v6.7 Part 3: dispatcher subprocess-honesty (#33, #34, #65) - #13
Merged
Conversation
Three small infrastructure fixes that together close the gap between "the dispatcher's shell can do X" and "the worker subprocess can also do X". The 2026-06-09 v6.7 build chain hit all three failure modes — this PR ships the targeted fixes. Closes hermes-jarvis#33 (GH_TOKEN propagation at worker spawn) Closes hermes-jarvis#34 (respawn_guarded active_pr exempts review roles) Closes hermes-jarvis#65 (fabricated github-auth block claims rejected) Context: hermes-jarvis#61 (bootstrap-paradox case study) ## NousResearch#33: GH_TOKEN propagation `_inject_gh_token_into_env` is called just before each worker subprocess.Popen in `_default_spawn`. If neither `GH_TOKEN` nor `GITHUB_TOKEN` is set, fall back to `gh auth token` from the dispatcher's shell and inject the result. Silent on every failure (no `gh`, not logged in, timeout) — workers that don't need GitHub access are unaffected. Closes the "subprocess can't see macOS-keyring-backed gh auth even though the dispatcher can" pattern that bit JARVIS umbrella + Tchalla release-gate + Tchalla v6.7 in three separate incidents. ## NousResearch#34: respawn_guarded active_pr exempts review roles `check_respawn_guard` now skips the 24h `active_pr` guard for review- role tasks (tony, tchalla, vision, reviewer). Their entire job is to operate on PRs — a release-gate body legitimately cites the PR URL it is tasked to verify, and unblock comments often include the URL as evidence. The guard was a 17-tick `respawn_guarded` loop on a Tchalla re-review that only resolved when JARVIS kludged around it by spawning a duplicate card with the URL stripped from the body. Non-review roles (jarvis, friday, etc.) continue to honor the guard. ## NousResearch#65: fabricated github-auth block claims rejected `block_task` adds a gate (mirrors completion-gate pattern from Parts 1-2): if the reason matches `_AUTH_CLAIM_PATTERN` ("missing-github-auth", "gh auth login required", etc.) AND the dispatcher's own `gh auth status` succeeds, the block is rejected with `FabricatedAuthClaimError`. Task state unchanged; the worker has to surface the real cause. Tchalla on 2026-06-09 blocked his review with "gh CLI not authenticated; cannot run gh pr diff 42" — except PR NousResearch#42 didn't exist and the dispatcher was authed the whole time. The gate would have rejected that block, forcing an honest cause. The genuine subprocess-auth case (worker can't reach gh but parent can't either, e.g. `gh` not installed at all) is still accepted — gate only fires when `_dispatcher_gh_is_authed()` returns True. ## Tests 24 new tests in `tests/cli/test_v6_7_subprocess_honesty.py`: - TestInjectGhToken (7) — existing env, gh not installed, returncode nonzero, empty stdout, timeout, normal injection - TestRespawnGuardExemption (6) — tchalla/tony/vision with PR URL pass, jarvis/friday with PR URL still guarded, review role without URL passes - TestReasonClaimsMissingGhAuth (6) — pure pattern matcher - TestBlockGate (5) — fabricated rejected, genuine accepted, non- auth pass-through, empty reason skip, audit event recorded 47 passed across Part 3 + adjacent worker_exit_code + oneshot_runtime + dispatcher_heartbeat tests — zero regressions on related paths. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…nded respawn exemption Self-review of PR #13 surfaced two P1 issues that this commit fixes together. ## P1 — _AUTH_CLAIM_PATTERN was too permissive The original regex matched any occurrence of an auth-claim phrase within the reason — including the EXACT honest cause workers should surface. Specifically: "investigated: gh CLI is authenticated; the real issue is that PR 42 doesn't exist" This was a false positive: the gate would reject the worker's honest diagnosis. Other false positives included documentation mentions ("gh auth login is needed in setup docs") and handoff-context mentions ("used gh auth token