git:20261102.8d2abe5 to git:20261104.ccd4cf7

18 added, 4 removed. Audit A to A.

---

name: review-trades

description: Review the desks' closed trades - rebuild the trade ledger from public fills, test every slice for fluke or edge, list backtest proposals. Read only; never edits live config or places an order. Runs daily from /desk-loops and once after any close.

---

# review-trades

The learning loop (operator, 2026-11-02: "continuously evolve, learn from

what we did right, fix what we did wrong, see if mistakes and successes are

just flukes or a predictions"). One script call; no agent, no LLM, no MCP.

1. **Run it** from the repo root:

python3 -m services.review run

It fetches the wallet's fills (public Hyperliquid reads), updates

`$PATHIEL_STATE_DIR/review/ledger.jsonl`, writes

`docs/strategy/learning/REVIEW_<UTC date>.md` and prints one line that

starts `REVIEW <utc> trades=...`.

2. **Print that line.** Nothing else when it says `new_leaks=none` and

`new_proposals=0`.

3. **Something new?** If `new_leaks` names a slice or `new_proposals` is

above 0: say so in one short message (the slice, its label, n, and the

lever the report proposes, from the report's "What to try next"), and

append the summary line to `/tmp/oil-loop/progress.log` as

`REVIEW <utc> <rest of the line>`.

- 4. **If it fails** (network, a parse error): print the last line of the

+ 4. **A new LEAK goes to the operator as a yes/no, the same day**

+ (operator, 2026-11-04: "aren't we grading and growing... too many tests

+ not enough results"). When `new_leaks` names a slice, send ONE

+ `PushNotification` and print the same question:

+ LEAK <slice>: <n> trades, <wins> win(s), <net $>. Turn on <lever>? yes/no

+ `<lever>` is that slice's line under "What to try next" (its `Try` and

+ what it does in plain words). Ask once per new leak. A yes is the go:

+ make the change in the desk rules with a gate test and a real-fills

+ eval (`services/strategies/evals/no_break_real_fills.py` is the pattern),

+ commit, and say what to restart. A no: note it in the next report run and

+ do not ask again for that slice. The review itself still changes nothing.

+ 5. **If it fails** (network, a parse error): print the last line of the

error once and stop. Do not retry in a loop; the next scheduled run will.

## Rules

- - Proposals are backtested on the operator's OK, with the command the

- report gives. Adopt only what beats live on most datasets as a plateau,

- and then only on the operator's go.

+ - LEANING proposals are backtested on the operator's OK, with the command

+ the report gives, and adopted only on the operator's go. A LEAK (10+

+ trades, labeled LEAK) skips the backtest: it goes straight to the

+ operator as the yes/no in step 4.

- Never edit live config, `.claude/agents/desk-tick.md`, the desk-tick

skill, or a live constant from this skill. Never call a pathiel MCP tool

here. The review only reads.

- Never feed the report into a thesis prompt (SCORECARD_TEST_2026-10-31.md:

the desk's own record in its prompt made money worse).

- n is small: an EDGE or LEAK needs at least 10 trades in one slice. Say

LEANING, FLUKE or NOISE as the report does; never round them up.