scheduled-autonomy · diff
git:20260611.cf364c4 to git:20261030.2fdba0c
19 added, 67 removed. Audit A to A.
What this update changes in the checks
No check changed: every fixed check gives the same answer before and after this update.
A tripped check is a label, not a verdict: read the matched text and the lines below. This week's updates that newly trip a check.
---
name: scheduled-autonomy
- description: Run trading/DeFi strategies on a schedule — DCA every morning, grid rebalances, recurring health checks — using saved plans + the execution journal (due_steps/mark_step) + Claude Code scheduled agents, with the agent wallet executing inside hard policy caps. Triggers - run this daily, schedule a DCA, recurring buy, set and forget, autonomous strategy, cron strategy, execute my plan on a schedule, walk away.
+ description: Run scheduled research or prepare strategy steps with durable claims, stable operation IDs, receipt reconciliation and human confirmation. Triggers: daily DCA, recurring buy, scheduled strategy, due steps, execution journal.
---
- # Scheduled Autonomy — strategies that run while the user sleeps
-
- The full loop for hands-off strategy execution. Three layers make it safe:
- 1. **The plan** is explicit and saved to disk (what to do, when).
- 2. **The execution journal** makes runs idempotent (crashes and re-runs never double-buy).
- 3. **The agent-wallet policy gate** caps what any single run, or day, can spend — `maxTxValueWei` per tx, `maxDailySpendWei` + `maxDailyTxCount` per rolling 24h. The schedule cannot exceed the policy even if it fires a thousand times.
-
- ## The loop (memorize this shape)
-
- ```text
- ONE-TIME SETUP
- 1. chaingpt_strategy_dca_plan ... # produces steps + a JSON payload
- 2. chaingpt_risk_token / chaingpt_research_token # pre-flight the asset ONCE
- 3. chaingpt_strategy_save_plan name=eth-dca type=dca payload=<the JSON>
- 4. Agent wallet ready: chaingpt_agent_wallet_status
- → recommend the "DCA bot" policy template (single chain, single router, small caps)
- 5. Create the schedule (see below)
-
- EVERY SCHEDULED TICK (this is the prompt the schedule runs)
- 1. chaingpt_strategy_due_steps name=eth-dca # what's due? nothing → done, exit
- 2. For each due step:
- a. chaingpt_dex_quote ... # fresh quote, sanity-check price impact
- b. execute:
- - agent-wallet mode: chaingpt_agent_wallet_sign_and_send (policy gate decides)
- - user-signs mode: chaingpt_dex_build_swap_tx acknowledgeMainnet=true → present tx
- c. chaingpt_strategy_mark_step name=eth-dca stepId=<id> txHash=<hash> # IMMEDIATELY
- 3. Report: step(s) executed, spend vs daily cap, next step time
- ```
-
- ## Creating the schedule (Claude Code)
-
- Use Claude Code's scheduled agents (`/schedule`) or any cron that invokes `claude`:
-
- ```text
- /schedule create "eth-dca tick" --cron "0 9 * * *" --prompt
- "Run one tick of the saved plan 'eth-dca': call chaingpt_strategy_due_steps name=eth-dca;
- if nothing is due, stop. For each due step: fresh chaingpt_dex_quote (abort the step and
- mark it skipped with a note if price impact > 1%), execute via the agent wallet
- (chaingpt_agent_wallet_sign_and_send, memo 'eth-dca step <id>'), then immediately
- chaingpt_strategy_mark_step with the tx hash. Finish with a one-line summary."
- ```
-
- Cadence note: schedule ticks MORE often than the plan cadence if you want catch-up behavior (a daily plan + daily tick that fails once skips a day; a daily plan + 6-hourly tick self-heals). `due_steps` returns everything overdue, so catch-up is automatic — the daily-spend cap bounds how much catch-up can execute at once.
-
- ## Hard rules
-
- - **Mark immediately after execution.** An executed-but-unmarked step is returned again next tick. The journal is the only memory the schedule has.
