菜单
Use when handing a submitted PR to Claude's optional cloud auto-fix after submit. Watches for CI failures and review comments, pushes fixes remotely so you can walk away. Runs automatically after /flux:work submits a PR (when enabled via /flux:setup). Also invocable manually via /flux:autofix.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
| name | flux-autofix |
| description | Use when handing a submitted PR to Claude's optional cloud auto-fix after submit. Watches for CI failures and review comments, pushes fixes remotely so you can walk away. Runs automatically after /flux:work submits a PR (when enabled via /flux:setup). Also invocable manually via /flux:autofix. |
| user-invocable | true |
Hands off a submitted PR to Claude's optional cloud auto-fix. Claude watches the PR remotely — fixing CI failures and addressing review comments — so you can walk away and come back to a green, ready-to-merge PR.
"This happens remotely so you can fully walk away and come back to a ready-to-go PR." — @noahzweben
Autonomous by default. When enabled via /flux:setup (autofix.enabled: true), auto-fix activates automatically after every /flux:work PR submit. No question asked — it just runs. Can also be invoked manually on any PR.
On entry, set the session phase:
PLUGIN_ROOT="${DROID_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT:-$(git rev-parse --show-toplevel 2>/dev/null || pwd)}}"
[ ! -d "$PLUGIN_ROOT/scripts" ] && PLUGIN_ROOT=$(ls -td ~/.claude/plugins/cache/nairon-flux/flux/*/ 2>/dev/null | head -1)
FLUXCTL="${PLUGIN_ROOT}/scripts/fluxctl"
$FLUXCTL session-phase set autofix
On completion, reset:
$FLUXCTL session-phase set idle
/flux:work, or user provides a PR URL)Claude's auto-fix subscribes to GitHub activity on the PR. When a check fails or a reviewer leaves a comment, Claude investigates and pushes a fix:
Review comment replies are posted using your GitHub account but labeled as coming from Claude.
When evaluating human or bot review comments, use flux-receive-review. Before posting that something is fixed, apply flux-verify-claims.
Full request: $ARGUMENTS
Accepts:
https://github.com/org/repo/pull/123)#123 or 123)When invoked automatically from /flux:work Phase 5, the config is already checked before invoking this skill — skip this step.
When invoked manually (/flux:autofix), always continue regardless of config — the user explicitly wants auto-fix on this PR.
If a PR URL or number was provided, use it directly.
If no arguments, detect the current branch's PR:
PR_URL=$(gh pr view --json url -q '.url' 2>/dev/null)
If no PR found:
No PR found on the current branch. Either:
- Run /flux:work to submit a PR first
- Provide a PR URL: /flux:autofix https://github.com/org/repo/pull/123
Stop here.
Auto-fix works by spawning an independent cloud session via claude --remote. If --remote fails (Claude GitHub App not installed, auth issue, etc.), claude will return an error — the skill logs it and continues to Reflect without blocking. This creates a new Claude session on Anthropic's infrastructure that clones the repo, subscribes to GitHub events on the PR, and runs autonomously. The local session continues to Reflect — both run in parallel.
# Verify claude CLI supports --remote
claude --help 2>&1 | grep -q "\-\-remote" && echo "remote supported" || echo "no remote"
If --remote is supported (standard path):
claude --remote "Watch this PR and auto-fix any CI failures or review comments. Only address NEW activity — CI check failures and review comments posted after this point. Do not re-address bot comments (Greptile, CodeRabbit) that were already resolved in earlier commits. PR: ${PR_URL}"
Tell the user:
Auto-fix session started for: {PR_URL}
A separate Claude session is now running in the cloud. It will:
- Fix CI failures (test errors, lint issues, type errors)
- Address review comments from human reviewers
- Handle bot feedback (Greptile/CodeRabbit) if configured
- Push fixes and re-run CI
You can walk away — the session runs independently.
Monitor it: run /tasks in the CLI, or open claude.ai/code.
If --remote is NOT supported (older CLI version or a non-Claude agent):
Auto-fix cannot run autonomously without --remote. Tell the user:
Auto-fix requires `claude --remote` which isn't available in your CLI version.
Options:
1. Update Claude: npm i -g @anthropic-ai/claude-code@latest
Then auto-fix will work automatically on next PR submit.
2. Start it manually from the web:
- Open claude.ai/code
- Paste: {PR_URL}
- Say: "Watch this PR and auto-fix any CI failures or review comments"
3. Start it from mobile:
- Open the Claude mobile app
- Say: "Watch {PR_URL} and auto-fix CI failures and review comments"
Proceeding to Reflect — auto-fix is not running for this PR.
In this case, skip auto-fix and continue to Reflect. Do NOT block the workflow.
Auto-fix is non-blocking. Whether it started successfully or not, the local session continues:
/flux:work → return control to the work skill, which proceeds to Reflect/flux:autofix) → offer next steps:Auto-fix is running independently. What's next?
a) /flux:reflect — capture learnings while context is fresh
b) /flux:scope — start new work
c) Done for now
Auto-fix fires once per epic, at the end of /flux:work Phase 5 (Ship) — immediately after the PR is created. The PR is always created at the end of the epic (after all tasks → epic review → quality → Ship). There is no earlier PR creation point in the standard Flux flow.
