@kolisachint/hoocode-agent 0.5.15 → 0.5.17

package/CHANGELOG.md

@@ -1,5 +1,58 @@
 # Changelog
 
+## [0.5.17] - 2026-08-14
+
+### Fixed
+
+- The `Working...` line no longer strobes, the hardware cursor no longer
+  flickers at the end of it, and a long working message is truncated instead of
+  wrapping onto a second line. See the tui changelog for details — the same
+  fixes cover the compaction, retry, branch-summary and bash-command loaders.
+
+## [0.5.16] - 2026-08-13
+
+### Fixed
+
+- `/learn` no longer reports "no recent sessions in this directory" while a full
+  history sits in `~/.hoocode/sessions/`. It searched only the directory the
+  live session manager pointed at, which is the right one in the ordinary case
+  and wrong in three: an in-memory session (`--no-session`) reports no directory
+  at all and was taken literally, `--session <path>` points at wherever that
+  file lives, and a custom `sessionDir` setting points at one directory shared
+  by every project. The per-cwd session directory is now always searched
+  alongside it, and a session reachable from both is counted once.
+- `/learn` no longer discards sessions whose recorded working directory is the
+  same directory spelled differently — reached through a symlink, in different
+  case on the case-insensitive filesystems macOS and Windows ship by default, or
+  with a trailing separator. Any of these silently dropped the entire history the
+  command exists to read.
+- `/learn`'s per-directory memory is keyed on the working directory rather than
+  on the live session's directory, so a session with no directory of its own no
+  longer lands every project on one shared state file, and projects configured
+  to share a `sessionDir` no longer share one bookmark.
+- Warnings and errors in the TUI are rendered as one filled block instead of two
+  different shapes. A warning previously had neither the blank line nor the left
+  padding an error had, so it collided with whatever preceded it and hung off the
+  margin, and neither had a background to separate it from ordinary output.
+
+### Added
+
+- `/learn` explains an empty result instead of asserting there is no history. It
+  names the directories it searched, how many transcripts they hold, and why each
+  was passed over — outside the age window, beyond the session cap, or recorded
+  under a different working directory. Those have different fixes, and the single
+  sentence it printed before could not tell them apart.
+- `/learn settings` — the five thresholds with the values in force, the user and
+  project `settings.json` paths to set them in, the session directory being read,
+  and the state file. The settings existed but had no discoverable home; every
+  message that now reports a disappointing result points at them.
+- A **Learning** category in the `/settings` pane holding the same five
+  thresholds. Top level rather than folded into Advanced, since these are what
+  decide whether `/learn` finds anything and burying them is what made them
+  undiscoverable. A value set by hand in `settings.json` that is not one of the
+  presets stays in the cycle, so a keypress cannot silently snap the window
+  down to the first preset.
+
 ## [0.5.15] - 2026-08-13
 
 ### Fixed

package/dist/core/learn/extract.d.ts

@@ -84,9 +84,35 @@ export interface WorkflowCandidate extends Proposable {
     count: number;
     sessions: number;
 }
+/**
+ * Why a session file on disk did not make it into the digest.
+ *
+ * "No recent sessions" is the one outcome a user cannot act on without this:
+ * an empty session directory, a directory full of month-old sessions, and a
+ * directory full of sessions belonging to another checkout all produce the same
+ * sentence, and the fix differs in each case.
+ */
+export interface SessionScanReport {
+    /** Directories actually searched, in order. */
+    dirs: string[];
+    /** Directories that do not exist on disk. */
+    missingDirs: string[];
+    /** `.jsonl` files found across all searched directories. */
+    files: number;
+    /** Skipped for being older than the age window. */
+    tooOld: number;
+    /** Skipped because the session header records a different working directory. */
+    otherCwd: number;
+    /** Skipped for being beyond `maxSessions`. */
+    overLimit: number;
+    /** Skipped for being unreadable, unparseable, or empty. */
+    unreadable: number;
+}
 export interface LearnDigest {
     scannedSessions: number;
     skippedSessions: number;
+    /** Where the sessions came from, and what was passed over. */
+    scan: SessionScanReport;
     oldestSession?: string;
     newestSession?: string;
     agentsFilePath?: string;
@@ -106,7 +132,11 @@ export interface LearnDigest {
 export interface ExtractOptions {
     cwd: string;
     agentDir: string;
-    /** Override the directory scanned. Defaults to the per-cwd session dir. */
+    /**
+     * An extra directory to scan, normally the live session manager's. The
+     * per-cwd default directory is always scanned as well, so a session manager
+     * pointing somewhere unusual cannot hide this directory's history.
+     */
     sessionDir?: string;
     maxSessions?: number;
     maxAgeDays?: number;
@@ -134,6 +164,23 @@ export interface ExtractOptions {
     /** Injectable clock, for tests. */
     now?: Date;
 }
+/**
+ * Every directory this cwd's sessions could be sitting in.
+ *
+ * The caller passes the live session manager's directory, which is the right
+ * answer almost always — but not quite always, and each exception silently
+ * emptied the digest. An in-memory session (`--no-session`) reports `""`; an
+ * explicit `--session <path>` reports wherever that file lives; a custom
+ * `sessionDir` setting points at one shared directory. In every one of those
+ * cases the per-cwd default directory still holds the history worth mining, so
+ * search both and let the header check sort out what belongs to this cwd.
+ */
+export declare function candidateSessionDirs(options: Pick<ExtractOptions, "cwd" | "agentDir" | "sessionDir">): string[];
+/**
+ * Where this cwd's sessions were found and what was passed over, without
+ * ranking anything. `/learn stats` reports on the window without re-mining it.
+ */
+export declare function scanSessions(options: ExtractOptions): SessionScanReport;
 /**
  * Everything a proposal could already have been written into.
  *

package/dist/core/learn/extract.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"extract.d.ts","sourceRoot":"","sources":["../../../src/core/learn/extract.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAqBH,OAAO,EAAS,KAAK,UAAU,EAAE,MAAM,YAAY,CAAC;AAEpD;;;;GAIG;AACH,eAAO,MAAM,mBAAmB,mBAAmB,CAAC;AAoCpD;;;;;;;;;;;;;GAaG;AACH,MAAM,MAAM,eAAe,GAAG,KAAK,GAAG,UAAU,GAAG,WAAW,CAAC;AAE/D,oFAAoF;AACpF,UAAU,UAAU;IACnB,qEAAmE;IACnE,GAAG,EAAE,MAAM,CAAC;IACZ,4CAA4C;IAC5C,QAAQ,EAAE,MAAM,CAAC;CACjB;AAED,MAAM,WAAW,gBAAiB,SAAQ,UAAU;IACnD,gEAAgE;IAChE,IAAI,EAAE,MAAM,CAAC;IACb,UAAU,EAAE,MAAM,CAAC;IACnB,wBAAwB;IACxB,KAAK,EAAE,MAAM,CAAC;IACd,2EAAyE;IACzE,QAAQ,EAAE,MAAM,CAAC;IACjB,MAAM,EAAE,eAAe,CAAC;IACxB,iEAAiE;IACjE,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,sEAAsE;IACtE,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB;;;;OAIG;IACH,kBAAkB,EAAE,OAAO,CAAC;CAC5B;AAED,MAAM,WAAW,YAAa,SAAQ,UAAU;IAC/C,kCAAkC;IAClC,OAAO,EAAE,MAAM,CAAC;IAChB,kDAAkD;IAClD,SAAS,EAAE,MAAM,CAAC;IAClB,gEAAgE;IAChE,YAAY,EAAE,MAAM,CAAC;IACrB,qDAAqD;IACrD,mBAAmB,EAAE,MAAM,EAAE,CAAC;IAC9B,qDAAqD;IACrD,WAAW,EAAE,MAAM,EAAE,CAAC;IACtB,sEAAsE;IACtE,KAAK,EAAE,MAAM,CAAC;IACd,QAAQ,EAAE,MAAM,CAAC;CACjB;AAED,MAAM,WAAW,iBAAkB,SAAQ,UAAU;IACpD,qCAAqC;IACrC,KAAK,EAAE,MAAM,EAAE,CAAC;IAChB,KAAK,EAAE,MAAM,CAAC;IACd,QAAQ,EAAE,MAAM,CAAC;CACjB;AAED,MAAM,WAAW,WAAW;IAC3B,eAAe,EAAE,MAAM,CAAC;IACxB,eAAe,EAAE,MAAM,CAAC;IACxB,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,gBAAgB,CAAC,EAAE,MAAM,CAAC;IAC1B,UAAU,EAAE,gBAAgB,EAAE,CAAC;IAC/B,KAAK,EAAE,YAAY,EAAE,CAAC;IACtB,SAAS,EAAE,iBAAiB,EAAE,CAAC;IAC/B,mFAAmF;IACnF,UAAU,EAAE,MAAM,CAAC;IACnB,oEAAoE;IACpE,QAAQ,EAAE,KAAK,CAAC;QAAE,GAAG,EAAE,MAAM,CAAC;QAAC,QAAQ,EAAE,MAAM,CAAC;QAAC,OAAO,EAAE,OAAO,CAAA;KAAE,CAAC,CAAC;CACrE;AAED,MAAM,WAAW,cAAc;IAC9B,GAAG,EAAE,MAAM,CAAC;IACZ,QAAQ,EAAE,MAAM,CAAC;IACjB,2EAA2E;IAC3E,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,kFAAkF;IAClF,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,sFAAsF;IACtF,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,sCAAsC;IACtC,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB;;;OAGG;IACH,KAAK,CAAC,EAAE,UAAU,CAAC;IACnB,kFAAkF;IAClF,WAAW,CAAC,EAAE,OAAO,CAAC;IACtB;;;OAGG;IACH,MAAM,CAAC,EAAE,KAAK,CAAC;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,WAAW,EAAE,MAAM,CAAA;KAAE,CAAC,CAAC;IACtD,mCAAmC;IACnC,GAAG,CAAC,EAAE,IAAI,CAAC;CACX;AA2jBD;;;;;;;GAOG;AACH,MAAM,WAAW,aAAa;IAC7B,4EAA4E;IAC5E,SAAS,EAAE,MAAM,EAAE,CAAC;IACpB,MAAM,EAAE,KAAK,CAAC;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,WAAW,EAAE,MAAM,CAAA;KAAE,CAAC,CAAC;CACrD;AAED,MAAM,WAAW,aAAa;IAC7B,sDAAsD;IACtD,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,yEAAyE;IACzE,KAAK,CAAC,EAAE,MAAM,CAAC;CACf;AAED;;;;;GAKG;AACH,wBAAgB,aAAa,CAAC,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,aAAa,GAAG,aAAa,CA2B/E;AAED,mDAAmD;AACnD,wBAAgB,kBAAkB,CAAC,OAAO,EAAE;IAC3C,GAAG,EAAE,MAAM,CAAC;IACZ,QAAQ,EAAE,MAAM,CAAC;IACjB,MAAM,CAAC,EAAE,KAAK,CAAC;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,WAAW,EAAE,MAAM,CAAA;KAAE,CAAC,CAAC;CACtD,GAAG,aAAa,CAShB;AA4CD,0EAA0E;AAC1E,wBAAgB,kBAAkB,CAAC,OAAO,EAAE,cAAc,GAAG,WAAW,CA+DvE","sourcesContent":["/**\n * Session mining for `/learn`.\n *\n * Reads session `.jsonl` files straight off disk rather than the live context.\n * That is the whole point: the on-disk transcript is complete even when the\n * in-context one has been compacted away, and it spans every past session\n * instead of only this one. Cross-session repetition is the signal that decides\n * whether something is a durable rule or a one-off, and it is the one thing a\n * prompt reading its own context cannot see.\n *\n * The split of labour is deliberate. This module is entirely deterministic: it\n * parses, filters, normalizes, counts and ranks. Judgement — is this a rule, how\n * should it be phrased, which scope owns it — belongs to the model reading the\n * digest, which is why the output carries evidence (counts, sessions, dates)\n * rather than conclusions.\n */\n\nimport { existsSync, readdirSync, readFileSync, statSync } from \"node:fs\";\nimport { dirname, join, resolve } from \"node:path\";\nimport type { AgentMessage } from \"@kolisachint/hoocode-agent-core\";\nimport type { TextContent, ToolCall } from \"@kolisachint/hoocode-ai\";\nimport { getUserAgentsDir } from \"../../config.js\";\nimport { getDefaultSessionDir } from \"../session-manager.js\";\nimport { loadSkills } from \"../skills.js\";\nimport {\n\tcommandHead,\n\tcontentWords,\n\textractErrorRegion,\n\tisBenignFailure,\n\tisRuleShapedDirective,\n\tisUninformativeFailure,\n\tnormalizeCommand,\n\tnormalizeDirective,\n\tnormalizeErrorSignature,\n\twordOverlap,\n} from \"./normalize.js\";\nimport { judge, type LearnState } from \"./state.js\";\n\n/**\n * Prefix on the message `/learn` injects. The digest is persisted like any user\n * turn, so without this marker the next `/learn` would mine its own output and\n * every proposal would compound its own count.\n */\nexport const LEARN_DIGEST_MARKER = \"[learn-digest]\";\n\n/** Sessions considered, newest first. */\nconst DEFAULT_MAX_SESSIONS = 20;\n/** Sessions older than this are ignored — a pattern that stopped is not a rule. */\nconst DEFAULT_MAX_AGE_DAYS = 30;\n/** Entries parsed per session file, as a guard against pathological transcripts. */\nconst MAX_ENTRIES_PER_SESSION = 8000;\n/** Tool calls per session fed to the workflow detector. */\nconst MAX_TOOL_CALLS_PER_SESSION = 400;\n/** How far forward the fix extractor looks for the same command succeeding. */\nconst FIX_LOOKAHEAD = 40;\n/** Word overlap against an existing rule above which a directive counts as covered. */\nconst COVERED_OVERLAP = 0.6;\n/**\n * The same bar for skills, set higher on purpose.