Two-step pipeline. Reads past sessions, never your repo.

Step 1: Extract

Per-session Claude sub-agent pulls bug snippets from session JSONLs. Costs tokens. Idempotent. Default cap 200/run.

Step 2: Infer

Local fingerprinting groups sessions by error shape. No LLM cost. Add --semantic for fuzzy merging via local embeddings.

Output

Clusters of recurring bugs. Mark "this fix worked" once; future occurrences surface the linked resolution.

Bug Patterns

Pipeline above. Cost guidance:

#Step 1 burns tokens

bash

recall infer outputs --project <name>             # canonical
recall infer outputs --project <name> --limit 100 # tighter cap
recall infer outputs --project <name> --yes       # skip 50-session warning
recall infer outputs --project <name> --force     # re-extract done sessions

Default cap: 200 sessions/run. Safety guard: > 50 sessions requires --yes.

#Or: let the daemon nibble passively

bash

recall semantic auto-extract on

One un-extracted session per hour, ~5 per 5-hour Claude window, leaving 95%+ of your plan rate budget free. Patterns and Galaxy populate over days without you running extract per project.

#Step 2 is local, free, automatic

bash

recall infer bug-patterns                          # cluster by fingerprint
recall infer bug-patterns --semantic               # fuzzy match wording
recall infer bug-patterns --min-cluster-size 2     # surface pairs (default 3)

The daemon also runs this on a 30-minute timer in the background. Re-running absorbs new matches into existing clusters.

#Strategy

Don't sweep every project. Pick the 3–5 you mine bugs from regularly. Archive projects rarely repay the token cost.