Files
makelore/.project-docs/30-worklog/tasks/20260824-pi-runtime-unavailable-hotfix-4e9c7a31.md
T

204 lines
11 KiB
Markdown

# Task: Fix packaged Pi proxy token initialization and recovery
## Identity
- Task ID: 20260824-pi-runtime-unavailable-hotfix-4e9c7a31
- Mode: Feature
- Branch: codex/20260824-pi-runtime-unavailable-hotfix-4e9c7a31-pi-runtime-unavailable-hotfix
- Worktree: D:\Datas\OthersProjects\makelore-pi-runtime-unavailable-hotfix-4e9c7a31
- Base commit: fa510c49761485cbd323801923dbe2472d6981bb
- Owner: codex-root
- Status: Ready for Integration
## Scope
- Fix the Windows packaged Pi startup failure for the
`works_square_ai_gateway_proxy` account without changing the OpenCode-to-Pi
architecture or starting PI-160.
- Resolve the Main Host API credential lazily for persistent parent open,
rebuild/recover, and ephemeral child open; preserve Main ownership and prove
token rotation does not reuse a captured value.
- Map Pi's typed `PROVIDER_AUTH_REQUIRED` failure through the existing single
refresh/reopen boundary, and make Renderer Conversation recovery failure
leave `recovering` for a retryable per-Conversation error without losing
user state.
- Produce and verify a Windows x64 NSIS installer from the final clean
candidate, including a real packaged Main-composition proxy first-
Conversation proof and token non-disclosure checks.
## Intent And Constraints
- Base exactly on `fa510c49761485cbd323801923dbe2472d6981bb`, which is the
common HEAD of the PI-150 release-proof and diagnostic worktrees and contains
code candidate `09841c8`; do not rewrite or drop cumulative PI-150 history.
- Do not modify the old PI-150 worktree, diagnostic worktree, occupied/dirty
`main`, or the user's installed application under
`D:\Tools\泥土\niancode\Makelore`.
- Keep the Host token Main-owned and memory-only. Never write it to argv,
`models.json`, logs, diagnostics, task evidence, or Renderer state.
- Do not upgrade Pi, add a fallback/compatibility layer, refactor Host auth, or
change direct API-key, OAuth, or local Provider behavior.
- Real Provider verification remains `Explicitly Waived / Accepted Risk` with
`realTurnVerified=false`; macOS and native non-WSL Linux gate status is
unchanged.
- Use the exact package-manager version pinned by `packageManager`, frozen
install semantics, focused regressions, full repository verification, and
the formal Windows packaging path.
## Project Context Loaded
- Concurrent Task Gate: Passed in the isolated worktree above; ownership is
`codex-root`, feature mode, and the task record matches the registry.
- Planning Gate: Passed after reading the diagnostic record, PI-150 source
record, memory index, positioning/current-state/decision/architecture/domain/
evidence/reflection/commitment/stale records, and all active peer scopes.
- The shared canonical snapshot is older than the Pi feature chain. The
diagnostic record, cumulative PI-150 source record, current source, and the
delegated repair boundary are authoritative for this task.
- Other active tasks use separate worktrees. The main-worktree integration task
concerns the older OpenCode model-switch path and explicitly does not package;
no semantic conflict requires changing this hotfix plan.
- No subagents are used by explicit user direction.
## Plan
1. Inspect the exact composition, opener, auth-classifier, recovery-store, test,
Windows Electron E2E, and packaged proof seams; identify the smallest
production and test surface.
2. Add red-capable focused regressions for lazy/current token reads across
parent/rebuild/recover/child, typed auth mapping and bounded refresh, recovery
failure/retry, and unaffected direct/OAuth/local Provider paths.
3. Implement the lazy Main-owned credential getter, typed error classification,
and retryable recovery-state transition with surgical changes only.
4. Run focused tests, typecheck, lint, the full unit suite, production build,
and Windows Electron E2E; fix only failures caused by this hotfix.
5. Commit the clean implementation candidate, run formal `package:win`, execute
Windows artifact/runtime closure checks, and run a final packaged Main-
composition proxy first-Conversation proof including token-secrecy and clean
process-exit assertions.
6. Record exact commits, changed files, verification results, NSIS size/hash,
structured packaged evidence and unchanged release waivers; pass the Task
Documentation Gate and mark the task ready for integration.
## Outcome
- Implemented a lazy `getLocalProxyCredential` seam from Electron Main through
the single coding composition. Composition construction no longer reads or
captures the pre-server token; persistent parent first-open, recover/rebuild,
and ephemeral child open all resolve the getter at the worker-open boundary.
- Extended the existing typed Provider-auth classifier to recognize only
`PiProviderConfigError.code === PROVIDER_AUTH_REQUIRED`. Initial prepare and
prompt authentication failures now use the existing maximum-one refresh and
reopen coordinator and project as `CODING_PROVIDER_AUTH_REQUIRED`.
- Made Renderer recovery failures leave `recovering` for the target
Conversation's retryable `error` state while preserving the selected
Conversation, last good Snapshot, draft, and attachments; a later retry can
recover normally.
- Added focused rotation/non-disclosure regressions for parent first-open,
recover, rebuild, and child open while retaining the pre-existing direct
API-key/OAuth/local credential coverage.
- Extended the final packaged proof entry so the actual global Main composition
can exercise `works_square_ai_gateway_proxy` through the current authenticated
Host proxy, create the first Conversation from the real UI, establish a Pi
binding, accept input, dispatch a real packaged parent and child, and verify
argv/models/log/diagnostic token non-disclosure. Its upstream is controlled
loopback and the report remains `realTurnVerified=false`.
