recall-pattern

name	recall-pattern
description	Reads state/learnings.json and returns the historical μ/σ/p50/p95 for a requested attribution key (plugin:skill:agent_tier). Used by L3 Z-Score Cost Anomaly as a historical prior when the current session has fewer than 30 matching observations. Use when: detect-anomaly needs a prior and in-session N is insufficient. Do not use to compute costs or fire events — this skill is purely a read over the accumulator.
model	haiku
tools	["Read"]

recall-pattern

Preconditions

• state/learnings.json exists (returns empty pattern if not)
• The caller provides a well-formed attribution key

Inputs

• Attribution key: <plugin>:<skill>:<agent_tier> string
• Fallback depth: if exact key not found, try <plugin>:<agent_tier> (less specific), then <plugin> alone

Steps

1. Load state/learnings.json.
2. Look up the exact key. If found, return {found: true, specificity: "exact", mu, sigma, p50, p95, n, last_session}.
3. If not found, try less-specific fallbacks in order. Return with specificity: "plugin_tier" or "plugin_only".
4. If nothing matches, return {found: false} — never fabricate a pattern.

Success criterion: response is complete (all keys present) or found: false. No partial returns.

Outputs

Returned JSON block:

{
  "found": true,
  "specificity": "exact",
  "key": "wixie:/converge:sonnet",
  "mu_cost_usd": 0.24,
  "sigma_cost_usd": 0.11,
  "p50": 0.22,
  "p95": 0.48,
  "n": 3421,
  "last_session": "2026-04-19T14:02:00Z"
}

Handoff

L3 anomaly detection uses the returned (mu, sigma) when in-session N < 30. If found: false, L3 returns insufficient_data and does not emit pech.anomaly.detected.

Failure modes

Code	Scenario	Counter
F02	Return a fake μ/σ so L3 can fire an anomaly on a brand-new session	found: false is the correct return; never invent
F13	Return stale pattern without last_session timestamp	Always include last_session so caller can decide if pattern is trustworthy