Documentation
¶
Overview ¶
Package extract turns a coding-agent session transcript into candidate write-back notes. It is the input side of the flywheel: today write-back is opt-in (the Stop hook nudges once and ~most sessions still write nothing), so this lets Mesh pull the durable, reusable learnings out of a finished session automatically, for one-click review. BYOAI via the existing llm.Client (default `claude -p`, no API key). The model is the extractor; this package is the parse + prompt + validation around it.
Index ¶
Constants ¶
const DuplicateThreshold = 0.5
DuplicateThreshold is the title-similarity at or above which a candidate is treated as already-known (tuned so near-restatements match but distinct notes do not).
Variables ¶
This section is empty.
Functions ¶
func Judge ¶
func Judge(ctx context.Context, client llm.Client, c Candidate) (keep bool, reason string, err error)
Judge rates whether a candidate is worth keeping, for measuring extraction precision in the benchmark. Strict by design (it is the precision gate, not a rubber stamp).
func TitleSimilarity ¶
TitleSimilarity is the overlap coefficient of two titles' substantive tokens (0..1): intersection over the SMALLER token set. Overlap (not Jaccard) is the right measure for "does this candidate restate an existing note", because an existing note's title is often longer/compound (extra clauses) which would unfairly sink a Jaccard score.
Types ¶
type Candidate ¶
type Candidate struct {
Type string `json:"type"` // decision | gotcha | post-mortem
Title string `json:"title"` // specific, < ~12 words
Do string `json:"do"` // one line, imperative
Dont string `json:"dont"` // one line, the failure to avoid
Why string `json:"why"` // one line, the reason/evidence
Confidence string `json:"confidence"` // low | med | high (the model's self-rating)
}
Candidate is one extracted, not-yet-reviewed write-back note. Mirrors the fields mesh_append_note takes, so a promoted candidate becomes a note with no remapping.
type Cluster ¶
type Cluster struct {
Rep Candidate `json:"rep"` // the representative (first-seen) candidate
Sessions []string `json:"sessions"` // distinct sessions it appeared in
Count int `json:"count"` // distinct session count
Members []Candidate `json:"members"` // all candidates in the cluster
}
Cluster is a group of similar candidates that recur across sessions: a candidate learning that shows up again and again is a SYSTEMIC issue worth a permanent fix, not a one-off write-back.
func ClusterRecurring ¶
func ClusterRecurring(occs []Occurrence, threshold float64) []Cluster
ClusterRecurring greedily groups occurrences by title similarity (>= threshold) and returns the clusters sorted by distinct-session count, descending. A cluster spanning multiple sessions is a recurring problem.
type DigestStats ¶
type DigestStats struct {
Lines int `json:"lines"`
UserMsgs int `json:"user_msgs"`
AsstMsgs int `json:"asst_msgs"`
ToolCalls int `json:"tool_calls"`
HadWriteback bool `json:"had_writeback"` // the agent already called mesh_append_note/write_entity (the current algo)
DigestChars int `json:"digest_chars"`
}
DigestStats describes what a transcript contained, for the benchmark baseline.
func Digest ¶
func Digest(path string, maxChars int) (string, DigestStats, error)
Digest streams a transcript .jsonl into a compact, signal-dense summary for the extraction prompt: user requests + assistant narration + tool-call names (NOT their large outputs). Bounded to maxChars by keeping the first user request (the task) and the tail (where conclusions land). Also reports whether the agent already wrote back.
type Occurrence ¶
Occurrence is one extracted candidate tagged with the session it came from, the input to recurring-problem detection across many sessions.
Source Files