handoff

Orchestrate Handoff

STOP — Check triggers before proceeding

Do NOT run this skill unless one of these is true:

• The user explicitly asked for a checkpoint/handoff/resume
• CC displayed an actual system usage-limit warning in this turn
• /orchestrate is auto-checkpointing after a completed wave
• A fresh session is resuming a prior orchestration

If none apply, exit this skill immediately. "The session is long" or "context might be getting large" are NOT triggers. Trust the system to signal real limits.

---

Solves: when CC runs the /orchestrate orchestrator and hits a 5h usage limit mid-run, the in-flight task table, wave state, and dispatched job IDs are stranded inside CC's conversation. Workers (Codex/Cursor daemons) keep running on disk, but no agent can pick up the orchestrator role without losing them.

This skill writes orchestrator state to disk after every wave so any agent (CC, Cursor, Codex via codex-rescue) can re-attach.

Three modes

Mode	When	Who calls it
checkpoint	After every wave completes, or on demand	/orchestrate (auto) or user
prepare-handoff [cursor|codex]	When usage limit is imminent	User (or CC reactively)
resume	Fresh session picking up an orchestration	Receiving agent (CC/Cursor/Codex)

State location

Repo-local: <repo-root>/.orc/ — gitignored, follows the work, no global discovery step needed. Files:

• state.json — structured task table, wave history, in-flight jobs, routing context paths
• HANDOFF.md — human-readable narrative with paste-ready resume prompts for Cursor/Codex
• tasks.json — last TaskList snapshot (orchestrator pipes this in before checkpoint)

Add .orc/ to .gitignore on first checkpoint if missing.

Mode 1: checkpoint

Invocation (from /orchestrate or user):

bash ~/.claude/skills/orchestrate-handoff/scripts/checkpoint.sh

The script reads from environment variables. The orchestrator MUST export these before calling:

Env var	Contents
ORCH_PLAN_REF	Path to plan file, or freeform:<short hash> if no file
ORCH_WAVE	Current wave number (integer)
ORCH_WAVE_STATUS	dispatched / polling / complete
ORCH_INFLIGHT_JOBS	Newline-separated agent:job_id:label triples (poll-wave.sh format). Empty if no jobs in flight.
ORCH_ROUTING_FILES	Newline-separated .mex/context/*.md paths loaded in Phase 0
ORCH_NEXT_ACTION	One-line description of next step
ORCH_NOTES	Optional: open decisions, blockers, recent reroutes

For task table: orchestrator can optionally write .orc/tasks.json (e.g., echoing TaskList output as JSON) before invoking the script. If absent, checkpoint still runs but state.json's task table will be empty.

What the script does:

1. Creates .orc/ if missing, adds entry to repo .gitignore
2. Writes state.json from env vars + tasks.json (if present) + ISO timestamp
3. Renders HANDOFF.md with current state and paste-ready Cursor/Codex resume prompts
4. Echoes the path to HANDOFF.md

The script is idempotent and overwrites previous checkpoints (not append-only — the latest state is always authoritative).

Where /orchestrate calls this: at the end of Phase 2's per-wave loop, after the review step but before dispatching the next wave.

Mode 2: prepare-handoff

When usage limit is imminent, run:

bash ~/.claude/skills/orchestrate-handoff/scripts/prepare-handoff.sh cursor
# or
bash ~/.claude/skills/orchestrate-handoff/scripts/prepare-handoff.sh codex

Both targets read the existing state.json and HANDOFF.md (call checkpoint.sh first to refresh them).

Cursor target: prints a complete prompt to paste into Cursor agent mode. The prompt:

• Names the absolute repo path
• Tells Cursor to read .orc/HANDOFF.md and .orc/state.json
• Instructs Cursor to invoke resume.sh
• Emphasizes that Codex/Cursor worker daemons may still be running

Codex target: prints a codex-rescue agent prompt. Codex-rescue already knows the runtime — the prompt just declares "this is a resume, not a fresh start; load .orc/HANDOFF.md first."

Mode 3: resume

Receiving agent invokes:

bash ~/.claude/skills/orchestrate-handoff/scripts/resume.sh

The script:

1. Reads state.json — fails clearly if missing
2. Re-attaches polling on in-flight jobs by running poll-wave.sh with the saved ORCH_INFLIGHT_JOBS triples (only if non-empty)
3. Prints a status table: which jobs finished while orchestrator was offline, which are still running, which failed
4. Echoes the routing file paths so the agent can Read them directly
5. Prints ORCH_NEXT_ACTION so the agent knows what to do next

After this, the receiving agent has full state and can continue /orchestrate's Phase 2 loop, calling checkpoint.sh again after the next wave.

Integration with /orchestrate

Two minimal edits to ~/.claude/skills/orchestrate/SKILL.md:

1. Phase 0, after routing probe: export ORCH_ROUTING_FILES.
2. Phase 2, end of per-wave loop: export the wave env vars and call checkpoint.sh.

These edits are non-breaking — checkpoint.sh exits cleanly if env vars are partial.

What does NOT get checkpointed

• Worker outputs — already on disk (files written, git diffs, codex/cursor logs). State only references paths and job IDs.
• CC conversation transcript — irrelevant; the orchestrator's decisions are captured in ORCH_NOTES and ORCH_NEXT_ACTION.
• CLAUDE.md, ROUTER.md content — paths only; resumer reads them fresh.

Keeps state.json small (~5–20KB).

Failure modes

Symptom	Cause	Fix
resume shows job as "unknown"	Job ID expired (codex 24h retention, cursor pid died)	Treat as failed; re-dispatch task
state.json missing	Checkpoint never ran	Reconstruct manually from plan file + git log
Routing files moved	Repo restructured between checkpoint and resume	Re-run Phase 0 of /orchestrate