accumulate-pattern

name	accumulate-pattern
description	Updates per-developer cost-pattern accumulator in state/learnings.json with the current session's observations. Rolls μ and σ per (skill, session_type, model) axis using a slow exponential update (α = 0.05) so one noisy session doesn't skew learned patterns. Exports a cross-plugin-readable snapshot to shared/learnings.json. Use when: PreCompact hook fires (end-of-session or compaction trigger). Do not use during an active session — runtime invocation mid-session would distort the pattern.
model	sonnet
tools	["Read","Write"]

accumulate-pattern

Preconditions

• The session has at least 5 ledger rows (otherwise there's nothing useful to accumulate)
• state/learnings.json exists or can be initialized (first run)

Inputs

• Session ledger: plugins/cost-tracker/state/ledger-YYYY-MM.jsonl (current month)
• Current session_id: from ENCHANTED_ATTRIBUTION.session_id
• Prior learnings: state/learnings.json (or empty {patterns: {}} on first run)

Steps

1. Group current session's rows by attribution key <plugin>:<skill>:<agent_tier> (or fall back to <plugin>:<agent_tier> if skill is null, e.g. for hook-driven observations).
2. For each key, compute session-local stats: session_mean, session_stdev, session_n.
3. Load state/learnings.json. For each key:
   • New key: initialize {mu = session_mean, sigma = session_stdev, n = session_n, p50 = session_median, p95 = session_p95}.
   • Existing key: update via slow exponential:

     μ_new = 0.95 × μ_old + 0.05 × session_mean
     σ_new = 0.95 × σ_old + 0.05 × session_stdev
     n_new = n_old + session_n
     

     (p50 and p95 use Welford-style online percentile estimators over the full history — they're informational, not used in L3 anomaly math.)
4. Increment n_sessions_accumulated. Update last_updated timestamp.
5. Write state/learnings.json atomically (write-to-tmp + fsync + rename — Emu-A4 pattern).
6. Append a cross-plugin snapshot to shared/learnings.json with origin: "pech" and the full patterns map.
7. Emit pech.learning.pattern.updated event (one per PreCompact, not per key).

Success criterion: atomic write succeeded (no partial file on crash); shared/learnings.json appended (not overwritten — peers own their own sections); exactly one event fired regardless of how many keys updated.

Outputs

• state/learnings.json (updated)
• shared/learnings.json (appended with Pech's section)
• Event: pech.learning.pattern.updated

Handoff

• budget-watcher/detect-anomaly reads state/learnings.json for the historical prior when current-session observations are < 30.
• Peer plugins (Wixie, Sylph) read shared/learnings.json#pech for their own cost-aware decisions.

Failure modes

Code	Scenario	Counter
F09	Concurrent PreCompact from parallel sessions clobbering learnings.json	Atomic rename is mandatory; last-writer-wins is acceptable for slow-update semantics
F11	Tune α up to 0.2 to "learn faster" — session-specific noise leaks into patterns	α = 0.05 is the contract; faster learning defeats the purpose of cross-session stability
F13	Accumulate orphan rows (missing attribution) into a "misc" bucket	Orphans are never accumulated — they're a bug signal, not a data point