\n *\n * A rule is one line, so overlap against it is a sharp signal. A skill is a name\n * plus a description written to attract matches, which is a far larger haystack\n * — a short directive's words turn up in it by chance much more readily. The\n * higher bar and the truncation below keep \"you already have a skill for this\"\n * from being said on a coincidence.\n */\nconst SKILL_COVERED_OVERLAP = 0.75;\n/** Description characters considered. The opening says what a skill does; the rest is trigger bait. */\nconst SKILL_DESCRIPTION_CHARS = 300;\n/** Directives must reach this many occurrences to be reported at all. */\nconst DEFAULT_MIN_DIRECTIVE_COUNT = 2;\n/** Tool sequence lengths considered as workflow candidates. */\nconst WORKFLOW_MIN_LEN = 3;\nconst WORKFLOW_MAX_LEN = 5;\n/** Repeats before a tool sequence is worth proposing as a skill. */\nconst DEFAULT_MIN_WORKFLOW_COUNT = 3;\n/** Cap on each list in the digest, so the model's budget goes to the top signals. */\nconst DEFAULT_MAX_PER_CATEGORY = 8;\n\n/**\n * Where a repeated directive already lives, if anywhere.\n *\n * A directive covered by a rule and said only once is simply dropped — the rule\n * exists and is working. What survives is one of three cases, and they want\n * different responses:\n *\n * - `new` — not written down anywhere. Propose it.\n * - `restated` — a context-file rule covers it and you said it anyway, so the\n *   rule is not working. Rewrite it; do not add a second one.\n * - `has-skill` — a *skill* covers it and you asked by hand anyway, which\n *   usually means the skill's `description` is not triggering. Sharpen the\n *   description rather than writing a rule that duplicates the skill.\n */\nexport type DirectiveStatus = \"new\" | \"restated\" | \"has-skill\";\n\n/** Fields every proposable item shares, so suppression can be applied uniformly. */\ninterface Proposable {\n\t/** Stable identity across runs — what the state file remembers. */\n\tkey: string;\n\t/** Newest occurrence in the window, ISO. */\n\tlastSeen: string;\n}\n\nexport interface DirectiveCluster extends Proposable {\n\t/** Representative raw text, the longest seen in the cluster. */\n\ttext: string;\n\tnormalized: string;\n\t/** Total times said. */\n\tcount: number;\n\t/** Distinct sessions it was said in — the stronger of the two counts. */\n\tsessions: number;\n\tstatus: DirectiveStatus;\n\t/** The existing rule line matched, when status is `restated`. */\n\texistingRule?: string;\n\t/** The skill that already covers this, when status is `has-skill`. */\n\texistingSkill?: string;\n\t/**\n\t * Shown before and still not written down anywhere — neither as a rule nor as\n\t * a skill — so you saw this proposal and passed on it. Only meaningful for\n\t * directives, which are the only items with a real coverage signal.\n\t */\n\tpreviouslyDeclined: boolean;\n}\n\nexport interface FixCandidate extends Proposable {\n\t/** Normalized failing command. */\n\tcommand: string;\n\t/** Normalized error signature, the dedupe key. */\n\tsignature: string;\n\t/** Short raw excerpt, so the model sees the real error text. */\n\terrorExcerpt: string;\n\t/** Commands run between the failure and the pass. */\n\tinterveningCommands: string[];\n\t/** Files edited between the failure and the pass. */\n\teditedFiles: string[];\n\t/** Times this signature failed and was resolved across the window. */\n\tcount: number;\n\tsessions: number;\n}\n\nexport interface WorkflowCandidate extends Proposable {\n\t/** Tool-call signatures in order. */\n\tsteps: string[];\n\tcount: number;\n\tsessions: number;\n}\n\nexport interface LearnDigest {\n\tscannedSessions: number;\n\tskippedSessions: number;\n\toldestSession?: string;\n\tnewestSession?: string;\n\tagentsFilePath?: string;\n\tagentsFileTokens?: number;\n\tdirectives: DirectiveCluster[];\n\tfixes: FixCandidate[];\n\tworkflows: WorkflowCandidate[];\n\t/** Items held back because nothing new has happened since they were last shown. */\n\tsuppressed: number;\n\t/** Everything this run put on screen, for the caller to persist. */\n\tsurfaced: Array<{ key: string; lastSeen: string; covered: boolean }>;\n}\n\nexport interface ExtractOptions {\n\tcwd: string;\n\tagentDir: string;\n\t/** Override the directory scanned. Defaults to the per-cwd session dir. */\n\tsessionDir?: string;\n\tmaxSessions?: number;\n\tmaxAgeDays?: number;\n\t/** Occurrences a directive needs before it is proposed. The signal/noise dial. */\n\tminRepeats?: number;\n\t/** Non-overlapping repeats a tool sequence needs before it is proposed as a skill. */\n\tminWorkflowRepeats?: number;\n\t/** Cap on each list in the digest. */\n\tmaxProposals?: number;\n\t/**\n\t * What previous runs already showed. Items with no new occurrences since are\n\t * held back. Omit (or pass `ignoreState`) to propose everything in the window.\n\t */\n\tstate?: LearnState;\n\t/** Re-propose everything, ignoring what previous runs surfaced (`/learn all`). */\n\tignoreState?: boolean;\n\t/**\n\t * Skills a directive can already be covered by. Defaults to the ones loaded\n\t * from disk; injectable so tests do not read the developer's real skills.\n\t */\n\tskills?: Array<{ name: string; description: string }>;\n\t/** Injectable clock, for tests. */\n\tnow?: Date;\n}\n\ninterface SessionHeaderLike {\n\ttype: \"session\";\n\tid?: string;\n\ttimestamp?: string;\n\tcwd?: string;\n}\n\ninterface EntryLike {\n\ttype: string;\n\tid?: string;\n\tparentId?: string | null;\n\ttimestamp?: string;\n\tmessage?: AgentMessage;\n}\n\n/** One session, reduced to the branch that was actually taken. */\ninterface ParsedSession {\n\tfile: string;\n\tid: string;\n\ttimestamp: string;\n\tentries: EntryLike[];\n}\n\nfunction textOf(content: unknown): string {\n\tif (typeof content === \"string\") return content;\n\tif (!Array.isArray(content)) return \"\";\n\treturn content\n\t\t.map((block) =>\n\t\t\tblock && typeof block === \"object\" && (block as TextContent).type === \"text\"\n\t\t\t\t? ((block as TextContent).text ?? \"\")\n\t\t\t\t: \"\",\n\t\t)\n\t\t.join(\"\\n\")\n\t\t.trim();\n}\n\nfunction isToolCall(block: unknown): block is ToolCall {\n\treturn !!block && typeof block === \"object\" && (block as ToolCall).type === \"toolCall\";\n}\n\n/**\n * Reduce a session's raw entries to the branch that was actually taken.\n *\n * Session files are trees — forks and clones append entries that were never\n * part of the same conversation. Walking parent links back from the last entry\n * keeps the extractor from stitching a \"fix\" out of two turns that never\n * happened in sequence. Sessions written before entry ids existed are flat, and\n * for those file order *is* the branch.\n */\nfunction activeBranch(entries: EntryLike[]): EntryLike[] {\n\tconst withIds = entries.filter((e) => typeof e.id === \"string\");\n\tif (withIds.length === 0) return entries;\n\n\tconst byId = new Map<string, EntryLike>();\n\tfor (const entry of withIds) byId.set(entry.id as string, entry);\n\n\tconst branch: EntryLike[] = [];\n\tconst seen = new Set<string>();\n\tlet cursor: EntryLike | undefined = withIds[withIds.length - 1];\n\twhile (cursor?.id && !seen.has(cursor.id)) {\n\t\tseen.add(cursor.id);\n\t\tbranch.push(cursor);\n\t\tcursor = cursor.parentId ? byId.get(cursor.parentId) : undefined;\n\t}\n\treturn branch.reverse();\n}\n\nfunction parseSessionFile(file: string, cwd: string): ParsedSession | undefined {\n\tlet raw: string;\n\ttry {\n\t\traw = readFileSync(file, \"utf-8\");\n\t} catch {\n\t\treturn undefined;\n\t}\n\n\tconst lines = raw.split(\"\\n\");\n\tlet header: SessionHeaderLike | undefined;\n\tconst entries: EntryLike[] = [];\n\tfor (const line of lines) {\n\t\tif (!line.trim()) continue;\n\t\tif (entries.length >= MAX_ENTRIES_PER_SESSION) break;\n\t\tlet parsed: EntryLike | SessionHeaderLike;\n\t\ttry {\n\t\t\tparsed = JSON.parse(line);\n\t\t} catch {\n\t\t\t// A partially-flushed final line is normal for a live session.\n\t\t\tcontinue;\n\t\t}\n\t\tif (parsed.type === \"session\") {\n\t\t\theader ??= parsed as SessionHeaderLike;\n\t\t\tcontinue;\n\t\t}\n\t\tentries.push(parsed as EntryLike);\n\t}\n\n\t// An explicit `--session` path can put a session for another directory in\n\t// this directory, so trust the header over the file's location.\n\tif (header?.cwd && resolve(header.cwd) !== resolve(cwd)) return undefined;\n\tif (entries.length === 0) return undefined;\n\n\treturn {\n\t\tfile,\n\t\tid: header?.id ?? file,\n\t\ttimestamp: header?.timestamp ?? statSync(file).mtime.toISOString(),\n\t\tentries: activeBranch(entries),\n\t};\n}\n\nfunction listSessions(options: ExtractOptions): { sessions: ParsedSession[]; skipped: number } {\n\tconst dir = options.sessionDir ?? getDefaultSessionDir(options.cwd, options.agentDir);\n\tif (!existsSync(dir)) return { sessions: [], skipped: 0 };\n\n\tconst maxSessions = options.maxSessions ?? DEFAULT_MAX_SESSIONS;\n\tconst maxAgeDays = options.maxAgeDays ?? DEFAULT_MAX_AGE_DAYS;\n\tconst now = options.now ?? new Date();\n\tconst cutoff = now.getTime() - maxAgeDays * 24 * 60 * 60 * 1000;\n\n\tlet files: string[];\n\ttry {\n\t\tfiles = readdirSync(dir)\n\t\t\t.filter((f) => f.endsWith(\".jsonl\"))\n\t\t\t.map((f) => join(dir, f));\n\t} catch {\n\t\treturn { sessions: [], skipped: 0 };\n\t}\n\n\tconst dated = files\n\t\t.map((file) => {\n\t\t\ttry {\n\t\t\t\treturn { file, mtime: statSync(file).mtime.getTime() };\n\t\t\t} catch {\n\t\t\t\treturn undefined;\n\t\t\t}\n\t\t})\n\t\t.filter((f): f is { file: string; mtime: number } => !!f)\n\t\t.sort((a, b) => b.mtime - a.mtime);\n\n\tconst sessions: ParsedSession[] = [];\n\tlet skipped = 0;\n\tfor (const { file, mtime } of dated) {\n\t\tif (sessions.length >= maxSessions) {\n\t\t\tskipped++;\n\t\t\tcontinue;\n\t\t}\n\t\tif (mtime < cutoff) {\n\t\t\tskipped++;\n\t\t\tcontinue;\n\t\t}\n\t\tconst parsed = parseSessionFile(file, options.cwd);\n\t\tif (parsed) sessions.push(parsed);\n\t\telse skipped++;\n\t}\n\treturn { sessions, skipped };\n}\n\n/**\n * Hold back items already shown that have not recurred since, then cap the rest.\n *\n * Order matters: suppression runs *before* the cap, or an item you already\n * decided on would occupy one of the few slots the digest has and push a live\n * signal off the list.\n */\nfunction applySuppression<T extends Proposable>(\n\titems: T[],\n\tstate: LearnState | undefined,\n\tmaxProposals: number,\n\tcovered: (item: T) => boolean,\n\tonDeclined?: (item: T) => void,\n): { kept: T[]; suppressed: number } {\n\tif (!state) return { kept: items.slice(0, maxProposals), suppressed: 0 };\n\n\tconst kept: T[] = [];\n\tlet suppressed = 0;\n\tfor (const item of items) {\n\t\tconst verdict = judge(state, { key: item.key, lastSeen: item.lastSeen, covered: covered(item) });\n\t\tif (verdict.suppressed) {\n\t\t\tsuppressed++;\n\t\t\tcontinue;\n\t\t}\n\t\tif (verdict.previouslyDeclined) onDeclined?.(item);\n\t\tkept.push(item);\n\t}\n\treturn { kept: kept.slice(0, maxProposals), suppressed };\n}\n\n/** Nearest AGENTS.md walking up from cwd, so proposals can be checked against it. */\nfunction findAgentsFile(cwd: string): string | undefined {\n\tlet dir = resolve(cwd);\n\twhile (true) {\n\t\tfor (const name of [\"AGENTS.md\", \"AGENTS.MD\", \"CLAUDE.md\", \"CLAUDE.MD\"]) {\n\t\t\tconst candidate = join(dir, name);\n\t\t\tif (existsSync(candidate)) return candidate;\n\t\t}\n\t\tconst parent = dirname(dir);\n\t\tif (parent === dir) return undefined;\n\t\tdir = parent;\n\t}\n}\n\ninterface ToolEvent {\n\tname: string;\n\targs: Record<string, any>;\n\t/** Set once the matching result is seen. */\n\tisError?: boolean;\n\toutput?: string;\n}\n\n/** Pair tool calls with their results along one branch, in call order. */\nfunction toolEvents(entries: EntryLike[]): ToolEvent[] {\n\tconst byCallId = new Map<string, ToolEvent>();\n\tconst ordered: ToolEvent[] = [];\n\n\tfor (const entry of entries) {\n\t\tconst message = entry.type === \"message\" ? entry.message : undefined;\n\t\tif (!message) continue;\n\t\tif (message.role === \"assistant\") {\n\t\t\tfor (const block of (message.content ?? []) as unknown[]) {\n\t\t\t\tif (!isToolCall(block)) continue;\n\t\t\t\tconst event: ToolEvent = { name: block.name, args: block.arguments ?? {} };\n\t\t\t\tbyCallId.set(block.id, event);\n\t\t\t\tordered.push(event);\n\t\t\t}\n\t\t} else if (message.role === \"toolResult\") {\n\t\t\tconst event = byCallId.get(message.toolCallId);\n\t\t\tif (!event) continue;\n\t\t\tevent.isError = message.isError;\n\t\t\tevent.output = textOf(message.content);\n\t\t}\n\t}\n\treturn ordered;\n}\n\n/** User turns worth mining, in order, with the digest's own output excluded. */\nfunction userDirectives(entries: EntryLike[]): string[] {\n\tconst out: string[] = [];\n\tfor (const entry of entries) {\n\t\tconst message = entry.type === \"message\" ? entry.message : undefined;\n\t\tif (!message || message.role !