/ops:deploy-fix — Auto-fix subsystem control surface

This skill is the operator console for the post-merge + build-failure auto-fix loop installed by /ops:setup Step 6.5a. The underlying daemons, hooks, and prompts live in ${CLAUDE_PLUGIN_ROOT}/scripts/, ${CLAUDE_PLUGIN_ROOT}/hooks/, and ${CLAUDE_PLUGIN_ROOT}/prompts/. State lives in ~/.claude/state/ops-deploy-fix/. Logs live in ~/.claude/logs/ops-deploy-fix/.

Plugin rules apply (see claude-ops/CLAUDE.md). In particular:

• Rule 0 — never echo personal data (slugs are fine, but redact tokens, webhooks, emails)
• Rule 1 — max 4 options per AskUserQuestion
• Rule 4 — background by default during configure
• Rule 5 — destructive ops (clearing locks, wiping state) require explicit per-action confirmation
• Rule 6 — this skill never sends outbound comms; if a future flow needs to, use the universal send gate

Arguments

The first positional argument selects the subcommand. If absent, present:

/ops:deploy-fix — what do you want to do?
  [status]
  [tail]
  [configure]
  [test]

Subcommands:

Subcommand: status

Print a compact dashboard. Read all data via Bash; favor parallel reads.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 OPS ► DEPLOY-FIX STATUS
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 Master switch:    <on|off>          (deploy_fix_enabled)
 Auto-dispatch:    <on|off>          (auto_dispatch_fixer)
 Danger flag:      <on|off>          (allow_dangerous)
 Budget cap:       <N>/hour/repo     (max_fixes_per_hour)
 Notify channel:   <macos|ntfy|discord|none>
 Fix model:        <haiku|sonnet|opus>

 Last 5 monitor runs:
   2026-04-26T08:14Z  owner/repo-a:dev   merge-watch         deploy-success
   2026-04-26T08:11Z  owner/repo-b:main  build-watch         transient → rerun
   ...

 Last 5 fixer dispatches:
   2026-04-26T08:09Z  owner/repo-b:main  build-fix.md   in-flight   pid 41123
   2026-04-26T07:42Z  owner/repo-a:dev   deploy-fix.md  done        log:fix-...log

 Active locks:
   owner/repo-b:main:build  pid=41123 (alive)

 Hourly budget remaining (this hour, <YYYYMMDD-HH>):
   owner/repo-a   2 / 3
   owner/repo-b   1 / 3
──────────────────────────────────────────────────────

Data sources:

PREFS=~/.claude/plugins/data/ops-ops-marketplace/preferences.json
STATE=~/.claude/state/ops-deploy-fix
LOGS=~/.claude/logs/ops-deploy-fix

# Config snapshot (all keys read with defaults from plugin.json userConfig)
jq -r '{
  deploy_fix_enabled, auto_dispatch_fixer, allow_dangerous,
  max_fixes_per_hour, notify_channel, fix_model
}' "$PREFS" 2>/dev/null

# Monitor runs — last 5 across all repos
ls -t "$LOGS"/monitor-*.log 2>/dev/null | head -5 | while read f; do
  printf '%s  %s\n' "$(stat -f '%Sm' -t '%FT%RZ' "$f" 2>/dev/null || stat --format='%y' "$f" 2>/dev/null | awk '{print $1"T"$2"Z"}')" "$(basename "$f")"
done

# Fixer dispatches — last 5
ls -t "$LOGS"/fix-*.log 2>/dev/null | head -5 | while read f; do
  pid_file="$STATE/lock-$(basename "$f" | sed 's/^fix-//; s/-[0-9]*\.log$//')"
  status="done"
  [ -f "$pid_file" ] && kill -0 "$(cat "$pid_file")" 2>/dev/null && status="in-flight"
  printf '%s  %s  %s\n' "$(stat -f '%Sm' -t '%FT%RZ' "$f" 2>/dev/null || stat --format='%y' "$f" 2>/dev/null | awk '{print $1"T"$2"Z"}')" "$status" "$(basename "$f")"
done

# Active locks
for f in "$STATE"/lock-*; do
  [ -f "$f" ] || continue
  pid=$(cat "$f")
  if kill -0 "$pid" 2>/dev/null; then
    echo "$(basename "$f" | sed 's/^lock-//')  pid=$pid (alive)"
  fi
done

# Hourly budget — read all budget files for current hour
hour=$(date +%Y%m%d-%H)
cap=$(jq -r '.max_fixes_per_hour // 3' "$PREFS")
for f in "$STATE"/budget-*-"$hour"; do
  [ -f "$f" ] || continue
  slug=$(basename "$f" | sed "s/^budget-//; s/-$hour\$//")
  used=$(cat "$f")
  echo "$slug  $((cap - used)) / $cap"
done

After the dashboard, if any lock is older than the configured watcher_timeout_seconds, surface a ⚠️ Stale lock: line and offer:

Stale lock detected for <repo>. Action?
  [Inspect log]
  [Clear lock (Rule 5 — destructive)]
  [Leave it]
  [Skip]

Only on Clear lock should you delete the file (per-action confirmation per Rule 5).