- The first packaged proxy run additionally proved that Host authentication was
accepted but exposed a second Windows release blocker: Pi's Node fetch adds
`sec-fetch-mode: cors`, and forwarding that transport-owned header into
Electron `net.fetch` fails with `net::ERR_INVALID_ARGUMENT` before the
upstream receives the request. The Host proxy now drops only that header;
direct API-key/OAuth/local Provider behavior and the no-mutation-replay rule
are unchanged.
- Generated and verified the final Windows x64 NSIS from clean candidate
`34a4434cfb3b6d083259f27dde065e78d4e0054f`. The final packaged global-Main
proxy proof passed and all proof Electron/Pi processes exited.
## Commits
- `f902edefd0026fc8b003d4769c4eb1c064243880` — lazy Main Host token,
typed Provider auth mapping, retryable recovery state, focused regressions,
and the packaged global-Main proxy proof path.
- `34a4434cfb3b6d083259f27dde065e78d4e0054f` — drop Pi/Undici's
`sec-fetch-mode` transport header before Electron upstream forwarding and use
an actionable real-UI send in the packaged proof harness.
## Changed Files
- `.project-docs/30-worklog/tasks/20260824-pi-runtime-unavailable-hotfix-4e9c7a31.md`
- `electron/api/coding-composition.ts`
- `electron/api/coding-provider-auth.ts`
- `electron/api/routes/ai-proxy.ts`
- `electron/coding-runtime/pi/release-proof.ts`
- `electron/main/index.ts`
- `scripts/run-pi-subagent-packaged-smoke.mjs`
- `src/stores/coding-conversations.ts`
- `tests/unit/ai-proxy-routes.test.ts`
- `tests/unit/coding-conversations-store.test.tsx`
- `tests/unit/coding-core-routes.test.ts`
- `tests/unit/coding-provider-auth.test.ts`
- `tests/unit/pi-managed-worker-opener.test.ts`
- `tests/unit/pi-runtime-auth-recovery.test.ts`
- `tests/unit/pi-subagent-child.test.ts`
## Verification
- `corepack pnpm install --frozen-lockfile` with pinned pnpm `10.33.4`: passed,
997 packages linked from the existing store and the lockfile remained
unchanged.
- Focused Vitest run for proxy forwarding, composition, auth, Renderer recovery,
parent opener, child opener, and auth recovery: 7 files / 66 tests passed.
The initial hotfix red
run failed exactly on typed auth classification, auth-required projection,
and permanent `recovering` state.
- `corepack pnpm run typecheck`: passed after the final proof changes.
- `corepack pnpm run lint:check`: passed with 0 errors and 5 pre-existing React
warnings in `src/pages/Home/index.tsx` and `src/pages/Makelore/index.tsx`.
- `corepack pnpm test`: passed; primary run 178 files / 1512 passed / 2 skipped,
serialized pressure run 1 file / 1 passed.
- `corepack pnpm run build:vite`: passed; Renderer, Main, Preload, and release
utility bundles built. Existing dynamic-import and large-chunk warnings remain.
- `corepack pnpm run test:electron:windows`: passed; 2 files / 4 tests.
- `corepack pnpm run package:win`: passed from clean candidate
`34a4434cfb3b6d083259f27dde065e78d4e0054f`; produced
`Makelore-2.0.0-win-x64.exe`, 211,888,141 bytes, SHA-256
`3A1335A6DEFDEBE53A84EA558E6A000DBD538C568EB9AFB202893AA4D82858B2`.
- `corepack pnpm run verify:artifact:win`: passed. The unpacked product runs
Electron `43.4.0` / Node `24.18.1`; bundled Python, pip, sqlite3, SSL,
`uv 0.10.0`, npm `11.6.2`, msgpackr, and x64 canvas load from the product.
- `corepack pnpm run verify:artifact:pi -- --samples 2`: passed. Final
`app.asar`, Pi `0.84.2` CLI, all 130 expected production packages, 6 runtime
assets, 5 native assets outside ASAR, 4 managed Skills, and the materialized
extension/subagent contract are present and executable. The nested runtime
report remains `partial-pass` only for the pre-existing explicit real-
Provider/macOS/Linux waivers; the artifact verifier result is `pass`.
- `corepack pnpm run test:pi-subagent:packaged`: passed against the final
unpacked executable. The actual packaged `app.asar` Main composition used
`works_square_ai_gateway_proxy`, created exactly one first Conversation,
established its Pi binding, projected the submitted user input and
`REAL_PARENT_COMPLETE`, and settled at worker `ready` / run `idle`.
Parent and child each reached the controlled upstream through the current
authenticated Host proxy; the token was absent from argv, `models.json`,
logs, and diagnostics. The UI remained editable with neither runtime-
unavailable nor permanent-recovering state. Cleanup reported zero retained
workers, and an independent Windows process query found zero proof Electron/
Pi processes.
- Pre-final packaged attempts were not counted as passes: the first used an
unsupported optional embedded-commit assertion, and the real proxy run then
failed on `net::ERR_INVALID_ARGUMENT`; that failure directly produced the
`sec-fetch-mode` regression and the final package/proof rerun above.
## Follow-ups
- Real external Provider turn verification remains explicitly waived and its
concurrency, credential-isolation, and protocol-compatibility risk remains
accepted; the loopback Host-proxy proof must not be reported as a real
Provider pass.
- macOS and native non-WSL Linux release gates remain unchanged by this Windows
hotfix.
## Promotion Candidates
- None. This feature task records the hotfix and evidence locally; it does not
promote stale shared OpenCode-era canonical documents.