== \"user\") continue;\n\t\tconst text = textOf(message.content);\n\t\tif (!text || text.startsWith(LEARN_DIGEST_MARKER)) continue;\n\t\tif (!isRuleShapedDirective(text)) continue;\n\t\tout.push(text.trim());\n\t}\n\treturn out;\n}\n\nfunction clusterDirectives(\n\tperSession: Array<{ session: ParsedSession; directives: string[] }>,\n\tcoverage: CoverageIndex,\n\tminRepeats: number,\n): DirectiveCluster[] {\n\tinterface Acc {\n\t\ttext: string;\n\t\tnormalized: string;\n\t\tcount: number;\n\t\tsessions: Set<string>;\n\t\tlastSeen: string;\n\t}\n\tconst acc = new Map<string, Acc>();\n\n\tfor (const { session, directives } of perSession) {\n\t\tfor (const text of directives) {\n\t\t\tconst normalized = normalizeDirective(text);\n\t\t\tif (!normalized) continue;\n\t\t\tconst existing = acc.get(normalized);\n\t\t\tif (existing) {\n\t\t\t\texisting.count++;\n\t\t\t\texisting.sessions.add(session.id);\n\t\t\t\tif (session.timestamp > existing.lastSeen) existing.lastSeen = session.timestamp;\n\t\t\t\tif (text.length > existing.text.length) existing.text = text;\n\t\t\t} else {\n\t\t\t\tacc.set(normalized, {\n\t\t\t\t\ttext,\n\t\t\t\t\tnormalized,\n\t\t\t\t\tcount: 1,\n\t\t\t\t\tsessions: new Set([session.id]),\n\t\t\t\t\tlastSeen: session.timestamp,\n\t\t\t\t});\n\t\t\t}\n\t\t}\n\t}\n\n\tconst clusters: DirectiveCluster[] = [];\n\tfor (const entry of acc.values()) {\n\t\tif (entry.count < minRepeats) continue;\n\n\t\t// Everything reaching here cleared the repeat threshold. Suppression handles\n\t\t// the case that used to make these labels lie — a proposal accepted from a\n\t\t// previous run coming back as \"not working\" when nothing had happened\n\t\t// since. By the time an item survives that filter, a match genuinely means\n\t\t// you repeated yourself after the rule or skill already existed.\n\t\tconst match = matchCoverage(entry.text, coverage);\n\t\tclusters.push({\n\t\t\tkey: `directive:${entry.normalized}`,\n\t\t\ttext: entry.text,\n\t\t\tnormalized: entry.normalized,\n\t\t\tcount: entry.count,\n\t\t\tsessions: entry.sessions.size,\n\t\t\tlastSeen: entry.lastSeen,\n\t\t\tstatus: match.rule ? \"restated\" : match.skill ? \"has-skill\" : \"new\",\n\t\t\texistingRule: match.rule,\n\t\t\texistingSkill: match.skill,\n\t\t\tpreviouslyDeclined: false,\n\t\t});\n\t}\n\n\treturn clusters.sort((a, b) => b.sessions - a.sessions || b.count - a.count || a.text.localeCompare(b.text));\n}\n\n/** Files a mutating tool touched, for the resolution summary. */\nfunction editedFile(event: ToolEvent): string | undefined {\n\tif (![\"edit\", \"write\", \"multi_edit\", \"apply_patch\"].includes(event.name)) return undefined;\n\tconst path = event.args?.path ?? event.args?.file_path ?? event.args?.filePath;\n\treturn typeof path === \"string\" ? path : undefined;\n}\n\nfunction extractFixes(perSession: Array<{ session: ParsedSession; events: ToolEvent[] }>): FixCandidate[] {\n\tinterface Acc {\n\t\tcandidate: FixCandidate;\n\t\tsessions: Set<string>;\n\t}\n\tconst acc = new Map<string, Acc>();\n\n\tfor (const { session, events } of perSession) {\n\t\tfor (let i = 0; i < events.length; i++) {\n\t\t\tconst failure = events[i]!;\n\t\t\tif (failure.name !== \"bash\" || !failure.isError) continue;\n\t\t\tconst command = typeof failure.args?.command === \"string\" ? failure.args.command : \"\";\n\t\t\tif (!command || isBenignFailure(command)) continue;\n\n\t\t\tconst normalized = normalizeCommand(command);\n\t\t\tconst interveningCommands: string[] = [];\n\t\t\tconst editedFiles: string[] = [];\n\t\t\tlet resolved = false;\n\n\t\t\tfor (let j = i + 1; j < Math.min(events.length, i + 1 + FIX_LOOKAHEAD); j++) {\n\t\t\t\tconst next = events[j]!;\n\t\t\t\tconst file = editedFile(next);\n\t\t\t\tif (file) editedFiles.push(file);\n\n\t\t\t\tif (next.name !== \"bash\") continue;\n\t\t\t\tconst nextCommand = typeof next.args?.command === \"string\" ? next.args.command : \"\";\n\t\t\t\tif (!nextCommand) continue;\n\n\t\t\t\t// The same command later succeeding is the only evidence that the\n\t\t\t\t// problem was actually fixed. A *different* command passing says\n\t\t\t\t// nothing, and neither does the model moving on.\n\t\t\t\tif (normalizeCommand(nextCommand) === normalized && !next.isError) {\n\t\t\t\t\tresolved = true;\n\t\t\t\t\tbreak;\n\t\t\t\t}\n\t\t\t\tinterveningCommands.push(nextCommand.trim());\n\t\t\t}\n\n\t\t\tif (!resolved) continue;\n\n\t\t\tconst output = failure.output ?? \"\";\n\t\t\t// An abort is the user changing their mind, not a problem that was\n\t\t\t// solved, and empty output carries nothing to sign or show.\n\t\t\tif (isUninformativeFailure(output)) continue;\n\n\t\t\t// Sign the error region, not the whole output: build tools lead with an\n\t\t\t// identical banner, so signing everything makes unrelated failures of\n\t\t\t// the same command collide on their shared preamble.\n\t\t\tconst errorRegion = extractErrorRegion(output);\n\t\t\tconst signature = normalizeErrorSignature(errorRegion);\n\t\t\tif (!signature) continue;\n\n\t\t\tconst key = `${normalized}\u0000${signature}`;\n\t\t\tconst existing = acc.get(key);\n\t\t\tif (existing) {\n\t\t\t\texisting.candidate.count++;\n\t\t\t\texisting.sessions.add(session.id);\n\t\t\t\tif (session.timestamp > existing.candidate.lastSeen) existing.candidate.lastSeen = session.timestamp;\n\t\t\t} else {\n\t\t\t\tacc.set(key, {\n\t\t\t\t\tsessions: new Set([session.id]),\n\t\t\t\t\tcandidate: {\n\t\t\t\t\t\tkey: `fix:${key}`,\n\t\t\t\t\t\tcommand: normalized,\n\t\t\t\t\t\tsignature,\n\t\t\t\t\t\terrorExcerpt: errorRegion.replace(/\\s+/g, \" \").trim().slice(0, 240),\n\t\t\t\t\t\tinterveningCommands: [...new Set(interveningCommands)].slice(0, 5),\n\t\t\t\t\t\teditedFiles: [...new Set(editedFiles)].slice(0, 5),\n\t\t\t\t\t\tcount: 1,\n\t\t\t\t\t\tsessions: 1,\n\t\t\t\t\t\tlastSeen: session.timestamp,\n\t\t\t\t\t},\n\t\t\t\t});\n\t\t\t}\n\t\t}\n\t}\n\n\tconst out: FixCandidate[] = [];\n\tfor (const { candidate, sessions } of acc.values()) {\n\t\tcandidate.sessions = sessions.size;\n\t\tout.push(candidate);\n\t}\n\treturn out.sort((a, b) => b.count - a.count || b.sessions - a.sessions || a.signature.localeCompare(b.signature));\n}\n\n/** A tool call reduced to a comparable step: the tool, plus what a bash call runs. */\nfunction stepSignature(event: ToolEvent): string {\n\tif (event.name === \"bash\") {\n\t\tconst command = typeof event.args?.command === \"string\" ? event.args.command : \"\";\n\t\tconst head = commandHead(command);\n\t\treturn head ? `bash:${head}` : \"bash\";\n\t}\n\treturn event.name;\n}\n\n/**\n * Commands that are how an agent looks around rather than what the user was\n * doing. A sequence built only from these plus file edits describes \"coding\",\n * not a workflow, and no useful skill has ever come out of one.\n */\nconst PLUMBING_COMMANDS = new Set([\n\t\"cd\",\n\t\"ls\",\n\t\"pwd\",\n\t\"cat\",\n\t\"head\",\n\t\"tail\",\n\t\"wc\",\n\t\"echo\",\n\t\"which\",\n\t\"find\",\n\t\"fd\",\n\t\"grep\",\n\t\"rg\",\n\t\"sed\",\n\t\"awk\",\n\t\"git status\",\n\t\"git diff\",\n\t\"git log\",\n\t\"git show\",\n]);\n\n/**\n * Whether a sequence is a procedure rather than the rhythm of editing code.\n *\n * Two distinct doing-commands is the bar, and it was set by looking at real\n * transcripts. One command is not enough: the edit/test loop\n * (`edit → edit → bash:npm run`) satisfies it, and because a sliding window\n * over a long alternating run produces every rotation of that cycle, it alone\n * filled all eight slots with `edit → npm run → edit`, `npm run → edit → edit`\n * and so on — one habit described eight ways.\n *\n * A procedure worth a skill chains *different* actions: test then commit then\n * push, build then tag then publish. Requiring two distinct ones keeps those and\n * drops the rhythm. The cost is real — a genuine one-command routine with setup\n * is missed — and that is the intended trade, since a missed skill costs nothing\n * while a digest full of noise costs the reader's attention every run.\n */\nfunction isProcedure(steps: string[]): boolean {\n\tconst commands = new Set<string>();\n\tfor (const step of steps) {\n\t\tif (!step.startsWith(\"bash:\")) continue;\n\t\tconst head = step.slice(\"bash:\".length);\n\t\tif (PLUMBING_COMMANDS.has(head) || PLUMBING_COMMANDS.has(head.split(\" \")[0] ?? \"\")) continue;\n\t\tcommands.add(head);\n\t}\n\treturn commands.size >= 2;\n}\n\n/** True when `needle` appears as a contiguous run inside `haystack`. */\nfunction containsSequence(haystack: string[], needle: string[]): boolean {\n\tif (needle.length > haystack.length) return false;\n\tfor (let i = 0; i + needle.length <= haystack.length; i++) {\n\t\tif (needle.every((step, offset) => haystack[i + offset] === step)) return true;\n\t}\n\treturn false;\n}\n\nfunction extractWorkflows(\n\tperSession: Array<{ session: ParsedSession; events: ToolEvent[] }>,\n\tminRepeats: number,\n): WorkflowCandidate[] {\n\tinterface Acc {\n\t\tsteps: string[];\n\t\tcount: number;\n\t\tsessions: Set<string>;\n\t\tlastSeen: string;\n\t}\n\tconst acc = new Map<string, Acc>();\n\n\tfor (const { session, events } of perSession) {\n\t\tconst steps = events.slice(0, MAX_TOOL_CALLS_PER_SESSION).map(stepSignature);\n\n\t\tfor (let len = WORKFLOW_MIN_LEN; len <= WORKFLOW_MAX_LEN; len++) {\n\t\t\t// Collect every position first, then count greedily without overlap.\n\t\t\t// Counting each sliding position separately treats one long stretch of\n\t\t\t// edit/read churn as dozens of repeats: an `edit > read > edit` run of\n\t\t\t// length 12 scores 10 occurrences when it is really one stretch of work.\n\t\t\tconst positions = new Map<string, number[]>();\n\t\t\tfor (let i = 0; i + len <= steps.length; i++) {\n\t\t\t\tconst window = steps.slice(i, i + len);\n\t\t\t\t// A run of one repeated tool is a loop, not a workflow.\n\t\t\t\tif (new Set(window).size < 2) continue;\n\t\t\t\tif (!isProcedure(window)) continue;\n\t\t\t\tconst key = window.join(\" > \");\n\t\t\t\tconst list = positions.get(key);\n\t\t\t\tif (list) list.push(i);\n\t\t\t\telse positions.set(key, [i]);\n\t\t\t}\n\n\t\t\tfor (const [key, occurrences] of positions) {\n\t\t\t\tlet count = 0;\n\t\t\t\tlet nextFree = -1;\n\t\t\t\tfor (const start of occurrences) {\n\t\t\t\t\tif (start < nextFree) continue;\n\t\t\t\t\tcount++;\n\t\t\t\t\tnextFree = start + len;\n\t\t\t\t}\n\n\t\t\t\tconst existing = acc.get(key);\n\t\t\t\tif (existing) {\n\t\t\t\t\texisting.count += count;\n\t\t\t\t\texisting.sessions.add(session.id);\n\t\t\t\t\tif (session.timestamp > existing.lastSeen) existing.lastSeen = session.timestamp;\n\t\t\t\t} else {\n\t\t\t\t\tacc.set(key, {\n\t\t\t\t\t\tsteps: key.split(\" > \"),\n\t\t\t\t\t\tcount,\n\t\t\t\t\t\tsessions: new Set([session.id]),\n\t\t\t\t\t\tlastSeen: session.timestamp,\n\t\t\t\t\t});\n\t\t\t\t}\n\t\t\t}\n\t\t}\n\t}\n\n\tconst ranked = [...acc.values()]\n\t\t.filter((entry) => entry.count >= minRepeats)\n\t\t.map((entry) => ({\n\t\t\tkey: `workflow:${entry.steps.join(\" > \")}`,\n\t\t\tsteps: entry.steps,\n\t\t\tcount: entry.count,\n\t\t\tsessions: entry.sessions.size,\n\t\t\tlastSeen: entry.lastSeen,\n\t\t}))\n\t\t// Sessions first, matching directives: a sequence seen in three sessions is\n\t\t// a workflow, while one repeated ten times in a single session is usually\n\t\t// just the shape of that one task.