Subcommand: tail

Tail the most-recent fixer log. Resolve via ls -t:

LATEST=$(ls -t ~/.claude/logs/ops-deploy-fix/fix-*.log 2>/dev/null | head -1)
if [ -z "$LATEST" ]; then
  echo "No fixer logs yet. Trigger one with /ops:deploy-fix test or wait for a real failure."
  exit 0
fi
echo "▸ Tailing $LATEST (Ctrl-C to stop)"
tail -f "$LATEST"

Run via Bash with run_in_background: true and a long timeout so the user can read the stream as it grows. Print the resolved log path and the spawned shell PID up-front so the user knows what's running.

If a second positional arg --lines N is passed, do tail -n N -f. If --no-follow, drop -f.

Subcommand: configure

Re-run the wizard from /ops:setup Step 6.5a, plus optionally 6.5b/c/d. Implementation: route into the setup skill with section filter.

Flow:

1. Read current state from $PREFS_PATH. Print compact summary (same fields as status, no logs).
2. Ask:

   What do you want to reconfigure?
     [Deploy auto-fix wizard (6.5a)]
     [Recap marquee (6.5b)]
     [Task* reminder (6.5c)]
     [Account rotation toggle (6.5d)]
   

3. Execute the corresponding sub-flow inline by following Step 6.5 of skills/setup/SKILL.md — same questions, same persistence, same Rule-3 "never silently skip" semantics.
4. After persistence, print:

   ✓ Reconfigured. Daemons & hooks pick up the new prefs on the next event — no restart required.
   

Use run_in_background: true (Rule 4) for any CLI install / brew / curl / tmux source-file triggered along the way.

Subcommand: test

Synthetic dry-run through the pipeline. No real fix dispatch — no agent runs. Confirms wiring: hooks fire → monitor classifies → notify channel pings → state files written → would-dispatch path logged.

Steps:

1. Pre-flight check:

   PLUGIN_ROOT="${CLAUDE_PLUGIN_ROOT:-}"
   [ -z "$PLUGIN_ROOT" ] && PLUGIN_ROOT=$(find ~/.claude -type d -name claude-ops 2>/dev/null | head -1)
   COMMON="$PLUGIN_ROOT/scripts/lib/deploy-fix-common.sh"
   [ -f "$COMMON" ] || { echo "✗ deploy-fix-common.sh missing — re-install the plugin"; exit 1; }
   

2. Ask:

   What should the synthetic failure look like?
     [Build failure (npm run build:* exit 1)]
     [Deploy workflow failure (gh actions check_run)]
     [Health check failure (curl /health 503)]
     [Version mismatch (served SHA != merged SHA)]
   

3. Generate a fake event payload matching the chosen kind. Set env var OPS_DEPLOY_FIX_DRY_RUN=1 and invoke the relevant trigger script:

   export OPS_DEPLOY_FIX_DRY_RUN=1
   export OPS_DEPLOY_FIX_TEST_REPO="your-org/your-repo"   # placeholder per Rule 0
   export OPS_DEPLOY_FIX_TEST_BASE="dev"
   case "$kind" in
     build)   bash "$PLUGIN_ROOT/bin/ops-deploy-fix-build-trigger" --synthetic ;;
     deploy)  bash "$PLUGIN_ROOT/bin/ops-deploy-fix-merge-trigger" --synthetic ;;
     health)  bash "$PLUGIN_ROOT/scripts/ops-deploy-monitor.sh" --synthetic-health ;;
     version) bash "$PLUGIN_ROOT/scripts/ops-deploy-monitor.sh" --synthetic-version ;;
   esac
   

   Trigger scripts must honor OPS_DEPLOY_FIX_DRY_RUN=1 by:

   • Skipping the actual claude -p ... dispatch in dispatch_fix_agent
   • Writing a would-dispatch-<id>-<ts>.log to $LOGS_DIR instead
   • Still incrementing the budget counter (so the test exercises the cap)
   • Logging [DRY-RUN] would notify <channel> instead of firing real notify (no actual outbound messages during test — Rule 6)

   If the trigger scripts don't yet support --synthetic / OPS_DEPLOY_FIX_DRY_RUN, surface that as a known TODO in the output and degrade to "would-have-dispatched" log-only.

4. Print the resulting would-dispatch-*.log path and a one-line classification result (transient? dedup-hit? budget-exhausted? would-dispatch-template=<file>).

5. Offer:

   Test complete. Next?
     [Tail the would-dispatch log]
     [Run another test]
     [Show status dashboard]
     [Done]
   

Failure modes / hand-offs

• No prefs file → tell the user to run /ops:setup (or /ops:deploy-fix configure) first.
• claude CLI missing (status / test) → instruct via output, do not auto-install.
• jq missing → run brew install jq in background (Rule 4) and retry.
• Stale lock detection → see status flow above, requires Rule-5 confirmation to clear.
• Notify channel mis-set (e.g. discord selected but no webhook) → on status, surface as ⚠️ notify_channel=discord but no webhook URL configured and offer [Reconfigure now].

Files this skill reads / writes

• Read-only: ~/.claude/plugins/data/ops-ops-marketplace/preferences.json, ~/.claude/state/ops-deploy-fix/*, ~/.claude/logs/ops-deploy-fix/*, ${CLAUDE_PLUGIN_ROOT}/scripts/lib/deploy-fix-common.sh, ${CLAUDE_PLUGIN_ROOT}/prompts/{build-fix,deploy-fix}.md
• Write (with explicit confirmation only): lock files in ~/.claude/state/ops-deploy-fix/lock-* (clear on stale)
• Write (via configure subcommand): the prefs file, via the merge pattern in Step 6.5