODELS } from '../../../providers/openai/chat/models.js';
+import { XAI_MODELS } from '../../../providers/xai/chat/models.js';
+import { GEMINI_MODELS } from '../../../providers/gemini/chat/models.js';
+import { CLAUDE_MODELS } from '../../../providers/claude/chat/models.js';

 const model = (provider: string, id: string): IChatModel => ({
    provider,
@@ -216,6 +216,24 @@ describe('estimatePromptTokens', () => {
    });
 });

+describe('estimatePromptTokens compaction blocks', () => {
+    const withCompaction = (encrypted_content: unknown) => [
+        { role: 'assistant', content: [{ type: 'compaction', encrypted_content }] },
+    ];
+
+    it('counts the encrypted content as text', () => {
+        expect(estimatePromptTokens(withCompaction('a'.repeat(400)))).toBe(
+            estimateTextTokens('a'.repeat(400)),
+        );
+    });
+
+    it('counts content that is not a string by its JSON', () => {
+        expect(
+            estimatePromptTokens(withCompaction({ toString: null, blob: 'abcdefgh' })),
+        ).toBe(estimateTextTokens('{"toString":null,"blob":"abcdefgh"}'));
+    });
+});
+
 describe('estimateToolTokens', () => {
    it('is zero for no tools', () => {
        expect(estimateToolTokens(undefined)).toBe(0);
@@ -33,10 +33,10 @@ import {
    createWordScanBudget,
    estimateTextTokens,
    type WordScanBudget,
-} from '../../util/tokenEstimate.js';
+} from '../../../shared/tokenEstimate.js';

 // Re-exported so existing consumers keep one import site for all estimators.
-export { estimateTextTokens } from '../../util/tokenEstimate.js';
+export { estimateTextTokens } from '../../../shared/tokenEstimate.js';

 /** Tokens per byte of an inline (base64) attachment; see `attachmentTokens`. */
 const ATTACHMENT_BYTES_PER_TOKEN = 150;
@@ -154,7 +154,12 @@ const partTokens = (part: unknown, budget: WordScanBudget): number => {
        );
    }
    if (p.type === 'compaction') {
-        return estimateTextTokens(String(p.encrypted_content ?? ''), budget);
+        return estimateTextTokens(
+            typeof p.encrypted_content === 'string'
+                ? p.encrypted_content
+                : JSON.stringify(p.encrypted_content ?? ''),
+            budget,
+        );
    }

    return UNMEASURED_ATTACHMENT_TOKENS;
@@ -41,23 +41,22 @@ import {
 } from 'vitest';

 import { v4 as uuidv4 } from 'uuid';
-import { runWithContext } from '../../core/context.js';
-import { type Actor, SYSTEM_ACTOR } from '../../core/actor.js';
-import { PuterServer } from '../../server.js';
-import { setupTestServer } from '../../testUtil.js';
-import { withAiCostFactorHook } from '../integrationTestUtil.js';
-import { generateDefaultFsentries } from '../../util/userProvisioning.js';
-import { CLOUDFLARE_IMAGE_GENERATION_MODELS } from './providers/cloudflare/models.js';
-import cloudflareModelSearchFixture from './providers/cloudflare/fixtures/model-search.json';
-import { GEMINI_IMAGE_GENERATION_MODELS } from './providers/gemini/models.js';
-import { OPEN_AI_IMAGE_GENERATION_MODELS } from './providers/openai/models.js';
-import { REPLICATE_IMAGE_GENERATION_MODELS } from './providers/replicate/models.js';
-import { TOGETHER_IMAGE_GENERATION_MODELS } from './providers/together/models.js';
-import { XAI_IMAGE_GENERATION_MODELS } from './providers/xai/models.js';
-import xaiLiveCatalog from './providers/xai/fixtures/image-generation-models.json';
+import { runWithContext } from '../../../../core/context.js';
+import { type Actor, SYSTEM_ACTOR } from '../../../../core/actor.js';
+import { PuterServer } from '../../../../server.js';
+import { setupTestServer, withAiCostFactorHook } from '../../../../testUtil.js';
+import { generateDefaultFsentries } from '../../../../util/userProvisioning.js';
+import { CLOUDFLARE_IMAGE_GENERATION_MODELS } from '../../providers/cloudflare/image/models.js';
+import cloudflareModelSearchFixture from '../../providers/cloudflare/image/fixtures/model-search.json';
+import { GEMINI_IMAGE_GENERATION_MODELS } from '../../providers/gemini/image/models.js';
+import { OPEN_AI_IMAGE_GENERATION_MODELS } from '../../providers/openai/image/models.js';
+import { REPLICATE_IMAGE_GENERATION_MODELS } from '../../providers/replicate/image/models.js';
+import { TOGETHER_IMAGE_GENERATION_MODELS } from '../../providers/together/image/models.js';
+import { XAI_IMAGE_GENERATION_MODELS } from '../../providers/xai/image/models.js';
+import xaiLiveCatalog from '../../providers/xai/image/fixtures/image-generation-models.json';
 import { ImageGenerationDriver } from './ImageGenerationDriver.js';
-import { CATALOG_TTL_MS } from '../ai-chat/utils/cachedRemoteCatalog.js';
-import { ReplicateImageGenerationProvider } from './providers/replicate/ReplicateImageGenerationProvider.js';
+import { CATALOG_TTL_MS } from '../../shared/cachedRemoteCatalog.js';
+import { ReplicateImageGenerationProvider } from '../../providers/replicate/image/ReplicateImageGenerationProvider.js';