\n\t\t.sort(\n\t\t\t(a, b) =>\n\t\t\t\tb.sessions - a.sessions ||\n\t\t\t\tb.count - a.count ||\n\t\t\t\tb.steps.length - a.steps.length ||\n\t\t\t\ta.steps.join().localeCompare(b.steps.join()),\n\t\t);\n\n\t// Every n-gram overlaps its own extensions and prefixes, so without this the\n\t// list is one workflow described five slightly different ways. The test runs\n\t// both directions on purpose: a shorter sequence always outranks the longer\n\t// one containing it (it occurs at least as often), so checking only\n\t// shorter-inside-kept would never fire. Keep the best-ranked member of each\n\t// family and drop the rest.\n\tconst distinct: typeof ranked = [];\n\tfor (const candidate of ranked) {\n\t\tconst overlapsKept = distinct.some(\n\t\t\t(kept) => containsSequence(kept.steps, candidate.steps) || containsSequence(candidate.steps, kept.steps),\n\t\t);\n\t\tif (overlapsKept) continue;\n\t\tdistinct.push(candidate);\n\t}\n\treturn distinct;\n}\n\n/**\n * Everything a proposal could already have been written into.\n *\n * Built once and shared, because the same question — is this already written\n * down? — is asked while ranking a run *and* afterwards by `/learn stats`,\n * which reconstructs adoption by comparing coverage now against coverage when\n * the item was shown.\n */\nexport interface CoverageIndex {\n\t/** Candidate rule lines from the repo context file and both user scopes. */\n\truleLines: string[];\n\tskills: Array<{ name: string; description: string }>;\n}\n\nexport interface CoverageMatch {\n\t/** The context-file line that covers this, if any. */\n\trule?: string;\n\t/** The skill that covers this, if any. Only set when no rule matched. */\n\tskill?: string;\n}\n\n/**\n * Where a piece of text is already written down, if anywhere.\n *\n * A rule wins over a skill when both match: it is the more specific answer, and\n * \"rewrite this line\" is more actionable than \"sharpen a description\".\n */\nexport function matchCoverage(text: string, index: CoverageIndex): CoverageMatch {\n\tconst words = contentWords(text);\n\n\tlet bestLine: string | undefined;\n\tlet bestOverlap = 0;\n\tfor (const line of index.ruleLines) {\n\t\tconst overlap = wordOverlap(words, line);\n\t\tif (overlap > bestOverlap) {\n\t\t\tbestOverlap = overlap;\n\t\t\tbestLine = line;\n\t\t}\n\t}\n\tif (bestOverlap >= COVERED_OVERLAP) return { rule: bestLine };\n\n\tlet bestSkill: string | undefined;\n\tlet bestSkillOverlap = 0;\n\tfor (const skill of index.skills) {\n\t\tconst haystack = `${skill.name} ${skill.description.slice(0, SKILL_DESCRIPTION_CHARS)}`;\n\t\tconst overlap = wordOverlap(words, haystack);\n\t\tif (overlap > bestSkillOverlap) {\n\t\t\tbestSkillOverlap = overlap;\n\t\t\tbestSkill = skill.name;\n\t\t}\n\t}\n\tif (bestSkillOverlap >= SKILL_COVERED_OVERLAP) return { skill: bestSkill };\n\n\treturn {};\n}\n\n/** Assemble the coverage index for a directory. */\nexport function buildCoverageIndex(options: {\n\tcwd: string;\n\tagentDir: string;\n\tskills?: Array<{ name: string; description: string }>;\n}): CoverageIndex {\n\tconst corpus = coverageCorpus(options.agentDir, findAgentsFile(options.cwd));\n\treturn {\n\t\truleLines: corpus\n\t\t\t.split(\"\\n\")\n\t\t\t.map((line) => line.trim())\n\t\t\t.filter((line) => line.length > 0 && !line.startsWith(\"#\")),\n\t\tskills: options.skills ?? loadSkillIndex(options.cwd, options.agentDir),\n\t};\n}\n\n/**\n * Text a proposal is checked against to decide whether it is already written\n * down — the nearest repo context file plus both user scopes.\n *\n * All three matter for suppression, because `/learn` can route a rule to the\n * user scope. Checking only the repo file would report a rule you accepted into\n * `~/.agents/AGENTS.md` as declined.\n */\nfunction coverageCorpus(agentDir: string, repoFile: string | undefined): string {\n\tconst parts: string[] = [];\n\tfor (const candidate of [repoFile, join(getUserAgentsDir(), \"AGENTS.md\"), join(agentDir, \"AGENTS.md\")]) {\n\t\tif (!candidate || !existsSync(candidate)) continue;\n\t\ttry {\n\t\t\tparts.push(readFileSync(candidate, \"utf-8\"));\n\t\t} catch {\n\t\t\t// Unreadable context file: treat as absent rather than failing the run.\n\t\t}\n\t}\n\treturn parts.join(\"\\n\");\n}\n\n/**\n * Skills a proposal could already have become.\n *\n * `/learn` routes long or conditional guidance to a skill rather than a rule, so\n * without this a proposal you adopted *as a skill* would read as declined —\n * looking only at context files sees an unchanged `AGENTS.md` and concludes you\n * passed. Reuses the real loader rather than a second SKILL.md scanner so the\n * set of locations cannot drift from what the session actually loads.\n */\nfunction loadSkillIndex(cwd: string, agentDir: string): Array<{ name: string; description: string }> {\n\ttry {\n\t\treturn loadSkills({ cwd, agentDir, skillPaths: [], includeDefaults: true }).skills.map((skill) => ({\n\t\t\tname: skill.name,\n\t\t\tdescription: skill.description ?? \"\",\n\t\t}));\n\t} catch {\n\t\t// Skills are an enrichment here, not the point of the command.\n\t\treturn [];\n\t}\n}\n\n/** Mine the recent sessions for this cwd and return the ranked digest. */\nexport function extractLearnDigest(options: ExtractOptions): LearnDigest {\n\tconst { sessions, skipped } = listSessions(options);\n\n\tconst agentsFilePath = findAgentsFile(options.cwd);\n\tlet agentsContent: string | undefined;\n\tif (agentsFilePath) {\n\t\ttry {\n\t\t\tagentsContent = readFileSync(agentsFilePath, \"utf-8\");\n\t\t} catch {\n\t\t\tagentsContent = undefined;\n\t\t}\n\t}\n\tconst coverage = buildCoverageIndex({ cwd: options.cwd, agentDir: options.agentDir, skills: options.skills });\n\n\tconst withDirectives = sessions.map((session) => ({ session, directives: userDirectives(session.entries) }));\n\tconst withEvents = sessions.map((session) => ({ session, events: toolEvents(session.entries) }));\n\n\tconst timestamps = sessions.map((s) => s.timestamp).sort();\n\tconst state = options.ignoreState ? undefined : options.state;\n\n\t// Directives carry a real coverage signal — is this written down as a rule or\n\t// a skill right now? — which is what separates an adopted proposal from a\n\t// declined one. Fixes and workflows do not: a fix may have become a rule, a\n\t// skill, or a habit, and which one is not recoverable here, so they get\n\t// suppression only and are never labelled declined.\n\tconst maxProposals = options.maxProposals ?? DEFAULT_MAX_PER_CATEGORY;\n\tconst directives = applySuppression(\n\t\tclusterDirectives(withDirectives, coverage, options.minRepeats ?? DEFAULT_MIN_DIRECTIVE_COUNT),\n\t\tstate,\n\t\tmaxProposals,\n\t\t(item) => item.status !== \"new\",\n\t\t(item) => {\n\t\t\titem.previouslyDeclined = true;\n\t\t},\n\t);\n\tconst fixes = applySuppression(extractFixes(withEvents), state, maxProposals, () => false);\n\tconst workflows = applySuppression(\n\t\textractWorkflows(withEvents, options.minWorkflowRepeats ?? DEFAULT_MIN_WORKFLOW_COUNT),\n\t\tstate,\n\t\tmaxProposals,\n\t\t() => false,\n\t);\n\n\tconst surfaced = [\n\t\t...directives.kept.map((d) => ({ key: d.key, lastSeen: d.lastSeen, covered: d.status !== \"new\" })),\n\t\t...fixes.kept.map((f) => ({ key: f.key, lastSeen: f.lastSeen, covered: false })),\n\t\t...workflows.kept.map((w) => ({ key: w.key, lastSeen: w.lastSeen, covered: false })),\n\t];\n\n\treturn {\n\t\tscannedSessions: sessions.length,\n\t\tskippedSessions: skipped,\n\t\toldestSession: timestamps[0],\n\t\tnewestSession: timestamps[timestamps.length - 1],\n\t\tagentsFilePath,\n\t\tagentsFileTokens:\n\t\t\tagentsContent === undefined ? undefined : Math.round(Buffer.byteLength(agentsContent, \"utf-8\") / 4),\n\t\tdirectives: directives.kept,\n\t\tfixes: fixes.kept,\n\t\tworkflows: workflows.kept,\n\t\tsuppressed: directives.suppressed + fixes.suppressed + workflows.suppressed,\n\t\tsurfaced,\n\t};\n}\n"]}
+{"version":3,"file":"extract.d.ts","sourceRoot":"","sources":["../../../src/core/learn/extract.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAqBH,OAAO,EAAS,KAAK,UAAU,EAAE,MAAM,YAAY,CAAC;AAEpD;;;;GAIG;AACH,eAAO,MAAM,mBAAmB,mBAAmB,CAAC;AAoCpD;;;;;;;;;;;;;GAaG;AACH,MAAM,MAAM,eAAe,GAAG,KAAK,GAAG,UAAU,GAAG,WAAW,CAAC;AAE/D,oFAAoF;AACpF,UAAU,UAAU;IACnB,qEAAmE;IACnE,GAAG,EAAE,MAAM,CAAC;IACZ,4CAA4C;IAC5C,QAAQ,EAAE,MAAM,CAAC;CACjB;AAED,MAAM,WAAW,gBAAiB,SAAQ,UAAU;IACnD,gEAAgE;IAChE,IAAI,EAAE,MAAM,CAAC;IACb,UAAU,EAAE,MAAM,CAAC;IACnB,wBAAwB;IACxB,KAAK,EAAE,MAAM,CAAC;IACd,2EAAyE;IACzE,QAAQ,EAAE,MAAM,CAAC;IACjB,MAAM,EAAE,eAAe,CAAC;IACxB,iEAAiE;IACjE,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,sEAAsE;IACtE,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB;;;;OAIG;IACH,kBAAkB,EAAE,OAAO,CAAC;CAC5B;AAED,MAAM,WAAW,YAAa,SAAQ,UAAU;IAC/C,kCAAkC;IAClC,OAAO,EAAE,MAAM,CAAC;IAChB,kDAAkD;IAClD,SAAS,EAAE,MAAM,CAAC;IAClB,gEAAgE;IAChE,YAAY,EAAE,MAAM,CAAC;IACrB,qDAAqD;IACrD,mBAAmB,EAAE,MAAM,EAAE,CAAC;IAC9B,qDAAqD;IACrD,WAAW,EAAE,MAAM,EAAE,CAAC;IACtB,sEAAsE;IACtE,KAAK,EAAE,MAAM,CAAC;IACd,QAAQ,EAAE,MAAM,CAAC;CACjB;AAED,MAAM,WAAW,iBAAkB,SAAQ,UAAU;IACpD,qCAAqC;IACrC,KAAK,EAAE,MAAM,EAAE,CAAC;IAChB,KAAK,EAAE,MAAM,CAAC;IACd,QAAQ,EAAE,MAAM,CAAC;CACjB;AAED;;;;;;;GAOG;AACH,MAAM,WAAW,iBAAiB;IACjC,+CAA+C;IAC/C,IAAI,EAAE,MAAM,EAAE,CAAC;IACf,6CAA6C;IAC7C,WAAW,EAAE,MAAM,EAAE,CAAC;IACtB,4DAA4D;IAC5D,KAAK,EAAE,MAAM,CAAC;IACd,mDAAmD;IACnD,MAAM,EAAE,MAAM,CAAC;IACf,gFAAgF;IAChF,QAAQ,EAAE,MAAM,CAAC;IACjB,8CAA8C;IAC9C,SAAS,EAAE,MAAM,CAAC;IAClB,2DAA2D;IAC3D,UAAU,EAAE,MAAM,CAAC;CACnB;AAED,MAAM,WAAW,WAAW;IAC3B,eAAe,EAAE,MAAM,CAAC;IACxB,eAAe,EAAE,MAAM,CAAC;IACxB,8DAA8D;IAC9D,IAAI,EAAE,iBAAiB,CAAC;IACxB,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,gBAAgB,CAAC,EAAE,MAAM,CAAC;IAC1B,UAAU,EAAE,gBAAgB,EAAE,CAAC;IAC/B,KAAK,EAAE,YAAY,EAAE,CAAC;IACtB,SAAS,EAAE,iBAAiB,EAAE,CAAC;IAC/B,mFAAmF;IACnF,UAAU,EAAE,MAAM,CAAC;IACnB,oEAAoE;IACpE,QAAQ,EAAE,KAAK,CAAC;QAAE,GAAG,EAAE,MAAM,CAAC;QAAC,QAAQ,EAAE,MAAM,CAAC;QAAC,OAAO,EAAE,OAAO,CAAA;KAAE,CAAC,CAAC;CACrE;AAED,MAAM,WAAW,cAAc;IAC9B,GAAG,EAAE,MAAM,CAAC;IACZ,QAAQ,EAAE,MAAM,CAAC;IACjB;;;;OAIG;IACH,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,kFAAkF;IAClF,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,sFAAsF;IACtF,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,sCAAsC;IACtC,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB;;;OAGG;IACH,KAAK,CAAC,EAAE,UAAU,CAAC;IACnB,kFAAkF;IAClF,WAAW,CAAC,EAAE,OAAO,CAAC;IACtB;;;OAGG;IACH,MAAM,CAAC,EAAE,KAAK,CAAC;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,WAAW,EAAE,MAAM,CAAA;KAAE,CAAC,CAAC;IACtD,mCAAmC;IACnC,GAAG,CAAC,EAAE,IAAI,CAAC;CACX;AAmJD;;;;;;;;;;GAUG;AACH,wBAAgB,oBAAoB,CAAC,OAAO,EAAE,IAAI,CAAC,cAAc,EAAE,KAAK,GAAG,UAAU,GAAG,YAAY,CAAC,GAAG,MAAM,EAAE,CAW/G;AAqHD;;;GAGG;AACH,wBAAgB,YAAY,CAAC,OAAO,EAAE,cAAc,GAAG,iBAAiB,CAEvE;AAkYD;;;;;;;GAOG;AACH,MAAM,WAAW,aAAa;IAC7B,4EAA4E;IAC5E,SAAS,EAAE,MAAM,EAAE,CAAC;IACpB,MAAM,EAAE,KAAK,CAAC;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,WAAW,EAAE,MAAM,CAAA;KAAE,CAAC,CAAC;CACrD;AAED,MAAM,WAAW,aAAa;IAC7B,sDAAsD;IACtD,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,yEAAyE;IACzE,KAAK,CAAC,EAAE,MAAM,CAAC;CACf;AAED;;;;;GAKG;AACH,wBAAgB,aAAa,CAAC,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,aAAa,GAAG,aAAa,CA2B/E;AAED,mDAAmD;AACnD,wBAAgB,kBAAkB,CAAC,OAAO,EAAE;IAC3C,GAAG,EAAE,MAAM,CAAC;IACZ,QAAQ,EAAE,MAAM,CAAC;IACjB,MAAM,CAAC,EAAE,KAAK,CAAC;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,WAAW,EAAE,MAAM,CAAA;KAAE,CAAC,CAAC;CACtD,GAAG,aAAa,CAShB;AA4CD,0EAA0E;AAC1E,wBAAgB,kBAAkB,CAAC,OAAO,EAAE,cAAc,GAAG,WAAW,CAgEvE","sourcesContent":["/**\n * Session mining for `/learn`.\n *\n * Reads session `.jsonl` files straight off disk rather than the live context.