- - **Never mark before the tx is sent.** A marked-but-unexecuted step is silently lost money on the table.
- - **Already-marked refusal means double-fire.** If `mark_step` says the step is already recorded, do NOT overwrite — investigate why the tick ran twice.
- - **The policy gate's refusal is final.** If `sign_and_send` is refused (daily cap reached, kill switch on), mark NOTHING, report the refusal, and stop. Never coach the user into widening caps mid-run; that decision belongs to a non-scheduled session.
- - **Pre-flight the asset once at setup, and re-check risk weekly** (`chaingpt_risk_token`) for long-running plans — tokens get compromised after launch too.
- - Plans + journal are local files (`~/.chaingpt-mcp/plans/`). Nothing is uploaded.
-
- ## Failure handling
-
- | Situation | Behavior |
- |---|---|
- | RPC/quote upstream down | Skip the tick (steps stay due; next tick catches up) |
- | Price impact above the user's threshold | `mark_step status=skipped note="impact 2.3%"` — a skipped step is a decision, not a failure |
- | Policy refusal (caps/kill switch) | Stop the tick, report, leave steps due |
- | Plan complete | `due_steps` says COMPLETE — tell the user to remove the schedule |
+ # Scheduled strategy execution
- ## Modes
+ Read `docs/wallet-safety.md`. Saved plans and execution claims are local. Scheduling does not
+ create spending authority or guarantee exactly-once execution of arbitrary external actions.
- - **Agent-wallet mode (fully autonomous):** the wallet signs inside policy caps. Pair with the `dca-base` policy template. The PreToolUse mainnet guard will ask for confirmation in interactive sessions; for headless scheduled runs, the policy gate + velocity caps are the (deliberate) safety boundary — set `CHAINGPT_GUARD=off` ONLY in the scheduled environment, never globally.
- - **User-signs mode (semi-autonomous):** the tick prepares unsigned transactions and the summary asks the user to sign. Zero custody risk; the user is the scheduler's hands.
+ 1. Generate a strategy plan and review its assets, chain, sizes and timing.
+ 2. Save it with `chaingpt_strategy_save_plan`.
+ 3. Prefer scheduled research/unsigned preparation with a human confirming and executing transfers.
+ 4. For an explicitly authorized agent-wallet action, call `chaingpt_strategy_due_steps claim=true`.
+ Pass the returned operationId unchanged to the signer. Keep the confirmation hook enabled.
+ 5. Re-run risk/quote/health checks as appropriate. Restricted signers reject opaque protocol calls;
+ use the unsigned builder and external wallet for those operations.
+ 6. Broadcast once and immediately mark the result with its transaction hash. Report pending versus
+ confirmed accurately. Completed journal entries are immutable.
- Solana plans (v1.19+): same loop with `chaingpt_dex_jupiter_quote` → `chaingpt_dex_jupiter_build_swap_tx` → `chaingpt_agent_wallet_solana_sign_and_send` → `chaingpt_strategy_mark_step`. The lamport velocity caps (`policy.solana.maxDailySpendLamports` / `maxDailyTxCount`) are the third safety layer, exactly like the wei caps on EVM.
+ A crash leaves the step IN FLIGHT. Reconcile with the same operationId; never execute it using a
+ new id. Pending transactions keep their native and token reservations until resolved. A policy
+ refusal is final for this tick. Do not widen caps, delete a claim or disable the guard to continue.
- Related: `skills/strategy/SKILL.md` (planners), `skills/agent-wallet/SKILL.md` (policy gate), `reference/scheduled-autonomy.md` (10-minute end-to-end walkthrough).
+ Unattended fund movement requires a separately reviewed executor and approval boundary. This skill
+ ships no instruction to disable per-action confirmation. The generic scheduler cannot infer that
+ new operation IDs represent duplicate human intentions, or make external wallet execution atomic.