@kolisachint/hoocode-agent 0.5.16 → 0.5.18

package/dist/core/learn/digest.js

@@ -31,6 +31,19 @@ export function renderLearnDigest(digest, options) {
         (digest.oldestSession ? `, ${shortDate(digest.oldestSession)} to ${shortDate(digest.newestSession)}` : "") +
         (digest.suppressed > 0 ? `. ${digest.suppressed} item(s) held back — already shown and unchanged since` : "") +
         ".");
+    // Naming the mode keeps two very different empty results from reading alike:
+    // "nothing new since last time" and "nothing here at all" are not the same
+    // answer, and the reader cannot tell them apart from the counts.
+    if (options.mode === "all") {
+        lines.push("Mode: all — suppression is off, so items you have already seen and decided on are included.");
+    }
+    // The model reads every transcript in full, which costs real tokens. Saying
+    // what was re-read versus reused keeps that price visible rather than hidden.
+    lines.push(`Read by the model this run: ${digest.mining.mined}; reused from cache: ${digest.mining.cached}` +
+        (digest.mining.failed > 0
+            ? `; failed: ${digest.mining.failed} (their signals are missing from the counts below)`
+            : "") +
+        ".");
     lines.push("");
     lines.push("The counts below are computed from session transcripts on disk, not from this conversation. " +
         "Treat them as evidence, not conclusions — your job is to decide what deserves to be written down, " +
@@ -43,6 +56,13 @@ export function renderLearnDigest(digest, options) {
         for (const cluster of digest.directives) {
             lines.push(`- **${cluster.status}** — "${cluster.text.replace(/\s+/g, " ").trim()}"`);
             lines.push(`  - ${evidence(cluster.count, cluster.sessions, cluster.lastSeen)}`);
+            // Occurrences were grouped by meaning, not by wording, so the quote above
+            // is one phrasing of several. Naming the shared point keeps a count of 5
+            // from looking like five copies of one sentence.
+            lines.push(`  - grouped as: ${cluster.label}`);
+            if (cluster.rationale) {
+                lines.push(`  - why it may be durable: ${cluster.rationale}`);
+            }
             if (cluster.existingRule) {
                 lines.push(`  - already covered by: "${cluster.existingRule.slice(0, 160)}"`);
             }
@@ -59,12 +79,16 @@ export function renderLearnDigest(digest, options) {
     if (digest.fixes.length > 0) {
         lines.push("## Failures you resolved");
         lines.push("");
-        lines.push("Each is a command that failed, then later succeeded unchanged after intervening work — " +
-            "so something in between was the fix.");
+        lines.push("Each is a command that failed and later succeeded, where something done in between was the fix. " +
+            "Recurring ones are worth writing down; a one-off is not.");
         lines.push("");
         for (const fix of digest.fixes) {
             lines.push(`- \`${fix.command}\` — ${evidence(fix.count, fix.sessions, fix.lastSeen)}`);
-            lines.push(`  - error: ${fix.errorExcerpt}`);
+            lines.push(`  - grouped as: ${fix.label}`);
+            // The excerpt comes from the model now, which may not have quoted one.
+            if (fix.errorExcerpt) {
+                lines.push(`  - error: ${fix.errorExcerpt}`);
+            }
             if (fix.interveningCommands.length > 0) {
                 lines.push(`  - commands in between: ${fix.interveningCommands.map((c) => `\`${c}\``).join(", ")}`);
             }
@@ -79,7 +103,10 @@ export function renderLearnDigest(digest, options) {
         lines.push("## Repeated tool sequences");
         lines.push("");
         for (const workflow of digest.workflows) {
-            lines.push(`- \`${workflow.steps.join(" → ")}\` — ${evidence(workflow.count, workflow.sessions, workflow.lastSeen)}`);
+            // A workflow the model named but did not enumerate still has a label
+            // worth showing; rendering an empty backtick pair instead would not.
+            const steps = workflow.steps.length > 0 ? `\`${workflow.steps.join(" → ")}\`` : workflow.label;
+            lines.push(`- ${steps} — ${evidence(workflow.count, workflow.sessions, workflow.lastSeen)}`);
         }
         lines.push("");
     }

package/dist/core/learn/digest.js.map

@@ -1 +1 @@
-{"version":3,"file":"digest.js","sourceRoot":"","sources":["../../../src/core/learn/digest.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAGH,OAAO,EAAE,mBAAmB,EAAE,MAAM,cAAc,CAAC;AAEnD,SAAS,SAAS,CAAC,GAAuB,EAAU;IACnD,IAAI,CAAC,GAAG;QAAE,OAAO,SAAS,CAAC;IAC3B,MAAM,IAAI,GAAG,IAAI,IAAI,CAAC,GAAG,CAAC,CAAC;IAC3B,OAAO,MAAM,CAAC,KAAK,CAAC,IAAI,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,IAAI,CAAC,WAAW,EAAE,CAAC,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;AAAA,CAClF;AAED,SAAS,QAAQ,CAAC,KAAa,EAAE,QAAgB,EAAE,QAAgB,EAAU;IAC5E,MAAM,KAAK,GAAG,KAAK,KAAK,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,GAAG,KAAK,GAAG,CAAC;IACjD,MAAM,KAAK,GAAG,QAAQ,KAAK,CAAC,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,CAAC,GAAG,QAAQ,WAAW,CAAC;IACpE,OAAO,GAAG,KAAK,WAAW,KAAK,UAAU,SAAS,CAAC,QAAQ,CAAC,EAAE,CAAC;AAAA,CAC/D;AAED,oEAAoE;AACpE,MAAM,UAAU,aAAa,CAAC,MAAmB,EAAW;IAC3D,OAAO,MAAM,CAAC,UAAU,CAAC,MAAM,KAAK,CAAC,IAAI,MAAM,CAAC,KAAK,CAAC,MAAM,KAAK,CAAC,IAAI,MAAM,CAAC,SAAS,CAAC,MAAM,KAAK,CAAC,CAAC;AAAA,CACpG;AAED,MAAM,UAAU,iBAAiB,CAAC,MAAmB,EAAE,OAAkC,EAAU;IAClG,MAAM,KAAK,GAAa,EAAE,CAAC;IAE3B,KAAK,CAAC,IAAI,CACT,GAAG,mBAAmB,UAAU,MAAM,CAAC,eAAe,+BAA+B;QACpF,CAAC,MAAM,CAAC,eAAe,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,MAAM,CAAC,eAAe,wCAAwC,CAAC,CAAC,CAAC,EAAE,CAAC;QACvG,CAAC,MAAM,CAAC,aAAa,CAAC,CAAC,CAAC,KAAK,SAAS,CAAC,MAAM,CAAC,aAAa,CAAC,OAAO,SAAS,CAAC,MAAM,CAAC,aAAa,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