\n * That is the whole point: the on-disk transcript is complete even when the\n * in-context one has been compacted away, and it spans every past session\n * instead of only this one. Cross-session repetition is the signal that decides\n * whether something is a durable rule or a one-off, and it is the one thing a\n * prompt reading its own context cannot see.\n *\n * The split of labour is deliberate. This module is entirely deterministic: it\n * parses, filters, normalizes, counts and ranks. Judgement — is this a rule, how\n * should it be phrased, which scope owns it — belongs to the model reading the\n * digest, which is why the output carries evidence (counts, sessions, dates)\n * rather than conclusions.\n */\n\nimport { existsSync, readdirSync, readFileSync, realpathSync, statSync } from \"node:fs\";\nimport { dirname, join, resolve, sep } from \"node:path\";\nimport type { AgentMessage } from \"@kolisachint/hoocode-agent-core\";\nimport type { TextContent, ToolCall } from \"@kolisachint/hoocode-ai\";\nimport { getUserAgentsDir } from \"../../config.js\";\nimport { getSessionDirPath } from \"../session-manager.js\";\nimport { loadSkills } from \"../skills.js\";\nimport {\n\tcommandHead,\n\tcontentWords,\n\textractErrorRegion,\n\tisBenignFailure,\n\tisRuleShapedDirective,\n\tisUninformativeFailure,\n\tnormalizeCommand,\n\tnormalizeDirective,\n\tnormalizeErrorSignature,\n\twordOverlap,\n} from \"./normalize.js\";\nimport { judge, type LearnState } from \"./state.js\";\n\n/**\n * Prefix on the message `/learn` injects. The digest is persisted like any user\n * turn, so without this marker the next `/learn` would mine its own output and\n * every proposal would compound its own count.\n */\nexport const LEARN_DIGEST_MARKER = \"[learn-digest]\";\n\n/** Sessions considered, newest first. */\nconst DEFAULT_MAX_SESSIONS = 20;\n/** Sessions older than this are ignored — a pattern that stopped is not a rule. */\nconst DEFAULT_MAX_AGE_DAYS = 30;\n/** Entries parsed per session file, as a guard against pathological transcripts. */\nconst MAX_ENTRIES_PER_SESSION = 8000;\n/** Tool calls per session fed to the workflow detector. */\nconst MAX_TOOL_CALLS_PER_SESSION = 400;\n/** How far forward the fix extractor looks for the same command succeeding. */\nconst FIX_LOOKAHEAD = 40;\n/** Word overlap against an existing rule above which a directive counts as covered. */\nconst COVERED_OVERLAP = 0.6;\n/**\n * The same bar for skills, set higher on purpose.\n *\n * A rule is one line, so overlap against it is a sharp signal. A skill is a name\n * plus a description written to attract matches, which is a far larger haystack\n * — a short directive's words turn up in it by chance much more readily. The\n * higher bar and the truncation below keep \"you already have a skill for this\"\n * from being said on a coincidence.\n */\nconst SKILL_COVERED_OVERLAP = 0.75;\n/** Description characters considered. The opening says what a skill does; the rest is trigger bait. */\nconst SKILL_DESCRIPTION_CHARS = 300;\n/** Directives must reach this many occurrences to be reported at all. */\nconst DEFAULT_MIN_DIRECTIVE_COUNT = 2;\n/** Tool sequence lengths considered as workflow candidates. */\nconst WORKFLOW_MIN_LEN = 3;\nconst WORKFLOW_MAX_LEN = 5;\n/** Repeats before a tool sequence is worth proposing as a skill. */\nconst DEFAULT_MIN_WORKFLOW_COUNT = 3;\n/** Cap on each list in the digest, so the model's budget goes to the top signals. */\nconst DEFAULT_MAX_PER_CATEGORY = 8;\n\n/**\n * Where a repeated directive already lives, if anywhere.\n *\n * A directive covered by a rule and said only once is simply dropped — the rule\n * exists and is working. What survives is one of three cases, and they want\n * different responses:\n *\n * - `new` — not written down anywhere. Propose it.\n * - `restated` — a context-file rule covers it and you said it anyway, so the\n *   rule is not working. Rewrite it; do not add a second one.\n * - `has-skill` — a *skill* covers it and you asked by hand anyway, which\n *   usually means the skill's `description` is not triggering. Sharpen the\n *   description rather than writing a rule that duplicates the skill.\n */\nexport type DirectiveStatus = \"new\" | \"restated\" | \"has-skill\";\n\n/** Fields every proposable item shares, so suppression can be applied uniformly. */\ninterface Proposable {\n\t/** Stable identity across runs — what the state file remembers. */\n\tkey: string;\n\t/** Newest occurrence in the window, ISO. */\n\tlastSeen: string;\n}\n\nexport interface DirectiveCluster extends Proposable {\n\t/** Representative raw text, the longest seen in the cluster. */\n\ttext: string;\n\tnormalized: string;\n\t/** Total times said. */\n\tcount: number;\n\t/** Distinct sessions it was said in — the stronger of the two counts. */\n\tsessions: number;\n\tstatus: DirectiveStatus;\n\t/** The existing rule line matched, when status is `restated`. */\n\texistingRule?: string;\n\t/** The skill that already covers this, when status is `has-skill`. */\n\texistingSkill?: string;\n\t/**\n\t * Shown before and still not written down anywhere — neither as a rule nor as\n\t * a skill — so you saw this proposal and passed on it. Only meaningful for\n\t * directives, which are the only items with a real coverage signal.\n\t */\n\tpreviouslyDeclined: boolean;\n}\n\nexport interface FixCandidate extends Proposable {\n\t/** Normalized failing command. */\n\tcommand: string;\n\t/** Normalized error signature, the dedupe key. */\n\tsignature: string;\n\t/** Short raw excerpt, so the model sees the real error text. */\n\terrorExcerpt: string;\n\t/** Commands run between the failure and the pass. */\n\tinterveningCommands: string[];\n\t/** Files edited between the failure and the pass. */\n\teditedFiles: string[];\n\t/** Times this signature failed and was resolved across the window. */\n\tcount: number;\n\tsessions: number;\n}\n\nexport interface WorkflowCandidate extends Proposable {\n\t/** Tool-call signatures in order. */\n\tsteps: string[];\n\tcount: number;\n\tsessions: number;\n}\n\n/**\n * Why a session file on disk did not make it into the digest.\n *\n * \"No recent sessions\" is the one outcome a user cannot act on without this:\n * an empty session directory, a directory full of month-old sessions, and a\n * directory full of sessions belonging to another checkout all produce the same\n * sentence, and the fix differs in each case.\n */\nexport interface SessionScanReport {\n\t/** Directories actually searched, in order. */\n\tdirs: string[];\n\t/** Directories that do not exist on disk. */\n\tmissingDirs: string[];\n\t/** `.jsonl` files found across all searched directories. */\n\tfiles: number;\n\t/** Skipped for being older than the age window. */\n\ttooOld: number;\n\t/** Skipped because the session header records a different working directory. */\n\totherCwd: number;\n\t/** Skipped for being beyond `maxSessions`. */\n\toverLimit: number;\n\t/** Skipped for being unreadable, unparseable, or empty. */\n\tunreadable: number;\n}\n\nexport interface LearnDigest {\n\tscannedSessions: number;\n\tskippedSessions: number;\n\t/** Where the sessions came from, and what was passed over. */\n\tscan: SessionScanReport;\n\toldestSession?: string;\n\tnewestSession?: string;\n\tagentsFilePath?: string;\n\tagentsFileTokens?: number;\n\tdirectives: DirectiveCluster[];\n\tfixes: FixCandidate[];\n\tworkflows: WorkflowCandidate[];\n\t/** Items held back because nothing new has happened since they were last shown. */\n\tsuppressed: number;\n\t/** Everything this run put on screen, for the caller to persist. */\n\tsurfaced: Array<{ key: string; lastSeen: string; covered: boolean }>;\n}\n\nexport interface ExtractOptions {\n\tcwd: string;\n\tagentDir: string;\n\t/**\n\t * An extra directory to scan, normally the live session manager's. The\n\t * per-cwd default directory is always scanned as well, so a session manager\n\t * pointing somewhere unusual cannot hide this directory's history.\n\t */\n\tsessionDir?: string;\n\tmaxSessions?: number;\n\tmaxAgeDays?: number;\n\t/** Occurrences a directive needs before it is proposed. The signal/noise dial. */\n\tminRepeats?: number;\n\t/** Non-overlapping repeats a tool sequence needs before it is proposed as a skill. */\n\tminWorkflowRepeats?: number;\n\t/** Cap on each list in the digest. */\n\tmaxProposals?: number;\n\t/**\n\t * What previous runs already showed. Items with no new occurrences since are\n\t * held back. Omit (or pass `ignoreState`) to propose everything in the window.\n\t */\n\tstate?: LearnState;\n\t/** Re-propose everything, ignoring what previous runs surfaced (`/learn all`). */\n\tignoreState?: boolean;\n\t/**\n\t * Skills a directive can already be covered by. Defaults to the ones loaded\n\t * from disk; injectable so tests do not read the developer's real skills.\n\t */\n\tskills?: Array<{ name: string; description: string }>;\n\t/** Injectable clock, for tests. */\n\tnow?: Date;\n}\n\ninterface SessionHeaderLike {\n\ttype: \"session\";\n\tid?: string;\n\ttimestamp?: string;\n\tcwd?: string;\n}\n\ninterface EntryLike {\n\ttype: string;\n\tid?: string;\n\tparentId?: string | null;\n\ttimestamp?: string;\n\tmessage?: AgentMessage;\n}\n\n/** One session, reduced to the branch that was actually taken. */\ninterface ParsedSession {\n\tfile: string;\n\tid: string;\n\ttimestamp: string;\n\tentries: EntryLike[];\n}\n\nfunction textOf(content: unknown): string {\n\tif (typeof content === \"string\") return content;\n\tif (!Array.isArray(content)) return \"\";\n\treturn content\n\t\t.map((block) =>\n\t\t\tblock && typeof block === \"object\" && (block as TextContent).type === \"text\"\n\t\t\t\t? ((block as TextContent).text ?? \"\")\n\t\t\t\t: \"\",\n\t\t)\n\t\t.join(\"\\n\")\n\t\t.trim();\n}\n\nfunction isToolCall(block: unknown): block is ToolCall {\n\treturn !!block && typeof block === \"object\" && (block as ToolCall).type === \"toolCall\";\n}\n\n/**\n * Reduce a session's raw entries to the branch that was actually taken.\n *\n * Session files are trees — forks and clones append entries that were never\n * part of the same conversation. Walking parent links back from the last entry\n * keeps the extractor from stitching a \"fix\" out of two turns that never\n * happened in sequence. Sessions written before entry ids existed are flat, and\n * for those file order *is* the branch.\n */\nfunction activeBranch(entries: EntryLike[]): EntryLike[] {\n\tconst withIds = entries.filter((e) => typeof e.id === \"string\");\n\tif (withIds.length === 0) return entries;\n\n\tconst byId = new Map<string, EntryLike>();\n\tfor (const entry of withIds) byId.set(entry.id as string, entry);\n\n\tconst branch: EntryLike[] = [];\n\tconst seen = new Set<string>();\n\tlet cursor: EntryLike | undefined = withIds[withIds.length - 1];\n\twhile (cursor?.id && !seen.has(cursor.id)) {\n\t\tseen.add(cursor.id);\n\t\tbranch.push(cursor);\n\t\tcursor = cursor.parentId ? byId.get(cursor.parentId) : undefined;\n\t}\n\treturn branch.reverse();\n}\n\n/**\n * Compare two directory paths the way the filesystem does.\n *\n * A session header stores the cwd as it was typed, and the same directory can\n * be spelled several ways: through a symlink (`/tmp` is `/private/tmp` on\n * macOS), with a trailing separator, or in different case on the\n * case-insensitive filesystems that macOS and Windows ship by default. String\n * equality on `resolve()` alone rejects every one of those, and rejecting them\n * here means silently discarding the whole history the command exists to read.\n */\nfunction normalizeDirPath(path: string): string {\n\tlet resolved = resolve(path);\n\ttry {\n\t\tresolved = realpathSync.native(resolved);\n\t} catch {\n\t\t// Deleted or never-created directory: the textual form is all we have.\n\t}\n\t// `resolve` already drops a trailing separator except at a filesystem root,\n\t// where dropping it would turn \"/\" into \"\".