 // ── SDK mocks ──────────────────────────────────────────────────────
 //
@@ -141,8 +140,8 @@ vi.mock('replicate', () => {

 const { secureFetchMock } = vi.hoisted(() => ({ secureFetchMock: vi.fn() }));

-vi.mock('../../util/secureHttp.js', async (importOriginal) => ({
-    ...(await importOriginal<typeof import('../../util/secureHttp.js')>()),
+vi.mock('../../../../util/secureHttp.js', async (importOriginal) => ({
+    ...(await importOriginal<typeof import('../../../../util/secureHttp.js')>()),
    secureFetch: secureFetchMock,
 }));

@@ -483,6 +482,33 @@ describe('ImageGenerationDriver.generate provider routing', () => {
        },
    );

+    it('rejects an input image type that is not a string before any provider runs', async () => {
+        await expect(withActor(() =>
+            driver.generate({
+                model: 'gpt-image-2',
+                prompt: 'make it blue',
+                input_image: 'data:image/png;base64,aGVsbG8=',
+                input_image_mime_type: 5,
+            } as never),
+        )).rejects.toMatchObject({
+            statusCode: 400,
+            message: 'input_image_mime_type must be a string',
+        });
+    });
+
+    it('ignores an input image type that is not a string when there is no input image', async () => {
+        openaiImagesGenerateMock.mockResolvedValueOnce({
+            data: [{ url: 'https://oai/image.png' }],
+        });
+        await expect(withDriverName('ai-image', () =>
+            driver.generate({
+                model: 'gpt-image-2',
+                prompt: 'a landscape',
+                input_image_mime_type: 5,
+            } as never),
+        )).resolves.toBe('https://oai/image.png');
+    });
+
    it('rejects an excluded model regardless of caller casing', async () => {
        await expect(withActor(() =>
            driver.generate({
@@ -1273,7 +1299,7 @@ describe('ImageGenerationDriver.generate audit log', () => {
 // ── puter_output_path ─────────────────────────────────────────────

 describe('ImageGenerationDriver.generate puter_output_path', () => {
-    const TEST_ACTOR: import('../../core/actor.js').Actor = {
+    const TEST_ACTOR: import('../../../../core/actor.js').Actor = {
        user: { uuid: 'a1b2c3d4-e5f6-4a7b-8c9d-0e1f2a3b4c5d', id: 42, username: 'testuser' },
    };

