sigma-cost

sigma-cost

A closed cost loop for sigma's heavy ops: estimate before → measure after → sharpen next. The pure logic lives in cli/cost.py; this skill is the in-session side that reads the same ledger (sigma/costs.jsonl) and advises.

Boundaries (compose, don't duplicate)

• RTK cuts token overhead at the proxy layer (transparent command rewriting).
• caveman trims output verbosity.
• sigma-cost estimates/measures/routes sigma's own operations. It may recommend enabling RTK or caveman when it detects waste — it does not replace them.

Before a heavy op (estimate + route)

1. Size the work in units: review = axes × changed files; profile = files walked; loop = cycles; research = models.
2. Estimate tokens: units × tokens-per-unit. Use the calibrated factor from sigma/costs.jsonl when present; otherwise the static fallback.
3. Route by reasoning load (performance.md strategy):
   • code axis, mechanical edits → haiku (cheap).
   • system-logic, orchestration → sonnet.
   • ml-logic, hardest reasoning → opus.
4. Skip work that adds no value: a stale-but-cheap profile refresh, an axis with no relevant files.

After a heavy op (measure)

Append one row to sigma/costs.jsonl (append-only, like events.jsonl): {"ts", "op", "units", "tokens", "estimated", "models"}. The caller supplies the timestamp (deterministic projection — never generated in pure code).

Report

sigma cost renders per-op totals, biggest token sinks, and est-vs-actual drift (which shrinks as calibration kicks in). Use it to decide where routing or RTK pays off most.

Efficiency (in-session, no CLI needed)

sigma trajectory --efficiency (or read trajectory.jsonl directly in-session, the same way this skill already reads sigma/costs.jsonl without shelling out) answers a DIFFERENT question than the cost ledger: not "how many tokens did this burn" but "is the loop actually working, and is escalation earning its keep."

Read the current spec workspace's trajectory.jsonl and report:

• Cycle pass rate — the fraction of role: "cycle" steps with ok: true (one such step is appended per completed execute_cycle, real verified/not-verified data, not a token guess).
• Escalation rate — (logic + advisor + test-writer + simplifier steps) / implementer steps. High and climbing → the expensive axes are firing often; worth checking whether --advisor-rounds or --logic are pulling their weight.
• Crash rate — subprocess exit failures across all roles. Not the same as a verify-fail: a verifier returning VERDICT: FAIL still exits 0, so this measures agent-CLI crashes, not verification quality.

Do NOT report a token-per-cycle number here. claude -p never surfaces real token usage, so any per-cycle token figure would be units × static_factor wearing a measured metric's clothes — it never calibrates (see calibrate() above) and collapses to a restatement of the pass rate. If asked "how many tokens per successful cycle," say plainly that sigma has no real per-run token signal today and point to the pass-rate/escalation-rate numbers instead, or to sigma cost's explicitly-labeled estimate if a rough order-of-magnitude is wanted.

Fail-safe

A missing or corrupt ledger falls back to static estimates and never blocks the op (the inverse of a hard gate). Estimates are advisory — the operator keeps the wheel (loop-engineering principle). Empty/missing trajectory.jsonl → "no data yet", never a crash — same discipline.