\n\tif (resolved.length > 1 && resolved.endsWith(sep)) resolved = resolved.slice(0, -1);\n\treturn process.platform === \"win32\" || process.platform === \"darwin\" ? resolved.toLowerCase() : resolved;\n}\n\nfunction sameDirectory(a: string, b: string): boolean {\n\treturn normalizeDirPath(a) === normalizeDirPath(b);\n}\n\n/** Reason a candidate file produced no session, for the scan report. */\ntype SkipReason = \"otherCwd\" | \"unreadable\";\n\nfunction parseSessionFile(file: string, cwd: string, onSkip: (reason: SkipReason) => void): ParsedSession | undefined {\n\tlet raw: string;\n\ttry {\n\t\traw = readFileSync(file, \"utf-8\");\n\t} catch {\n\t\tonSkip(\"unreadable\");\n\t\treturn undefined;\n\t}\n\n\tconst lines = raw.split(\"\\n\");\n\tlet header: SessionHeaderLike | undefined;\n\tconst entries: EntryLike[] = [];\n\tfor (const line of lines) {\n\t\tif (!line.trim()) continue;\n\t\tif (entries.length >= MAX_ENTRIES_PER_SESSION) break;\n\t\tlet parsed: EntryLike | SessionHeaderLike;\n\t\ttry {\n\t\t\tparsed = JSON.parse(line);\n\t\t} catch {\n\t\t\t// A partially-flushed final line is normal for a live session.\n\t\t\tcontinue;\n\t\t}\n\t\tif (parsed.type === \"session\") {\n\t\t\theader ??= parsed as SessionHeaderLike;\n\t\t\tcontinue;\n\t\t}\n\t\tentries.push(parsed as EntryLike);\n\t}\n\n\t// An explicit `--session` path can put a session for another directory in\n\t// this directory, so trust the header over the file's location.\n\tif (header?.cwd && !sameDirectory(header.cwd, cwd)) {\n\t\tonSkip(\"otherCwd\");\n\t\treturn undefined;\n\t}\n\tif (entries.length === 0) {\n\t\tonSkip(\"unreadable\");\n\t\treturn undefined;\n\t}\n\n\treturn {\n\t\tfile,\n\t\tid: header?.id ?? file,\n\t\ttimestamp: header?.timestamp ?? statSync(file).mtime.toISOString(),\n\t\tentries: activeBranch(entries),\n\t};\n}\n\n/**\n * Every directory this cwd's sessions could be sitting in.\n *\n * The caller passes the live session manager's directory, which is the right\n * answer almost always — but not quite always, and each exception silently\n * emptied the digest. An in-memory session (`--no-session`) reports `\"\"`; an\n * explicit `--session <path>` reports wherever that file lives; a custom\n * `sessionDir` setting points at one shared directory. In every one of those\n * cases the per-cwd default directory still holds the history worth mining, so\n * search both and let the header check sort out what belongs to this cwd.\n */\nexport function candidateSessionDirs(options: Pick<ExtractOptions, \"cwd\" | \"agentDir\" | \"sessionDir\">): string[] {\n\tconst dirs: string[] = [];\n\tconst seen = new Set<string>();\n\tfor (const dir of [options.sessionDir, getSessionDirPath(options.cwd, options.agentDir)]) {\n\t\tif (!dir) continue;\n\t\tconst key = normalizeDirPath(dir);\n\t\tif (seen.has(key)) continue;\n\t\tseen.add(key);\n\t\tdirs.push(dir);\n\t}\n\treturn dirs;\n}\n\nfunction listSessions(options: ExtractOptions): {\n\tsessions: ParsedSession[];\n\tskipped: number;\n\tscan: SessionScanReport;\n} {\n\tconst dirs = candidateSessionDirs(options);\n\tconst scan: SessionScanReport = {\n\t\tdirs,\n\t\tmissingDirs: [],\n\t\tfiles: 0,\n\t\ttooOld: 0,\n\t\totherCwd: 0,\n\t\toverLimit: 0,\n\t\tunreadable: 0,\n\t};\n\n\tconst maxSessions = options.maxSessions ?? DEFAULT_MAX_SESSIONS;\n\tconst maxAgeDays = options.maxAgeDays ?? DEFAULT_MAX_AGE_DAYS;\n\tconst now = options.now ?? new Date();\n\tconst cutoff = now.getTime() - maxAgeDays * 24 * 60 * 60 * 1000;\n\n\tconst files: string[] = [];\n\tfor (const dir of dirs) {\n\t\tif (!existsSync(dir)) {\n\t\t\tscan.missingDirs.push(dir);\n\t\t\tcontinue;\n\t\t}\n\t\ttry {\n\t\t\tfor (const name of readdirSync(dir)) {\n\t\t\t\tif (name.endsWith(\".jsonl\")) files.push(join(dir, name));\n\t\t\t}\n\t\t} catch {\n\t\t\tscan.missingDirs.push(dir);\n\t\t}\n\t}\n\tscan.files = files.length;\n\n\t// Newest first across all directories, so `maxSessions` keeps the most recent\n\t// history rather than whichever directory happened to be searched first.\n\tconst dated = files\n\t\t.map((file) => {\n\t\t\ttry {\n\t\t\t\treturn { file, mtime: statSync(file).mtime.getTime() };\n\t\t\t} catch {\n\t\t\t\tscan.unreadable++;\n\t\t\t\treturn undefined;\n\t\t\t}\n\t\t})\n\t\t.filter((f): f is { file: string; mtime: number } => !!f)\n\t\t.sort((a, b) => b.mtime - a.mtime);\n\n\tconst sessions: ParsedSession[] = [];\n\tconst seenIds = new Set<string>();\n\tlet skipped = 0;\n\tfor (const { file, mtime } of dated) {\n\t\tif (sessions.length >= maxSessions) {\n\t\t\tscan.overLimit++;\n\t\t\tskipped++;\n\t\t\tcontinue;\n\t\t}\n\t\tif (mtime < cutoff) {\n\t\t\tscan.tooOld++;\n\t\t\tskipped++;\n\t\t\tcontinue;\n\t\t}\n\t\tconst parsed = parseSessionFile(file, options.cwd, (reason) => {\n\t\t\tscan[reason]++;\n\t\t});\n\t\tif (!parsed) {\n\t\t\tskipped++;\n\t\t\tcontinue;\n\t\t}\n\t\t// Searching two directories can turn up the same session twice (an explicit\n\t\t// `--session` path inside the default directory). Counting it twice would\n\t\t// inflate the cross-session repetition that decides what gets proposed.\n\t\tif (seenIds.has(parsed.id)) {\n\t\t\tskipped++;\n\t\t\tcontinue;\n\t\t}\n\t\tseenIds.add(parsed.id);\n\t\tsessions.push(parsed);\n\t}\n\treturn { sessions, skipped, scan };\n}\n\n/**\n * Hold back items already shown that have not recurred since, then cap the rest.\n *\n * Order matters: suppression runs *before* the cap, or an item you already\n * decided on would occupy one of the few slots the digest has and push a live\n * signal off the list.\n */\nfunction applySuppression<T extends Proposable>(\n\titems: T[],\n\tstate: LearnState | undefined,\n\tmaxProposals: number,\n\tcovered: (item: T) => boolean,\n\tonDeclined?: (item: T) => void,\n): { kept: T[]; suppressed: number } {\n\tif (!state) return { kept: items.slice(0, maxProposals), suppressed: 0 };\n\n\tconst kept: T[] = [];\n\tlet suppressed = 0;\n\tfor (const item of items) {\n\t\tconst verdict = judge(state, { key: item.key, lastSeen: item.lastSeen, covered: covered(item) });\n\t\tif (verdict.suppressed) {\n\t\t\tsuppressed++;\n\t\t\tcontinue;\n\t\t}\n\t\tif (verdict.previouslyDeclined) onDeclined?.(item);\n\t\tkept.push(item);\n\t}\n\treturn { kept: kept.slice(0, maxProposals), suppressed };\n}\n\n/**\n * Where this cwd's sessions were found and what was passed over, without\n * ranking anything. `/learn stats` reports on the window without re-mining it.\n */\nexport function scanSessions(options: ExtractOptions): SessionScanReport {\n\treturn listSessions(options).scan;\n}\n\n/** Nearest AGENTS.md walking up from cwd, so proposals can be checked against it. */\nfunction findAgentsFile(cwd: string): string | undefined {\n\tlet dir = resolve(cwd);\n\twhile (true) {\n\t\tfor (const name of [\"AGENTS.md\", \"AGENTS.MD\", \"CLAUDE.md\", \"CLAUDE.MD\"]) {\n\t\t\tconst candidate = join(dir, name);\n\t\t\tif (existsSync(candidate)) return candidate;\n\t\t}\n\t\tconst parent = dirname(dir);\n\t\tif (parent === dir) return undefined;\n\t\tdir = parent;\n\t}\n}\n\ninterface ToolEvent {\n\tname: string;\n\targs: Record<string, any>;\n\t/** Set once the matching result is seen. */\n\tisError?: boolean;\n\toutput?: string;\n}\n\n/** Pair tool calls with their results along one branch, in call order. */\nfunction toolEvents(entries: EntryLike[]): ToolEvent[] {\n\tconst byCallId = new Map<string, ToolEvent>();\n\tconst ordered: ToolEvent[] = [];\n\n\tfor (const entry of entries) {\n\t\tconst message = entry.type === \"message\" ? entry.message : undefined;\n\t\tif (!message) continue;\n\t\tif (message.role === \"assistant\") {\n\t\t\tfor (const block of (message.content ?? []) as unknown[]) {\n\t\t\t\tif (!isToolCall(block)) continue;\n\t\t\t\tconst event: ToolEvent = { name: block.name, args: block.arguments ?? {} };\n\t\t\t\tbyCallId.set(block.id, event);\n\t\t\t\tordered.push(event);\n\t\t\t}\n\t\t} else if (message.role === \"toolResult\") {\n\t\t\tconst event = byCallId.get(message.toolCallId);\n\t\t\tif (!event) continue;\n\t\t\tevent.isError = message.isError;\n\t\t\tevent.output = textOf(message.content);\n\t\t}\n\t}\n\treturn ordered;\n}\n\n/** User turns worth mining, in order, with the digest's own output excluded. */\nfunction userDirectives(entries: EntryLike[]): string[] {\n\tconst out: string[] = [];\n\tfor (const entry of entries) {\n\t\tconst message = entry.type === \"message\" ? entry.message : undefined;\n\t\tif (!message || message.role !== \"user\") continue;\n\t\tconst text = textOf(message.content);\n\t\tif (!text || text.startsWith(LEARN_DIGEST_MARKER)) continue;\n\t\tif (!isRuleShapedDirective(text)) continue;\n\t\tout.push(text.trim());\n\t}\n\treturn out;\n}\n\nfunction clusterDirectives(\n\tperSession: Array<{ session: ParsedSession; directives: string[] }>,\n\tcoverage: CoverageIndex,\n\tminRepeats: number,\n): DirectiveCluster[] {\n\tinterface Acc {\n\t\ttext: string;\n\t\tnormalized: string;\n\t\tcount: number;\n\t\tsessions: Set<string>;\n\t\tlastSeen: string;\n\t}\n\tconst acc = new Map<string, Acc>();\n\n\tfor (const { session, directives } of perSession) {\n\t\tfor (const text of directives) {\n\t\t\tconst normalized = normalizeDirective(text);\n\t\t\tif (!normalized) continue;\n\t\t\tconst existing = acc.get(normalized);\n\t\t\tif (existing) {\n\t\t\t\texisting.count++;\n\t\t\t\texisting.sessions.add(session.id);\n\t\t\t\tif (session.timestamp > existing.lastSeen) existing.lastSeen = session.timestamp;\n\t\t\t\tif (text.length > existing.text.length) existing.text = text;\n\t\t\t} else {\n\t\t\t\tacc.set(normalized, {\n\t\t\t\t\ttext,\n\t\t\t\t\tnormalized,\n\t\t\t\t\tcount: 1,\n\t\t\t\t\tsessions: new Set([session.id]),\n\t\t\t\t\tlastSeen: session.timestamp,\n\t\t\t\t});\n\t\t\t}\n\t\t}\n\t}\n\n\tconst clusters: DirectiveCluster[] = [];\n\tfor (const entry of acc.values()) {\n\t\tif (entry.count < minRepeats) continue;\n\n\t\t// Everything reaching here cleared the repeat threshold. Suppression handles\n\t\t// the case that used to make these labels lie — a proposal accepted from a\n\t\t// previous run coming back as \"not working\" when nothing had happened\n\t\t// since. By the time an item survives that filter, a match genuinely means\n\t\t// you repeated yourself after the rule or skill already existed.\n\t\tconst match = matchCoverage(entry.text, coverage);\n\t\tclusters.push({\n\t\t\tkey: `directive:${entry.normalized}`,\n\t\t\ttext: entry.text,\n\t\t\tnormalized: entry.normalized,\n\t\t\tcount: entry.count,\n\t\t\tsessions: entry.sessions.size,\n\t\t\tlastSeen: entry.lastSeen,\n\t\t\tstatus: match.rule ? \"restated\" : match.skill ? \"has-skill\" : \"new\",\n\t\t\texistingRule: match.rule,\n\t\t\texistingSkill: match.skill,\n\t\t\tpreviouslyDeclined: false,\n\t\t});\n\t}\n\n\treturn clusters.sort((a, b) => b.sessions - a.sessions || b.count - a.count || a.text.localeCompare(b.text));\n}\n\n/** Files a mutating tool touched, for the resolution summary. */\nfunction editedFile(event: ToolEvent): string | undefined {\n\tif (![\"edit\", \"write\", \"multi_edit\", \"apply_patch\"].includes(event.name)) return undefined;\n\tconst path = event.args?.path ?? event.args?.file_path ?? event.args?.filePath;\n\treturn typeof path === \"string\" ? path : undefined;\n}\n\nfunction extractFixes(perSession: Array<{ session: ParsedSession; events: ToolEvent[] }>): FixCandidate[] {\n\tinterface Acc {\n\t\tcandidate: FixCandidate;\n\t\tsessions: Set<string>;\n\t}\n\tconst acc = new Map<string, Acc>();\n\n\tfor (const { session, events } of perSession) {\n\t\tfor (let i = 0; i < events.length; i++) {\n\t\t\tconst failure = events[i]!;\n\t\t\tif (failure.name !== \"bash\" || !failure.isError) continue;\n\t\t\tconst command = typeof failure.args?.command === \"string\" ? failure.args.command : \"\";\n\t\t\tif (!command || isBenignFailure(command)) continue;\n\n\t\t\tconst normalized = normalizeCommand(command);\n\t\t\tconst interveningCommands: string[] = [];\n\t\t\tconst editedFiles: string[] = [];\n\t\t\tlet resolved = false;\n\n\t\t\tfor (let j = i + 1; j < Math.min(events.length, i + 1 + FIX_LOOKAHEAD); j++) {\n\t\t\t\tconst next = events[j]!;\n\t\t\t\tconst file = editedFile(next);\n\t\t\t\tif (file) editedFiles.push(file);\n\n\t\t\t\tif (next.name !