QAC1G,CAAC,MAAM,CAAC,UAAU,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,MAAM,CAAC,UAAU,0DAAwD,CAAC,CAAC,CAAC,EAAE,CAAC;QAC7G,GAAG,CACJ,CAAC;IACF,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACf,KAAK,CAAC,IAAI,CACT,8FAA8F;QAC7F,sGAAoG;QACpG,2CAA2C,CAC5C,CAAC;IACF,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IAEf,sMAA4E;IAC5E,IAAI,MAAM,CAAC,UAAU,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAClC,KAAK,CAAC,IAAI,CAAC,iCAAiC,CAAC,CAAC;QAC9C,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;QACf,KAAK,MAAM,OAAO,IAAI,MAAM,CAAC,UAAU,EAAE,CAAC;YACzC,KAAK,CAAC,IAAI,CAAC,OAAO,OAAO,CAAC,MAAM,WAAS,OAAO,CAAC,IAAI,CAAC,OAAO,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC,IAAI,EAAE,GAAG,CAAC,CAAC;YACtF,KAAK,CAAC,IAAI,CAAC,OAAO,QAAQ,CAAC,OAAO,CAAC,KAAK,EAAE,OAAO,CAAC,QAAQ,EAAE,OAAO,CAAC,QAAQ,CAAC,EAAE,CAAC,CAAC;YACjF,IAAI,OAAO,CAAC,YAAY,EAAE,CAAC;gBAC1B,KAAK,CAAC,IAAI,CAAC,4BAA4B,OAAO,CAAC,YAAY,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,GAAG,CAAC,CAAC;YAC/E,CAAC;YACD,IAAI,OAAO,CAAC,aAAa,EAAE,CAAC;gBAC3B,KAAK,CAAC,IAAI,CAAC,gCAAgC,OAAO,CAAC,aAAa,UAAU,CAAC,CAAC;YAC7E,CAAC;YACD,IAAI,OAAO,CAAC,kBAAkB,EAAE,CAAC;gBAChC,KAAK,CAAC,IAAI,CAAC,mFAAiF,CAAC,CAAC;YAC/F,CAAC;QACF,CAAC;QACD,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IAChB,CAAC;IAED,gNAA4E;IAC5E,IAAI,MAAM,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAC7B,KAAK,CAAC,IAAI,CAAC,0BAA0B,CAAC,CAAC;QACvC,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;QACf,KAAK,CAAC,IAAI,CACT,2FAAyF;YACxF,sCAAsC,CACvC,CAAC;QACF,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;QACf,KAAK,MAAM,GAAG,IAAI,MAAM,CAAC,KAAK,EAAE,CAAC;YAChC,KAAK,CAAC,IAAI,CAAC,OAAO,GAAG,CAAC,OAAO,UAAQ,QAAQ,CAAC,GAAG,CAAC,KAAK,EAAE,GAAG,CAAC,QAAQ,EAAE,GAAG,CAAC,QAAQ,CAAC,EAAE,CAAC,CAAC;YACxF,KAAK,CAAC,IAAI,CAAC,cAAc,GAAG,CAAC,YAAY,EAAE,CAAC,CAAC;YAC7C,IAAI,GAAG,CAAC,mBAAmB,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;gBACxC,KAAK,CAAC,IAAI,CAAC,4BAA4B,GAAG,CAAC,mBAAmB,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;YACrG,CAAC;YACD,IAAI,GAAG,CAAC,WAAW,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;gBAChC,KAAK,CAAC,IAAI,CAAC,qBAAqB,GAAG,CAAC,WAAW,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;YAC/D,CAAC;QACF,CAAC;QACD,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IAChB,CAAC;IAED,wMAA4E;IAC5E,IAAI,MAAM,CAAC,SAAS,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QACjC,KAAK,CAAC,IAAI,CAAC,4BAA4B,CAAC,CAAC;QACzC,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;QACf,KAAK,MAAM,QAAQ,IAAI,MAAM,CAAC,SAAS,EAAE,CAAC;YACzC,KAAK,CAAC,IAAI,CACT,OAAO,QAAQ,CAAC,KAAK,CAAC,IAAI,CAAC,OAAK,CAAC,UAAQ,QAAQ,CAAC,QAAQ,CAAC,KAAK,EAAE,QAAQ,CAAC,QAAQ,EAAE,QAAQ,CAAC,QAAQ,CAAC,EAAE,CACzG,CAAC;QACH,CAAC;QACD,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IAChB,CAAC;IAED,kMAA4E;IAC5E,KAAK,CAAC,IAAI,CAAC,eAAe,CAAC,CAAC;IAC5B,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACf,KAAK,CAAC,IAAI,CAAC,gFAAgF,CAAC,CAAC;IAC7F,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACf,KAAK,CAAC,IAAI,CACT,0GAA0G;QACzG,4GAA0G;QAC1G,6DAA6D,CAC9D,CAAC;IACF,KAAK,CAAC,IAAI,CACT,6GAA6G;QAC5G,wGAAsG;QACtG,6GAA2G;QAC3G,mDAAmD,CACpD,CAAC;IACF,KAAK,CAAC,IAAI,CACT,0GAAwG;QACvG,yGAAyG;QACzG,0BAAwB,OAAO,CAAC,aAAa,6DAA6D;QAC1G,kBAAkB,CACnB,CAAC;IACF,KAAK,CAAC,IAAI,CACT,4GAA4G;QAC3G,4GAA0G;QAC1G,gFAAgF,CACjF,CAAC;IACF,KAAK,CAAC,IAAI,CACT,2GAA2G;QAC1G,wGAAwG;QACxG,2GAA2G;QAC3G,8BAA8B,CAC/B,CAAC;IACF,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACf,KAAK,CAAC,IAAI,CAAC,+CAA+C,CAAC,CAAC;IAC5D,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACf,KAAK,CAAC,IAAI,CACT,yGAAyG,CACzG,CAAC;IACF,KAAK,CAAC,IAAI,CAAC,qGAAqG,CAAC,CAAC;IAClH,KAAK,CAAC,IAAI,CACT,8GAA4G,CAC5G,CAAC;IACF,KAAK,CAAC,IAAI,CACT,8GAA8G;QAC7G,kDAAkD,CACnD,CAAC;IACF,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IAEf,IAAI,MAAM,CAAC,cAAc,EAAE,CAAC;QAC3B,KAAK,CAAC,IAAI,CACT,8BAA8B,MAAM,CAAC,cAAc,IAAI;YACtD,CAAC,MAAM,CAAC,gBAAgB,CAAC,CAAC,CAAC,MAAM,MAAM,CAAC,gBAAgB,iCAAiC,CAAC,CAAC,CAAC,EAAE,CAAC;YAC/F,4GAA4G,CAC7G,CAAC;IACH,CAAC;SAAM,CAAC;QACP,KAAK,CAAC,IAAI,CACT,wGAAwG,CACxG,CAAC;IACH,CAAC;IACD,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACf,KAAK,CAAC,IAAI,CACT,8GAA4G;QAC3G,mGAAmG,CACpG,CAAC;IAEF,OAAO,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAAA,CACxB","sourcesContent":["/**\n * Renders the extractor's output into the message `/learn` injects.\n *\n * The digest is evidence plus instructions, and the split matters: the numbers\n * come from {@link extractLearnDigest} and are not negotiable, while everything\n * the model does with them — phrasing, routing, deciding a pattern is not worth\n * a rule — is judgement it has to exercise. Counts are printed on every item\n * because \"said in 5 of your last 12 sessions\" is a decision the reader can\n * make in one keystroke, where \"extracted from your session\" is not.\n */\n\nimport type { LearnDigest } from \"./extract.js\";\nimport { LEARN_DIGEST_MARKER } from \"./extract.js\";\n\nfunction shortDate(iso: string | undefined): string {\n\tif (!iso) return \"unknown\";\n\tconst date = new Date(iso);\n\treturn Number.isNaN(date.getTime()) ? \"unknown\" : date.toISOString().slice(0, 10);\n}\n\nfunction evidence(count: number, sessions: number, lastSeen: string): string {\n\tconst times = count === 1 ? \"once\" : `${count}x`;\n\tconst where = sessions === 1 ? \"1 session\" : `${sessions} sessions`;\n\treturn `${times} across ${where}, last ${shortDate(lastSeen)}`;\n}\n\n/** True when there is nothing worth asking the model to look at. */\nexport function isEmptyDigest(digest: LearnDigest): boolean {\n\treturn digest.directives.length === 0 && digest.fixes.length === 0 && digest.workflows.length === 0;\n}\n\nexport function renderLearnDigest(digest: LearnDigest, options: { userScopePath: string }): string {\n\tconst lines: string[] = [];\n\n\tlines.push(\n\t\t`${LEARN_DIGEST_MARKER} Mined ${digest.scannedSessions} session(s) in this directory` +\n\t\t\t(digest.skippedSessions > 0 ? ` (${digest.skippedSessions} skipped: out of window or unreadable)` : \"\") +\n\t\t\t(digest.oldestSession ? `, ${shortDate(digest.oldestSession)} to ${shortDate(digest.newestSession)}` : \"\") +\n\t\t\t(digest.suppressed > 0 ? `. ${digest.suppressed} item(s) held back — already shown and unchanged since` : \"\") +\n\t\t\t\".