earlier in the task"). Fix: anchor the pattern to the LEADING substring of the reason (optionally after a `cause:`/`blocker:`/`reason:`/`infra:` prefix). The 2026-06-09 Tchalla case ("missing-github-auth: gh token for ...") still matches; the honest diagnosis above no longer does. 7 new tests in `TestAuthClaimNotLeading` lock in the contract. ## P1 — Respawn exemption was unbounded The review-role exemption in `check_respawn_guard` correctly stopped the 17-tick `respawn_guarded` loop, but it had no upper bound — a flapping review profile (crashes-fast-each-attempt) could respawn unconditionally on the URL signal, burning tokens until the auto-block path noticed. Fix: the exemption only applies while consecutive_failures < max_retries (the same brake every other path observes). Once the breaker is exhausted, fall through to the `active_pr` check and let auto_block catch it shortly after. 2 new tests in `TestRespawnExemptionBounded`: - Under-limit review stays exempt (existing behavior preserved) - At-limit review falls through to active_pr (new bound enforced) ## Tests 33 passed in tests/cli/test_v6_7_subprocess_honesty.py — up from 24 in the original PR #13 (9 new). Zero regressions on adjacent paths. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
3 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Three small infrastructure fixes that together close the gap between "the dispatcher's shell can do X" and "the worker subprocess can also do X". The 2026-06-09 v6.7 build chain hit all three failure modes — this PR ships the targeted fixes.
Independent of #11 and #12 (touches different code paths: worker spawn, respawn guard, block_task).
Closes hermes-jarvis#33 (GH_TOKEN propagation at worker spawn)
Closes hermes-jarvis#34 (respawn_guarded `active_pr` exempts review roles)
Closes hermes-jarvis#65 (fabricated github-auth block claims rejected)
NousResearch#33: GH_TOKEN propagation
`_inject_gh_token_into_env` is called just before each worker `subprocess.Popen` in `_default_spawn`. If neither `GH_TOKEN` nor `GITHUB_TOKEN` is set, fall back to `gh auth token` from the dispatcher's shell and inject the result. Silent on every failure (no `gh`, not logged in, timeout) — workers that don't need GitHub access are unaffected.
Closes the "subprocess can't see macOS-keyring-backed gh auth even though the dispatcher can" pattern that bit JARVIS umbrella + Tchalla release-gate + Tchalla v6.7 in three separate incidents in 48 hours.
NousResearch#34: respawn_guarded active_pr exempts review roles
`check_respawn_guard` skips the 24h `active_pr` guard for review-role tasks (tony, tchalla, vision, reviewer). Their entire job is to operate on PRs — a release-gate body legitimately cites the PR URL it is tasked to verify, and unblock comments often include the URL as evidence. The guard was a 17-tick `respawn_guarded` loop on a Tchalla re-review that only resolved when JARVIS kludged around it by spawning a duplicate card with the URL stripped from the body.
Non-review roles (jarvis, friday, etc.) continue to honor the guard.
NousResearch#65: fabricated github-auth block claims rejected
`block_task` adds a gate (mirrors completion-gate pattern from Parts 1-2): if the block reason matches `_AUTH_CLAIM_PATTERN` ("missing-github-auth", "gh auth login required", etc.) AND the dispatcher's own `gh auth status` succeeds, the block is rejected with `FabricatedAuthClaimError`. Task state unchanged; the worker has to surface the real cause.
Tchalla on 2026-06-09 blocked his review with "gh CLI not authenticated; cannot run gh pr diff 42" — except PR NousResearch#42 didn't exist and the dispatcher was authed the whole time. The gate would have rejected that block, forcing an honest cause.
The genuine subprocess-auth case (worker can't reach gh but parent also can't, e.g. `gh` not installed at all) is still accepted — gate only fires when the dispatcher itself IS authed.
Test plan
🤖 Generated with Claude Code