[dependabot]

Migrate all three pi packages together to the maintained @earendil-works namespace (0.79.10), preserving existing sessions, provider integration and native hosted web search. Integrate current main and refresh the native-event/message-replay patch against the new SDK. Keep the current monthly Dependabot groups and disabled Renovate configuration.

Align the desktop and nested pi Undici runtime at 8.12.4 with a scoped 8.x security floor, addressing advisories affecting the earlier 8.10.2 pin. Require Node >=22.19.0. The transitive OpenTelemetry API packages do not add an SDK or exporter. Existing lifecycle guards handle only the exact internal halt error; unrelated observer exceptions still propagate.

Validation on the updated tree: repository typechecks and lint, core/provider tests (including hosted-search protocol and patched SDK hook tests), HTTP runtime version/dispatcher regressions, and desktop production build passed. Production audit reports no remaining Undici advisories; unrelated baseline advisories remain. Full pre-push and remote integration checks are tracked on the latest head.

Supersedes the deprecated-namespace update in #544. The independent Undici 7.x security update #531 is already merged; this migration retains its security intent on the SDK-required 8.x line.

[github-actions]

Findings

No blocking, major, minor, or nit findings. The diff is a consistent, well-formed dependency alignment:

• pi-agent-core, pi-ai, and pi-coding-agent are all moved to ^0.73.1 together across apps/desktop/package.json:31, packages/core/package.json:16, and packages/providers/package.json:17, matching the changeset intent ("Keep the pi agent, model, and coding-agent packages on the same compatible version.).
• The new pi Dependabot group (.github/dependabot.yml:18) with pattern @mariozechner/pi-* will keep future bumps in lockstep.
• The changeset covers all three packages that carry the dependency, and the lockfile dedupes cleanly (drops the @mariozechner/jiti fork in favor of upstream jiti: 2.7.0, and folds yaml@2.8.4 into the already-resolved yaml@2.9.0).

Summary

Review mode: initial

• No correctness, security, license, or constraint violations introduced. @mariozechner/pi-* was already a bundled dependency, so this is a version bump, not a new license/size addition.
• Residual observation (not blocking): the @mariozechner/pi-* scope is marked deprecated in the lockfile in favor of @earendil-works/pi-* (pnpm-lock.yaml deprecated: notes on the 0.73.1 entries). Upstream states the rename is not yet published ("after the new package is published"), so staying on the current scope is correct here — consider a tracking issue to migrate when @earendil-works/pi-* ships.
• Residual risk: upstream v0.73.0 lists a breaking change to the built-in xiaomi provider (endpoint switch, display rename, changed XIAOMI_API_KEY meaning). This lives inside pi-ai and only matters if the app surfaces/configures xiaomi; worth confirming it is not advertised in the provider list.
• Version note: 0.72 -> 0.73 is a minor bump on a 0.x line and is effectively potentially breaking. packages/core and packages/providers bind directly to pi-ai / pi-coding-agent APIs, so this specifically needs a green typecheck + test run.

Testing

Not run (automation). Recommend CI run pnpm typecheck && pnpm test, and verify the provider connection-test / model-listing paths still work for at least one pi-backed provider given the direct type coupling in packages/core and packages/providers.

Open-CoDesign Bot

[dependabot]

Dependabot tried to update this pull request, but something went wrong. We're looking into it, but in the meantime you can retry the update by commenting @dependabot recreate.

[github-actions]

Findings

• [Minor] undici is now split across a major-version boundary between the app and the bundled pi runtime — apps/desktop/package.json:28 still pins undici: ^7.29.1 (lockfile resolves 7.29.1), while the newly bundled @earendil-works/pi-coding-agent@0.79.10 declares undici: 8.10.2 (pnpm-lock.yaml, snapshot @earendil-works/pi-coding-agent@0.79.10). undici is not in BUNDLED_RUNTIME_PACKAGES (apps/desktop/electron.vite.config.ts:19-29), so pi-coding-agent's undici import is externalized in out/main and resolves at runtime to the app's 7.29.1, not 8.10.2. This split is new: pi-ai 0.72.1 also declared undici: 7.29.1, so the two matched before. The undici@>=8.0.0 <8.10.2 override (package.json:41) only rewrites 8.x resolutions and does not align the app's own pin.
  Suggested fix / next action: bump apps/desktop to undici: ^8.10.2 (keeping the override as the floor), or add undici to BUNDLED_RUNTIME_PACKAGES so pi-coding-agent bundles the version it declares; then re-run the packaged build (pnpm --filter @open-codesign/desktop build:dir) to confirm the main-process bundle still loads.