\",\n\t);\n\tlines.push(\"\");\n\tlines.push(\n\t\t\"The counts below are computed from session transcripts on disk, not from this conversation. \" +\n\t\t\t\"Treat them as evidence, not conclusions — your job is to decide what deserves to be written down, \" +\n\t\t\t\"phrase it, and put it in the right place.\",\n\t);\n\tlines.push(\"\");\n\n\t// ── Directives ───────────────────────────────────────────────────────────\n\tif (digest.directives.length > 0) {\n\t\tlines.push(\"## Directives you have repeated\");\n\t\tlines.push(\"\");\n\t\tfor (const cluster of digest.directives) {\n\t\t\tlines.push(`- **${cluster.status}** — \"${cluster.text.replace(/\\s+/g, \" \").trim()}\"`);\n\t\t\tlines.push(`  - ${evidence(cluster.count, cluster.sessions, cluster.lastSeen)}`);\n\t\t\tif (cluster.existingRule) {\n\t\t\t\tlines.push(`  - already covered by: \"${cluster.existingRule.slice(0, 160)}\"`);\n\t\t\t}\n\t\t\tif (cluster.existingSkill) {\n\t\t\t\tlines.push(`  - already covered by the \\`${cluster.existingSkill}\\` skill`);\n\t\t\t}\n\t\t\tif (cluster.previouslyDeclined) {\n\t\t\t\tlines.push(\"  - proposed before and not written down — you have already passed on this once\");\n\t\t\t}\n\t\t}\n\t\tlines.push(\"\");\n\t}\n\n\t// ── Fixes ────────────────────────────────────────────────────────────────\n\tif (digest.fixes.length > 0) {\n\t\tlines.push(\"## Failures you resolved\");\n\t\tlines.push(\"\");\n\t\tlines.push(\n\t\t\t\"Each is a command that failed, then later succeeded unchanged after intervening work — \" +\n\t\t\t\t\"so something in between was the fix.\",\n\t\t);\n\t\tlines.push(\"\");\n\t\tfor (const fix of digest.fixes) {\n\t\t\tlines.push(`- \\`${fix.command}\\` — ${evidence(fix.count, fix.sessions, fix.lastSeen)}`);\n\t\t\tlines.push(`  - error: ${fix.errorExcerpt}`);\n\t\t\tif (fix.interveningCommands.length > 0) {\n\t\t\t\tlines.push(`  - commands in between: ${fix.interveningCommands.map((c) => `\\`${c}\\``).join(\", \")}`);\n\t\t\t}\n\t\t\tif (fix.editedFiles.length > 0) {\n\t\t\t\tlines.push(`  - files edited: ${fix.editedFiles.join(\", \")}`);\n\t\t\t}\n\t\t}\n\t\tlines.push(\"\");\n\t}\n\n\t// ── Workflows ────────────────────────────────────────────────────────────\n\tif (digest.workflows.length > 0) {\n\t\tlines.push(\"## Repeated tool sequences\");\n\t\tlines.push(\"\");\n\t\tfor (const workflow of digest.workflows) {\n\t\t\tlines.push(\n\t\t\t\t`- \\`${workflow.steps.join(\" → \")}\\` — ${evidence(workflow.count, workflow.sessions, workflow.lastSeen)}`,\n\t\t\t);\n\t\t}\n\t\tlines.push(\"\");\n\t}\n\n\t// ── Instructions ─────────────────────────────────────────────────────────\n\tlines.push(\"## What to do\");\n\tlines.push(\"\");\n\tlines.push(\"Work through the items above and propose concrete edits. For each one, decide:\");\n\tlines.push(\"\");\n\tlines.push(\n\t\t\"1. **Is it durable?** A rule that will still be true next month belongs somewhere. A one-off preference \" +\n\t\t\t\"about the task you happened to be doing does not. When in doubt, drop it — a wrong rule costs more than \" +\n\t\t\t\"a missing one, because it is paid on every request forever.\",\n\t);\n\tlines.push(\n\t\t\"2. **Rule or skill?** This is the most important call. A context file is loaded on **every** turn; a skill \" +\n\t\t\t\"is loaded **on demand**. So: short, always-true, unconditional → a one-line rule. Long, procedural, \" +\n\t\t\t'or conditional (a sequence of steps, a runbook, anything starting \"when X, do Y\") → a skill, not a rule. ' +\n\t\t\t\"Repeated tool sequences are almost always skills.\",\n\t);\n\tlines.push(\n\t\t`3. **Which scope?** Project-specific (this repo's tests, build, architecture, conventions) → the repo ` +\n\t\t\t`\\`AGENTS.md\\`. Personal habits that travel with you across every repo (style preferences, how you like ` +\n\t\t\t`commits written) → \\`${options.userScopePath}\\`. If it names this repo's files or commands, it is not a ` +\n\t\t\t`user-scope rule.`,\n\t);\n\tlines.push(\n\t\t\"4. **Restated items are rewrites, not additions.** An item marked `restated` is already covered by a rule \" +\n\t\t\t\"that is not working — too vague, buried, or contradicted elsewhere. Rewrite the existing line or delete \" +\n\t\t\t\"it in favour of a sharper one. Do not add a second rule saying the same thing.\",\n\t);\n\tlines.push(\n\t\t\"5. **`has-skill` items are a triggering problem, not a missing rule.** A skill already covers it and you \" +\n\t\t\t\"asked by hand anyway, which usually means the skill's `description` frontmatter does not describe the \" +\n\t\t\t\"situation you were in. Sharpen that description so it matches, rather than adding a rule that duplicates \" +\n\t\t\t\"what the skill already does.\",\n\t);\n\tlines.push(\"\");\n\tlines.push(\"Then, while you have the file open, audit it:\");\n\tlines.push(\"\");\n\tlines.push(\n\t\t\"- **Delete rules that no longer match the code.** Check a sample against the repo before trusting them.\",\n\t);\n\tlines.push(\"- **Delete rules that restate default behaviour.** Guidance the agent already follows is pure cost.\");\n\tlines.push(\n\t\t\"- **Collapse duplicates**, including any rule stated at both repo and user scope — that one is paid twice.\",\n\t);\n\tlines.push(\n\t\t\"- **One line per rule.** No rationale, no examples, no preamble, unless the example *is* the rule. Prose is \" +\n\t\t\t\"the single biggest source of context-file bloat.\",\n\t);\n\tlines.push(\"\");\n\n\tif (digest.agentsFilePath) {\n\t\tlines.push(\n\t\t\t`The repo context file is \\`${digest.agentsFilePath}\\`` +\n\t\t\t\t(digest.agentsFileTokens ? ` (~${digest.agentsFileTokens} tokens, re-sent every request)` : \"\") +\n\t\t\t\t\". Report the token delta of your proposed changes before applying them; a net reduction is a good outcome.\",\n\t\t);\n\t} else {\n\t\tlines.push(\n\t\t\t\"No repo context file exists yet. Create one only if at least one durable project rule survives step 1.\",\n\t\t);\n\t}\n\tlines.push(\"\");\n\tlines.push(\n\t\t\"Show what you propose, then apply it with edits — do not ask a separate approval question first, the edit \" +\n\t\t\t\"prompt is the approval. If nothing here is worth writing down, say so plainly and change nothing.\",\n\t);\n\n\treturn lines.join(\"\\n\");\n}\n"]}