@@ -1462,7 +1488,7 @@ describe('ImageGenerationDriver.generate puter_output_path', () => {
    });

    it('throws 400 when actor has no user ID but puter_output_path is set', async () => {
-        const noIdActor: import('../../core/actor.js').Actor = {
+        const noIdActor: import('../../../../core/actor.js').Actor = {
            user: { uuid: 'f0e1d2c3-b4a5-4968-8777-0a1b2c3d4e5f', username: 'noone' },
        };
        await expect(
@@ -19,32 +19,32 @@

 import { assertImagePrompt } from './imageValidation.js';
 import crypto from 'node:crypto';
-import { Context } from '../../core/context.js';
-import { HttpError } from '../../core/http/HttpError.js';
-import type { Actor } from '../../core/actor.js';
-import { PuterDriver } from '../types.js';
+import { Context } from '../../../../core/context.js';
+import { HttpError } from '../../../../core/http/HttpError.js';
+import type { Actor } from '../../../../core/actor.js';
+import { PuterDriver } from '../../../types.js';
 import {
    type AiMeteringService,
    withAiCostFactor,
-} from '../util/aiCostFactor.js';
-import { AI_CONCURRENT, AI_RATE_LIMIT } from '../util/aiLimits.js';
+} from '../../../../util/aiCostFactor.js';
+import { AI_CONCURRENT, AI_RATE_LIMIT } from '../../shared/aiLimits.js';
 import {
    resolveOutputPath,
    saveGeneratedMediaToFS,
-} from '../util/generatedMedia.js';
-import { readPuterImageInputs } from '../util/mediaInput.js';
-import { ModelCatalog } from '../util/modelCatalog.js';
-import { CATALOG_TTL_MS } from '../ai-chat/utils/cachedRemoteCatalog.js';
-import { readProviderKey } from '../util/providerRegistry.js';
-import { BytePlusImageProvider } from './providers/byteplus/BytePlusImageProvider.js';
-import { CloudflareImageProvider } from './providers/cloudflare/CloudflareImageProvider.js';
-import { GeminiImageProvider } from './providers/gemini/GeminiImageProvider.js';
-import { KlingImageProvider } from './providers/kling/KlingImageProvider.js';
-import { OpenAiImageProvider } from './providers/openai/OpenAiImageProvider.js';
-import { ReplicateImageGenerationProvider } from './providers/replicate/ReplicateImageGenerationProvider.js';
+} from '../../shared/generatedMedia.js';
+import { readPuterImageInputs } from '../../shared/mediaInput.js';
+import { ModelCatalog } from '../../shared/modelCatalog.js';
+import { CATALOG_TTL_MS } from '../../shared/cachedRemoteCatalog.js';
+import { readProviderKey } from '../../../../util/aiProviderKey.js';
+import { BytePlusImageProvider } from '../../providers/byteplus/image/BytePlusImageProvider.js';
+import { CloudflareImageProvider } from '../../providers/cloudflare/image/CloudflareImageProvider.js';
+import { GeminiImageProvider } from '../../providers/gemini/image/GeminiImageProvider.js';
+import { KlingImageProvider } from '../../providers/kling/image/KlingImageProvider.js';
+import { OpenAiImageProvider } from '../../providers/openai/image/OpenAiImageProvider.js';
+import { ReplicateImageGenerationProvider } from '../../providers/replicate/image/ReplicateImageGenerationProvider.js';
 import { resolveImageSize } from './imageDimensions.js';
-import { TogetherImageProvider } from './providers/together/TogetherImageProvider.js';
-import { XAIImageProvider } from './providers/xai/XAIImageProvider.js';
+import { TogetherImageProvider } from '../../providers/together/image/TogetherImageProvider.js';
+import { XAIImageProvider } from '../../providers/xai/image/XAIImageProvider.js';
 import type { IGenerateParams, IImageModel, IImageProvider } from './types.js';
 import { assertInputImagesShape } from './inputImage.js';

@@ -78,7 +78,7 @@ export class ImageGenerationDriver extends PuterDriver {
    ];
    readonly isDefault = true;

-    // Shared AI policy — see `drivers/util/aiLimits.ts` for the tier table.
+    // Shared AI policy — see `drivers/ai/shared/aiLimits.ts` for the tier table.
    readonly rateLimit = AI_RATE_LIMIT;
    readonly concurrent = AI_CONCURRENT;