== \"bash\") continue;\n\t\t\t\tconst nextCommand = typeof next.args?.command === \"string\" ? next.args.command : \"\";\n\t\t\t\tif (!nextCommand) continue;\n\n\t\t\t\t// The same command later succeeding is the only evidence that the\n\t\t\t\t// problem was actually fixed. A *different* command passing says\n\t\t\t\t// nothing, and neither does the model moving on.\n\t\t\t\tif (normalizeCommand(nextCommand) === normalized && !next.isError) {\n\t\t\t\t\tresolved = true;\n\t\t\t\t\tbreak;\n\t\t\t\t}\n\t\t\t\tinterveningCommands.push(nextCommand.trim());\n\t\t\t}\n\n\t\t\tif (!resolved) continue;\n\n\t\t\tconst output = failure.output ?? \"\";\n\t\t\t// An abort is the user changing their mind, not a problem that was\n\t\t\t// solved, and empty output carries nothing to sign or show.\n\t\t\tif (isUninformativeFailure(output)) continue;\n\n\t\t\t// Sign the error region, not the whole output: build tools lead with an\n\t\t\t// identical banner, so signing everything makes unrelated failures of\n\t\t\t// the same command collide on their shared preamble.\n\t\t\tconst errorRegion = extractErrorRegion(output);\n\t\t\tconst signature = normalizeErrorSignature(errorRegion);\n\t\t\tif (!signature) continue;\n\n\t\t\tconst key = `${normalized}\u0000${signature}`;\n\t\t\tconst existing = acc.get(key);\n\t\t\tif (existing) {\n\t\t\t\texisting.candidate.count++;\n\t\t\t\texisting.sessions.add(session.id);\n\t\t\t\tif (session.timestamp > existing.candidate.lastSeen) existing.candidate.lastSeen = session.timestamp;\n\t\t\t} else {\n\t\t\t\tacc.set(key, {\n\t\t\t\t\tsessions: new Set([session.id]),\n\t\t\t\t\tcandidate: {\n\t\t\t\t\t\tkey: `fix:${key}`,\n\t\t\t\t\t\tcommand: normalized,\n\t\t\t\t\t\tsignature,\n\t\t\t\t\t\terrorExcerpt: errorRegion.replace(/\\s+/g, \" \").trim().slice(0, 240),\n\t\t\t\t\t\tinterveningCommands: [...new Set(interveningCommands)].slice(0, 5),\n\t\t\t\t\t\teditedFiles: [...new Set(editedFiles)].slice(0, 5),\n\t\t\t\t\t\tcount: 1,\n\t\t\t\t\t\tsessions: 1,\n\t\t\t\t\t\tlastSeen: session.timestamp,\n\t\t\t\t\t},\n\t\t\t\t});\n\t\t\t}\n\t\t}\n\t}\n\n\tconst out: FixCandidate[] = [];\n\tfor (const { candidate, sessions } of acc.values()) {\n\t\tcandidate.sessions = sessions.size;\n\t\tout.push(candidate);\n\t}\n\treturn out.sort((a, b) => b.count - a.count || b.sessions - a.sessions || a.signature.localeCompare(b.signature));\n}\n\n/** A tool call reduced to a comparable step: the tool, plus what a bash call runs. */\nfunction stepSignature(event: ToolEvent): string {\n\tif (event.name === \"bash\") {\n\t\tconst command = typeof event.args?.command === \"string\" ? event.args.command : \"\";\n\t\tconst head = commandHead(command);\n\t\treturn head ? `bash:${head}` : \"bash\";\n\t}\n\treturn event.name;\n}\n\n/**\n * Commands that are how an agent looks around rather than what the user was\n * doing. A sequence built only from these plus file edits describes \"coding\",\n * not a workflow, and no useful skill has ever come out of one.\n */\nconst PLUMBING_COMMANDS = new Set([\n\t\"cd\",\n\t\"ls\",\n\t\"pwd\",\n\t\"cat\",\n\t\"head\",\n\t\"tail\",\n\t\"wc\",\n\t\"echo\",\n\t\"which\",\n\t\"find\",\n\t\"fd\",\n\t\"grep\",\n\t\"rg\",\n\t\"sed\",\n\t\"awk\",\n\t\"git status\",\n\t\"git diff\",\n\t\"git log\",\n\t\"git show\",\n]);\n\n/**\n * Whether a sequence is a procedure rather than the rhythm of editing code.\n *\n * Two distinct doing-commands is the bar, and it was set by looking at real\n * transcripts. One command is not enough: the edit/test loop\n * (`edit → edit → bash:npm run`) satisfies it, and because a sliding window\n * over a long alternating run produces every rotation of that cycle, it alone\n * filled all eight slots with `edit → npm run → edit`, `npm run → edit → edit`\n * and so on — one habit described eight ways.\n *\n * A procedure worth a skill chains *different* actions: test then commit then\n * push, build then tag then publish. Requiring two distinct ones keeps those and\n * drops the rhythm. The cost is real — a genuine one-command routine with setup\n * is missed — and that is the intended trade, since a missed skill costs nothing\n * while a digest full of noise costs the reader's attention every run.\n */\nfunction isProcedure(steps: string[]): boolean {\n\tconst commands = new Set<string>();\n\tfor (const step of steps) {\n\t\tif (!step.startsWith(\"bash:\")) continue;\n\t\tconst head = step.slice(\"bash:\".length);\n\t\tif (PLUMBING_COMMANDS.has(head) || PLUMBING_COMMANDS.has(head.split(\" \")[0] ?? \"\")) continue;\n\t\tcommands.add(head);\n\t}\n\treturn commands.size >= 2;\n}\n\n/** True when `needle` appears as a contiguous run inside `haystack`. */\nfunction containsSequence(haystack: string[], needle: string[]): boolean {\n\tif (needle.length > haystack.length) return false;\n\tfor (let i = 0; i + needle.length <= haystack.length; i++) {\n\t\tif (needle.every((step, offset) => haystack[i + offset] === step)) return true;\n\t}\n\treturn false;\n}\n\nfunction extractWorkflows(\n\tperSession: Array<{ session: ParsedSession; events: ToolEvent[] }>,\n\tminRepeats: number,\n): WorkflowCandidate[] {\n\tinterface Acc {\n\t\tsteps: string[];\n\t\tcount: number;\n\t\tsessions: Set<string>;\n\t\tlastSeen: string;\n\t}\n\tconst acc = new Map<string, Acc>();\n\n\tfor (const { session, events } of perSession) {\n\t\tconst steps = events.slice(0, MAX_TOOL_CALLS_PER_SESSION).map(stepSignature);\n\n\t\tfor (let len = WORKFLOW_MIN_LEN; len <= WORKFLOW_MAX_LEN; len++) {\n\t\t\t// Collect every position first, then count greedily without overlap.\n\t\t\t// Counting each sliding position separately treats one long stretch of\n\t\t\t// edit/read churn as dozens of repeats: an `edit > read > edit` run of\n\t\t\t// length 12 scores 10 occurrences when it is really one stretch of work.\n\t\t\tconst positions = new Map<string, number[]>();\n\t\t\tfor (let i = 0; i + len <= steps.length; i++) {\n\t\t\t\tconst window = steps.slice(i, i + len);\n\t\t\t\t// A run of one repeated tool is a loop, not a workflow.\n\t\t\t\tif (new Set(window).size < 2) continue;\n\t\t\t\tif (!isProcedure(window)) continue;\n\t\t\t\tconst key = window.join(\" > \");\n\t\t\t\tconst list = positions.get(key);\n\t\t\t\tif (list) list.push(i);\n\t\t\t\telse positions.set(key, [i]);\n\t\t\t}\n\n\t\t\tfor (const [key, occurrences] of positions) {\n\t\t\t\tlet count = 0;\n\t\t\t\tlet nextFree = -1;\n\t\t\t\tfor (const start of occurrences) {\n\t\t\t\t\tif (start < nextFree) continue;\n\t\t\t\t\tcount++;\n\t\t\t\t\tnextFree = start + len;\n\t\t\t\t}\n\n\t\t\t\tconst existing = acc.get(key);\n\t\t\t\tif (existing) {\n\t\t\t\t\texisting.count += count;\n\t\t\t\t\texisting.sessions.add(session.id);\n\t\t\t\t\tif (session.timestamp > existing.lastSeen) existing.lastSeen = session.timestamp;\n\t\t\t\t} else {\n\t\t\t\t\tacc.set(key, {\n\t\t\t\t\t\tsteps: key.split(\" > \"),\n\t\t\t\t\t\tcount,\n\t\t\t\t\t\tsessions: new Set([session.id]),\n\t\t\t\t\t\tlastSeen: session.timestamp,\n\t\t\t\t\t});\n\t\t\t\t}\n\t\t\t}\n\t\t}\n\t}\n\n\tconst ranked = [...acc.values()]\n\t\t.filter((entry) => entry.count >= minRepeats)\n\t\t.map((entry) => ({\n\t\t\tkey: `workflow:${entry.steps.join(\" > \")}`,\n\t\t\tsteps: entry.steps,\n\t\t\tcount: entry.count,\n\t\t\tsessions: entry.sessions.size,\n\t\t\tlastSeen: entry.lastSeen,\n\t\t}))\n\t\t// Sessions first, matching directives: a sequence seen in three sessions is\n\t\t// a workflow, while one repeated ten times in a single session is usually\n\t\t// just the shape of that one task.\n\t\t.sort(\n\t\t\t(a, b) =>\n\t\t\t\tb.sessions - a.sessions ||\n\t\t\t\tb.count - a.count ||\n\t\t\t\tb.steps.length - a.steps.length ||\n\t\t\t\ta.steps.join().localeCompare(b.steps.join()),\n\t\t);\n\n\t// Every n-gram overlaps its own extensions and prefixes, so without this the\n\t// list is one workflow described five slightly different ways. The test runs\n\t// both directions on purpose: a shorter sequence always outranks the longer\n\t// one containing it (it occurs at least as often), so checking only\n\t// shorter-inside-kept would never fire. Keep the best-ranked member of each\n\t// family and drop the rest.\n\tconst distinct: typeof ranked = [];\n\tfor (const candidate of ranked) {\n\t\tconst overlapsKept = distinct.some(\n\t\t\t(kept) => containsSequence(kept.steps, candidate.steps) || containsSequence(candidate.steps, kept.steps),\n\t\t);\n\t\tif (overlapsKept) continue;\n\t\tdistinct.push(candidate);\n\t}\n\treturn distinct;\n}\n\n/**\n * Everything a proposal could already have been written into.\n *\n * Built once and shared, because the same question — is this already written\n * down? — is asked while ranking a run *and* afterwards by `/learn stats`,\n * which reconstructs adoption by comparing coverage now against coverage when\n * the item was shown.\n */\nexport interface CoverageIndex {\n\t/** Candidate rule lines from the repo context file and both user scopes. */\n\truleLines: string[];\n\tskills: Array<{ name: string; description: string }>;\n}\n\nexport interface CoverageMatch {\n\t/** The context-file line that covers this, if any. */\n\trule?: string;\n\t/** The skill that covers this, if any. Only set when no rule matched. */\n\tskill?: string;\n}\n\n/**\n * Where a piece of text is already written down, if anywhere.\n *\n * A rule wins over a skill when both match: it is the more specific answer, and\n * \"rewrite this line\" is more actionable than \"sharpen a description\".\n */\nexport function matchCoverage(text: string, index: CoverageIndex): CoverageMatch {\n\tconst words = contentWords(text);\n\n\tlet bestLine: string | undefined;\n\tlet bestOverlap = 0;\n\tfor (const line of index.ruleLines) {\n\t\tconst overlap = wordOverlap(words, line);\n\t\tif (overlap > bestOverlap) {\n\t\t\tbestOverlap = overlap;\n\t\t\tbestLine = line;\n\t\t}\n\t}\n\tif (bestOverlap >= COVERED_OVERLAP) return { rule: bestLine };\n\n\tlet bestSkill: string | undefined;\n\tlet bestSkillOverlap = 0;\n\tfor (const skill of index.skills) {\n\t\tconst haystack = `${skill.name} ${skill.description.slice(0, SKILL_DESCRIPTION_CHARS)}`;\n\t\tconst overlap = wordOverlap(words, haystack);\n\t\tif (overlap > bestSkillOverlap) {\n\t\t\tbestSkillOverlap = overlap;\n\t\t\tbestSkill = skill.name;\n\t\t}\n\t}\n\tif (bestSkillOverlap >= SKILL_COVERED_OVERLAP) return { skill: bestSkill };\n\n\treturn {};\n}\n\n/** Assemble the coverage index for a directory. */\nexport function buildCoverageIndex(options: {\n\tcwd: string;\n\tagentDir: string;\n\tskills?: Array<{ name: string; description: string }>;\n}): CoverageIndex {\n\tconst corpus = coverageCorpus(options.agentDir, findAgentsFile(options.cwd));\n\treturn {\n\t\truleLines: corpus\n\t\t\t.split(\"\\n\")\n\t\t\t.map((line) => line.trim())\n\t\t\t.filter((line) => line.length > 0 && !line.startsWith(\"#\")),\n\t\tskills: options.skills ?? loadSkillIndex(options.cwd, options.agentDir),\n\t};\n}\n\n/**\n * Text a proposal is checked against to decide whether it is already written\n * down — the nearest repo context file plus both user scopes.\n *\n * All three matter for suppression, because `/learn` can route a rule to the\n * user scope. Checking only the repo file would report a rule you accepted into\n * `~/.agents/AGENTS.md` as declined.\n */\nfunction coverageCorpus(agentDir: string, repoFile: string | undefined): string {\n\tconst parts: string[] = [];\n\tfor (const candidate of [repoFile, join(getUserAgentsDir(), \"AGENTS.md\"), join(agentDir, \"AGENTS.md\")]) {\n\t\tif (!candidate || !existsSync(candidate)) continue;\n\t\ttry {\n\t\t\tparts.push(readFileSync(candidate, \"utf-8\"));\n\t\t} catch {\n\t\t\t// Unreadable context file: treat as absent rather than failing the run.\n\t\t}\n\t}\n\treturn parts.join(\"\\n\");\n}\n\n/**\n * Skills a proposal could already have become.\n *\n * `/learn` routes long or conditional guidance to a skill rather than a rule, so\n * without this a proposal you adopted *as a skill* would read as declined —\n * looking only at context files sees an unchanged `AGENTS.md` and concludes you\n * passed. Reuses the real loader rather than a second SKILL.md scanner so the\n * set of locations cannot drift from what the session actually loads.