implementer-agent: raise timeout to 20min + add npm/pip cache guidance to save ~2-3min per run

[verkyyi]

Summary

The first real dispatch of the newly-installed implementer-agent on verkyyi/agentfolio hit the 15-minute cap with the implementation fully written and tests passing, but timed out 9 seconds before the draft PR could be created. Dogfood finding — two small changes to the catalog source would have let the run succeed and would save ~2-3min on every subsequent run.

Run: verkyyi/agentfolio run 24680708086 — job duration 15m44s, status failure (timeout).

Timeline from the failed run

Parsed from the Claude Code CLI event stream:

Window	Duration	What the agent was doing
17:29:27 – 17:30:36	69s	19 tool calls hunting for inputs.issue_number (dispatch inputs weren't propagated — separate user-side issue, fixed)
17:30:36 – 17:38:47	491s	Silent — single thinking block over ChatPanel.tsx (329 LOC) + test file (~500 LOC), planning 5 edits
17:38:47 – 17:39:28	41s	5 Edit calls + kicked off npm install + tests
17:39:28 – 17:43:00	211s	npm install --prefer-offline + first vitest run on cold node_modules
17:43:00 – 17:43:58	58s	Analyzed test failure, imported act, re-ran, 108/108 green
17:43:58 – 17:44:08	10s	git add + commit + branch check + ToolSearch for PR-create tool
17:44:08 – 17:44:17	9s	TIMEOUT before PR create executed

Two leverage points

1. Bump timeout-minutes: 15 → 20

The agent completed the implementation, fixed the test bug, got all 108 tests green, and started staging the commit. It missed the PR creation call by ~10 seconds. A 5-minute budget bump is the cheapest possible fix for this class of failure and directly handles the 8-minute silent-thinking case (big model on large files) without requiring a prompt or model change.

Diff: catalog/agent-team/implementer-agent.md

-timeout-minutes: 15
+timeout-minutes: 20

2. Add actions/setup-node cache guidance to the implementer's default steps (or prompt)

npm install on a cold runner burned 3.5 minutes. actions/setup-node@v4 with cache: 'npm' brings this to near-zero when package-lock.json hasn't changed. The same pattern would help pip/cargo/go mod/maven projects.

Two ways to ship this, in order of preference:

(a) Bake cache setup into the checkout/runner config block. If gh-aw exposes a hook to run setup steps before the Claude Code CLI, add a language-detection step that runs the appropriate setup-* action with caching. The implementer-agent.md frontmatter already has network: allowed: [defaults, node, python, rust, dotnet, java] — leverage that signal to conditionally enable caches.

(b) Tell the agent about it in the prompt body. Cheaper. Add a section to the implementer prompt like:

## Runner performance

Before running tests, check for cached dependency directories:
- Node: if `~/.npm` / `node_modules` exist, prefer `npm ci --prefer-offline` over `npm install`.
- Python: if `~/.cache/pip` exists, prefer `pip install -r requirements.txt --no-index --find-links`.

If no cache exists, runner setup cost is unavoidable. Don't spend more than 4 minutes on test setup — if tests are still building after 4 min, fall back to running only the tests you added (`npx vitest run <path-to-new-test>`), commit, and let CI run the full suite.

The 4-minute fallback is the important bit: the agent today doesn't know there's a wall clock, so it happily waits 3.5 min for a slow install. Telling it to bail to a narrower test run preserves the discipline (TDD green before commit) without eating the whole budget.

Why not reduce the 8-min silent-thinking block

That's genuine model-time for claude-sonnet-4-6 reasoning over the files. Workarounds (smaller model, forced edit-decomposition) trade depth for speed and are probably not worth it for most tasks. The timeout bump covers this case without behavior change.

Suggested acceptance

• catalog/agent-team/implementer-agent.md frontmatter: timeout-minutes: 20.
• Prompt body gets the "Runner performance" note (or cache setup gets baked in if feasible).
• README or catalog/agent-team/README.md documents the budget: spec ~4min, planner ~7min, implementer ~20min, reviewer ~10min — so users know what to expect.

Tangential

gh-aw#27407 covers the unrelated SHA-pinning bug that caused the first dispatch to fail for a different reason. That issue is independent of this one, but both surfaced in the same dogfood session.