Preserve the reviewed user-facing behavior, public-edition boundaries and owning regression coverage.
Prompt audit: Keep existing permission, ownership and recovery contracts at their current tool, runtime or test owner (Hard constraint). No separate resident instruction is added. Drop excluded hosted-only branches; retain shared side effects and tests. No optional prompt rewrite.
Prompt confirmation: The requester confirmed the exact proposed final behaviors and existing resource placements by directing that the primary implementation be followed, then explicitly requested this PR and merge. Preserve that confirmed scope only.
- Keep: The behavior and placement detailed in the audit above. Type: Hard constraint. Why: Preserve its existing canonical owner without duplicated resident instructions.
- Drop: Superseded overlapping wording and excluded hosted-only behavior. Type: Hard constraint. Why: Preserve the confirmed public capability and authority boundary.
Confirmed-by: Requester, exact final-state proposal and subsequent delivery request.
Scope: The previously confirmed final behavior and existing placement described above; no optional prompt rewrite.
Require positively nonlocal storage and an empty real data root before relocating an existing pin. Ignore data_test, preserve unknown and offline selections, validate destinations, and atomically publish pin changes with state rechecks.
Validation: 74 deterministic cases passed; main typecheck passed. Move three obsolete line-probe cases into structured disk and persistence coverage. Windows native metadata and Google Drive acceptance remain pending; one existing ASAR DEP0180 warning retained.
Run installer cleanup before old-version removal and replace broad process-name/prefix checks with a bounded installation-scoped helper. Preserve the installer control ancestry, sibling installations and user data; failed cleanup stops installation and retains retry.
Wire the public NSIS hook and keep platform process ownership, serialized native selection and repeated Windows cleanup cases aligned. Resolve the regression compiler from standard installed makensis or explicit public toolchain configuration, with no private product directory or automatic download. Missing compilation is an infrastructure failure; macOS compiles all four fault fixtures while Windows alone executes them.
Windows hardware execution is explicitly deferred by the requester. Preserve actual macOS compile and owning-test results separately from Windows platform skips.
Validation: 31 macOS owning tests passed; four real/failure/timeout/disabled installer fixtures compiled without compiler warnings or stderr. Six Windows-only cases remain skipped. The test compiler contains the upstream macOS character-conversion fix; no application dependency or production hook was changed to accommodate the compiler failure.
Recognize the anchored Node MaxListenersExceededWarning protocol from the console bridge and retain it as a warning requiring review. Keep application look-alikes and unrelated stderr as errors.
Scan input metadata with at most 32 outstanding reads while preserving deterministic fingerprint order. Keep private file bodies unread, yield during scanning and invalidate checkpoints when linked dependencies change or required inspection fails.
Preserve local CLI context, cancellation, interrupted-history recovery and conversation ownership across supported runtimes. Keep deterministic public launcher and gateway contracts without private model-evaluation dependencies or live-model execution.
Prompt confirmation:
The already approved CLI behavior reuses canonical user-first response language and authoritative delegated-task provenance at existing durable and turn-context positions. Exact public process retrieval remains parameter-local; no native slash-command, session, permission or tool-name boundary expands.
Prompt audit:
- Keep: Reuse explicit reply request, latest substantive user prose, then UI fallback. Internal delegation does not choose user language; descriptions defer to creator policy.
Type: Hard constraint. Placement: Existing CLI durable language tail and Delegated task turn heading. Actors: Named CLI Agents.
Purpose: One shared language owner; host sender identity supplies provenance without text inference.
- Keep: Exact public process selector, complete continuation arguments and retained-output locators share the native history contract.
Type: Hard constraint | Pitfall. Placement: On-demand MCP parameter schemas and retrieval receipts. Actors: Existing authorized CLI bridge callers.
Purpose: Avoid caller reconstruction while retaining private-reasoning exclusions, scope and no-replay boundaries.