\n */\nfunction loadSkillIndex(cwd: string, agentDir: string): Array<{ name: string; description: string }> {\n\ttry {\n\t\treturn loadSkills({ cwd, agentDir, skillPaths: [], includeDefaults: true }).skills.map((skill) => ({\n\t\t\tname: skill.name,\n\t\t\tdescription: skill.description ?? \"\",\n\t\t}));\n\t} catch {\n\t\t// Skills are an enrichment here, not the point of the command.\n\t\treturn [];\n\t}\n}\n\n/** Mine the recent sessions for this cwd and return the ranked digest. */\nexport function extractLearnDigest(options: ExtractOptions): LearnDigest {\n\tconst { sessions, skipped, scan } = listSessions(options);\n\n\tconst agentsFilePath = findAgentsFile(options.cwd);\n\tlet agentsContent: string | undefined;\n\tif (agentsFilePath) {\n\t\ttry {\n\t\t\tagentsContent = readFileSync(agentsFilePath, \"utf-8\");\n\t\t} catch {\n\t\t\tagentsContent = undefined;\n\t\t}\n\t}\n\tconst coverage = buildCoverageIndex({ cwd: options.cwd, agentDir: options.agentDir, skills: options.skills });\n\n\tconst withDirectives = sessions.map((session) => ({ session, directives: userDirectives(session.entries) }));\n\tconst withEvents = sessions.map((session) => ({ session, events: toolEvents(session.entries) }));\n\n\tconst timestamps = sessions.map((s) => s.timestamp).sort();\n\tconst state = options.ignoreState ? undefined : options.state;\n\n\t// Directives carry a real coverage signal — is this written down as a rule or\n\t// a skill right now? — which is what separates an adopted proposal from a\n\t// declined one. Fixes and workflows do not: a fix may have become a rule, a\n\t// skill, or a habit, and which one is not recoverable here, so they get\n\t// suppression only and are never labelled declined.\n\tconst maxProposals = options.maxProposals ?? DEFAULT_MAX_PER_CATEGORY;\n\tconst directives = applySuppression(\n\t\tclusterDirectives(withDirectives, coverage, options.minRepeats ?? DEFAULT_MIN_DIRECTIVE_COUNT),\n\t\tstate,\n\t\tmaxProposals,\n\t\t(item) => item.status !== \"new\",\n\t\t(item) => {\n\t\t\titem.previouslyDeclined = true;\n\t\t},\n\t);\n\tconst fixes = applySuppression(extractFixes(withEvents), state, maxProposals, () => false);\n\tconst workflows = applySuppression(\n\t\textractWorkflows(withEvents, options.minWorkflowRepeats ?? DEFAULT_MIN_WORKFLOW_COUNT),\n\t\tstate,\n\t\tmaxProposals,\n\t\t() => false,\n\t);\n\n\tconst surfaced = [\n\t\t...directives.kept.map((d) => ({ key: d.key, lastSeen: d.lastSeen, covered: d.status !== \"new\" })),\n\t\t...fixes.kept.map((f) => ({ key: f.key, lastSeen: f.lastSeen, covered: false })),\n\t\t...workflows.kept.map((w) => ({ key: w.key, lastSeen: w.lastSeen, covered: false })),\n\t];\n\n\treturn {\n\t\tscannedSessions: sessions.length,\n\t\tskippedSessions: skipped,\n\t\tscan,\n\t\toldestSession: timestamps[0],\n\t\tnewestSession: timestamps[timestamps.length - 1],\n\t\tagentsFilePath,\n\t\tagentsFileTokens:\n\t\t\tagentsContent === undefined ? undefined : Math.round(Buffer.byteLength(agentsContent, \"utf-8\") / 4),\n\t\tdirectives: directives.kept,\n\t\tfixes: fixes.kept,\n\t\tworkflows: workflows.kept,\n\t\tsuppressed: directives.suppressed + fixes.suppressed + workflows.suppressed,\n\t\tsurfaced,\n\t};\n}\n"]}

package/dist/core/learn/extract.js

@@ -14,10 +14,10 @@
  * digest, which is why the output carries evidence (counts, sessions, dates)
  * rather than conclusions.
  */
-import { existsSync, readdirSync, readFileSync, statSync } from "node:fs";
-import { dirname, join, resolve } from "node:path";
+import { existsSync, readdirSync, readFileSync, realpathSync, statSync } from "node:fs";
+import { dirname, join, resolve, sep } from "node:path";
 import { getUserAgentsDir } from "../../config.js";
-import { getDefaultSessionDir } from "../session-manager.js";
+import { getSessionDirPath } from "../session-manager.js";
 import { loadSkills } from "../skills.js";
 import { commandHead, contentWords, extractErrorRegion, isBenignFailure, isRuleShapedDirective, isUninformativeFailure, normalizeCommand, normalizeDirective, normalizeErrorSignature, wordOverlap, } from "./normalize.js";
 import { judge } from "./state.js";
@@ -101,12 +101,40 @@ function activeBranch(entries) {
     }
     return branch.reverse();
 }
-function parseSessionFile(file, cwd) {
+/**
+ * Compare two directory paths the way the filesystem does.
+ *
+ * A session header stores the cwd as it was typed, and the same directory can
+ * be spelled several ways: through a symlink (`/tmp` is `/private/tmp` on
+ * macOS), with a trailing separator, or in different case on the
+ * case-insensitive filesystems that macOS and Windows ship by default. String
+ * equality on `resolve()` alone rejects every one of those, and rejecting them
+ * here means silently discarding the whole history the command exists to read.
+ */
+function normalizeDirPath(path) {
+    let resolved = resolve(path);
+    try {
+        resolved = realpathSync.native(resolved);
+    }
+    catch {
+        // Deleted or never-created directory: the textual form is all we have.
+    }
+    // `resolve` already drops a trailing separator except at a filesystem root,
+    // where dropping it would turn "/" into "".
+    if (resolved.length > 1 && resolved.endsWith(sep))
+        resolved = resolved.slice(0, -1);
+    return process.platform === "win32" || process.platform === "darwin" ? resolved.toLowerCase() : resolved;
+}
+function sameDirectory(a, b) {
+    return normalizeDirPath(a) === normalizeDirPath(b);
+}
+function parseSessionFile(file, cwd, onSkip) {
     let raw;
     try {
         raw = readFileSync(file, "utf-8");
     }
     catch {
+        onSkip("unreadable");
         return undefined;
     }
     const lines = raw.split("\n");
@@ -133,10 +161,14 @@ function parseSessionFile(file, cwd) {
     }
     // An explicit `--session` path can put a session for another directory in
     // this directory, so trust the header over the file's location.
-    if (header?.cwd && resolve(header.cwd) !== resolve(cwd))
+    if (header?.cwd && !sameDirectory(header.cwd, cwd)) {
+        onSkip("otherCwd");
         return undefined;
-    if (entries.length === 0)
+    }
+    if (entries.length === 0) {
+        onSkip("unreadable");
         return undefined;
+    }
     return {
         file,
         id: header?.id ?? file,
@@ -144,52 +176,109 @@ function parseSessionFile(file, cwd) {
         entries: activeBranch(entries),
     };
 }
+/**
+ * Every directory this cwd's sessions could be sitting in.
+ *
+ * The caller passes the live session manager's directory, which is the right
+ * answer almost always — but not quite always, and each exception silently
+ * emptied the digest. An in-memory session (`--no-session`) reports `""`; an
+ * explicit `--session <path>` reports wherever that file lives; a custom
+ * `sessionDir` setting points at one shared directory. In every one of those
+ * cases the per-cwd default directory still holds the history worth mining, so
+ * search both and let the header check sort out what belongs to this cwd.
+ */
+export function candidateSessionDirs(options) {
+    const dirs = [];
+    const seen = new Set();
+    for (const dir of [options.sessionDir, getSessionDirPath(options.cwd, options.agentDir)]) {
+        if (!dir)
+            continue;
+        const key = normalizeDirPath(dir);
+        if (seen.has(key))
+            continue;
+        seen.add(key);
+        dirs.push(dir);
+    }
+    return dirs;
+}
 function listSessions(options) {
-    const dir = options.sessionDir ?? getDefaultSessionDir(options.cwd, options.agentDir);
-    if (!existsSync(dir))
-        return { sessions: [], skipped: 0 };
+    const dirs = candidateSessionDirs(options);
+    const scan = {
+        dirs,
+        missingDirs: [],
+        files: 0,
+        tooOld: 0,
+        otherCwd: 0,
+        overLimit: 0,
+        unreadable: 0,
+    };
     const maxSessions = options.maxSessions ?? DEFAULT_MAX_SESSIONS;
     const maxAgeDays = options.maxAgeDays ?? DEFAULT_MAX_AGE_DAYS;
     const now = options.now ?? new Date();
     const cutoff = now.getTime() - maxAgeDays * 24 * 60 * 60 * 1000;
-    let files;
-    try {
-        files = readdirSync(dir)
-            .filter((f) => f.endsWith(".jsonl"))
-            .map((f) => join(dir, f));
-    }
-    catch {
-        return { sessions: [], skipped: 0 };
+    const files = [];
+    for (const dir of dirs) {
+        if (!existsSync(dir)) {
+            scan.missingDirs.push(dir);
+            continue;
+        }
+        try {
+            for (const name of readdirSync(dir)) {
+                if (name.endsWith(".jsonl"))
+                    files.push(join(dir, name));
+            }
+        }
+        catch {
+            scan.missingDirs.push(dir);
+        }
     }
+    scan.files = files.length;
+    // Newest first across all directories, so `maxSessions` keeps the most recent
+    // history rather than whichever directory happened to be searched first.
     const dated = files
         .map((file) => {
         try {
             return { file, mtime: statSync(file).mtime.getTime() };
         }
         catch {
+            scan.unreadable++;
             return undefined;
         }
     })
         .filter((f) => !!f)
         .sort((a, b) => b.mtime - a.mtime);
     const sessions = [];
+    const seenIds = new Set();
     let skipped = 0;
     for (const { file, mtime } of dated) {
         if (sessions.length >= maxSessions) {
+            scan.overLimit++;
             skipped++;
             continue;
         }
         if (mtime < cutoff) {
+            scan.tooOld++;
             skipped++;
             continue;
         }
-        const parsed = parseSessionFile(file, options.cwd);
-        if (parsed)
-            sessions.push(parsed);
-        else
+        const parsed = parseSessionFile(file, options.cwd, (reason) => {
+            scan[reason]++;
+        });
+        if (!parsed) {
             skipped++;
+            continue;
+        }
+        // Searching two directories can turn up the same session twice (an explicit
+        // `--session` path inside the default directory). Counting it twice would
+        // inflate the cross-session repetition that decides what gets proposed.
+        if (seenIds.has(parsed.id)) {
+            skipped++;
+            continue;
+        }
+        seenIds.add(parsed.id);
+        sessions.push(parsed);
     }
-    return { sessions, skipped };
+    return { sessions, skipped, scan };
 }
 /**
  * Hold back items already shown that have not recurred since, then cap the rest.
@@ -215,6 +304,13 @@ function applySuppression(items, state, maxProposals, covered, onDeclined) {
     }
     return { kept: kept.slice(0, maxProposals), suppressed };
 }
+/**
+ * Where this cwd's sessions were found and what was passed over, without
+ * ranking anything. `/learn stats` reports on the window without re-mining it.
+ */
+export function scanSessions(options) {
+    return listSessions(options).scan;
+}
 /** Nearest AGENTS.md walking up from cwd, so proposals can be checked against it. */
 function findAgentsFile(cwd) {
     let dir = resolve(cwd);
@@ -656,7 +752,7 @@ function loadSkillIndex(cwd, agentDir) {
 }
 /** Mine the recent sessions for this cwd and return the ranked digest. */
 export function extractLearnDigest(options) {
-    const { sessions, skipped } = listSessions(options);
+    const { sessions, skipped, scan } = listSessions(options);
     const agentsFilePath = findAgentsFile(options.cwd);
     let agentsContent;
     if (agentsFilePath) {
@@ -691,6 +787,7 @@ export function extractLearnDigest(options) {
     return {
         scannedSessions: sessions.length,
         skippedSessions: skipped,
+        scan,
         oldestSession: timestamps[0],
         newestSession: timestamps[timestamps.length - 1],
         agentsFilePath,