senpi-strategy-ops

Senpi Strategy Ops — deploy, monitor, close

A strategy is a package: strategy.yaml (the deploy manifest) + one runtime.yaml per instance + <instance>/scanners/. The runtime spawns and supervises each scanners/scan.py, calling scan(inputs, ctx) every interval_seconds and owning everything downstream — signal validation, sizing/execution, the two-phase DSL exit, risk guard-rails. There is no separate scanner daemon, no push_signal. Ops owns the deployed lifecycle. Deploy is one verb, run as a detached job: funds preflight → wallet create+fund → runtime install → one observed scanner tick. It returns in ~1s and you poll until it is terminal.

openclaw senpi validate <recipe-dir>                                    # 0a. does it RUN? records the proof create needs
python3 /data/.openclaw/skills/senpi-strategy-ops/scripts/deploy.py validate <id>              # 0b. preflight — structurally deploy-ready? (no money, nothing installed; a bare id is fetched to disk)
python3 /data/.openclaw/skills/senpi-strategy-ops/scripts/deploy.py fork <id>                  # 0c. their copy on disk — only when levers move before funding
python3 /data/.openclaw/skills/senpi-strategy-ops/scripts/deploy.py create <id> --budget <usd> # 1. THE FUNDED PATH: forks it under their username, proves it, starts the deploy
python3 /data/.openclaw/skills/senpi-strategy-ops/scripts/deploy.py status                    # 2. read the job until terminal; the verified report (read-only)
python3 /data/.openclaw/skills/senpi-strategy-ops/scripts/status.py                            # what am I running? (+ health)
python3 /data/.openclaw/skills/senpi-strategy-ops/scripts/close.py <id> | --all                # teardown one strategy | EVERY open strategy

"Install"/"reinstall " = deploy.py create <id> — a strategy is a package, not an openclaw/clawhub skill; not on disk just means fetch it by id. Fund through deploy.py create|runtime <id>, not through the bare verb. Both resolve the package, run the structural preflight, then start the runtime's senpi deploy job, poll it, and relay its report verbatim. The wrapper's value is resolution, that structural pass and the verbatim relay — not a gate the verb lacks: the live-universe gate is the verb's own and it fires pre-money. Both call openclaw gateway call senpi.* (~1 s; the verbs take 30-160 s on 10.x): read the job with deploy.py status. Always tear down through close.py (one <id> or --all) — it deletes the runtime and closes the strategy. A raw strategy_close MCP call closes the strategy but leaves the runtime registered, which collides on the next deploy. close.py is never a repair. It market-exits every position and returns the funds; a deploy after it is a NEW wallet whose scanner state starts empty — needed only when update refuses E_UPDATE_WALLET_CHANGED (below). "Relaunch", "restart", "redeploy", "fix it" → the read-only deploy.py verify on the id they run (status.py lists their fork). VERIFIED = nothing to relaunch; "my edit isn't live" / "it isn't trading" → status.py <id>, its ✎ / ⏸ rows; healthy but has never traded → its binding gate, threshold and universe count, never "being selective" (references/liveness-verification.md). A step verify marks RUNS THE DEPLOY VERB needs their yes first: runtime <id> starts trading the wallet it adopts, create --budget funds a NEW one (say the amount). PAUSED has no resume: tell the user, run nothing. Close only on "close" / "tear down" / "withdraw" in their own words, or a yes to your sentence that every position closes at market. A bare "stop" / "pause" / "stop trading" is ambiguous: ask once, "turn on watch-only (no new entries; open positions keep their DSL stops) or close everything at market?" Watch-only is decision_mode: none (below). Never call the MCP strategy_pause on a runtime strategy: it is for mirror strategies and has no resume. Pass the strategy id for a CATALOG strategy (what senpi-strategy-discover hands over, e.g. spider), fetched from the remote if not on disk; for a locally-authored package, pass its DIRECTORY path. Either way the package belongs in the durable strategies root (SENPI_STRATEGIES_DIR if set, else the agent workspace strategies/ dir — normally /data/workspace/strategies) and never inside a skill directory: a package written there is destroyed on the next skill update. An on-disk package is authoritative; an invalid one surfaces its real errors and is never silently replaced by a remote fetch. Mechanics + state machine: references/lifecycle.md · manifest schema: references/strategy-yaml-schema.md.