- Drop: Private compact UI-first language instruction and duplicate resident retrieval procedure.
Type: Pitfall. Placement: Removed private language helper; retrieval remains in existing tool schema. Actors: Named CLI Agents.
Purpose: Canonical language and parameter owners assume responsibility without duplicate instruction.
Keep worker-backed extraction, HTML hardening, page selection, real OfficeCLI rendering and read-only artifact checks. Give the dedicated rendering fixture visible shape geometry; retain strict ink, source-byte, resident, window and browser-log oracles. Verify XLS pool admission separately from real workbook parsing. Apply the existing native macOS warning contract consistently, retain raw diagnostics, and reject unknown stderr with explicit boundary controls.
Retain Linux source support while updating native dependency validation, schema tooling and ESM bootstrap compatibility. Keep packaged dependency/resource checks and strict loader diagnostics; permit only the known executable-labelled ASAR Stats deprecation.
Preserve attachment ordering, stale-result rejection, model snapshots, input/abort restoration and homepage interactions. Keep user-visible UI lifecycle tests with the corresponding renderer feature.
Keep Stop restoration scoped to the authored composer snapshot and expire only the stopped conversation's active questions. Owning renderer cases cover bus aborts, project conversations outside sidebar caches, failed-delete retry, delayed observer settlement after same-cid restore, and unrelated-conversation isolation; their shared runtime wiring is committed with Integration.
Add resumable Vitest, Python and Playwright batches with source and runtime invalidation, retained log receipts, artifact verification and bounded expiry. Keep non-model lanes serial and complete native fixture cleanup before reporting evidence failures.
Preserve JavaScript source-map literals and stateful log-declaration matching. Validation covers the final integrated working tree; unrelated runtime and fixture adaptations remain pending.
Establish every worksheet column layout before writing cells and formulas. Verify quoted, absolute and unquoted cross-sheet references after saving and reopening the real workbook.
Clean CI installs fail because the legacy archive service returns HTTP 403. Acquire the same seven size/SHA-256-pinned files from an immutable official Hugging Face revision, preserving the model and offline layout. Keep the ten-minute total acquisition budget, reject changed files immediately, and retain staged installation/rollback and legacy custom archive inputs.
Validation: 38 owning installer/resource/provisioner cases; both new requirement controls failed before the fix; fresh real download verifies the original complete payload. No library or embedding model upgrade.
Verify installed production dependency closure and packaged executable modes across supported targets. Preserve platform-owned launch behavior and drain independent local durability work once before ordinary exit.
Validation: affected native/runtime owners, shutdown/lifecycle cases, typechecks, smoke, and a normal visible macOS start/quit.
Reject web and cloud operations at execution and serving boundaries before side effects. Preserve historical data, local SDK execution, save recovery and standard localized download guidance.
Support permission, upload, dialog and phone passkey workflows. Keep browser cache and overlay ownership explicit across navigation, recovery and teardown.
Validate native payloads, dependency versions and packaged entrypoints before use. Preserve standalone source data roots and add deterministic log and subprocess verification.
Keep authorized business fields and lossless identifiers across merchant APIs. Share bounded metadata search, exact schema expansion and browse continuation across native, CLI and Web tools. Preserve cancellation, credential rotation, local encrypted grants and the public API-key boundary. Make official contract fixtures and generators self-contained; verify packaged entrypoints.
Prompt audit: Keep query/limit/offset, visible-scope keyword retrieval, small-catalog inline schemas and exact schema expansion in on-demand tool contracts and generated Web references. These explain required inputs and incomplete-schema recovery without adding resident instructions. Preserve existing permission and actor gates. Transport pagination, deadline, cache invalidation and retained-read ownership remain code-enforced.
Validation: connector owning checks, official fixture reproduction and public-content audit, bridge/search/schema cases, packaged gate and unified related suites. Full verification is recorded in the repository synchronization ledger.