@@ -177,7 +177,15 @@ export class ImageGenerationDriver extends PuterDriver {
        // settled here — once — before any provider runs.
        assertImagePrompt(args.prompt);
        assertInputImagesShape(args, 'image generation');
-        for (const option of ['quality', 'resolution'] as const) {
+        // The type hint describes an input image, so it is only read, and
+        // only has to be a string, when the request carries one.
+        const hasInputImage =
+            args.input_image != null || (args.input_images?.length ?? 0) > 0;
+        for (const option of [
+            'quality',
+            'resolution',
+            ...(hasInputImage ? (['input_image_mime_type'] as const) : []),
+        ] as const) {
            if (args[option] != null && typeof args[option] !== 'string') {
                throw new HttpError(400, `${option} must be a string`, {
                    legacyCode: 'bad_request',
@@ -189,6 +197,7 @@ export class ImageGenerationDriver extends PuterDriver {
        // caller's `args` is re-emitted by reference in `.after`/`.error`
        // payloads, so it is never mutated here.
        const { puter_output_path: puterOutputPath, ...request } = args;
+        if (!hasInputImage) delete request.input_image_mime_type;

        // Validate the output path early — before spending credits.
        const resolvedOutputPath = puterOutputPath
@@ -17,7 +17,7 @@
 * along with this program.  If not, see <https://www.gnu.org/licenses/>.
 */

