mirror of
https://github.com/Heretek-AI/GDevelop-Unlocked.git
synced 2026-11-07 10:55:58 -10:00
BillyOutlast
30da3f8b4e
Replaces the 205-file fork delivery vehicle with patches/ + launcher/ + versioned packages, per the implementation plan: Phase 1: seam tripwires (27 tests) freezing every surgical target in upstream (Usage predicate, %16 gate, SubscriptionChecker, debugger, LoadingScreen branding, Electron hook surface, no-plugin-registry verdict). Phase 2: 5 generated Tier-1 hunks (round-trip verified; 3 files reach byte-identical fork parity) + apply driver (dry-run gate, marker verify, asar extract/repack). Phase 3: launcher hook (preload fetch/XHR spoof of the two usage GETs, BrowserWindow wrap, webRequest fallback, updater pin) + 06 loader patch. Module._load interception dropped: renderers run a webpack bundle, so the spoof rides the axios/XHR path the provider hydrates from instead. Phase 4: Tier-2 verbatim extraction (CustomAIClient + 12 Studio modules + offline mock harness, manifest-pinned) + 07/08 verbatim glue patches + pure generation shim with fail-closed leak guard. Phase 5: web userscript demo (Tier-1 only, documented CORS/LAN/churn limits). Phase 6 runbook + upstream provider-interface RFC draft included; fork archiving left to owners (destructive). Suite: 66 tests green (seams, patcher incl. re-apply refusal, launcher incl. vm-sandbox fallback eval, tier-2 provenance, web).
2.2 KiB
2.2 KiB
packages/ai-studio — Tier-2 extraction (Phase 4, issue #4)
~8.9k-line BYOK AI + multi-agent Studio, moved verbatim out of the
fork. The fork is a donor, not a delivery vehicle. Never hand-edit
files under AI/ or AiGeneration/ here — re-run the extractor:
node scripts/extract-tier2.mjs --fork <fork-checkout>
MANIFEST.json pins every file (sha256); tests/tier2/extract.test.mjs
fails if any byte drifts from the donor.
Contents
| Path | Donor | Notes |
|---|---|---|
AI/CustomAIClient.js |
newIDE/app/src/AI/CustomAIClient.js (5870 lines) |
OpenAI-compatible client, 39-tool GDEVELOP_OPENAI_TOOLS, SSE streaming, gd-custom-ai-* localStorage keys. Zero api.gdevelop.io references (pinned by test) |
AI/CustomAIClient.spec.js |
co-located spec | Run inside newIDE/app (react-app-rewired test), not here |
AiGeneration/Studio/*.js (12 prod) |
AiGeneration/Studio/ (2371 LOC) |
Roles, PlanStore, LoopGuard, SpawnSubAgents, FinalizeSubAgents, UseStudioRuntime, UseStudioNudge, NudgePolicy, RequestWriteGate, RoleToolPolicy, SafeTruncation, EditApprovalLabel |
dev-harness/mock-ai-provider.js |
scripts/dev/mock-ai-provider.js |
Committed OpenAI-compatible mock for offline E2E (+ mock-script*.json fixtures) |
Wiring
Import rewiring is limited to packages/unlock-shim/generation-shim.mjs
(routing predicates + fail-closed leak guard) plus the generated glue
patches 07-generation-glue.patch (Generation.js branches) and
08-preferences-ai.patch (ai prefs tab + storage). The extracted
modules themselves are byte-identical to the donor.
Desktop E2E (requires a built app — procedure, not automation)
- Apply all patches to a checkout (
node scripts/apply-patches.mjs --target …). - Start the offline harness:
node packages/ai-studio/dev-harness/mock-ai-provider.js(or real Ollama athttp://localhost:11434/v1; setORIGINSfor CORS). - In Preferences → AI, set Base URL to the harness, complete a generation.
- Assert: subagent spawn/finalize loop runs (Roles/PlanStore/LoopGuard
paths); zero requests to
api.gdevelop.io/generationin BYOK mode (the hook logs spoofed usage calls inwindow.__HERETEK__.spoofedRequests; generation traffic is observable in the harness log).