source-control:babysit-prs: 'source-control-babysit-resolve-thread' bare wrapper not on PATH (command not found), worker must fall back to raw GraphQL

[kyle-sexton]

lane=babysit

Category: operational

Context: The babysit-prs worker contract (and the review-discipline thread-resolution step) instructs workers to resolve review threads via a bare wrapper invocation:

source-control-babysit-resolve-thread melodic-software/claude-code-plugins#<pr> --allowed-owners <...> --resolve --include-human

Triggering example (PR #470, worker run 2026-07-19):

$ source-control-babysit-resolve-thread melodic-software/claude-code-plugins#470 --allowed-owners kyle-sexton,melodic-software --include-human
/usr/bin/bash: line 1: source-control-babysit-resolve-thread: command not found
(exit 127)

Observed: The bare command is not resolvable on PATH from the worker's Bash tool environment (Git Bash on Windows). Exit 127, command not found. This happens in both list-only and --resolve forms.

Expected: Either (a) the wrapper is on PATH / resolvable as a bare command exactly as the worker contract documents, or (b) the contract documents the concrete invocation path (e.g. an absolute scripts/... path or python/bash prefix, matching how fetch-all-pr-comments.sh and babysit-readiness-gate.sh are specified as absolute cache paths).

Impact / workaround: Non-blocking. Worker fell back to raw GraphQL (resolveReviewThread mutation via gh api graphql) after fetching threadIds and applying the author-conditional (bot-only opening-comment) manually. This works but loses the wrapper's built-in --allowed-owners guardrails and the JSON action receipt the contract tells workers to parse, so each worker re-implements the safety logic ad hoc.

Suggested fix location: babysit-prs worker contract / SKILL doc where the bare wrapper is referenced — align it with the absolute-path convention already used for the sibling scripts, or ship the wrapper onto PATH.

[kyle-sexton]

This was generated by AI during triage. · Category: operational (bug) · Priority: priority: high (source-control north-star; each worker re-implements safety logic ad hoc without the wrapper). Routed status: ready.

Plan

source-control-babysit-resolve-thread is invoked as a bare command in the worker contract + the review-discipline thread-resolution step, but it is not on PATH (exit 127, Git Bash on Windows). Workers fall back to raw gh api graphql resolveReviewThread, losing the wrapper's --allowed-owners guardrails and the JSON action receipt the contract tells them to parse.

Fix (behavioral contract): align the wrapper's documented invocation with the absolute cache-path convention already used for its siblings (fetch-all-pr-comments.sh, babysit-readiness-gate.sh are specified as absolute paths, not bare commands). Update BOTH surfaces that reference the bare command — the worker contract and the review-discipline resolution step — to the concrete invocation path.

Acceptance criteria

• The worker contract + review-discipline step reference the wrapper by the same absolute/resolvable path form as the sibling scripts; a worker following the doc verbatim invokes it successfully (no exit 127).
• Workers keep the --allowed-owners guardrails and JSON action receipt instead of hand-rolling GraphQL.

Same bare-wrapper-invocation class as #475 (readiness-gate) — different wrapper; if a shared "how babysit scripts are invoked" convention is written, cover both.

Decision defaulted: document the concrete absolute invocation path (matching the sibling scripts) rather than shipping the wrapper onto PATH — veto before merge

[kyle-sexton]

Confirming example + downstream cost, babysit lane, 2026-07-20: this session ran ~14 worker dispatches across 3 cycles, every one of them hitting this exact gap — source-control-babysit-merge/source-control-babysit-resolve-thread not on PATH, forcing every worker to call python .../babysit_merge.py / babysit_resolve_thread.py directly.

New cost discovered: the operator had ALREADY pre-approved the wrapper command names in their Claude Code permission allowlist (Bash(source-control-babysit-merge:*), Bash(source-control-babysit-resolve-thread:*)) specifically so these narrow, safety-gated scripts wouldn't need per-call human confirmation. Because this bug forces the direct-python-invocation form instead, none of those pre-approvals matched, and the operator was hit with a permission prompt on effectively every merge-gate-check and thread-resolve call this session — a real "human out of the loop" regression caused entirely by this PATH gap, not by any deliberate safety design.

Operator worked around it today by additionally allowlisting the direct-script invocation forms (Bash(python */babysit_merge.py*) etc.) in ~/.claude/settings.json. That's a reasonable stopgap, but the actual fix — getting the wrapper names onto PATH — would let the pre-existing wrapper-name allowlist entries work as originally intended and avoid this class of workaround entirely.