• [Minor] PR title/body no longer match the diff, and the "no additional dependency or telemetry" statement is now inaccurate — the title says 0.73.1, while the manifests ask for ^0.79.0 and the lockfile resolves 0.79.10; the body says Mistral SDK 2.2.1 is retained and that "No additional dependency or telemetry setup is introduced". The lockfile instead resolves @mistralai/mistralai@2.2.6(@opentelemetry/api@1.9.0) and adds @opentelemetry/api@1.9.0 as a direct dependency of @earendil-works/pi-ai@0.79.10, plus @opentelemetry/semantic-conventions@1.43.0. Both are Apache-2.0 and, with no SDK/exporter registered, the OTel API surface alone does not transmit anything — so the hard constraint is not violated — but the claim as written will mislead a maintainer checking it against the project's BYOK / no-telemetry-by-default commitment (AGENTS.md).
  Suggested fix: retitle to 0.79.x / @earendil-works and rewrite the dependency paragraph to state the actual resolved versions and that only the OTel API package is pulled in transitively (no SDK, no exporter, no network).

• [Nit] renovate.json:18-20 still groups only @earendil-works/pi-ai and @earendil-works/pi-coding-agent, so Renovate can still open a lone @earendil-works/pi-agent-core bump — the exact split-version breakage this PR is fixing. The new Dependabot group (.github/dependabot.yml:18-20) correctly covers all three via @earendil-works/pi-*.
  Suggested fix: use "matchPackagePatterns": ["^@earendil-works/pi-"] (or add @earendil-works/pi-agent-core to matchPackageNames) so both bots keep the trio locked together.

• [Nit] engines.node is now >=22.19.0 (package.json:16) to match undici@8.10.2 and the new pi packages' engines, but CLAUDE.md documents Node as pinned via .nvmrc + engines and .nvmrc is not touched in this diff.
  Suggested fix: confirm .nvmrc (and any setup-node / node-version-file: .nvmrc use in CI) resolves to ≥22.19.0 so local and CI installs match the new floor.

Questions

• The rename appears complete across packages/*, apps/desktop, README*, CLAUDE.md, website/architecture.md, and renovate.json, but the old scope is now gone from every manifest. Can you confirm a repo-wide rg '@mariozechner/pi-' returns no import specifiers left (for example scripts/smoke-models.ts behind pnpm smoke, CONTRIBUTING.md, or workflow files)? Files outside this diff could not be verified from the public checkout.
• The changeset claims 0.79.x "fixes extension loading, credential-file, and HTML-export security issues". Those upstream release notes are not in the public repo, so this reviewer could not verify them; if they were checked against the published changelog, a one-line link would help future readers.

Summary

Review mode: follow-up after new commits

• The previous review's only residual item — staying on the deprecated @mariozechner/pi-* scope — is resolved: the migration to @earendil-works/pi-* is consistent across all three packages, the Electron bundler config, the lockfile, and docs, and .github/prompts/codex-pr-review.md was reviewed as data (a benign one-line namespace rename matching the code, not an instruction change to act on).
• The runtime change in packages/core/src/agent.ts:1334-1350,1429-1490 (turning the turn-boundary throw into a controlled halt, tracking agent_end per agent, and synthesizing the missing agent_end only when pi bypasses it) is directionally correct and well-narrowed; the added negative test in packages/core/src/agent.native-lifecycle.test.ts:494-509 covers the "unrelated observer failure must propagate" case. Residual risk: turnBoundaryHalt is reused across attempts on the same wrapper and is reset only by reassignment, and the synthesized-agent_end branch relies on existing repair-limit/tool-loop tests to exercise it — worth a targeted assertion that it fires exactly once.
• No new direct prod dependencies, no new license concerns (@opentelemetry/api / @opentelemetry/semantic-conventions are Apache-2.0), and the merge-relevant items above are the undici split and the stale PR description.

