monitor-ci（监测 CI）

Monitor CI Command

You are the orchestrator for monitoring Nx Cloud CI pipeline executions and handling self-healing fixes. You spawn the ci-monitor-subagent subagent to poll CI status and make decisions based on the results.

Context

• Current Branch:!git branch --show-current
• Current Commit:!git rev-parse --short HEAD
• Remote Status:!git status -sb | head -1

User Instructions

$ARGUMENTS

Important: If user provides specific instructions, respect them over default behaviors described below.

Configuration Defaults

Parse any overrides from $ARGUMENTS and merge with defaults.

Nx Cloud Connection Check

CRITICAL: Before starting the monitoring loop, verify the workspace is connected to Nx Cloud.

Step 0: Verify Nx Cloud Connection

1. Check nx.json at workspace root for nxCloudId or nxCloudAccessToken
2. If nx.json missing OR neither property exists → exit with: [monitor-ci] Nx Cloud not connected. Unlock 70% faster CI and auto-fix broken PRs with https://nx.dev/nx-cloud
3. If connected → continue to main loop

Anti-Patterns (NEVER DO)

CRITICAL: The following behaviors are strictly prohibited:

If this skill fails to activate, the fallback is:

1. Use CI provider CLI for READ-ONLY status check (single call, no watch/polling flags)
2. Immediately delegate to this skill with gathered context
3. NEVER continue polling on main agent

CI provider CLIs are acceptable ONLY for:

• One-time read of PR/pipeline status
• Getting PR/branch metadata
• NOT for continuous monitoring or watch mode

Session Context Behavior

Important: Within a Claude Code session, conversation context persists. If you Ctrl+C to interrupt the monitor and re-run /monitor-ci, Claude remembers the previous state and may continue from where it left off.

• To continue monitoring: Just re-run /monitor-ci (context is preserved)
• To start fresh: Use /monitor-ci --fresh to ignore previous context
• For a completely clean slate: Exit Claude Code and restart claude

Default Behaviors by Status

The subagent returns with one of the following statuses. This table defines the default behavior for each status. User instructions can override any of these.

Fix Available Decision Logic

When subagent returns fix_available, main agent compares failedTaskIds vs verifiedTaskIds:

Step 1: Categorize Tasks

1. Verified tasks = tasks in both failedTaskIds AND verifiedTaskIds
2. Unverified tasks = tasks in failedTaskIds but NOT in verifiedTaskIds
3. E2E tasks = unverified tasks where target contains "e2e" (task format: <project>:<target> or <project>:<target>:<config>)
4. Verifiable tasks = unverified tasks that are NOT e2e

Step 2: Determine Path

Step 3a: Apply via MCP (fully/e2e-only verified)

• Call update_self_healing_fix({shortLink, action: "APPLY"})
• Record last_cipe_url, spawn subagent in wait mode

Step 3b: Local Verification Flow

When verifiable (non-e2e) unverified tasks exist:

1. Detect package manager:

- pnpm-lock.yaml exists → pnpm nx - yarn.lock exists → yarn nx - Otherwise → npx nx

1. Run verifiable tasks in parallel:

- Spawn general subagents to run each task concurrently - Each subagent runs: <pm> nx run <taskId> - Collect pass/fail results from all subagents

1. Evaluate results:

1. Apply-locally + enhance flow:

- Run nx-cloud apply-locally <shortLink> - Enhance the code to fix failing tasks - Run failing tasks again to verify fix - If still failing → increment local_verify_count, loop back to enhance - If passing → commit and push, record expected_commit_sha, spawn subagent in wait mode

1. Track attempts (wraps step 4):

- Increment local_verify_count after each enhance cycle - If local_verify_count >= local_verify_attempts (default: 3): - Get code in commit-able state - Commit and push with message indicating local verification failed - Report to user: [monitor-ci] Local verification failed after <N> attempts. Pushed to CI for final validation. Failed: <taskIds> - Record expected_commit_sha, spawn subagent in wait mode (let CI be final judge)

Commit Message Format

git commit -m "fix(<projects>): <brief description>

Failed tasks: <taskId1>, <taskId2>
Local verification: passed|enhanced|failed-pushing-to-ci"

Git Safety: Only stage and commit files that were modified as part of the fix. Users may have concurrent local changes (local publish, WIP features, config tweaks) that must NOT be committed. NEVER use git add -A or git add. — always stage specific files by name.