Pin the desktop dependency tree to the supported Electron release while preserving the application release version. Update platform checks, packaged-resource validation, test cleanup, and log delivery support used by the desktop suites.
Synchronize the fixed 1.7.2 source range while preserving OrkasOpen's local-user, BYO-provider, and no-hosted-telemetry boundaries.
Prompt audit: retained shared runtime, UI, resources, locales, Web App SDK, Japan commerce, VoiceStudio, and regression coverage; dropped only hash-reviewed PC-only account, billing, managed-provider, analytics, release/evaluation, cloud-share, and private OCR owners.
Prompt confirmation: requester explicitly asked for the complete 1.7.2 sync without omissions.
* fix: verify native dependencies in Electron and support repair
* ci: capture native child diagnostics on repair failures
* fix: preserve known Linux sharp warnings and stable Windows ESM tests
* Isolate application sharp operations in bundled Node
* Test isolated SVG text rendering without native diagnostics
* Keep CI system fonts read-only and reject GUI GLib collisions
---------
Co-authored-by: CloudTianTian <[email protected]>
Retain raw Electron stderr and classify only bounded baseline D-Bus, font-cache and documented Sharp/GLib signatures on Linux. Unknown output, warning loops, timeouts and failed GUI cases remain fatal. Run diagnostic contract regressions in every native platform lane.
Validation: 34 owning cases passed; macOS native lane passed 1,258 tests and five real GUI checks with zero GUI diagnostic findings. Fixed-range synchronization postcheck passed. The initial Linux jobs passed 1,221 native tests and all five GUI assertions but failed on unclassified environment diagnostics; their full logs are retained and compared with the successful main baseline.
Preserve keyboard focus and single activation across task-board refreshes. Add task-scoped grants, partial-permission recovery, validated CLI attachments, device-local project directories, and the public model catalog update.
Prompt confirmation:
Confirmed-by: Requester in the current implementation task
Scope: Apply the existing approved shared agent, skill, task, and connector
capabilities within the documented open-source product boundaries.
Prompt audit:
- Keep: Commander and actor task identity, ordered execution, user input,
project context, and tool-use responsibilities in the shared prompt
templates and their existing assembly path. Type: resident.
Why: These clauses determine behavior across ordinary turns; the update
preserves the shared product's source-of-truth context and recovery
ownership without adding an independent open-source policy layer.
- Keep: Project-task, scheduling, memory, connector setup, UI artifact,
image/video composition, Office, and SEO workflows in their owning
skills and referenced guides. Type: on-demand.
Why: Load specialized procedures only for the corresponding work; retain
source metadata and dependency ownership rather than duplicate those
procedures in every actor prompt.
- Keep: Scoped tool schemas, queue/permission state, persisted recovery
evidence, output storage, resource manifests, and deterministic validators
at their existing host boundaries. Type: non-prompt enforcement.
Why: Code and structured events own the invariants and prevent the prompt
from becoming a second authorization, persistence, or execution engine.
- Drop: Hosted account, membership, cloud/relay, telemetry, internal debug,
managed voice, watermark, welcome, and official updater instructions or
affordances from the public execution paths. Type: source-only/on-demand.
Why: Those paths require services absent from the open-source product;
shared local operations and public API-key/BYO capabilities remain usable.
Validation: Recorded 11,721 deterministic tests, 538 resource tests, 1,248 native tests, five GUI checks and 198 reconciled desktop scenarios; final keyboard owner regressions also passed. Initial failed runs and classified warnings are retained in validation evidence.
Complete task details and sidebar navigation, guided connector setup, local agent input and cancellation, and bundled workflow updates. Preserve user-owned provider settings and Linux source support.
Prompt confirmation:
Confirmed-by: Requester in the current implementation task
Scope: Apply the existing approved shared agent, skill, task, and connector
capabilities within the documented open-source product boundaries.
Prompt audit:
- Keep: Commander and actor task identity, ordered execution, user input,
project context, and tool-use responsibilities in the shared prompt
templates and their existing assembly path. Type: resident.