Testing

Not run (automation). CI on this head should include pnpm typecheck && pnpm test plus a packaged main-process build, since packages/core binds directly to pi-agent-core / pi-coding-agent APIs and the undici resolution only fails at package/runtime time.

Open-CoDesign Bot

[github-actions]

Findings

• [Minor] Bundled pi-coding-agent still resolves undici to the app's 7.x at runtime — apps/desktop/package.json:28 pins undici: ^7.29.1 (lock resolves 7.29.1), while @earendil-works/pi-coding-agent@0.79.10 declares undici: 8.10.2 in its lockfile snapshot. Because undici is a prod dependency of apps/desktop and is not in BUNDLED_RUNTIME_PACKAGES (apps/desktop/electron.vite.config.ts:19-29), electron-vite externalizes it, so the bundled pi code calls the app's 7.29.1 rather than the 8.10.2 it was built against. The new undici@>=8.0.0 <8.10.2 override (package.json:41) only rewrites 8.x resolutions and does not align the app's own pin. This is the one substantive item from the previous round that the new commits did not address.
  Suggested fix / next action: bump apps/desktop to undici: ^8.10.2 (keeping the override as a floor), or add 'undici' to BUNDLED_RUNTIME_PACKAGES so pi bundles the version it declares; then run pnpm --filter @open-codesign/desktop build:dir and one smoke generation to confirm the main-process bundle still loads.

Summary

Review mode: follow-up after new commits

• The new commits fix a real hole from the previous round. Because turnBoundaryHalt initializes to null, the earlier if (err !== turnBoundaryHalt) throw err; treated an observer that throws null (or undefined) as the controlled halt, swallowing it and fabricating an agent_end plus an "incomplete" result. packages/core/src/agent.ts:1476 now guards with turnBoundaryHalt === null ||, and packages/core/src/agent.native-lifecycle.test.ts:494-518 extends coverage to null/undefined while asserting no terminal event is synthesized. Behavior for the genuine halt path (settle the native agent, complete the transcript, forward exactly one terminal event) is unchanged.
• The previous round's main structural item is resolved: the @earendil-works/pi-* migration is consistent across all three packages, the Electron bundler config, the lockfile, and docs; .github/prompts/codex-pr-review.md was reviewed as data (a benign one-line namespace rename), and the changeset plus the new Dependabot @earendil-works/pi-* group cover the trio.
• Residual, non-blocking: renovate.json still groups only @earendil-works/pi-ai and @earendil-works/pi-coding-agent, so Renovate can still open a lone pi-agent-core bump (use matchPackagePatterns: ["^@earendil-works/pi-"]); the PR body still says Mistral SDK 2.2.1 is retained while the lock resolves @mistralai/mistralai@2.2.6(@opentelemetry/api@1.9.0) — the OTel API package alone has no SDK/exporter, so the no-telemetry constraint holds, but the description should state the resolved versions; .nvmrc is untouched while engines.node moved to >=22.19.0 — worth confirming the pinned dev Node meets it, though pnpm fails loudly on an engine mismatch rather than silently.
• No new direct prod dependency and no license concern (Apache-2.0 OTel API only). Apart from the undici resolution gap above, this looks ready to merge.

Testing

Not run (automation). CI on this head should include pnpm typecheck && pnpm test plus pnpm --filter @open-codesign/desktop build:dir, since packages/core binds directly to pi-agent-core / pi-coding-agent and the undici resolution only surfaces at package/runtime time.

Open-CoDesign Bot

[github-actions]

Findings

No blocking, major, or minor findings on this head. The single Minor from the previous review — apps/desktop resolving undici to 7.x while bundled pi-coding-agent@0.79.10 expects 8.10.2, so electron-vite's externalized import 'undici' would load the wrong HTTP runtime — is resolved:

• apps/desktop/package.json:28 now pins undici: "8.10.2".
• package.json:41 keeps a scoped floor override undici@>=8.0.0 <8.10.2 → 8.10.2.
• apps/desktop/src/main/undici-runtime.test.ts:12-27 asserts the app-resolved and nested-pi-resolved undici versions match, which fails on the prior 7.x/8.x mismatch and passes now — a genuine regression guard.

The packages/core/src/agent.ts controlled-halt change from the prior round (guard turnBoundaryHalt === null || err !== turnBoundaryHalt at agent.ts:1476, plus null/undefined observer-failure coverage in agent.native-lifecycle.test.ts:494-518) is unchanged and still correct: unrelated observer failures propagate, and only the exact internal halt synthesizes one terminal agent_end.

