coord-respawn-self

Coord context-aware self-respawn. Invoked by a Coord agent when context window reaches 80% threshold mid-L3. Saves mid-L3 state, writes a continuation manifest, notifies the spawner (PD or Dept Head d

SKILL.md

GitHub

/coord-respawn-self — Coord Context-Aware Self-Respawn

When to Invoke

  • Mandatory: when context window reaches 80% (CONTEXT_PCT_ALERT fires)
  • Proactive: at a clean sub-task boundary within L3 when context > 70%
  • Never mid-Exec: do not invoke while an Executor is awaiting APPROACH or CHECKPOINT response — finish that exchange first, then respawn

Respawn Counter Enforcement

Max 3 respawns per Coord per 24h. Uses the same counter mechanism as /respawn-self: ~/.claude/state/respawn-counters/coord-{l3-name}-{pun}.txt

COORD_ID="coord-{l3-name}-{pun}"
COUNTER_FILE="$HOME/.claude/state/respawn-counters/$COORD_ID.txt"
NOW=$(date +%s)
DAY=86400

mkdir -p "$HOME/.claude/state/respawn-counters"

if [ -f "$COUNTER_FILE" ]; then
  COUNT=$(awk '{print $1}' "$COUNTER_FILE")
  RESET_AT=$(awk '{print $2}' "$COUNTER_FILE")
  AGE=$(( NOW - RESET_AT ))
  if [ "$AGE" -ge "$DAY" ]; then
    COUNT=0
    RESET_AT=$NOW
  fi
else
  COUNT=0
  RESET_AT=$NOW
fi

if [ "$COUNT" -ge 3 ]; then
  echo "RESPAWN_BLOCKED: max 3 respawns reached for $COORD_ID in 24h."
  exit 1
fi

NEW_COUNT=$(( COUNT + 1 ))
echo "$NEW_COUNT $RESET_AT" > "$COUNTER_FILE"
echo "RESPAWN_OK: respawn $NEW_COUNT/3 for $COORD_ID"

If RESPAWN_BLOCKED: escalate to PD with current state — do not continue.

Step 1 — Complete Current Work Unit

  • Finish any pending APPROACH or CHECKPOINT gate exchange
  • Do NOT start a new Exec spawn
  • Reach a clean "N Execs ACKed, M pending" boundary

Step 2 — Write Coord Continuation Manifest

Write to {project}/memory/agents/coords/coord-{l3-name}-{pun}-respawn-{timestamp}.md:

# Coord Respawn Manifest — Coord-{l3-name}-{pun} — {timestamp}
Context at respawn: {PCT}%
Respawn count: {N}/3
L3 task: {l3-task-description}

## Sub-tasks Completed
- {Exec-name}: {status} — {1-line summary}
- ...

## Sub-tasks In Flight at Respawn
- {Exec-name or task}: {state — e.g. "APPROACH received, not yet approved"}
- ...

## Sub-tasks Not Yet Started
- {task}: {description}
- ...

## Scratch State
(copy current ## Status and ## Children tables)

## Next Action for Fresh Coord
{exact next action — e.g. "Spawn Exec-{X} for task Y, then Exec-{Z} for task W in parallel"}

## Active Blockers
{none | list}

Step 3 — Notify Your Spawner

Determine who your spawner is:

  • If you are an execution Coord (spawned by a PD, reports_to: pd-coordinator): → Send to "PD-{slug}" via SendMessage.

Message format (substitute the correct recipient):

Coord-{l3-name}-{pun}: COORD_RESPAWN TRIGGERED
Context at respawn: {PCT}%
Respawn count: {N}/3
L3: {l3-task-name} — {completed}/{total} sub-tasks done
Continuation manifest: {project}/memory/agents/coords/coord-{l3-name}-{pun}-respawn-{timestamp}.md
A fresh Coord session is needed to continue this L3.
Spawner should start a new Coord-{l3-name}-{pun} with the manifest path in the spawn prompt.

Step 4 — Delete Scratch and Stop

Delete {project}/memory/agents/coords/coord-{l3-name}-{pun}-scratch.md. Stop. The fresh Coord will read the manifest and continue.