+{"version":3,"file":"digest.js","sourceRoot":"","sources":["../../../src/core/learn/digest.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAGH,OAAO,EAAE,mBAAmB,EAAE,MAAM,cAAc,CAAC;AAEnD,SAAS,SAAS,CAAC,GAAuB,EAAU;IACnD,IAAI,CAAC,GAAG;QAAE,OAAO,SAAS,CAAC;IAC3B,MAAM,IAAI,GAAG,IAAI,IAAI,CAAC,GAAG,CAAC,CAAC;IAC3B,OAAO,MAAM,CAAC,KAAK,CAAC,IAAI,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,IAAI,CAAC,WAAW,EAAE,CAAC,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;AAAA,CAClF;AAED,SAAS,QAAQ,CAAC,KAAa,EAAE,QAAgB,EAAE,QAAgB,EAAU;IAC5E,MAAM,KAAK,GAAG,KAAK,KAAK,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,GAAG,KAAK,GAAG,CAAC;IACjD,MAAM,KAAK,GAAG,QAAQ,KAAK,CAAC,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,CAAC,GAAG,QAAQ,WAAW,CAAC;IACpE,OAAO,GAAG,KAAK,WAAW,KAAK,UAAU,SAAS,CAAC,QAAQ,CAAC,EAAE,CAAC;AAAA,CAC/D;AAED,oEAAoE;AACpE,MAAM,UAAU,aAAa,CAAC,MAAmB,EAAW;IAC3D,OAAO,MAAM,CAAC,UAAU,CAAC,MAAM,KAAK,CAAC,IAAI,MAAM,CAAC,KAAK,CAAC,MAAM,KAAK,CAAC,IAAI,MAAM,CAAC,SAAS,CAAC,MAAM,KAAK,CAAC,CAAC;AAAA,CACpG;AAED,MAAM,UAAU,iBAAiB,CAChC,MAAmB,EACnB,OAAgE,EACvD;IACT,MAAM,KAAK,GAAa,EAAE,CAAC;IAE3B,KAAK,CAAC,IAAI,CACT,GAAG,mBAAmB,UAAU,MAAM,CAAC,eAAe,+BAA+B;QACpF,CAAC,MAAM,CAAC,eAAe,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,MAAM,CAAC,eAAe,wCAAwC,CAAC,CAAC,CAAC,EAAE,CAAC;QACvG,CAAC,MAAM,CAAC,aAAa,CAAC,CAAC,CAAC,KAAK,SAAS,CAAC,MAAM,CAAC,aAAa,CAAC,OAAO,SAAS,CAAC,MAAM,CAAC,aAAa,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QAC1G,CAAC,MAAM,CAAC,UAAU,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,MAAM,CAAC,UAAU,0DAAwD,CAAC,CAAC,CAAC,EAAE,CAAC;QAC7G,GAAG,CACJ,CAAC;IACF,6EAA6E;IAC7E,2EAA2E;IAC3E,iEAAiE;IACjE,IAAI,OAAO,CAAC,IAAI,KAAK,KAAK,EAAE,CAAC;QAC5B,KAAK,CAAC,IAAI,CAAC,+FAA6F,CAAC,CAAC;IAC3G,CAAC;IACD,4EAA4E;IAC5E,8EAA8E;IAC9E,KAAK,CAAC,IAAI,CACT,+BAA+B,MAAM,CAAC,MAAM,CAAC,KAAK,wBAAwB,MAAM,CAAC,MAAM,CAAC,MAAM,EAAE;QAC/F,CAAC,MAAM,CAAC,MAAM,CAAC,MAAM,GAAG,CAAC;YACxB,CAAC,CAAC,aAAa,MAAM,CAAC,MAAM,CAAC,MAAM,oDAAoD;YACvF,CAAC,CAAC,EAAE,CAAC;QACN,GAAG,CACJ,CAAC;IACF,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACf,KAAK,CAAC,IAAI,CACT,8FAA8F;QAC7F,sGAAoG;QACpG,2CAA2C,CAC5C,CAAC;IACF,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IAEf,sMAA4E;IAC5E,IAAI,MAAM,CAAC,UAAU,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAClC,KAAK,CAAC,IAAI,CAAC,iCAAiC,CAAC,CAAC;QAC9C,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;QACf,KAAK,MAAM,OAAO,IAAI,MAAM,CAAC,UAAU,EAAE,CAAC;YACzC,KAAK,CAAC,IAAI,CAAC,OAAO,OAAO,CAAC,MAAM,WAAS,OAAO,CAAC,IAAI,CAAC,OAAO,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC,IAAI,EAAE,GAAG,CAAC,CAAC;YACtF,KAAK,CAAC,IAAI,CAAC,OAAO,QAAQ,CAAC,OAAO,CAAC,KAAK,EAAE,OAAO,CAAC,QAAQ,EAAE,OAAO,CAAC,QAAQ,CAAC,EAAE,CAAC,CAAC;YACjF,0EAA0E;YAC1E,yEAAyE;YACzE,iDAAiD;YACjD,KAAK,CAAC,IAAI,CAAC,mBAAmB,OAAO,CAAC,KAAK,EAAE,CAAC,CAAC;YAC/C,IAAI,OAAO,CAAC,SAAS,EAAE,CAAC;gBACvB,KAAK,CAAC,IAAI,CAAC,8BAA8B,OAAO,CAAC,SAAS,EAAE,CAAC,CAAC;YAC/D,CAAC;YACD,IAAI,OAAO,CAAC,YAAY,EAAE,CAAC;gBAC1B,KAAK,CAAC,IAAI,CAAC,4BAA4B,OAAO,CAAC,YAAY,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,GAAG,CAAC,CAAC;YAC/E,CAAC;YACD,IAAI,OAAO,CAAC,aAAa,EAAE,CAAC;gBAC3B,KAAK,CAAC,IAAI,CAAC,gCAAgC,OAAO,CAAC,aAAa,UAAU,CAAC,CAAC;YAC7E,CAAC;YACD,IAAI,OAAO,CAAC,kBAAkB,EAAE,CAAC;gBAChC,KAAK,CAAC,IAAI,CAAC,mFAAiF,CAAC,CAAC;YAC/F,CAAC;QACF,CAAC;QACD,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IAChB,CAAC;IAED,gNAA4E;IAC5E,IAAI,MAAM,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAC7B,KAAK,CAAC,IAAI,CAAC,0BAA0B,CAAC,CAAC;QACvC,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;QACf,KAAK,CAAC,IAAI,CACT,kGAAkG;YACjG,0DAA0D,CAC3D,CAAC;QACF,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;QACf,KAAK,MAAM,GAAG,IAAI,MAAM,CAAC,KAAK,EAAE,CAAC;YAChC,KAAK,CAAC,IAAI,CAAC,OAAO,GAAG,CAAC,OAAO,UAAQ,QAAQ,CAAC,GAAG,CAAC,KAAK,EAAE,GAAG,CAAC,QAAQ,EAAE,GAAG,CAAC,QAAQ,CAAC,EAAE,CAAC,CAAC;YACxF,KAAK,CAAC,IAAI,CAAC,mBAAmB,GAAG,CAAC,KAAK,EAAE,CAAC,CAAC;YAC3C,uEAAuE;YACvE,IAAI,GAAG,CAAC,YAAY,EAAE,CAAC;gBACtB,KAAK,CAAC,IAAI,CAAC,cAAc,GAAG,CAAC,YAAY,EAAE,CAAC,CAAC;YAC9C,CAAC;YACD,IAAI,GAAG,CAAC,mBAAmB,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;gBACxC,KAAK,CAAC,IAAI,CAAC,4BAA4B,GAAG,CAAC,mBAAmB,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;YACrG,CAAC;YACD,IAAI,GAAG,CAAC,WAAW,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;gBAChC,KAAK,CAAC,IAAI,CAAC,qBAAqB,GAAG,CAAC,WAAW,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;YAC/D,CAAC;QACF,CAAC;QACD,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IAChB,CAAC;IAED,wMAA4E;IAC5E,IAAI,MAAM,CAAC,SAAS,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QACjC,KAAK,CAAC,IAAI,CAAC,4BAA4B,CAAC,CAAC;QACzC,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;QACf,KAAK,MAAM,QAAQ,IAAI,MAAM,CAAC,SAAS,EAAE,CAAC;YACzC,qEAAqE;YACrE,qEAAqE;YACrE,MAAM,KAAK,GAAG,QAAQ,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,QAAQ,CAAC,KAAK,CAAC,IAAI,CAAC,OAAK,CAAC,IAAI,CAAC,CAAC,CAAC,QAAQ,CAAC,KAAK,CAAC;YAC/F,KAAK,CAAC,IAAI,CAAC,KAAK,KAAK,QAAM,QAAQ,CAAC,QAAQ,CAAC,KAAK,EAAE,QAAQ,CAAC,QAAQ,EAAE,QAAQ,CAAC,QAAQ,CAAC,EAAE,CAAC,CAAC;QAC9F,CAAC;QACD,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IAChB,CAAC;IAED,kMAA4E;IAC5E,KAAK,CAAC,IAAI,CAAC,eAAe,CAAC,CAAC;IAC5B,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACf,KAAK,CAAC,IAAI,CAAC,gFAAgF,CAAC,CAAC;IAC7F,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACf,KAAK,CAAC,IAAI,CACT,0GAA0G;QACzG,4GAA0G;QAC1G,6DAA6D,CAC9D,CAAC;IACF,KAAK,CAAC,IAAI,CACT,6GAA6G;QAC5G,wGAAsG;QACtG,6GAA2G;QAC3G,mDAAmD,CACpD,CAAC;IACF,KAAK,CAAC,IAAI,CACT,0GAAwG;QACvG,yGAAyG;QACzG,0BAAwB,OAAO,CAAC,aAAa,6DAA6D;QAC1G,kBAAkB,CACnB,CAAC;IACF,KAAK,CAAC,IAAI,CACT,4GAA4G;QAC3G,4GAA0G;QAC1G,gFAAgF,CACjF,CAAC;IACF,KAAK,CAAC,IAAI,CACT,2GAA2G;QAC1G,wGAAwG;QACxG,2GAA2G;QAC3G,8BAA8B,CAC/B,CAAC;IACF,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACf,KAAK,CAAC,IAAI,CAAC,+CAA+C,CAAC,CAAC;IAC5D,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACf,KAAK,CAAC,IAAI,CACT,yGAAyG,CACzG,CAAC;IACF,KAAK,CAAC,IAAI,CAAC,qGAAqG,CAAC,CAAC;IAClH,KAAK,CAAC,IAAI,CACT,8GAA4G,CAC5G,CAAC;IACF,KAAK,CAAC,IAAI,CACT,8GAA8G;QAC7G,kDAAkD,CACnD,CAAC;IACF,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IAEf,IAAI,MAAM,CAAC,cAAc,EAAE,CAAC;QAC3B,KAAK,CAAC,IAAI,CACT,8BAA8B,MAAM,CAAC,cAAc,IAAI;YACtD,CAAC,MAAM,CAAC,gBAAgB,CAAC,CAAC,CAAC,MAAM,MAAM,CAAC,gBAAgB,iCAAiC,CAAC,CAAC,CAAC,EAAE,CAAC;YAC/F,4GAA4G,CAC7G,CAAC;IACH,CAAC;SAAM,CAAC;QACP,KAAK,CAAC,IAAI,CACT,wGAAwG,CACxG,CAAC;IACH,CAAC;IACD,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACf,KAAK,CAAC,IAAI,CACT,8GAA4G;QAC3G,mGAAmG,CACpG,CAAC;IAEF,OAAO,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAAA,CACxB","sourcesContent":["/**\n * Renders the extractor's output into the message `/learn` injects.\n *\n * The digest is evidence plus instructions, and the split matters: the numbers\n * come from {@link extractLearnDigest} and are not negotiable, while everything\n * the model does with them — phrasing, routing, deciding a pattern is not worth\n * a rule — is judgement it has to exercise. Counts are printed on every item\n * because \"said in 5 of your last 12 sessions\" is a decision the reader can\n * make in one keystroke, where \"extracted from your session\" is not.\n */\n\nimport type { LearnDigest } from \"./extract.js\";\nimport { LEARN_DIGEST_MARKER } from \"./extract.js\";\n\nfunction shortDate(iso: string | undefined): string {\n\tif (!iso) return \"unknown\";\n\tconst date = new Date(iso);\n\treturn Number.isNaN(date.getTime()) ? \"unknown\" : date.toISOString().slice(0, 10);\n}\n\nfunction evidence(count: number, sessions: number, lastSeen: string): string {\n\tconst times = count === 1 ? \"once\" : `${count}x`;\n\tconst where = sessions === 1 ? \"1 session\" : `${sessions} sessions`;\n\treturn `${times} across ${where}, last ${shortDate(lastSeen)}`;\n}\n\n/** True when there is nothing worth asking the model to look at. */\nexport function isEmptyDigest(digest: LearnDigest): boolean {\n\treturn digest.directives.length === 0 && digest.fixes.length === 0 && digest.workflows.length === 0;\n}\n\nexport function renderLearnDigest(\n\tdigest: LearnDigest,\n\toptions: { userScopePath: string; mode?: \"incremental\" | \"all\" },\n): string {\n\tconst lines: string[] = [];\n\n\tlines.push(\n\t\t`${LEARN_DIGEST_MARKER} Mined ${digest.scannedSessions} session(s) in this directory` +\n\t\t\t(digest.skippedSessions > 0 ? ` (${digest.skippedSessions} skipped: out of window or unreadable)` : \"\") +\n\t\t\t(digest.oldestSession ? `, ${shortDate(digest.oldestSession)} to ${shortDate(digest.newestSession)}` : \"\") +\n\t\t\t(digest.suppressed > 0 ? `. ${digest.suppressed} item(s) held back — already shown and unchanged since` : \"\") +\n\t\t\t\".\",\n\t);\n\t// Naming the mode keeps two very different empty results from reading alike:\n\t// \"nothing new since last time\" and \"nothing here at all\" are not the same\n\t// answer, and the reader cannot tell them apart from the counts.\n\tif (options.mode === \"all\") {\n\t\tlines.push(\"Mode: all — suppression is off, so items you have already seen and decided on are included.\");\n\t}\n\t// The model reads every transcript in full, which costs real tokens. Saying\n\t// what was re-read versus reused keeps that price visible rather than hidden.\n\tlines.push(\n\t\t`Read by the model this run: ${digest.mining.mined}; reused from cache: ${digest.mining.cached}` +\n\t\t\t(digest.mining.failed > 0\n\t\t\t\t? `; failed: ${digest.mining.failed} (their signals are missing from the counts below)`\n\t\t\t\t: \"\") +\n\t\t\t\".\",\n\t);\n\tlines.push(\"\");\n\tlines.push(\n\t\t\"The counts below are computed from session transcripts on disk, not from this conversation. \" +\n\t\t\t\"Treat them as evidence, not conclusions — your job is to decide what deserves to be written down, \" +\n\t\t\t\"phrase it, and put it in the right place.\",\n\t);\n\tlines.push(\"\");\n\n\t// ── Directives ───────────────────────────────────────────────────────────\n\tif (digest.directives.length > 0) {\n\t\tlines.push(\"## Directives you have repeated\");\n\t\tlines.push(\"\");\n\t\tfor (const cluster of digest.directives) {\n\t\t\tlines.push(`- **${cluster.status}** — \"${cluster.text.replace(/\\s+/g, \" \").trim()}\"`);\n\t\t\tlines.push(`  - ${evidence(cluster.count, cluster.sessions, cluster.lastSeen)}`);\n\t\t\t// Occurrences were grouped by meaning, not by wording, so the quote above\n\t\t\t// is one phrasing of several. Naming the shared point keeps a count of 5\n\t\t\t// from looking like five copies of one sentence.