-import { HttpError } from '../../core/http/HttpError.js';
+import { HttpError } from '../../../../core/http/HttpError.js';
 import type {
    IGenerateParams,
    IImageModel,
Author	SHA1	Message	Date
Daniel Salazar	8c31f9177c	ci: the AI provider workflow runs only the provider tests and passes every provider's key	2026-11-13 20:31:22 -07:00
Daniel Salazar	5531f381f5	docs(ai): the image reference and quotas page cover Replicate's safety reruns	2026-11-13 20:31:22 -07:00
Daniel Salazar	8a198b6f0a	style: restore the repo formatting in files the AI stack reflowed	2026-11-13 20:31:22 -07:00
Daniel Salazar	17dc879ff5	fix(ai): the native Gemini request type accepts the xhigh effort Hydrogen takes	2026-11-13 20:31:22 -07:00
Daniel Salazar	d3321b1d0b	fix(ai): a model id with a colon after its vendor keeps its name when it is qualified	2026-11-13 20:31:21 -07:00
Daniel Salazar	a52f32e0d2	test(ai): a fuzz case keeps its input and route when the catalog changes, and a suite that reaches no provider fails	2026-11-13 20:31:21 -07:00
Daniel Salazar	086e3a0028	fix(ai): OpenAI speech to text answers 400 to a prompt that is a list or an object	2026-11-13 20:31:21 -07:00
Daniel Salazar	aabbafdaab	fix(ai): the image type hint only has to be a string when a request carries an input image	2026-11-13 20:31:21 -07:00
404oops	97ab0ee12b	Forward Claude inference speed from SDK chat options	2026-11-13 17:12:31 +01:00
404oops	358b680b69	Qualify published chat model names while preserving routing aliases	2026-11-13 14:07:30 +01:00
404oops	e534a3d09a	Merge branch 'feat/gemini-carbon-kling-4' into filipkujundzic/put-2351-backend-deprecated-models-check-2026	2026-11-13 13:53:25 +01:00
404oops	5fd07e9ed7	Cap Mistral Large 4 output at the 262,144 tokens Mistral accepts The catalog listed `mistral-large-4` with a 1,000,000-token output limit, equal to its context window. Mistral rejects a request for more than 262,144 completion tokens, so a request that left `max_tokens` unset, or set it higher, failed with a 400 whenever the account's credit covered more than 262,144 output tokens. - `max_tokens` for `mistral-large-4` is 262,144. The context window stays at 1,000,000 tokens. - A test sends `mistral-large-4` and `mistral-large-latest` through the chat driver with ample credit and checks the cap that reaches Mistral.	2026-11-13 13:51:47 +01:00
404oops	e138f1e9d8	Move the reasoning_effort model list out of the chat options table The `reasoning_effort` row in chat.md listed the models that read the option. Model-specific details belong under Model-specific features, which already gives each family's accepted values. - The row no longer names models. - The Gemini and Moonshot Kimi sections say that their other models ignore the option.	2026-11-13 13:44:06 +01:00
404oops	7886907bdf	Add Gemini 4 Hydrogen and cap unary Gemini 4 Carbon output Google released `gemini-4-hydrogen` on 2026-11-12. It uses the same native request shape as Carbon and Argon. - Hydrogen is served natively at $1.50 input, $7.50 output and $0.075 cached input per million tokens, with fast service at 1.8 times those rates. It accepts `low`, `medium`, `high` and `xhigh` reasoning effort, ignores temperature, and takes text and images. - A Carbon response that isn't streamed is capped at 32,768 output tokens, as Argon's already is, and the credit hold is sized to that cap. Google reports this limit for both models. The cap is now a field of the native model options instead of a check on Argon's id. - chat.md and rate-limits-and-quotas.md list Hydrogen's reasoning levels and limits and Carbon's unary cap.	2026-11-13 11:31:10 +01:00
404oops	0eec08df98	Point Mistral's floating Large aliases at Large 4 and apply its GA price Mistral made Large 4 generally available on 2026-11-05, moved `mistral-large-latest` to it, and deprecated Large 3 and Medium 3.5, which both retire on 2027-04-30. - `mistral-large-latest`, `mistral-large` and `mistralai/mistral-large-latest` resolve to `mistral-large-4`. They resolved to `mistral-large-2512`, which stays callable by its own id until Mistral retires it. - Large 4 uses Mistral's current rates: $0.70 input, $0.07 cached input and $2.10 output per million tokens, in place of the preview rates. It is marked open-weight. - The Large 3 and Medium 3.5 entries carry their retirement date.	2026-11-13 11:30:34 +01:00
404oops	4b5d6e034c	Document the Kling 4 prompt limit for non-English text Kling 4.0 and 4.0 Flash take 26,000 characters of English but 11,000 of other languages, with mixed text sharing the limit in proportion. The video page and the quotas page said 26,000 for every prompt. - txt2vid.md and rate-limits-and-quotas.md give both limits, the rule for mixed text, and the error each one returns. - Puter still rejects only a prompt over 26,000 characters. Kling enforces the lower limit, and its refusal reaches the caller as `upstream_bad_request` with nothing charged. A test covers that path.	2026-11-13 11:00:50 +01:00
404oops	7d53cc7ed6	Include Kling 4 models in the video driver catalog test	2026-11-13 07:53:16 +01:00
404oops	86f871c9a8	Add Gemini 4 Carbon and Kling 4 video models	2026-11-13 07:53:16 +01:00