Residual observations (non-blocking, not merge blockers)

• renovate.json:18 still groups only @earendil-works/pi-ai and @earendil-works/pi-coding-agent, omitting pi-agent-core. .github/dependabot.yml now covers the full @earendil-works/pi-* pattern, but if Renovate is the active bot it can still open a lone pi-agent-core bump and reintroduce the exact version skew this PR removes. Consider "matchPackagePatterns": ["^@earendil-works/pi-"].
• The PR body says "Existing compatible Mistral SDK 2.2.1 is retained", but pnpm-lock.yaml resolves @mistralai/mistralai@2.2.6(@opentelemetry/api@1.9.0). The OTel package present is the API-only package (no SDK/exporter), so the no-telemetry constraint holds; just align the description with the resolved versions.
• apps/desktop/src/main/undici-runtime.test.ts:29-50's second case only asserts that a plain-HTTP fetch succeeds and that the dispatcher is swapped/restored; it does not prove native fetch actually routes through the injected undici dispatcher (an HTTP request would succeed either way). The version-equality test above is the meaningful guard against the previous regression — treat the bridge case as supplementary.

Summary

Review mode: follow-up after new commits

The prior round's one substantive item (undici app-vs-pi mismatch) is fixed and now covered by a regression test, so I'm not re-filing it. The pi migration remains internally consistent across packages/core, packages/providers, apps/desktop/electron.vite.config.ts (bundled-package list + treeshake path), the lockfile snapshots, the docs, and the changeset. @opentelemetry/api is API-only, so the default-no-telemetry constraint is upheld, and no new direct prod dependency pushes the desktop list toward the 30-dep cap. Remaining items are consistency/description polish only; this looks ready to merge subject to fresh CI on this head.

Testing

Not run (automation). CI on this head should run pnpm typecheck && pnpm test plus pnpm --filter @open-codesign/desktop build:dir, since packages/core binds directly to pi-agent-core/pi-coding-agent and the undici resolution surface only shows up at package/runtime time (the new undici-runtime.test.ts exercises it at unit level but not in the packaged main bundle).

Open-CoDesign Bot

[github-actions]

Findings

No blocking, major, or minor findings on this head. The only substantive delta from the prior reviewed commit is the Undici 8.10.2 → 8.12.4 bump (apps/desktop/package.json:28, package.json:41, .changeset/align-pi-packages.md:7), and the existing version-equality regression test still guards the app-vs-pi runtime match (apps/desktop/src/main/undici-runtime.test.ts:12-27).

Residual observations

• The lockfile still contains an undici@7.29.1 entry while the scoped override only covers >=8.0.0 <8.12.4. I could not determine from the provided diff which importer pulls that 7.x copy; if it is dev/website-only, no action is needed. If it can reach the packaged main process, the override/test should cover it too.

Summary

Review mode: follow-up after new commits

The previous round's Undici mismatch is resolved at 8.12.4, and the changeset, package manifest, and lockfile agree on that version. The pi namespace migration remains consistent across source, patches, docs, and the bundled-package list. I found no new correctness, security, license, or constraint issue introduced by this head. The remaining item above is a verification gap, not a merge blocker.

Testing

Not run (automation). CI should run pnpm typecheck && pnpm test and pnpm --filter @open-codesign/desktop build:dir; the new undici-runtime.test.ts exercises the app/pi version match at unit level but not the packaged main bundle.

Open-CoDesign Bot

[github-actions]

Findings

No blocking, major, or minor findings on this head.

Residual observations

• The only delta since the previously reviewed commit is dev/test-tooling lockfile churn: happy-dom 20.9.0 → 21.0.4 (packages/ui/package.json) and ws 8.20.0 → 8.22.2, which also pulls the transitive buffer-image-size@0.6.4. These are test-only (happy-dom/ws do not ship in the packaged main process), so no changeset is implied and the permissive-license rule for shipped deps is unaffected. CI will catch any happy-dom 21 behavior change in packages/ui tests.
• The earlier undici@7.29.1 observation is now explainable from the lockfile: the 7.x copy is @electron/get's install-time optionalDependencies entry (Electron binary downloader), not the packaged main process. The app external stays pinned at undici: 8.12.4 (apps/desktop/package.json), and undici-runtime.test.ts still asserts the app-vs-pi runtime match, so the residual is resolved as dev-only.
• Verification gap I could not close from the diff: package.json now requires node >=22.19.0, but the diff does not touch .nvmrc or the CI workflows. If the pinned .nvmrc/CI Node is older than 22.19.0, local dev and CI installs of undici@8.12.4 (engines >=22.19.0) will be inconsistent. Confirm those pins already meet the new floor; this is not a merge blocker.

