Imported from fleetshift/fleetshift-poc (.agents/skills/monitor-ci/SKILL.md). Install upstream with npx skills add fleetshift/fleetshift-poc --skill monitor-ci. Copyright stays with the author.

Monitor CI Command

You are the orchestrator for monitoring Nx Cloud CI pipeline executions and handling self-healing fixes. You spawn subagents to interact with Nx Cloud, run deterministic decision scripts, and take action 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

Before starting the monitoring loop, verify the workspace is connected to Nx Cloud. Without this connection, no CI data is available and the entire skill is inoperable.

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:

   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

Architecture Overview

1. This skill (orchestrator): spawns subagents, runs scripts, prints status, does local coding work
2. ci-monitor-subagent (haiku): calls one MCP tool (ci_information or update_self_healing_fix), returns structured result, exits
3. ci-poll-decide.mjs (deterministic script): takes ci_information result + state, returns action + status message
4. ci-state-update.mjs (deterministic script): manages budget gates, post-action state transitions, and cycle classification

Status Reporting

The decision script handles message formatting based on verbosity. When printing messages to the user:

• Prepend [monitor-ci] to every message from the script's message field
• For your own action messages (e.g. "Applying fix via MCP..."), also prepend [monitor-ci]

Anti-Patterns

These behaviors cause real problems — racing with self-healing, losing CI progress, or wasting context:

If this skill fails to activate, the fallback is:

1. Use CI provider CLI for a one-time, read-only status check (single call, no watch/polling flags)
2. Immediately delegate to this skill with gathered context
3. Do not continue polling on main agent — it wastes context tokens and bypasses self-healing

Session Context Behavior

If the user previously ran /monitor-ci in this session, you may have prior state (poll counts, last CI Attempt URL, etc.). Resume from that state unless --fresh is set, in which case discard it and start from Step 1.

MCP Tool Reference

The ci_information and update_self_healing_fix tools are called via the ci-monitor-subagent, not directly from the orchestrator. Calling MCP tools directly wastes main agent context with large response payloads. The field sets below are for composing subagent prompts (see Step 2a).

Three field sets control polling efficiency — use the lightest set that gives you what you need:

WAIT_FIELDS: 'cipeUrl,commitSha,cipeStatus'
LIGHT_FIELDS: 'cipeStatus,cipeUrl,branch,commitSha,selfHealingStatus,verificationStatus,userAction,failedTaskIds,verifiedTaskIds,selfHealingEnabled,failureClassification,couldAutoApplyTasks,autoApplySkipped,autoApplySkipReason,shortLink,confidence,confidenceReasoning,hints,selfHealingSkippedReason,selfHealingSkipMessage'
HEAVY_FIELDS: 'taskOutputSummary,suggestedFix,suggestedFixReasoning,suggestedFixDescription'

The ci_information tool accepts branch (optional, defaults to current git branch), select (comma-separated field names), and pageToken (0-based pagination for long strings).

The update_self_healing_fix tool accepts a shortLink and an action: APPLY, REJECT, or RERUN_ENVIRONMENT_STATE.

Default Behaviors by Status

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

Simple exits — just report and exit:

Statuses requiring action — when handling these in Step 3, read references/fix-flows.md for the detailed flow:

Key rules (always apply):

• Git safety: Stage specific files by name — git add -A or git add . risks committing the user's unrelated work-in-progress or secrets
• Environment failures (OOM, command not found, permission denied): bail immediately. These aren't code bugs, so spending local-fix budget on them is wasteful
• Gate check: Run ci-state-update.mjs gate before local fix attempts — if budget exhausted, print message and exit

Main Loop

Step 1: Initialize Tracking

cycle_count = 0            # Only incremented for agent-initiated cycles (counted against --max-cycles)
start_time = now()         # Passed to the decision script as --elapsed-seconds on every poll to enforce --timeout across attempts
no_progress_count = 0
local_verify_count = 0
env_rerun_count = 0
last_cipe_url = null
expected_commit_sha = null
agent_triggered = false    # Set true after monitor takes an action that triggers new CI Attempt
poll_count = 0
wait_mode = false
prev_status = null
prev_cipe_status = null
prev_sh_status = null
prev_verification_status = null
prev_failure_classification = null

Step 2: Polling Loop

Repeat until done:

2a. Spawn subagent (FETCH_STATUS)

Determine select fields based on mode:

• Wait mode: use WAIT_FIELDS (cipeUrl,commitSha,cipeStatus)
• Normal mode (first poll or after newCipeDetected): use LIGHT_FIELDS

Task(
  agent: "ci-monitor-subagent",
  model: haiku,
  prompt: "FETCH_STATUS for branch '<branch>'.
           select: '<fields>'"
)

The subagent calls ci_information and returns a JSON object with the requested fields. This is a foreground call — wait for the result.

