mr-watcher

name

mr-watcher

description

Activate when the user says "watch MR X until it lands", "babysit MR Y", "keep MR Z healthy", "keep an eye on this MR until it merges", or asks to run a long-running shepherd that survives across sessions. Complements `mr-shepherd` (single-pass) and `mr-submit` (end-to-end one-shot): mr-watcher is the persistent tail state that polls until GitLab marks the MR `merged` or `closed`, dispatching per-cycle shepherd work as new threads/CI events land.

MR Watcher Workflow

Persistent, checkpoint-resumable coverage of a single GitLab MR. Runs in-process with a configurable sleep between polls. Exits only when GitLab itself marks the MR merged or closed — no inactivity timeout, no auto-exit on red CI.

When to Use

Use mr-watcher when:

• The user wants an MR baby-sat from submission through merge.
• CI is flaky and needs persistent retry-on-infra-failure coverage.
• The MR is expected to collect new reviewer comments over time, and the operator wants them addressed without manual re-invocation of shepherd.
• A long-running poll survives across shell sessions — the checkpoint makes resume deterministic.

Do NOT use mr-watcher when:

• You want a single-pass review + fix. Use mr-shepherd for that.
• You want end-to-end one-shot (review → fix → push → resolve threads in one call). Use mr-submit / mr-shepherd-loop instead.
• You want multi-MR coverage. Use fan-out-mr-shepherd.
• The MR is not yours to modify. The watcher pushes commits and resolves threads — only dispatch when the operator has write rights.

Distinction from siblings

| Skill | Scope | Duration | Terminal | Composes | | --------------------- | ---------------- | --------------- | ------------------ | ------------------ | ------------------ | | mr-shepherd | One MR, one pass | Seconds–minutes | Operator exit | mr-shepherd-loop | | mr-submit | One MR, one shot | Minutes | Push + resolve | mr-shepherd-loop | | mr-shepherd-loop | One MR, bounded | Up to N minutes | max_poll_minutes | Primitive services | | mr-watcher | One MR, full | Unbounded | merged | closed | mr-shepherd-loop | | fan-out-mr-shepherd | N MRs, parallel | Bounded per-MR | Per-MR verdict | mr-shepherd-loop |

Workflow

1. Inputs

Input	Type	Default	Purpose
mr_id	int	required	GitLab MR IID
poll_interval	int	60	Seconds between polls. Floor 60 (rate-limit guard).
max_runtime	int?	null	Seconds. null = unlimited; stop only on merged/closed.
resume_from_checkpoint	bool	true	On start, read prior checkpoint if present.
terminal_states	list	["merged","closed"]	Override the terminal-state set (rare; keep default).

2. Checkpoint bootstrap

Checkpoint path: .claude/mr-watcher/<mr_id>/state.json

CKPT_DIR=".claude/mr-watcher/${MR_ID}"
CKPT_FILE="${CKPT_DIR}/state.json"
mkdir -p "${CKPT_DIR}"

if [ "${RESUME_FROM_CHECKPOINT}" = "true" ] && [ -f "${CKPT_FILE}" ]; then
  # Resume: load prior state. polls_total, fixes_applied, infra_retries
  # carry forward so the summary reflects total lifetime across restarts.
  prior_polls=$(jq -r .polls_total "${CKPT_FILE}")
  prior_fixes=$(jq -r .fixes_applied "${CKPT_FILE}")
  prior_retries=$(jq -r .infra_retries "${CKPT_FILE}")
else
  prior_polls=0
  prior_fixes=0
  prior_retries=0
fi

3. Poll loop

Each cycle:

1. Fetch terminal state. Call fetch-mr (not fetch-mr-status — only fetch-mr.mr.state is the authoritative opened|merged|closed|locked source).
   • If mr.state ∈ terminal_states → write final checkpoint, emit summary, exit.
2. Fetch CI + thread counts. Call fetch-mr-status for pipeline.state and threads.unresolved.
3. React by precedence (one action per cycle, stop after first match):
   • New unresolved threads (threads.unresolved > last_known_unresolved_count) → dispatch mr-shepherd-loop with max_review_iterations: 1, max_poll_minutes: 0 (single pass, no nested poll). Increment fixes_applied by its fixed_count.
   • Pipeline failed with infra pattern (match against the harness CI-triage ruleset — docker-pull timeouts, network 5xx, known-flaky stages) → retry the failing job via glab ci retry <pipeline_id>. Increment infra_retries.
   • Pipeline failed with code pattern → dispatch mr-shepherd-loop with stop_on_pipeline_fail: false so it diagnoses, fixes, and pushes. Increment fixes_applied.
   • Branch behind origin/main (detected by comparing mr.base_sha against origin/main tip) → dispatch sync-and-push (MR !27). Rebase in place, resolve conflicts, push with --force-with-lease.
   • No-op → nothing needed this cycle.
4. Write checkpoint (see schema below). Sleep poll_interval. Repeat.