404oops	c038a6983f	Hold the AI concurrency slot until a detached chat completion settles A chat completion keeps running for up to 10 minutes after the caller hangs up, so the upstream can report usage. It now counts against the caller's concurrency cap for as long as the upstream runs, instead of until the response closes. - The chat driver registers its upstream run with the request's concurrency slot, so `/drivers/call` keeps the slot until the upstream is done. The cleanup after it does not hold the slot. - The route concurrency gate now uses the same CallSlot as `/drivers/call`: the slot is held until the response is done and the work the handler registered has settled. The OpenAI- and Anthropic-compatible chat routes register the completion, so a non-streaming call counts until the upstream answers too. - CallSlot moves to core/http, since the route gate below the controllers uses it. The context key is renamed keepConcurrencySlot. - rate-limits-and-quotas.md says a completion counts until the model finishes it, even after a disconnect.	2026-11-13 07:53:16 +01:00
404oops	f044d8d071	Restore the CODEOWNERS default owner and wildcard patterns The default owner line read `- @Salazareo`, which only matches a file named `-`, so files without a specific owner had none. Three patterns had their `*` escaped, which GitHub rejects as invalid, so their lines assigned no owner.	2026-11-13 07:52:52 +01:00
404oops	aea6e662f6	Fuzz the AI drivers with seeded requests and upstream answers Two seeded suites send malformed input through the real drivers and providers. The vendors are played at the HTTP boundary, so each provider's request and response code runs. - ChatCompletionDriver.fuzz.test.ts covers the 17 chat providers that have a static catalog. Requests with a few values changed must resolve or reject below 500 without marking a route unhealthy. Valid requests against wrong statuses, bodies of the wrong shape and broken event streams must end with an answer or a classified error. A stream only has to end; the error line it carries is not checked. - mediaDrivers.fuzz.test.ts covers the image, video, speech and OCR drivers, called as a user with unlimited metering so a request gets past file input and the credit hold. Every vendor refuses the request, and every call must end with a classified error. shared/fuzzTestUtil.ts holds the seeded random source, the hostile JSON values, the mutation, and the stand-in for fetch, undici and axios. An SDK that calls https.request itself fails the suite. PUTER_FUZZ_CASES raises the case count for a longer run. doc/ai-drivers.md describes the suites.	2026-11-13 07:52:52 +01:00
404oops	31f6dc2d37	Answer malformed AI requests with a 400 where they enter A value that used to end in an unclassified error, in a request or in an upstream answer, now gets a classified one before provider code runs on it. A value a driver or provider ignored or refused before is handled as it was. Every driver: - shared/scalarOptions.ts answers 400 for a value that cannot be converted to text, such as { "toString": null }, in the options a driver reads as text, a number or a flag. The chat, video, text to speech, speech to text, voice changer and OCR drivers call it first. - An unknown provider name that is not text gets the same 400 as any other unknown name. Chat requests: - A falsy `tools` means no tools on every provider. - A tool call id or a content block type that cannot be converted to text is a 400. - A tool result that is an object is sent as its JSON, like the content of a `tool` message. - `include` that is not an array is a 400. - `anthropicBetas` that is not a list of names is dropped. - Gemini's native route answers 400 when no user or assistant message is left to send. - The prompt estimate reads a compaction block whose content is not a string. Image, video, speech and OCR requests: - The image driver checks `input_image_mime_type` with `quality` and `resolution`. - The video driver answers 400 for a model that is not a string. - The voice changer answers 400 for an `output_format` that is not a string. - Mistral OCR answers 400 when the SDK refuses the options before sending them, as Mistral chat does. Upstream answers to chat: - A choice without a message is the existing 400 bad_response, which still falls back. - A usage count that is not a number is reported as zero in the cost event, as pricing already skips it.	2026-11-13 07:52:52 +01:00
404oops	2b03bc53b1	Rerun Replicate predictions an output safety checker rejects Some community models check the generated image and fail the prediction with "Try running it again" for a share of seeds, whatever the prompt. adirik/realvisxl-v3.0-turbo failed 5 of 24 runs of "a small red cube on a white table" that way. A prediction that fails with that message is now run again, up to two more times, inside the same deadline and credit hold. A request that pins a seed is not rerun, because the same seed is rejected again. Only the prediction that succeeds is metered, as before.	2026-11-13 07:52:52 +01:00
404oops	3ea0010220	Remove the unused ImageProvider base class Every image provider implements IImageProvider directly; nothing extended or imported the abstract class in interfaces/image/ImageProvider.ts. The guide's import rules now say that only an interface's driver file imports providers, which holds since the tool helpers stopped importing the Claude policy.	2026-11-13 07:52:52 +01:00