\n\t\t\tlines.push(`  - grouped as: ${cluster.label}`);\n\t\t\tif (cluster.rationale) {\n\t\t\t\tlines.push(`  - why it may be durable: ${cluster.rationale}`);\n\t\t\t}\n\t\t\tif (cluster.existingRule) {\n\t\t\t\tlines.push(`  - already covered by: \"${cluster.existingRule.slice(0, 160)}\"`);\n\t\t\t}\n\t\t\tif (cluster.existingSkill) {\n\t\t\t\tlines.push(`  - already covered by the \\`${cluster.existingSkill}\\` skill`);\n\t\t\t}\n\t\t\tif (cluster.previouslyDeclined) {\n\t\t\t\tlines.push(\"  - proposed before and not written down — you have already passed on this once\");\n\t\t\t}\n\t\t}\n\t\tlines.push(\"\");\n\t}\n\n\t// ── Fixes ────────────────────────────────────────────────────────────────\n\tif (digest.fixes.length > 0) {\n\t\tlines.push(\"## Failures you resolved\");\n\t\tlines.push(\"\");\n\t\tlines.push(\n\t\t\t\"Each is a command that failed and later succeeded, where something done in between was the fix. \" +\n\t\t\t\t\"Recurring ones are worth writing down; a one-off is not.\",\n\t\t);\n\t\tlines.push(\"\");\n\t\tfor (const fix of digest.fixes) {\n\t\t\tlines.push(`- \\`${fix.command}\\` — ${evidence(fix.count, fix.sessions, fix.lastSeen)}`);\n\t\t\tlines.push(`  - grouped as: ${fix.label}`);\n\t\t\t// The excerpt comes from the model now, which may not have quoted one.\n\t\t\tif (fix.errorExcerpt) {\n\t\t\t\tlines.push(`  - error: ${fix.errorExcerpt}`);\n\t\t\t}\n\t\t\tif (fix.interveningCommands.length > 0) {\n\t\t\t\tlines.push(`  - commands in between: ${fix.interveningCommands.map((c) => `\\`${c}\\``).join(\", \")}`);\n\t\t\t}\n\t\t\tif (fix.editedFiles.length > 0) {\n\t\t\t\tlines.push(`  - files edited: ${fix.editedFiles.join(\", \")}`);\n\t\t\t}\n\t\t}\n\t\tlines.push(\"\");\n\t}\n\n\t// ── Workflows ────────────────────────────────────────────────────────────\n\tif (digest.workflows.length > 0) {\n\t\tlines.push(\"## Repeated tool sequences\");\n\t\tlines.push(\"\");\n\t\tfor (const workflow of digest.workflows) {\n\t\t\t// A workflow the model named but did not enumerate still has a label\n\t\t\t// worth showing; rendering an empty backtick pair instead would not.\n\t\t\tconst steps = workflow.steps.length > 0 ? `\\`${workflow.steps.join(\" → \")}\\`` : workflow.label;\n\t\t\tlines.push(`- ${steps} — ${evidence(workflow.count, workflow.sessions, workflow.lastSeen)}`);\n\t\t}\n\t\tlines.push(\"\");\n\t}\n\n\t// ── Instructions ─────────────────────────────────────────────────────────\n\tlines.push(\"## What to do\");\n\tlines.push(\"\");\n\tlines.push(\"Work through the items above and propose concrete edits. For each one, decide:\");\n\tlines.push(\"\");\n\tlines.push(\n\t\t\"1. **Is it durable?** A rule that will still be true next month belongs somewhere. A one-off preference \" +\n\t\t\t\"about the task you happened to be doing does not. When in doubt, drop it — a wrong rule costs more than \" +\n\t\t\t\"a missing one, because it is paid on every request forever.\",\n\t);\n\tlines.push(\n\t\t\"2. **Rule or skill?** This is the most important call. A context file is loaded on **every** turn; a skill \" +\n\t\t\t\"is loaded **on demand**. So: short, always-true, unconditional → a one-line rule. Long, procedural, \" +\n\t\t\t'or conditional (a sequence of steps, a runbook, anything starting \"when X, do Y\") → a skill, not a rule. ' +\n\t\t\t\"Repeated tool sequences are almost always skills.\",\n\t);\n\tlines.push(\n\t\t`3. **Which scope?** Project-specific (this repo's tests, build, architecture, conventions) → the repo ` +\n\t\t\t`\\`AGENTS.md\\`. Personal habits that travel with you across every repo (style preferences, how you like ` +\n\t\t\t`commits written) → \\`${options.userScopePath}\\`. If it names this repo's files or commands, it is not a ` +\n\t\t\t`user-scope rule.`,\n\t);\n\tlines.push(\n\t\t\"4. **Restated items are rewrites, not additions.** An item marked `restated` is already covered by a rule \" +\n\t\t\t\"that is not working — too vague, buried, or contradicted elsewhere. Rewrite the existing line or delete \" +\n\t\t\t\"it in favour of a sharper one. Do not add a second rule saying the same thing.\",\n\t);\n\tlines.push(\n\t\t\"5. **`has-skill` items are a triggering problem, not a missing rule.** A skill already covers it and you \" +\n\t\t\t\"asked by hand anyway, which usually means the skill's `description` frontmatter does not describe the \" +\n\t\t\t\"situation you were in. Sharpen that description so it matches, rather than adding a rule that duplicates \" +\n\t\t\t\"what the skill already does.\",\n\t);\n\tlines.push(\"\");\n\tlines.push(\"Then, while you have the file open, audit it:\");\n\tlines.push(\"\");\n\tlines.push(\n\t\t\"- **Delete rules that no longer match the code.** Check a sample against the repo before trusting them.\",\n\t);\n\tlines.push(\"- **Delete rules that restate default behaviour.** Guidance the agent already follows is pure cost.\");\n\tlines.push(\n\t\t\"- **Collapse duplicates**, including any rule stated at both repo and user scope — that one is paid twice.\",\n\t);\n\tlines.push(\n\t\t\"- **One line per rule.** No rationale, no examples, no preamble, unless the example *is* the rule. Prose is \" +\n\t\t\t\"the single biggest source of context-file bloat.\",\n\t);\n\tlines.push(\"\");\n\n\tif (digest.agentsFilePath) {\n\t\tlines.push(\n\t\t\t`The repo context file is \\`${digest.agentsFilePath}\\`` +\n\t\t\t\t(digest.agentsFileTokens ? ` (~${digest.agentsFileTokens} tokens, re-sent every request)` : \"\") +\n\t\t\t\t\". Report the token delta of your proposed changes before applying them; a net reduction is a good outcome.\",\n\t\t);\n\t} else {\n\t\tlines.push(\n\t\t\t\"No repo context file exists yet. Create one only if at least one durable project rule survives step 1.\",\n\t\t);\n\t}\n\tlines.push(\"\");\n\tlines.push(\n\t\t\"Show what you propose, then apply it with edits — do not ask a separate approval question first, the edit \" +\n\t\t\t\"prompt is the approval. If nothing here is worth writing down, say so plainly and change nothing.\",\n\t);\n\n\treturn lines.join(\"\\n\");\n}\n"]}

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

@@ -8,82 +8,31 @@
  * whether something is a durable rule or a one-off, and it is the one thing a
  * prompt reading its own context cannot see.