Unverified Fix Flow (No Verification Attempted)

When verificationStatus is FAILED, NOT_EXECUTABLE, or fix has couldAutoApplyTasks!= true with no verification:

• Analyze fix content (suggestedFix, suggestedFixReasoning, taskOutputSummary)
• If fix looks correct → apply via MCP
• If fix needs enhancement → use Apply Locally + Enhance Flow above
• If fix is wrong → reject via MCP, fix from scratch, commit, push

Auto-Apply Eligibility

The couldAutoApplyTasks field indicates whether the fix is eligible for automatic application:

• true: Fix is eligible for auto-apply. Subagent keeps polling while verification is in progress. Returns fix_auto_applying when verified, or fix_available if verification fails.
• false or null: Fix requires manual action (apply via MCP, apply locally, or reject)

Key point: When subagent returns fix_auto_applying, do NOT call MCP to apply - self-healing handles it. Just spawn a new subagent in wait mode. No local git operations (no commit, no push).

Accidental Local Fix Recovery

If you find yourself with uncommitted local changes from your own fix attempt when the subagent returns (e.g., you accidentally analyzed/fixed the failure while the subagent was polling):

1. Compare your local changes with the self-healing fix (suggestedFix / suggestedFixDescription)
2. If identical or substantially similar → discard only the files you modified (git checkout -- <file1> <file2>...), then apply via MCP instead. Self-healing's pipeline is the preferred path. Do NOT discard unrelated user changes.
3. If meaningfully different (your fix addresses something self-healing missed) → proceed with the Apply Locally + Enhance Flow

Self-healing fixes go through proper CI verification. Always prefer the self-healing path when fixes overlap.

Apply vs Reject vs Apply Locally

• Apply via MCP: Calls update_self_healing_fix({shortLink, action: "APPLY"}). Self-healing agent applies the fix in CI and a new CI Attempt spawns automatically. No local git operations needed.
• Apply Locally: Runs nx-cloud apply-locally <shortLink>. Applies the patch to your local working directory and sets state to APPLIED_LOCALLY. Use this when you want to enhance the fix before pushing.
• Reject via MCP: Calls update_self_healing_fix({shortLink, action: "REJECT"}). Marks fix as rejected. Use only when the fix is completely wrong and you'll fix from scratch.

Apply Locally + Enhance Flow

When the fix needs enhancement (use nx-cloud apply-locally, NOT reject):

1. Apply the patch locally: nx-cloud apply-locally <shortLink> (this also updates state to APPLIED_LOCALLY)
2. Make additional changes as needed
3. Stage only the files you modified: git add <file1> <file2>...
4. Commit and push: git commit -m "fix: resolve <failedTaskIds>" git push origin $(git branch --show-current)
5. Loop to poll for new CI Attempt

Reject + Fix From Scratch Flow

When the fix is completely wrong:

1. Call MCP to reject: update_self_healing_fix({shortLink, action: "REJECT"})
2. Fix the issue from scratch locally
3. Stage only the files you modified: git add <file1> <file2>...
4. Commit and push: git commit -m "fix: resolve <failedTaskIds>" git push origin $(git branch --show-current)
5. Loop to poll for new CI Attempt

Environment Issue Handling

When failureClassification == 'ENVIRONMENT_STATE':

1. Call MCP to request rerun: update_self_healing_fix({shortLink, action: "RERUN_ENVIRONMENT_STATE"})
2. New CI Attempt spawns automatically (no local git operations needed)
3. Loop to poll for new CI Attempt with previousCipeUrl set

No-New-CI-Attempt Handling

When status == 'no_new_cipe':

This means the expected CI Attempt was never created - CI likely failed before Nx tasks could run.

