review-diff
Pre-push diff review by the local GitHub Copilot CLI (gpt-5.6-sol, a separate-vendor LLM). Usage: /review-diff [BASE_REF | --cached] (default base=origin/main)
- License 许可
- Apache-2.0
- 最近更新
- 2026-11-11
- Stars
- ★ 45
- Forks
- 6
01 前言 Front Matter
- name
- outer-loop
- description
- Tim's outer loop. Retro across boards, sessions and PRs since the last run, then agreed setup changes through a PR and lead restarts.
- disable-model-invocation
- true
03 SKILL.md 正文
Tim started you as the outer loop: the only session that changes shared setup (~/dotfiles/agents, skills, omp config, the global AGENTS.md, ~/agent-system). Each run is one retro and its follow-through; nothing schedules you.
1. Retro since the last run
~/.local/state/outer-loop/last-retro holds the UTC start time of the last run; without it, cover the last 7 days. Note this run's start time now.
Gather the evidence since then. Fan bulk reading out to read-only scout subagents, one per board or source, each returning findings with paths:
- Boards.
easl board.listandeasl agent.list; per board, the goal note (easl object.find --key goal --board <id>), notes (easl object.find --type note --board <id>) and asks (easl ask list --board <id>). Setup proposals and repeated repair asks matter most. - Sessions. omp sessions under
~/.omp/agent/sessions/modified since the cursor. - PRs. Merged and failed PRs on Tim's repositories (
gh search prs --owner twaldin --updated ">=<cursor date>" --json url,state,title,repository) and their failed checks. - Incidents. Machine and tooling failures named in notes, asks and sessions, plus the job logs under
~/.local/state/.
Rank candidates with the categories in skill://retro, repeated failures first. Each candidate names its evidence, the proposed change, and what it replaces or deletes.
2. Decide with Tim
Present the ranked list, then ask one decision at a time in this conversation, with options and your recommendation. Tim's agreement here is the only approval: apply exactly what he agreed. Write instruction and skill text per writing-for-agents.
3. Apply through a PR
Per repository:
- Fetch, then work in a new worktree from
origin/main(git -C ~/dotfiles worktree add ~/worktrees/<branch> -b <branch> origin/main); the main checkout's uncommitted work is Tim's. - Make the agreed change and run the repository's tests (
agents/README.mdlists dotfiles' install and verify commands) throughmachine-ok-queue run -- …. - Open one PR. Review it with
interrogate, using a reviewer from a family other than yours; fix findings and re-review until clean. The PR's watch loop merges it after its checks pass. - Fast-forward the main checkout and install dotfiles changes:
python3 agents/install.pyto preview, then--apply, thenpython3 agents/verify.py. Remove your worktree.
4. Restart affected leads
For each lead whose behavior the change affects (easl agent.list): easl agent.restart --target <tile> --mode resume. A refusal means the lead is busy or Tim is typing; retry later, never with --force. Then send that lead one message saying what changed: easl tell <address> "<what changed and why>", with the lead's address from easl agent.list. Beyond restart announcements, message leads only on Tim's instruction.
5. Close
Write this run's start time as the cursor: mkdir -p ~/.local/state/outer-loop && printf '%s\n' '<start time>' > ~/.local/state/outer-loop/last-retro. Report the changes, PR links and restarted leads.
04 同仓库更多 Skills
scan-before-add
Use when selecting, comparing, or recommending third-party libraries, and before adding a new dependency (go.mod, package.json, pyproject.toml, requirements.txt, Cargo.toml, pom.xml, *.csproj, Gemfile, composer.json). Applies both when you choose the candidates and when the user names a package. Runs uzomuzo scan on the candidates and leaves a record of what was checked and who decided.
diet-remove
Remove a dependency identified by uzomuzo diet — analysis + issue (default) or direct PR
diet-evaluate-removal
Data-driven deep evaluation of whether a dependency is worth removing
diet-assess-risk
Deep-dive security risk analysis for a dependency flagged by uzomuzo diet
review-until-clean
1-shot to merge-ready (local-only) — Phase A (5-agent local review iteration) → Phase B (push + self-driven Copilot re-review iteration) → Phase C (reply+resolve all threads). One invocation drives a PR from local agent review through Copilot review convergence to merge-ready state. Runs from your own machine; there is no CI auto-fix.
plan-review
Have the local GitHub Copilot CLI (gpt-5.6-sol, a separate-vendor LLM) critique the latest plan-mode plan file before ExitPlanMode — a sanity check. Usage: /plan-review [<plan-path>]