  *
- * The split of labour is deliberate. This module is entirely deterministic: it
- * parses, filters, normalizes, counts and ranks. Judgement — is this a rule, how
- * should it be phrased, which scope owns it — belongs to the model reading the
- * digest, which is why the output carries evidence (counts, sessions, dates)
- * rather than conclusions.
- */
-import { type LearnState } from "./state.js";
-/**
- * Prefix on the message `/learn` injects. The digest is persisted like any user
- * turn, so without this marker the next `/learn` would mine its own output and
- * every proposal would compound its own count.
- */
-export declare const LEARN_DIGEST_MARKER = "[learn-digest]";
-/**
- * Where a repeated directive already lives, if anywhere.
+ * This module is the orchestrator, and the split of labour inside it is
+ * deliberate:
  *
- * A directive covered by a rule and said only once is simply dropped — the rule
- * exists and is working. What survives is one of three cases, and they want
- * different responses:
+ * - **Gathering** is deterministic. Finding session files, resolving which cwd
+ *   they belong to, walking the active branch of a forked session — all exact,
+ *   all cheap, all here.
+ * - **Judgement** is the model's, in `mine.ts` and `coverage.ts`. What counts as
+ *   a directive, what two phrasings have in common, whether a rule already
+ *   covers something — none of that survives contact with a regex, and it used
+ *   to be decided by one.
+ * - **Counting** is deterministic again, in `reduce.ts`. The number is the
+ *   product, and a model asked to count over a long context will be
+ *   approximately right.
  *
- * - `new` — not written down anywhere. Propose it.
- * - `restated` — a context-file rule covers it and you said it anyway, so the
- *   rule is not working. Rewrite it; do not add a second one.
- * - `has-skill` — a *skill* covers it and you asked by hand anyway, which
- *   usually means the skill's `description` is not triggering. Sharpen the
- *   description rather than writing a rule that duplicates the skill.
+ * The expensive step is memoized per session file (`cache.ts`), so a session is
+ * read by the model exactly once in its life and the counts are still computed
+ * over every session in the window on every run.
  */
-export type DirectiveStatus = "new" | "restated" | "has-skill";
-/** Fields every proposable item shares, so suppression can be applied uniformly. */
-interface Proposable {
-    /** Stable identity across runs — what the state file remembers. */
-    key: string;
-    /** Newest occurrence in the window, ISO. */
-    lastSeen: string;
-}
-export interface DirectiveCluster extends Proposable {
-    /** Representative raw text, the longest seen in the cluster. */
-    text: string;
-    normalized: string;
-    /** Total times said. */
-    count: number;
-    /** Distinct sessions it was said in — the stronger of the two counts. */
-    sessions: number;
-    status: DirectiveStatus;
-    /** The existing rule line matched, when status is `restated`. */
-    existingRule?: string;
-    /** The skill that already covers this, when status is `has-skill`. */
-    existingSkill?: string;
-    /**
-     * Shown before and still not written down anywhere — neither as a rule nor as
-     * a skill — so you saw this proposal and passed on it. Only meaningful for
-     * directives, which are the only items with a real coverage signal.
-     */
-    previouslyDeclined: boolean;
-}
-export interface FixCandidate extends Proposable {
-    /** Normalized failing command. */
-    command: string;
-    /** Normalized error signature, the dedupe key. */
-    signature: string;
-    /** Short raw excerpt, so the model sees the real error text. */
-    errorExcerpt: string;
-    /** Commands run between the failure and the pass. */
-    interveningCommands: string[];
-    /** Files edited between the failure and the pass. */
-    editedFiles: string[];
-    /** Times this signature failed and was resolved across the window. */
-    count: number;
-    sessions: number;
-}
-export interface WorkflowCandidate extends Proposable {
-    /** Tool-call signatures in order. */
-    steps: string[];
-    count: number;
-    sessions: number;
-}
+import type { CoverageIndex, CoverageJudge } from "./coverage.js";
+import type { Miner } from "./mine.js";
+import type { DirectiveCluster, FixCandidate, WorkflowCandidate } from "./reduce.js";
+import { type LearnState } from "./state.js";
+export type { CoverageIndex, CoverageMatch } from "./coverage.js";
+export { LEARN_DIGEST_MARKER } from "./mine.js";
+export type { DirectiveCluster, DirectiveStatus, FixCandidate, WorkflowCandidate } from "./reduce.js";
 /**
  * Why a session file on disk did not make it into the digest.
  *
@@ -108,11 +57,29 @@ export interface SessionScanReport {
     /** Skipped for being unreadable, unparseable, or empty. */
     unreadable: number;
 }
+/** What the run cost, so the price of an LLM-read pipeline is visible rather than hidden. */
+export interface MiningReport {
+    /** Sessions whose candidates came from cache, free. */
+    cached: number;
+    /** Sessions sent to the model this run. */
+    mined: number;
+    /** Sessions the model failed on. Their signals are missing from the counts. */
+    failed: number;
+}
 export interface LearnDigest {
     scannedSessions: number;
     skippedSessions: number;
     /** Where the sessions came from, and what was passed over. */
     scan: SessionScanReport;
+    /** What was read by the model versus reused. */
+    mining: MiningReport;
+    /**
+     * The run stopped before reading the whole window, so the counts below are
+     * computed from part of it. Callers must not record these as surfaced: a
+     * partial count can fall under the repeat threshold, and bookmarking it would
+     * hide the item on the next run, when the evidence is complete.