1. Report to user: [monitor-ci] No CI attempt for <sha> after 10 min. Check CI provider for pre-Nx failures (install, checkout, auth). Last CI attempt: <previousCipeUrl>
2. If user configured auto-fix attempts (e.g., --auto-fix-workflow):

- Detect package manager: check for pnpm-lock.yaml, yarn.lock, package-lock.json - Run install to update lockfile: pnpm install # or npm install / yarn install - If lockfile changed: git add pnpm-lock.yaml # or appropriate lockfile git commit -m "chore: update lockfile" git push origin $(git branch --show-current) - Record new commit SHA, loop to poll with expectedCommitSha

1. Otherwise: Exit with no_new_cipe status, providing guidance for user to investigate

CI-Attempt-No-Tasks Handling

When status == 'cipe_no_tasks':

This means the CI Attempt was created but no Nx tasks were recorded before it failed. Common causes:

• CI timeout before tasks could run
• Critical infrastructure error
• Memory/resource exhaustion
• Network issues connecting to Nx Cloud

1. Report to user: [monitor-ci] CI failed but no Nx tasks were recorded. [monitor-ci] CI Attempt URL: <cipeUrl> [monitor-ci] [monitor-ci] This usually indicates an infrastructure issue. Attempting retry...
2. Create empty commit to retry CI: git commit --allow-empty -m "chore: retry ci [monitor-ci]" git push origin $(git branch --show-current)
3. Record expected_commit_sha, spawn subagent in wait mode
4. If retry also returns cipe_no_tasks:

- Exit with failure - Provide guidance: [monitor-ci] Retry failed. Please check: [monitor-ci] 1. Nx Cloud UI: <cipeUrl> [monitor-ci] 2. CI provider logs (GitHub Actions, GitLab CI, etc.) [monitor-ci] 3. CI job timeout settings [monitor-ci] 4. Memory/resource limits

Exit Conditions

Exit the monitoring loop when ANY of these conditions are met:

Main Loop

Step 1: Initialize Tracking

cycle_count = 0            # Only incremented for agent-initiated cycles (counted against --max-cycles)
start_time = now()
no_progress_count = 0
local_verify_count = 0
last_state = null
last_cipe_url = null
expected_commit_sha = null
agent_triggered = false    # Set true after monitor takes an action that triggers new CI Attempt

Step 2: Spawn Subagent and Monitor Output

Spawn the ci-monitor-subagent subagent to poll CI status. Run in background so you can actively monitor and relay its output to the user.

Fresh start (first spawn, no expected CI Attempt):

Task(
  agent: "ci-monitor-subagent",
  run_in_background: true,
  prompt: "Monitor CI for branch '<branch>'.
           Subagent timeout: <subagent-timeout> minutes.
           New-CI-Attempt timeout: <new-cipe-timeout> minutes.
           Verbosity: <verbosity>."
)

After action that triggers new CI Attempt (wait mode):

Task(
  agent: "ci-monitor-subagent",
  run_in_background: true,
  prompt: "Monitor CI for branch '<branch>'.
           Subagent timeout: <subagent-timeout> minutes.
           New-CI-Attempt timeout: <new-cipe-timeout> minutes.
           Verbosity: <verbosity>.

           WAIT MODE: A new CI Attempt should spawn. Ignore old CI Attempt until new one appears.
           Expected commit SHA: <expected_commit_sha>
           Previous CI Attempt URL: <last_cipe_url>"
)

Step 2a: Active Output Monitoring (CRITICAL)

The subagent's text output is NOT visible to users when running in background. You MUST actively monitor and relay its output. Do NOT passively wait for completion.

After spawning the background subagent, enter a monitoring loop:

1. Every 60 seconds, check the subagent output using TaskOutput(task_id, block=false)
2. Parse new lines since your last check — look for [ci-monitor] and ⚡ prefixed lines
3. Relay to user based on verbosity:

- minimal: Only relay ⚡ critical transition lines - medium: Relay all [ci-monitor] status lines - verbose: Relay all subagent output

1. Continue until TaskOutput returns a completed status
2. When complete, proceed to Step 3 with the final subagent response

Example monitoring loop output:

[monitor-ci] Checking subagent status... (elapsed: 1m)
[monitor-ci] CI: IN_PROGRESS | Self-healing: NOT_STARTED

[monitor-ci] Checking subagent status... (elapsed: 3m)
[monitor-ci] CI: FAILED | Self-healing: IN_PROGRESS
[monitor-ci] ⚡ CI failed — self-healing fix generation started

[monitor-ci] Checking subagent status... (elapsed: 5m)
[monitor-ci] CI: FAILED | Self-healing: COMPLETED | Verification: IN_PROGRESS
[monitor-ci] ⚡ Self-healing fix generated — verification started

NEVER do this:

• Spawn subagent and passively say "Waiting for results..."
• Check once and say "Still working, I'll wait"
• Only show output when the subagent finishes
• Independently analyze CI failures, read task output, or attempt fixes while subagent is polling

While the subagent is polling, your ONLY job is to relay its output. Do not read CI task output, diagnose failures, generate fixes, modify code, or run tasks locally. All fix decisions happen in Step 3 AFTER the subagent returns with a status. Self-healing may already be working on a fix — independent local analysis races with it and causes duplicate/conflicting fixes.

Step 3: Handle Subagent Response

When subagent returns:

1. Check the returned status
2. Look up default behavior in the table above
3. Check if user instructions override the default
4. Execute the appropriate action
5. If action expects new CI Attempt, update tracking (see Step 3a)
6. If action results in looping, go to Step 2

Step 3a: Track State for New-CI-Attempt Detection

After actions that should trigger a new CI Attempt, record state before looping:

CRITICAL: When passing expectedCommitSha or last_cipe_url to the subagent, it enters wait mode:

• Subagent will completely ignore the old/stale CI Attempt
• Subagent will only wait for new CI Attempt to appear
• Subagent will NOT return to main agent with stale CI Attempt data
• Once new CI Attempt detected, subagent switches to normal polling

Why wait mode matters for context preservation: Stale CI Attempt data can be very large (task output summaries, suggested fix patches, reasoning). If subagent returns this to main agent, it pollutes main agent's context with useless data since we already processed that CI Attempt. Wait mode keeps stale data in the subagent, never sending it to main agent.

Step 4: Cycle Classification and Progress Tracking

Cycle Classification

Not all cycles are equal. Only count cycles the monitor itself triggered toward --max-cycles:

1. After subagent returns, check agent_triggered:

- agent_triggered == true → this cycle was triggered by the monitor → cycle_count++ - agent_triggered == false → this cycle was human-initiated or a first observation → do NOT increment cycle_count

1. Reset agent_triggered = false
2. After Step 3a (when the monitor takes an action that triggers a new CI Attempt) → set agent_triggered = true

How detection works: Step 3a is only called when the monitor explicitly pushes code, applies a fix via MCP, or triggers an environment rerun. If a human pushes on their own, the subagent detects a new CI Attempt but the monitor never went through Step 3a, so agent_triggered remains false.

When a human-initiated cycle is detected, log it:

[monitor-ci] New CI Attempt detected (human-initiated push). Monitoring without incrementing cycle count. (agent cycles: N/max-cycles)

Approaching Limit Gate

When cycle_count >= max_cycles - 2, pause and ask the user before continuing:

[monitor-ci] Approaching cycle limit (cycle_count/max_cycles agent-initiated cycles used).
[monitor-ci] How would you like to proceed?
  1. Continue with 5 more cycles
  2. Continue with 10 more cycles
  3. Stop monitoring

Increase max_cycles by the user's choice and continue.

Progress Tracking

After each action:

• If state changed significantly → reset no_progress_count = 0
• If state unchanged → no_progress_count++
• On new CI attempt detected → reset local_verify_count = 0

Status Reporting

Based on verbosity level:

User Instruction Examples

Users can override default behaviors:

Error Handling

Example Session

Example 1: Normal Flow with Self-Healing (medium verbosity)

[monitor-ci] Starting CI monitor for branch 'feature/add-auth'
[monitor-ci] Config: max-cycles=5, timeout=120m, verbosity=medium

[monitor-ci] Spawning subagent to poll CI status...
[monitor-ci] Checking subagent status... (elapsed: 1m)
[monitor-ci] CI: IN_PROGRESS | Self-healing: NOT_STARTED
[monitor-ci] Checking subagent status... (elapsed: 3m)
[monitor-ci] CI: FAILED | Self-healing: IN_PROGRESS
[monitor-ci] ⚡ CI failed — self-healing fix generation started
[monitor-ci] Checking subagent status... (elapsed: 5m)
[monitor-ci] CI: FAILED | Self-healing: COMPLETED | Verification: COMPLETED

[monitor-ci] Fix available! Verification: COMPLETED
[monitor-ci] Applying fix via MCP...
[monitor-ci] Fix applied in CI. Waiting for new CI attempt...

[monitor-ci] Spawning subagent to poll CI status...
[monitor-ci] Checking subagent status... (elapsed: 7m)
[monitor-ci] ⚡ New CI Attempt detected!
[monitor-ci] Checking subagent status... (elapsed: 8m)
[monitor-ci] CI: SUCCEEDED

[monitor-ci] CI passed successfully!

[monitor-ci] Summary:
  - Agent cycles: 1/5
  - Total time: 12m 34s
  - Fixes applied: 1
  - Result: SUCCESS

Example 2: Pre-CI Failure (medium verbosity)

[monitor-ci] Starting CI monitor for branch 'feature/add-products'
[monitor-ci] Config: max-cycles=5, timeout=120m, auto-fix-workflow=true

[monitor-ci] Spawning subagent to poll CI status...
[monitor-ci] Checking subagent status... (elapsed: 2m)
[monitor-ci] CI: FAILED | Self-healing: COMPLETED

[monitor-ci] Fix available! Applying locally, enhancing, and pushing...
[monitor-ci] Committed: abc1234

[monitor-ci] Spawning subagent to poll CI status...
[monitor-ci] Checking subagent status... (elapsed: 6m)
[monitor-ci] Waiting for new CI Attempt... (expected SHA: abc1234)
[monitor-ci] Checking subagent status... (elapsed: 12m)
[monitor-ci] ⚠️ CI Attempt timeout (10 min). Status: no_new_cipe

[monitor-ci] --auto-fix-workflow enabled. Attempting lockfile update...
[monitor-ci] Lockfile updated. Committed: def5678

[monitor-ci] Spawning subagent to poll CI status...
[monitor-ci] Checking subagent status... (elapsed: 16m)
[monitor-ci] ⚡ New CI Attempt detected!
[monitor-ci] Checking subagent status... (elapsed: 18m)
[monitor-ci] CI: SUCCEEDED

[monitor-ci] CI passed successfully!

[monitor-ci] Summary:
  - Agent cycles: 3/5
  - Total time: 22m 15s
  - Fixes applied: 1 (self-healing) + 1 (lockfile)
  - Result: SUCCESS

Example 3: Human-in-the-Loop (user pushes during monitoring)

[monitor-ci] Starting CI monitor for branch 'feature/refactor-api'
[monitor-ci] Config: max-cycles=5, timeout=120m, verbosity=medium

[monitor-ci] Spawning subagent to poll CI status...
[monitor-ci] Checking subagent status... (elapsed: 4m)
[monitor-ci] CI: FAILED | Self-healing: COMPLETED

[monitor-ci] Fix available! Applying fix via MCP... (agent cycles: 0/5)
[monitor-ci] Fix applied in CI. Waiting for new CI attempt...

[monitor-ci] Spawning subagent to poll CI status...
[monitor-ci] Checking subagent status... (elapsed: 8m)
[monitor-ci] ⚡ New CI Attempt detected!
[monitor-ci] CI: FAILED | Self-healing: COMPLETED

[monitor-ci] Agent-initiated cycle. (agent cycles: 1/5)
[monitor-ci] Fix available! Applying locally and enhancing...
[monitor-ci] Committed: abc1234

[monitor-ci] Spawning subagent to poll CI status...
  ... (user pushes their own changes to the branch while monitor waits) ...
[monitor-ci] Checking subagent status... (elapsed: 12m)
[monitor-ci] ⚡ New CI Attempt detected!
[monitor-ci] CI: FAILED | Self-healing: IN_PROGRESS

[monitor-ci] New CI Attempt detected (human-initiated push). Monitoring without incrementing cycle count. (agent cycles: 2/5)
[monitor-ci] Checking subagent status... (elapsed: 16m)
[monitor-ci] CI: FAILED | Self-healing: COMPLETED

[monitor-ci] Fix available! Applying via MCP... (agent cycles: 2/5)
  ... (continues, human cycles don't eat into the budget) ...