Repository navigation
Adapts Cursor Pstack 0.15.36 in PR179.
Orchestrate reads live PR state, head and checks before reports and matches verification to that head. Worker cancellation needs concrete failure evidence, and recovery remains under the saved provider policy. The public orch CLI, stores, parent Git ownership, model settings and README remain unchanged. Cursor's incompatible CLI deletion and Project-specific mechanisms are excluded.
Both real hosts passed supervised checkpoint and controlled-process cases. They rejected an obsolete verdict, refreshed the stored head/frontier, kept a quiet process active and recorded a controlled failing process's termination. The permission case used supplied policy state, not a real provider denial. Claude emitted a native Skill event; Codex loading evidence is indirect. These cases do not prove real provider-agent cancellation, full-program autonomy, automatic selection, universal liveness or OS confinement. Tests used real user shell permissions.
Local checks passed 1,008 Bun tests, 367 actual pinned-TLC Statechart tests, strict typechecks, 27 maintenance, 8 hook and 16 documentation tests, static invariants, provenance and plugin validation. Exact-head CI and the independent verdict are linked from PR179. Historical local failures in issue141 remain unexplained.
All 650 candidate package files match the native-tested checkpoint. A fresh final-head private installer transaction verified them. The release tag is a tested checkpoint and will not be moved. The tag-pinned adaptation record records decisions and exclusions.
Rollback checkpoint: v1.17.7, commit b654aebbacdf4f6ba24ee19e468588ee3334dc46. Claude's actual installation is preserved. The authorized Codex upgrade is verified separately after publication.
Tested release checkpoint: a80f632878066f0291db9e583e76daf09702d303; package tree 4e916716fb3cbbcec3c99ec5a45b1ecaf33f598b; all 650 installed files verified in a fresh private transaction. Merged-checkpoint CI passed. Independent exact-head verdict and validation detail are in PR179.
Validation method clarification
The native Codex case used per-test danger-full-access, -a never and --ignore-rules. Claude used per-test sandbox.enabled:false, acceptEdits and fixed Python commands. These overrides did not change personal host settings. The scope limits were instructions, without an OS-confinement guarantee.
Codex used a fresh private CLI installation with existing authentication through a temporary link, which was removed. Claude loaded the exact package with --plugin-dir; its global installation was preserved. The validation clarification records the review and lifecycle limits.