404oops	395e4bd54a	Split the tool and compaction helpers by dialect FunctionCalling.js and compaction.js each held Puter-facing parsing next to OpenAI and Anthropic request shapes. Each dialect now owns its mapping: interfaces/chat/utils normalize_tools_object, readCompaction and the other dialect-neutral checks protocols/openaiCompatible/openaiParams.ts toolsToWire (was make_openai_tools) and toOpenAiContextManagement providers/claude/chat/contextManagement.ts toAnthropicContextManagement toolsToWire takes the dialect as its second argument, like the other mappers in that module, and takes an array; every caller already guarded the empty case. make_claude_tools is removed. Nothing but its tests called it: the Claude provider uses claudeToolPolicy directly. The three cases only those tests had are added to the claudeToolPolicy tests. The chat interface helpers no longer import from providers/claude.	2026-11-13 07:52:52 +01:00
404oops	b55cb715ca	Move the AI test helpers into the AI tree drivers/integrationTestUtil.ts held helpers for AI provider tests, but a store test and the realtime relay test imported from it too. drivers/ai/shared/integrationTestUtil.ts the provider test helpers testEnv.ts optionalEnv and skipUnlessEnv, for any env-gated suite testUtil.ts withAiCostFactorHook, next to the server it hooks No test under services/ or stores/ imports from the driver layer's AI code now. The typecheck baseline entry for the moved file is renamed, not regenerated.	2026-11-13 07:52:52 +01:00
404oops	2e4d918900	Remove the aiUserIdentifier alias shared/aiUserIdentifier.ts re-exported upstreamUserIdentifier under a second name, and only its own test imported it. The alias is gone; its tests, which cover truncation and actor shapes the other file didn't, replace the four cases in upstreamIdentifier.test.ts.	2026-11-13 07:33:27 +01:00
404oops	8c799a433a	Move the AI billing helpers the realtime relay uses below the driver layer OpenAIRealtimeRelayService is a service, and it imported the cost factor, the provider key reader and a transcription price from drivers/ai. Those now live in src/backend/util, which both layers can import: util/aiCostFactor.ts withAiCostFactor, withAiCreditHold and the request cost helpers, unchanged util/aiProviderKey.ts readProviderKey, cut from providerRegistry.ts util/aiPrices.ts the gpt-transcribe per-second price, read by the OpenAI speech-to-text cost table and the relay withStreamedAiCharge returns a DriverStreamResult, so it stays in the driver layer as drivers/ai/shared/streamedAiCharge.ts with its tests. MeteringService is untouched. No prices or billing behavior change.	2026-11-13 07:33:27 +01:00
404oops	5f84d5897b	Move the ElevenLabs voice changer into a provider class VoiceChangerDriver built the ElevenLabs request, measured the converted audio and metered it, and imported the output formats from the ElevenLabs text-to-speech provider. That code is now in providers/elevenlabs: outputFormats.ts formats both speech interfaces use speechToSpeech/ElevenLabsSpeechToSpeechProvider.ts speechToSpeech/measuredOutput.ts WAV, Opus and byte-rate timing speechToSpeech/costs.ts The driver keeps provider resolution, test mode and the auth check. With one provider it holds the class directly; there is no provider interface. No behavior change: same validation order, error messages, usage keys, holds and charges.	2026-11-13 07:33:27 +01:00
404oops	c0aff857a4	Move the Textract and Mistral OCR calls into provider classes OCRDriver held both vendors' clients, request building, response normalizing and metering. That now lives in TextractOcrProvider (providers/awsTextract/ocr) and MistralOcrProvider (providers/mistral/ocr), both implementing IOcrProvider. The driver keeps the provider catalog, model resolution, auth and credit gates, and input resolution. The page estimate, document detection and credit hold both providers use are in interfaces/ocr/common.ts. No behavior change: same validation order, error messages, usage keys, holds and charges.	2026-11-13 07:33:26 +01:00
404oops	ff606abf3c	Move catalogPrice into ai/shared Chat providers and the Cloudflare and xAI image catalogs all parse vendor prices with catalogPrice, so it lives in shared/ instead of the chat interface's pricing helpers. The image catalogs no longer import from the chat interface.	2026-11-13 07:33:26 +01:00
404oops	8159998830	Document the AI driver layout doc/ai-drivers.md explains what interfaces, providers, protocols and shared each own, the import rules between them, where a new file goes, and the steps for adding a provider or an AI interface. It also lists the three places that don't follow the layout yet. doc/architecture.md and AGENTS.md link to it.	2026-11-13 07:33:26 +01:00
404oops	395f7f6adb	Move the AI drivers into drivers/ai The seven drivers/ai-* folders and the AI-only helpers in drivers/util become one tree: drivers/ai/interfaces/<interface>/ the Puter-facing drivers drivers/ai/providers/<vendor>/<interface>/ vendor adapters, catalogs, tests drivers/ai/protocols/openaiCompatible/ the shared OpenAI wire protocol drivers/ai/shared/ helpers several interfaces use Files are moved and their relative imports rewritten; no implementation changes. Driver names, aliases, interfaces, config keys, model ids and usage keys are unchanged. The chat catalog registration test scanned providers// next to itself. It now scans providers//chat/ and fails if the scan finds no catalog, so a wrong directory can't pass it. General helpers that non-AI code also uses stay in drivers/util: file input, data URIs, upstream errors, stream piping, CRUD limits.	2026-11-13 07:33:26 +01:00