fix(#914): add Skill to triage and prioritize tools

[fullsend-ai-coder]

Summary

Since Claude Code v2.1.119 (enforced on 2026-07-06), frontmatter tools: is a hard clamp on agent init tools. The triage and prioritize agents list skills in their skills: field (issue-labels and customer-research respectively), but their tools: field was Bash-only — the Skill tool was absent at runtime, making the declared skills silently inert.

This PR adds Skill to the tools: frontmatter in both agents/triage.md and agents/prioritize.md, restoring skill activation alongside the existing Bash scoping.

Not modified: agents/scribe.md — it has skills: [] (no skills to load) and is explicitly out of scope.

Changes

• agents/triage.md: tools: Bash(gh,curl,jq) → tools: Bash(gh,curl,jq),Skill
• agents/prioritize.md: tools: Bash(gh,curl,jq) → tools: Bash(gh,curl,jq),Skill

Testing

• Secret scan: passed
• make test: all tests pass (pre-existing failures in post-retro-test.sh are unrelated to this change)
• Pre-commit: could not run due to sandbox network restriction (exit 3); post-script runs it authoritatively

---

Closes #914

Post-script verification

• Branch is not main/master (agent/914-restore-skill-tool)
• Secret scan passed (gitleaks — 7ad7b5000373150ce1bf1a312dd6713045bea8f8..HEAD)
• PR body secret scan passed (gitleaks — no-git)

[fullsend-ai-review]

🤖 Finished Review · ✅ Success · Started 11:28 AM UTC · Completed 11:40 AM UTC

Commit: ab2d45d · View workflow run →

[fullsend-ai-review]

Review

Findings

Medium

• [protected-path] agents/triage.md, agents/prioritize.md — Both changed files are under the agents/ protected path. The PR is linked to issue Triage and prioritize cannot open listed skills (Skill tool missing from tools:) #914 and the rationale is well-documented (restoring Skill tool access for agents that declare skills in their frontmatter). Human approval is always required for changes to protected paths, regardless of context.

[ben-alkov]

LGTM

[fullsend-ai-retro]

🤖 Finished Retro · ✅ Success · Started 6:58 PM UTC · Completed 7:10 PM UTC

Commit: ab2d45d · View workflow run →

[fullsend-ai-retro]

Retro: PR #917 — Add Skill to triage and prioritize tools

Timeline

• 11:12Z — Issue #914 opened with detailed evidence that triage and prioritize agents had silently lost Skill tool access for 6 weeks (since 2026-07-06)
• 11:13Z — Both code and triage agents triggered simultaneously (ready-to-code and ready-for-triage labels both present at issue creation)
• 11:19Z — First code run cancelled after 6 minutes when triage re-applied ready-to-code, spawning a second code dispatch that superseded it via concurrency group
• 11:20–11:26Z — Second code run succeeded in 3m35s ($2.01), creating PR fix(#914): add Skill to triage and prioritize tools #917 with the correct 2-line fix
• 11:28–11:40Z — Review agent ran ($2.63), found 1 medium protected-path finding, applied requires-manual-review
• 18:23Z — ben-alkov approved; 18:43Z — ascerra approved; 18:56Z — merged

What went well

• Zero rework. Code agent produced the correct fix on the first attempt — added ,Skill to tools: in both agents/triage.md and agents/prioritize.md, correctly excluded scribe.md.
• Efficient execution. 44 turns, $2.01, 3m35s for a surgical 2-line change.
• Review quality. The review agent correctly identified the protected-path concern and found no functional bugs, matching the human reviewers' assessment.
• Excellent issue quality. Issue Triage and prioritize cannot open listed skills (Skill tool missing from tools:) #914 provided 22/22 affected runs as evidence, contrast with pre-breakage runs, and full root cause analysis, giving the code agent unambiguous context.

Evidence for existing issues

• fullsend#945 (tool-skill consistency lint): This bug persisted for 6 weeks because no automated check validates that agents declaring skills: also include Skill in tools:. A platform-level lint would catch this class of regression at commit time.
• agents#256 / agents#379 (model specifier in sub-agents): During review, two sonnet sub-agents failed because claude-sonnet-4-5@20250929 was unavailable, requiring re-dispatch without model override. Added latency but did not affect review quality.
• fullsend#766 (dispatch dedup): The first code run was cancelled after 6 minutes when triage re-applied ready-to-code, wasting compute. The race between simultaneous ready-for-triage and ready-to-code labels is a concrete instance of the dispatch dedup gap.

Proposals filed

• Add CI test for agent tool-skill declaration consistency (in fullsend-ai/agents)