Why: These clauses determine behavior across ordinary turns; the update
preserves the shared product's source-of-truth context and recovery
ownership without adding an independent open-source policy layer.
- Keep: Project-task, scheduling, memory, connector setup, UI artifact,
image/video composition, Office, and SEO workflows in their owning
skills and referenced guides. Type: on-demand.
Why: Load specialized procedures only for the corresponding work; retain
source metadata and dependency ownership rather than duplicate those
procedures in every actor prompt.
- Keep: Scoped tool schemas, queue/permission state, persisted recovery
evidence, output storage, resource manifests, and deterministic validators
at their existing host boundaries. Type: non-prompt enforcement.
Why: Code and structured events own the invariants and prevent the prompt
from becoming a second authorization, persistence, or execution engine.
- Drop: Hosted account, membership, cloud/relay, telemetry, internal debug,
managed voice, watermark, welcome, and official updater instructions or
affordances from the public execution paths. Type: source-only/on-demand.
Why: Those paths require services absent from the open-source product;
shared local operations and public API-key/BYO capabilities remain usable.
Validation: Recorded full desktop, deterministic, resource, native and GUI baseline for this exact content tree; retain classified environment diagnostics.
Add device-local CLI and merchant connector workflows with catalog-guided
setup, scoped authorization, encrypted credentials, and action confirmation.
Extend task queues, retry recovery, project to-dos, library processing, local
agent inputs, Office artifacts, and desktop notification/navigation handling.
Update shared agents, skills, tool contracts, resource fixtures, and coverage.
Preserve newer conversation drafts and reference identities across delayed
send acceptance and navigation. Publish process-test readiness atomically
so cleanup scenarios cannot observe a partially written PID record.
Retain the independent date version, Linux source runtime, existing BYO
provider policies, production API-key access, and open-source settings.
Prompt confirmation:
Confirmed-by: Requester in the current implementation task
Scope: Apply the existing approved shared agent, skill, task, and connector
capabilities within the documented open-source product boundaries.
Prompt audit:
- Keep: Commander and actor task identity, ordered execution, user input,
project context, and tool-use responsibilities in the shared prompt
templates and their existing assembly path. Type: resident.
Why: These clauses determine behavior across ordinary turns; the update
preserves the shared product's source-of-truth context and recovery
ownership without adding an independent open-source policy layer.
- Keep: Project-task, scheduling, memory, connector setup, UI artifact,
image/video composition, Office, and SEO workflows in their owning
skills and referenced guides. Type: on-demand.
Why: Load specialized procedures only for the corresponding work; retain
source metadata and dependency ownership rather than duplicate those
procedures in every actor prompt.
- Keep: Scoped tool schemas, queue/permission state, persisted recovery
evidence, output storage, resource manifests, and deterministic validators
at their existing host boundaries. Type: non-prompt enforcement.
Why: Code and structured events own the invariants and prevent the prompt
from becoming a second authorization, persistence, or execution engine.
- Drop: Hosted account, membership, cloud/relay, telemetry, internal debug,
managed voice, watermark, welcome, and official updater instructions or
affordances from the public execution paths. Type: source-only/on-demand.
Why: Those paths require services absent from the open-source product;
shared local operations and public API-key/BYO capabilities remain usable.
Validation: Full deterministic JavaScript and resource suites, platform
native and GUI checks, typechecks, startup smoke, public endpoint contracts,
Office artifact/fixture smoke, and desktop E2E are recorded with the frozen
tree with emitted-log review.
Cover both pinned Linux OfficeCLI targets and the Whisper degradation boundary. Add an Ubuntu x64 native OfficeCLI smoke and document the glibc-only source-build contract.
Improve agent orchestration, local integrations, persisted result handling, and renderer continuity.\n\nExpand deterministic coverage for packaged resources, recovery paths, and open desktop behavior.
wyt