[MarkusNeusinger]

Summary

• Jules lane narrowed after three test rounds (2026-10-27 to 10-29). Jules now takes mechanically verifiable work (codemods, migrations, tests, lint fixes); behaviour-preserving refactors go to /gemini-task, because only three of nine page-split variants kept their page unchanged and steering all nine took many Claude turns. CLAUDE.md, .claude/guardrails.md and the gemini-task skill follow (owner's agreement of 2026-10-29).
• Claude's share cut. Every Monitor event is a Claude turn, so the skill watches a round with one low-noise Monitor: watch-sessions.py gains --changes-only and repeatable --allow <prefix> (flags scratch files such as fix_*.py or get_lines.sh), settles only after its PR checks ran, and rejects unknown flags. Prompts now forbid confirmation questions and string-sliced code edits; stuck sessions get a budget (one nudge, about 60 minutes in the provisioning queue); a REST fallback covers the MCP send failures.
• How to prove unchanged behaviour for refactors: a jsdom snapshot with the existing test's mocks (0 vs 206 differing lines decided DebugPage), a browser comparison built against the production API, a verbatim check of canvas paint callbacks, and the Flash check on the PR into main for loading and error states. Also recorded: Jules names branches <base>-<session id>, and workflow triggers come from the pushed commit, so branches cut before a CI change keep old triggers.

Plan

N/A

Test plan

• tests/unit/test_agent_instructions.py: 36 passed
• ruff check and ruff format --check on watch-sessions.py
• Watcher against real sessions: --allow flagged get_lines.sh on refactor(app): split PalettePage into hooks and sections #12451, --changes-only stayed quiet for running sessions, -h and unknown flags exit 2
• REST :sendMessage fallback from the skill returned HTTP 200
• uv run python -m tools.changelog check --base origin/main
• Gemini 3.8 Flash pre-PR check: its five findings fixed (unset $D, settle before PR check, grep dropping PR lines, single-dash flags, missing kill)

[Copilot]

🟡 Changes recommended

Active critical and moderate findings remain in the watcher’s error-status handling, option parsing, and synchronized routing documentation, so the change is not ready for approval.

3 open issues

• Preserve watcher exit status through grep pipeline · New
• Reject missing or option-like parser values · New
• Synchronize companion guides for narrowed Jules routing · New

What changed in this PR

This PR narrows Jules assignments to mechanically verifiable work, routes larger behavior-preserving refactors to Gemini, and reduces monitoring overhead through watcher, prompt, fallback, and verification updates.

Changes:

• Updates Jules/Gemini routing guidance, guardrails, and operational documentation.
• Adds quieter session monitoring, path allowlists, stricter argument handling, and PR validation.
• Documents branch conventions, fallback messaging, and unchanged-behavior verification procedures.
• Final review findings remain unresolved: credential exposure in the REST fallback (moderate, 1 vote), lost watcher exit status (critical, 2 votes), option-value parsing accepting unknown flags (moderate, 3 votes), and an outdated synchronized routing guide (moderate, 2 votes).

File	Description
CLAUDE.md	Updates agent-selection guidance to restrict Jules to mechanically verifiable work.
changelog.d/​jules-skill-round3.md	Documents the Jules lane, watcher updates, and verification results.
.claude/​skills/​jules-task/​SKILL.md	Defines Jules task boundaries, monitoring, prompts, fallback handling, and review procedures.
.claude/​skills/​jules-task/​scripts/​watch-sessions.py	Adds quiet monitoring, path allowlists, argument parsing, and PR validation.
.claude/​skills/​gemini-task/​SKILL.md	Routes larger behavior-preserving refactors to Gemini.
.claude/​guardrails.md	Aligns delegation rationale and cost guidance with the revised Jules/Gemini boundaries.

_{Review effort: Lite}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!