Spike: empirically verify Task/Skill invocation + cp -a survival in claude-code-action@v1

[cbeaulieu-gt]

Why

Implementation planning for #259 (overlay reachability) surfaced two unknowns that no source-reading or documentation search could resolve:

1. Does claude-code-action@v1 recognize Task and Skill as valid tokens in --allowedTools? Investigation found zero references to either token in the action's source, test suite, or documentation. The grammar parses them as valid strings, but recognition by the underlying SDK (@anthropic-ai/claude-agent-sdk, npm-package black box) is unverified.
2. Are sub-agents dispatched via Task (if recognized) able to use tools beyond the parent's allowlist? If sub-agents inherit the parent's narrow review allowlist, inquisitor cannot use Bash or Write and is functionally useless — collapsing the entire "different eyes" architectural premise.

Without empirical answers to both, the overlay images deliver no real benefit beyond a baked CLI binary and persona file — at which point we have a separate conversation about whether the overlay infrastructure is justified at all. This spike answers the existential question before we commit to #259's implementation path.

A third question rides along cheaply because the same run will produce the data:

3. What does $HOME/.claude/ actually contain after a real claude-code-action@v1 run? Investigation A from Make baked overlay tree (agents/skills/plugins/hooks) reachable from CLI at job runtime #259 verified the action only writes settings.json during setup, but the CLI subprocess's writes are unobservable from existing logs. A diagnostic step lists the post-run state.

What

A throwaway diagnostic PR that:

1. Adds a temporary diagnostic step in pr-review/action.yml that runs ls -laR $HOME/.claude/ (a) before the persona-install step, (b) after cp -a runs, (c) after claude-code-action@v1 finishes. Output goes to the run log.
2. Adds cp -a /opt/claude/.claude/. $HOME/.claude/ (the proposed reachability fix) gated behind a SPIKE=true env var so the change is reversible.
3. Adds Task,Skill,Read,Grep,Glob to the review action's --allowedTools.
4. Modifies the persona prompt or system prompt for the spike run to explicitly request sub-agent dispatch — e.g., "use the Task tool to dispatch the inquisitor sub-agent and have it report its available tools."

The spike branch is spike-overlay-reachability-empirical off main. No tag moves; no @v2 changes. Triggered via a test PR in mom-bot (or self-dogfood here) whose caller workflow temporarily references the spike branch SHA via glitchwerks/github-actions/.github/workflows/claude-pr-review.yml@<spike-sha>.

Measurement plan

For each spike run, capture and report:

• Pre-cp $HOME/.claude/ listing — confirms the action's setup state before our copy
• Post-cp $HOME/.claude/ listing — confirms cp -a completed; counts agents/skills copied
• Post-action $HOME/.claude/ listing — identifies what claude-code-action's CLI subprocess wrote (collision check for Make baked overlay tree (agents/skills/plugins/hooks) reachable from CLI at job runtime #259 charge 2)
• Whether Task tool appears in the action's available-tools advertisement to the LLM
• Whether the LLM successfully invokes Task to dispatch inquisitor
• If dispatched, what tools inquisitor reports as available — proves or disproves sub-agent inheritance
• Final review comment posted by inquisitor (or absence thereof) — the user-visible signal

Acceptance

The spike is complete when one run of the diagnostic produces enough log output to answer:

1. Does Task work in non-interactive claude-code-action@v1? YES / NO / partial
2. Does sub-agent inheritance allow useful work, or does the parent allowlist starve the sub-agent? YES / NO / TBD
3. What paths under $HOME/.claude/ does the CLI subprocess write? List them.

The spike does not ship to consumers. After acceptance, the spike branch is closed without merging; findings are recorded as a comment on this issue and folded into #259's plan revision.

Decision tree on findings

Q1	Q2	Implication for #259 and runtime overlays
YES	YES	#259 plan revision proceeds; ALLOWED_TOOLS fix is real
YES	NO	#259 plan revision must add per-sub-agent allowlist mechanism (claude-code-action input or wrapper)
NO	n/a	Existential question for runtime overlays. If non-interactive claude-code-action@v1 cannot dispatch sub-agents, the overlay images deliver only persona + CLI caching. Re-evaluate whether overlay infrastructure is justified.

Out of scope

• Any production behavior change. Spike is throwaway.
• Plugins / hooks reachability — separate from this spike (also unresolved per Make baked overlay tree (agents/skills/plugins/hooks) reachable from CLI at job runtime #259 investigation, but doesn't block the existential question above).
• Smoke-test additions — those are part of Make baked overlay tree (agents/skills/plugins/hooks) reachable from CLI at job runtime #259's plan, not the spike.

Related

• Make baked overlay tree (agents/skills/plugins/hooks) reachable from CLI at job runtime #259 — overlay reachability remediation; this spike unblocks its plan revision.
• Verify whether baked agents/skills/plugins are reachable from the runtime CLI #245 — original investigation; closed.
• Evaluate bot-based enforcement for §10.2 'different eyes' guarantee (CODEOWNERS gap) #137 — "different eyes" enforcement; depends on Q1+Q2 resolving favorably.
• Adopt structured marker contract for quality-gate (replace grep) #185 — W3 cutover; same dependency.

🤖 Generated by Claude Code on behalf of @cbeaulieu-gt