use-statem

name	use-statem
description	Manage long Claude Code work with statem state-machine runbooks. Use when tasks need durable state, explicit transitions, redo loops, save/resume behavior, or cross-agent review.
argument-hint	["task-or-spec"]
allowed-tools	Bash(statem *) Bash(python3 -m statem *)

Use Statem

Use statem as a state-aware runbook for long agent runs. It should guide the work and preserve progress without turning the session into a rigid workflow harness.

Operating Loop

1. Inspect or create a spec such as statem.yaml.
2. Run statem validate statem.yaml --json.
3. Start or resume with statem start statem.yaml --run-id <id> --json.
4. Before work, run statem cur --run-id <id> --json.
5. When a state is complete, run statem goto <next> --run-id <id> --json.
6. Before pausing, run statem save --run-id <id> --json.
7. For handoff context, read recent history: statem history --run-id <id> --tail 10 --json.

If statem is not on PATH, use python3 -m statem from the repository that contains the Python package.

For dynamic servers, company machines, or disposable git checkouts, prefer a machine-local state directory. Set STATEM_STATE_DIR once, for example $HOME/.local/state/statem/<project>, so runtime state survives checkout replacement. After moving to a new checkout, run statem start <spec> --run-id <id> once to rebind the run to the current spec path.

Commit YAML runbooks with the repo. Do not commit runtime state; treat .statem/ or STATEM_STATE_DIR as local, copy-on-use execution data.

Context Clear

Avoid /clear in normal loops. It flushes the conversation, including any instructions that were supposed to run after it, and can lose useful intent that was not written to durable files. Prefer explicit transitions plus safe compaction. Before a hard clear, run statem prompt --run-id <id> and paste the generated prompt immediately after /clear. After a hard clear, recover from .statem with start, cur, and history instead of relying on previous chat context.

Loop Compaction

For cyclic runbooks, prefer an explicit session hygiene node after a full loop. When another cycle will continue and the context is noisy, run statem compact-prompt --run-id <id>, use the generated /compact instruction, then recover with statem cur and statem history --tail 10.

Agents may inspect the full graph with statem state; cur, next, and goto are for disciplined execution and attention anchoring, not for hiding the runbook.

Auto Loop Hook

When the host supports a Stop hook, users may opt into auto-loop behavior with integrations/hooks/statem_stop_hook.py. The hook runs when the agent is about to hand control back to the user. If a statem run is active, the current node is not a terminal/handoff node, and there are outgoing transitions, it returns a continuation prompt that tells the agent to inspect statem cur and keep working from the current node.

Registration examples live in:

• examples/hooks/README.md
• examples/hooks/codex-stop-autoloop.hooks.json
• examples/hooks/claude-stop-autoloop.settings.json

Merge the matching snippet into the host hook configuration and use an absolute script path if the hook is registered outside the statem repository.

Treat this as host-level glue. It must not advance state, run /clear, or hide the graph. The agent should still transition only with statem goto.

Spec Guidance

• Put runtime data under .statem/; do not mix it into the spec.
• Use in_hook for setup after entering a node.
• Use before_transfer for checks/redos while still in the current node. It is a spec field, not a CLI command; statem goto runs it automatically.
• Use out_hook to persist current-node progress before leaving.
• Use edge hook as prepare-transfer work after out_hook and before entering the target. If a blocking edge hook fails, the pointer stays at the source so the agent can retry.
• Use type: command for deterministic shell checks and hooks.
• Use type: predicate for file checks instead of shell when possible.
• Use type: llm_review to call Codex, Claude, or another reviewer command.
• Treat blocked transitions as useful feedback: stay in the node, fix the issue, then retry.