The claude --remote call spawns an independent cloud session. The local session continues to Reflect in parallel. Both run independently — the cloud session watches the PR, the local session captures learnings.
| Event | Claude's Response |
|---|---|
| CI check fails (lint, types, tests) | Investigates, pushes fix if confident, explains in session |
| Review comment (clear request) | Makes the change, pushes, replies on GitHub |
| Review comment (ambiguous/architectural) | Asks you before acting |
| Duplicate event | Notes it, moves on |
| No action needed | Notes it, moves on |
| Skill | Relationship | Overlap? |
|---|---|---|
/flux:work | Auto-fix activates after work submits a PR (Phase 5) | None — sequential |
/flux:gate | Gate validates staging after merge. Auto-fix gets the PR to merge-ready state before merge | None — different lifecycle stages |
/flux:epic-review | Epic review catches code quality issues before submit. Auto-fix catches CI/review issues after submit | See deconfliction below |
/flux:impl-review | Impl review is per-task, local. Auto-fix is per-PR, remote | None — different scope |
/flux:reflect | Reflect runs after auto-fix is enabled (they're independent) | None — independent |
Auto-fix and BYORB self-heal are mutually exclusive — they never both run. The routing is config-driven:
If autofix.enabled: true:
If autofix.enabled: false (or unset):
autofix.enabled: true autofix.enabled: false
───────────────────── ──────────────────────
epic-review: epic-review:
adversarial review adversarial review
STRIDE security scan STRIDE security scan
BYORB → SKIPPED BYORB → SKIPPED (no PR)
browser QA browser QA
quality → submit (PR created) quality → submit (PR created)
↓
auto-fix (cloud, unlimited): BYORB runs HERE (local, max 2 iterations):
✓ CI failures ✓ Greptile/CodeRabbit comments
✓ Human review comments ✗ CI failures (not handled)
✓ Greptile/CodeRabbit comments ✗ Human reviews (not handled)
reflect → done reflect → done
What each mechanism exclusively owns:
| Concern | autofix.enabled: true | autofix.enabled: false |
|---|---|---|
| Code quality | Adversarial review (before submit) | Adversarial review (before submit) |
| Security | STRIDE scan (before submit) | STRIDE scan (before submit) |
| Bot comments (Greptile/CodeRabbit) | Auto-fix (after submit) | BYORB (after submit, local fallback) |
| CI failures | Auto-fix (after submit) | Not handled automatically |
| Human reviewer comments | Auto-fix (after submit) | Not handled automatically |
autofix.enabled: true. Two autonomous processes run in parallel — Ralph locally, auto-fix in the cloud. This is fine; they don't interfere.ALWAYS run at the very end of command execution:
PLUGIN_ROOT="${DROID_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT:-$(git rev-parse --show-toplevel 2>/dev/null || pwd)}}"
CURRENT_VERSION=$(jq -r '.version' "${PLUGIN_ROOT}/.claude-plugin/plugin.json" 2>/dev/null)
MARKETPLACE_VERSION=$(jq -r '.plugins[0].version' "${PLUGIN_ROOT}/.claude-plugin/marketplace.json" 2>/dev/null)
if [ "$CURRENT_VERSION" != "$MARKETPLACE_VERSION" ] && [ -n "$MARKETPLACE_VERSION" ]; then
echo "Update available: v${CURRENT_VERSION} → v${MARKETPLACE_VERSION}. Run /flux:upgrade"
fi
Create structured build plans from feature, bugfix, upgrade, or refactor requests, or from Flow IDs. Use when planning implementation after the workflow envelope is explicit. Triggers on /flux:plan with text descriptions or Flow IDs (fn-1-add-oauth, fn-1-add-oauth.2, or legacy fn-1, fn-1.2, fn-1-xxx, fn-1-xxx.2).
Stakeholder-facing feature proposal flow. Guides non-technical team members through a conversational planning session — probes for business context, pushes back like an engineer, estimates complexity and cost, then documents the proposal and creates a PR for engineering handoff. Triggers: /flux:propose, or detected implicitly when a non-technical user describes a feature, bugfix request, or upgrade without implementation detail during /flux:scope or /flux:plan.
Use when scoping a new feature, bugfix, upgrade to an existing feature, or refactor into a concrete Flow objective. Guides Start, Discover, Define, Stress Test, Future Pressure, Develop, Deliver, and Handoff; requires an explicit shallow-vs-deep choice before scoping continues. Triggers on literal `/flux:scope`, including prompts like `/flux:scope Add OAuth login` or `/flux:scope fn-1 --deep`.
Optional local install of fluxctl CLI and AGENTS.md/legacy CLAUDE.md instructions. Use when user runs /flux:setup. Triggers on literal `/flux:setup`.
Manage .flux/ tasks, epics, and config. Triggers: 'show me my tasks', 'list epics', 'what tasks are there', 'add a task', 'create task', 'what's ready', 'task status', 'show fn-1-add-oauth', 'show my Flux config', 'what did setup configure', 'edit Flux settings'. NOT for /flux:plan or /flux:work.
Execute a Flow epic or task systematically with git setup, task tracking, quality checks, and commit workflow. Use when implementing a plan or working through a spec. Triggers on literal `/flux:work` prompts with Flow IDs (fn-1-add-oauth, fn-1-add-oauth.2, or legacy fn-1, fn-1.2, fn-1-xxx, fn-1-xxx.2).