4. Terminal exit

On mr.state == "merged" or mr.state == "closed":

1. Write final checkpoint with final_state set.

2. Emit summary:

   [MR_WATCHER_TERMINAL]
   mr_id: <int>
   final_state: merged | closed
   polls_total: <int>
   fixes_applied: <int>
   infra_retries: <int>
   elapsed_seconds: <int>
   checkpoint_path: .claude/mr-watcher/<mr_id>/state.json
   

3. Return. Do NOT continue polling.

Checkpoint schema

.claude/mr-watcher/<mr_id>/state.json:

{
  "mr_id": 42,
  "schema_version": 1,
  "started_at": "2026-05-12T14:03:11Z",
  "last_poll_at": "2026-05-12T14:18:11Z",
  "polls_total": 15,
  "fixes_applied": 3,
  "infra_retries": 1,
  "elapsed_seconds": 900,
  "last_known_pipeline": {
    "state": "success",
    "sha": "abc1234",
    "web_url": "https://gitlab.../pipelines/12345"
  },
  "last_known_unresolved_count": 0,
  "next_action_planned": "idle",
   
   
   
    
       
       
       
       
    
  

Field contracts

| Field | Type | Notes | | ----------------------------- | ------------ | ------------------------------------------------------------------------- | ------- | -------- | ------ | -------- | | mr_id | int | Constant across restarts; mismatch → refuse to resume. | | schema_version | int | Start at 1. Bump on breaking shape changes. | | started_at | ISO 8601 UTC | First started_at; preserved across resumes. | | last_poll_at | ISO 8601 UTC | Updated every cycle, including no-op cycles. | | polls_total | int | Cumulative across restarts. | | fixes_applied | int | Cumulative count of successful mr-shepherd-loop dispatches that pushed. | | infra_retries | int | Cumulative count of CI retries triggered on infra patterns. | | elapsed_seconds | int | now - started_at at last write. | | last_known_pipeline | object | Last non-null pipeline object; null until first poll. | | last_known_unresolved_count | int | Used to detect new threads between cycles. | | next_action_planned | string | One of idle | threads | ci_retry | ci_fix | rebase. | | final_state | string? | null while watching; set on terminal exit. | | poll_interval | int | Current effective interval (floor-adjusted). | | recent_events | array | Rolling buffer, max 20 entries. Oldest dropped on overflow. |

Resume contract

On resume:

• started_at, polls_total, fixes_applied, infra_retries carry forward.
• last_poll_at is overwritten on the first new cycle.
• If mr_id in checkpoint doesn't match the invocation input, refuse to resume (mismatched MR).
• If final_state is non-null (prior run already terminated), refuse to resume — terminal is terminal.

Composition with siblings

mr-watcher is a composer, not a reimplementer:

• Per-cycle fix work → mr-shepherd-loop. Dispatch with max_review_iterations: 1 + max_poll_minutes: 0 so it does one review-fix-push pass and returns. mr-watcher's outer loop handles the "keep going" part.
• Rebase-when-behind → sync-and-push (MR !27). Do not inline rebase logic here.
• Thread resolution → inherited via mr-shepherd-loop. Its Phase 5b match-threads-to-resolved is the contract for resolving threads post-fix.
• Terminal detection → fetch-mr. Authoritative source of mr.state.
• CI snapshot → fetch-mr-status. Pipeline + unresolved-thread count in one call.
• Polling cadence → self. wait-pipeline is for single-shot pipeline waits; mr-watcher owns its own poll cadence so it can interleave thread + CI + rebase checks.

Idempotence + safety

• Every dispatched service is itself idempotent (push-changes handles no-op trees, resolve-thread on already-resolved is non-fatal, fetch-* are read-only).
• A watcher killed mid-cycle loses at most the pending action for that cycle. Resume picks up on the next poll interval.
• Two watchers on the same mr_id race on the checkpoint file. Operator discipline: one watcher per MR. The skill does not enforce a lock — add that at the harness layer if needed.

Related

• Skill mr-shepherd — single-pass shepherd; mr-watcher dispatches it per cycle.
• Skill mr-status — operator dashboard for active watchers (reads the checkpoint directory).
• Service mr-shepherd-loop — per-cycle fix primitive.
• Service sync-and-push — rebase-when-behind primitive.
• Service mr-watch — the ProseScript orchestrator that implements this skill's contract.