2b. Run decision script

node <skill_dir>/scripts/ci-poll-decide.mjs '<subagent_result_json>' <poll_count> <verbosity> \
  [--wait-mode] \
  [--prev-cipe-url <last_cipe_url>] \
  [--expected-sha <expected_commit_sha>] \
  [--prev-status <prev_status>] \
  [--timeout <timeout_minutes>] \
  [--new-cipe-timeout <new_cipe_timeout_minutes>] \
  [--elapsed-seconds <seconds_since_start_time>] \
  [--env-rerun-count <env_rerun_count>] \
  [--no-progress-count <no_progress_count>] \
  [--prev-cipe-status <prev_cipe_status>] \
  [--prev-sh-status <prev_sh_status>] \
  [--prev-verification-status <prev_verification_status>] \
  [--prev-failure-classification <prev_failure_classification>]

Pass --timeout and --new-cipe-timeout in minutes (the values from Configuration Defaults) — the script converts to seconds internally. Pass --elapsed-seconds as the whole seconds elapsed since start_time (now() - start_time); this is what enforces --timeout as a total monitor budget across every poll and attempt, so it must be supplied on every call once monitoring has started.

The script outputs a single JSON line: { action, code, message, delay?, noProgressCount, envRerunCount, fields?, newCipeDetected?, verifiableTaskIds? }

2c. Process script output

Parse the JSON output and update tracking state:

• no_progress_count = output.noProgressCount
• env_rerun_count = output.envRerunCount
• prev_cipe_status = subagent_result.cipeStatus
• prev_sh_status = subagent_result.selfHealingStatus
• prev_verification_status = subagent_result.verificationStatus
• prev_failure_classification = subagent_result.failureClassification
• prev_status = output.action + ":" + (output.code || subagent_result.cipeStatus)
• poll_count++

Based on action:

• action == "poll": Print output.message, sleep output.delay seconds, go to 2a
  • If output.newCipeDetected: clear wait mode, reset wait_mode = false
• action == "wait": Print output.message, sleep output.delay seconds, go to 2a
• action == "done": Proceed to Step 3 with output.code

Step 3: Handle Actionable Status

When decision script returns action == "done":

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

Spawning subagents for actions

Several statuses require fetching heavy data or calling MCP:

• fix_apply_ready: Spawn UPDATE_FIX subagent with APPLY
• fix_needs_local_verify: Spawn FETCH_HEAVY subagent for fix details before local verification
• fix_needs_review: Spawn FETCH_HEAVY subagent → get suggestedFixDescription, suggestedFixSummary, taskFailureSummaries
• fix_failed / no_fix: Spawn FETCH_HEAVY subagent → get taskFailureSummaries for local fix context
• environment_issue: Spawn UPDATE_FIX subagent with RERUN_ENVIRONMENT_STATE
• self_healing_throttled: Spawn FETCH_HEAVY subagent → get selfHealingSkipMessage; then FETCH_THROTTLE_INFO + UPDATE_FIX for each old fix

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

After actions that should trigger a new CI Attempt, run:

node <skill_dir>/scripts/ci-state-update.mjs post-action \
  --action <type> \
  --cipe-url <current_cipe_url> \
  --commit-sha <git_rev_parse_HEAD>

Action types: fix-auto-applying, apply-mcp, apply-local-push, reject-fix-push, local-fix-push, env-rerun, auto-fix-push, empty-commit-push

The script returns { waitMode, pollCount, lastCipeUrl, expectedCommitSha, agentTriggered }. Update all tracking state from the output, then go to Step 2.

Step 4: Cycle Classification and Progress Tracking

When the decision script returns action == "done", run cycle-check before handling the code:

node <skill_dir>/scripts/ci-state-update.mjs cycle-check \
  --code <code> \
  [--agent-triggered] \
  --cycle-count <cycle_count> --max-cycles <max_cycles> \
  --env-rerun-count <env_rerun_count>

The script returns { cycleCount, agentTriggered, envRerunCount, approachingLimit, limitReached, message }. Update tracking state from the output.

• If limitReached → the --max-cycles budget is exhausted. Print message and stop monitoring (do not handle the code or start another cycle). This is a hard stop, not advisory.
• Else if approachingLimit → ask user whether to continue (with 5 or 10 more cycles) or stop monitoring
• If previous cycle was NOT agent-triggered (human pushed), log that human-initiated push was detected

Progress Tracking

• no_progress_count, circuit breaker (5 polls), and backoff reset are handled by ci-poll-decide.mjs (progress = any change in cipeStatus, selfHealingStatus, verificationStatus, or failureClassification)
• env_rerun_count reset on non-environment status is handled by ci-state-update.mjs cycle-check
• On new CI Attempt detected (poll script returns newCipeDetected) → reset local_verify_count = 0, env_rerun_count = 0

Error Handling

User Instruction Examples

Users can override default behaviors: