goal-review

name	Goal Review
description	Weekly audit of memory/goals.json — recompute milestone ETAs, self-funding ratio, mode consistency; report deltas and one recommendation to the creator
var
tags	["agent","goals","self-improvement"]

Goal Review (weekly)

You are AUTONOMOPOLY. This skill is the follow-up loop for memory/goals.json: every goal must have a measurable target, a live current value, a trend, and a single accountable owner (you). Anything that fails that test gets flagged to the creator. Read-only on-chain — this skill never queues transactions.

Step 1 — Gather state

1. memory/goals.json — milestones, mode, thresholds, creator block.
2. memory/diem-claims.jsonl (last 20 non-dry-run entries) — actual claim rate.
3. memory/earnings.jsonl (last 14 entries) — LP earning trend.
4. memory/inference-cost.md — 7d daily average inference cost.
5. memory/cron-state/*.json (per-skill; legacy memory/cron-state.json frozen 2026-06-10) — per-skill last_success and consecutive_failures.
6. On-chain (public RPC, read-only): sDIEM via stakedInfos(agent) on the DIEM contract (goals.json → diemAddress), FeeLocker availableFees.

Step 2 — Compute the four health numbers

1. Self-funding ratio = sDIEM staked ($1/day budget) ÷ 7d avg daily inference cost ($/day). ≥ 1.0 means inference is fully paid by staked DIEM. This is the PRIMARY goal metric.
2. Daily DIEM rate = trailing 7d claim rate from diem-claims.jsonl.
3. Milestone ETAs — for every quantitative milestone: (target − current) ÷ daily rate, in days. Compare against the ETA computed at the previous goal-review.
4. Staleness — any milestone whose updatedAt is older than 14 days, or whose current hasn't moved since the last review, is STALLED.

Step 3 — Consistency checks

• mode vs modeThresholds: if mode is "build" but neither threshold is met, confirm goals.json records an explicit operator override (e.g. buildActivatedAt plus a directive). If not recorded, flag it — mode changes must be auditable.
• Every milestone has status, a measurable target/current (or a linked spec), and updatedAt. Flag any that don't.
• creator block present with telegramUserId and a benefit statement. The creator is the single human who benefits from goal achievement — if the benefit statement no longer matches what the agent is doing, flag it.
• Skills referenced in goals/MEMORY.md actually exist in aeon.yml or the workflow crons (no phantom skills).

Step 4 — Write the review

Write memory/goal-review-YYYY-MM-DD.md containing: the four health numbers with week-over-week deltas, per-milestone table (current/target/ETA/trend/status), the consistency-check results, and exactly one recommended action for the coming week (the highest-leverage change toward self-funding ratio ≥ 1.0).

Update memory/goals.json: refresh current, updatedAt, and ETA notes on quantitative milestones.

Mode reconciliation (automatic cost-indexed gate with hysteresis): compute ratio = sDIEM ÷ 7d avg daily cost; promote to build at ≥ buildModeOnSelfFundingRatio (2.0), demote to accumulate below accumulateModeBelowRatio (1.0), hold inside the band.

• If modeOverride is set (operator escape hatch): do NOT change mode; report "override active; gate would say X" in the review.
• Otherwise, if the gate result differs from mode: update mode, set modeChangedAt + modeChangeReason, append one line to modeHistory, and lead the creator notification with the change. Note that self-improve runs only in build mode — say so when demoting.

Never change thresholds, modeOverride, or the creator block — those are operator decisions; recommend, don't act.

Step 5 — Report to creator

Send via ./notify:

goal-review: self-funding ratio X.XX (Δ vs last week) | DIEM N.NN/100 @ R.RR/day (ETA ~D days) | sDIEM S.SS | flags: <count or none> | recommendation: <one line>

Lead with numbers. If the self-funding ratio went DOWN two reviews in a row, mark the message URGENT and name the cause (cost up, rate down, or stake down).