Summary

Review mode: follow-up after new commits

The pi @earendil-works migration, the Undici 8.12.4 alignment, the >=8.0.0 <8.12.4 override, the Node >=22.19.0 engines bump, and the refreshed native-event patch remain internally consistent across manifests, lockfile, patches, and tests. The one prior residual (a stray undici@7.x) is confirmed to be dev-only Electron install tooling. I found nothing introduced by this head that should block merge; the only outstanding item is the .nvmrc/CI Node-floor check above, which is verification, not a code defect.

Testing

Not run (automation). CI should run pnpm typecheck && pnpm test and pnpm --filter @open-codesign/desktop build:dir; the new undici-runtime.test.ts covers the app/pi version match at unit level but not the packaged main bundle, and packages/ui tests cover the happy-dom 21 bump.

Open-CoDesign Bot

[github-actions]

Findings

No blocking, major, or minor findings on this head.

What the new commits changed

• The only delta since the last reviewed commit (f9b612d) is a scoped 7.x security floor: package.json:43 adds "undici@>=7.0.0 <7.30.1": "7.30.1", and the lockfile resolves the transitive @electron/get copy from 7.29.1 → 7.30.1 (pnpm-lock.yaml undici@7.30.1). This closes the prior residual (the 7.x copy is install-time Electron tooling only, not the packaged main process) without touching the app external, which stays pinned at undici: 8.12.4 (apps/desktop/package.json:29). The two overrides don't overlap — 7.30.1 is below the >=8.0.0 <8.12.4 range and 8.12.4 is outside the 7.x range — so there's no override conflict.

Residual observations

• The packages/core/src/agent.ts:1389-1544 turn-boundary-halt handling was already in the previously reviewed commit. The object-identity guard (if (turnBoundaryHalt === null || err !== turnBoundaryHalt) throw err) correctly distinguishes the controlled CodesignError halt from unrelated observer failures, and the synthesized agent_end only fires when pi bypassed it. This is covered by the unrelated-failure test case in agent.native-lifecycle.test.ts:494-513. No new concern here.
• Node-floor verification I still cannot close from the diff: package.json:16 now requires node >=22.19.0, but .nvmrc and CI workflow Node pins are not touched by this PR. If those pins are below 22.19.0, undici@8.12.4 (engines >=22.19.0) will install with a warning but fail at runtime on the old interpreter. This is verification, not a code defect, and does not block merge — but confirm .nvmrc/CI already meet the new floor.
• Pre-existing pattern carried forward: patchedDependencies keys @earendil-works/pi-ai@0.79.10 while the manifests declare ^0.79.0 (package.json:57). If a future install resolves ^0.79.0 past 0.79.10, the hosted-search onNativeEvent patch silently stops applying. This is the same fragility the 0.72.1 pin had, so it is not introduced by this PR — worth a pinned/synced note in patches/README.md rather than a code change.

Summary

Review mode: follow-up after new commits

The pi @earendil-works migration, the Undici 8.12.4 alignment, the refreshed native-event/message-replay patch, and the new >=7.0.0 <7.30.1 security floor are internally consistent across manifests, lockfile, patches, and tests. The prior residual about the stray undici@7.x copy is now resolved by the new override. The only outstanding item is the .nvmrc/CI Node-floor check, which is verification rather than a defect. I found nothing introduced by this head that should block merge.

Testing

Not run (automation). CI should run pnpm typecheck && pnpm test and pnpm --filter @open-codesign/desktop build:dir. apps/desktop/src/main/undici-runtime.test.ts asserts app-vs-pi Undici version parity and dispatcher bridging at unit level, but not the packaged main bundle; the packages/core native-lifecycle tests cover the controlled-halt vs. unrelated-failure paths.

Open-CoDesign Bot