v1.17.0 — Scheduled Autonomy: strategies that run while you sleep
Added — Scheduled Autonomy: strategies that run while the user sleeps
The missing connective tissue between the strategy planners, the agent wallet, and Claude Code's scheduled agents. Three independent safety layers: the saved plan + execution journal (what/when, idempotent), the policy gate (per-tx + rolling-24h velocity caps), and the schedule itself (which holds zero spending authority).
- 2 new MCP tools (129 → 131):
chaingpt_strategy_due_steps— which steps of a saved plan are due now (overdue + configurable lookahead) and not yet executed. The heart of crash-safe scheduled execution: re-runs never double-execute.chaingpt_strategy_mark_step— record execution (or a deliberate skip) in the plan's journal with tx hash. Idempotent: double-recording is refused (a refusal means something double-fired);overwrite: trueonly corrects wrong entries.
- Plan journal —
executionsmap persisted alongside (not inside) the payload, so the original plan stays pristine. Local files only, nothing uploaded. chaingpt_strategy_dca_plannow emits a machine-readable payload ({ kind, network, outToken, steps: [{id, atUnix, action, usd}] }) ready forsave_plan→ scheduled execution.- New
scheduled-autonomyskill (22 → 23) — the full loop: plan → save →/schedulecron → each tick: due steps → fresh quote pre-flight → policy-gated sign → mark. Hard rules (mark immediately, never before send, refusals are final), failure table, agent-wallet vs user-signs modes. - New reference walkthrough (
reference/scheduled-autonomy.md, 19 → 20) — zero to a running daily DCA in 10 minutes, including the dry-run journal test, thedca-basepolicy pairing, catch-up cadence design, and recipes beyond DCA (Aave health sentinel, funding monitor, grid rebalance, weekly risk re-check).
Tests
- Plans suite 6 → 11 (due/lookahead windowing, journal lifecycle, idempotency refusal, unknown-step rejection, skip + completion reporting, non-schedulable-plan guidance). Suite total 356 → 361.