+     */
+    aborted: boolean;
     oldestSession?: string;
     newestSession?: string;
     agentsFilePath?: string;
@@ -127,6 +94,7 @@ export interface LearnDigest {
         key: string;
         lastSeen: string;
         covered: boolean;
+        text?: string;
     }>;
 }
 export interface ExtractOptions {
@@ -142,7 +110,7 @@ export interface ExtractOptions {
     maxAgeDays?: number;
     /** Occurrences a directive needs before it is proposed. The signal/noise dial. */
     minRepeats?: number;
-    /** Non-overlapping repeats a tool sequence needs before it is proposed as a skill. */
+    /** Repeats a tool sequence needs before it is proposed as a skill. */
     minWorkflowRepeats?: number;
     /** Cap on each list in the digest. */
     maxProposals?: number;
@@ -164,6 +132,20 @@ export interface ExtractOptions {
     /** Injectable clock, for tests. */
     now?: Date;
 }
+/** Everything the async pipeline needs beyond the window settings. */
+export interface MineOptions extends ExtractOptions {
+    /** Reads one session and reports what it saw. */
+    miner: Miner;
+    /** Decides which proposals are already written down. Defaults to "none are". */
+    coverageJudge?: CoverageJudge;
+    /** Progress callback, so a cold-cache run is not a silent wait. */
+    onProgress?: (progress: {
+        done: number;
+        total: number;
+        cached: number;
+    }) => void;
+    signal?: AbortSignal;
+}
 /**
  * Every directory this cwd's sessions could be sitting in.
  *
@@ -178,38 +160,25 @@ export interface ExtractOptions {
 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.
+ * mining anything. `/learn settings` and `/learn stats` report on the window
+ * without paying for a model call.
  */
 export declare function scanSessions(options: ExtractOptions): SessionScanReport;
 /**
- * Everything a proposal could already have been written into.
- *
- * Built once and shared, because the same question — is this already written
- * down? — is asked while ranking a run *and* afterwards by `/learn stats`,
- * which reconstructs adoption by comparing coverage now against coverage when
- * the item was shown.
- */
-export interface CoverageIndex {
-    /** Candidate rule lines from the repo context file and both user scopes. */
-    ruleLines: string[];
-    skills: Array<{
-        name: string;
-        description: string;
-    }>;
-}
-export interface CoverageMatch {
-    /** The context-file line that covers this, if any. */
-    rule?: string;
-    /** The skill that covers this, if any. Only set when no rule matched. */
-    skill?: string;
-}
-/**
- * Where a piece of text is already written down, if anywhere.
+ * What a run would read, without reading it.
  *
- * A rule wins over a skill when both match: it is the more specific answer, and
- * "rewrite this line" is more actionable than "sharpen a description".
+ * Runs the real selection — the same age, cwd, cap and de-duplication rules
+ * `mineLearnDigest` applies — and then asks the cache about each survivor. It
+ * has to be the same selection: this number is what the confirmation prompt
+ * quotes, and a prompt that says twelve before reading three is worse than no
+ * prompt at all. Hashing the chosen files is cheap next to sending them to a
+ * model.
  */
-export declare function matchCoverage(text: string, index: CoverageIndex): CoverageMatch;
+export declare function planMining(options: ExtractOptions): {
+    total: number;
+    cached: number;
+    pending: number;
+};
 /** Assemble the coverage index for a directory. */
 export declare function buildCoverageIndex(options: {
     cwd: string;
@@ -220,6 +189,5 @@ export declare function buildCoverageIndex(options: {
     }>;
 }): CoverageIndex;
 /** Mine the recent sessions for this cwd and return the ranked digest. */
-export declare function extractLearnDigest(options: ExtractOptions): LearnDigest;
-export {};
+export declare function mineLearnDigest(options: MineOptions): Promise<LearnDigest>;
 //# sourceMappingURL=extract.d.ts.map

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;;;;;;;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"]}
+{"version":3,"file":"extract.d.ts","sourceRoot":"","sources":["../../../src/core/learn/extract.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;GA2BG;AASH,OAAO,KAAK,EAAE,aAAa,EAAE,aAAa,EAAiB,MAAM,eAAe,CAAC;AAEjF,OAAO,KAAK,EAAkB,KAAK,EAAE,MAAM,WAAW,CAAC;AACvD,OAAO,KAAK,EAAE,gBAAgB,EAAE,YAAY,EAA4B,iBAAiB,EAAE,MAAM,aAAa,CAAC;AAE/G,OAAO,EAAS,KAAK,UAAU,EAAE,MAAM,YAAY,CAAC;AAEpD,YAAY,EAAE,aAAa,EAAE,aAAa,EAAE,MAAM,eAAe,CAAC;AAClE,OAAO,EAAE,mBAAmB,EAAE,MAAM,WAAW,CAAC;AAChD,YAAY,EAAE,gBAAgB,EAAE,eAAe,EAAE,YAAY,EAAE,iBAAiB,EAAE,MAAM,aAAa,CAAC;AAetG;;;;;;;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,6FAA6F;AAC7F,MAAM,WAAW,YAAY;IAC5B,uDAAuD;IACvD,MAAM,EAAE,MAAM,CAAC;IACf,2CAA2C;IAC3C,KAAK,EAAE,MAAM,CAAC;IACd,+EAA+E;IAC/E,MAAM,EAAE,MAAM,CAAC;CACf;AAED,MAAM,WAAW,WAAW;IAC3B,eAAe,EAAE,MAAM,CAAC;IACxB,eAAe,EAAE,MAAM,CAAC;IACxB,8DAA8D;IAC9D,IAAI,EAAE,iBAAiB,CAAC;IACxB,gDAAgD;IAChD,MAAM,EAAE,YAAY,CAAC;IACrB;;;;;OAKG;IACH,OAAO,EAAE,OAAO,CAAC;IACjB,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,CAAC;QAAC,IAAI,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC,CAAC;CACpF;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,sEAAsE;IACtE,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;AAED,sEAAsE;AACtE,MAAM,WAAW,WAAY,SAAQ,cAAc;IAClD,iDAAiD;IACjD,KAAK,EAAE,KAAK,CAAC;IACb,gFAAgF;IAChF,aAAa,CAAC,EAAE,aAAa,CAAC;IAC9B,mEAAmE;IACnE,UAAU,CAAC,EAAE,CAAC,QAAQ,EAAE;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAC;QAAC,MAAM,EAAE,MAAM,CAAA;KAAE,KAAK,IAAI,CAAC;IACjF,MAAM,CAAC,EAAE,WAAW,CAAC;CACrB;AAkID;;;;;;;;;;GAUG;AACH,wBAAgB,oBAAoB,CAAC,OAAO,EAAE,IAAI,CAAC,cAAc,EAAE,KAAK,GAAG,UAAU,GAAG,YAAY,CAAC,GAAG,MAAM,EAAE,CAW/G;AAqHD;;;;GAIG;AACH,wBAAgB,YAAY,CAAC,OAAO,EAAE,cAAc,GAAG,iBAAiB,CAEvE;AAED;;;;;;;;;GASG;AACH,wBAAgB,UAAU,CAAC,OAAO,EAAE,cAAc,GAAG;IAAE,KAAK,EAAE,MAAM,CAAC;IAAC,MAAM,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,MAAM,CAAA;CAAE,CAUtG;AAgBD,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;AAwHD,0EAA0E;AAC1E,wBAAsB,eAAe,CAAC,OAAO,EAAE,WAAW,GAAG,OAAO,CAAC,WAAW,CAAC,CAsFhF","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 * This module is the orchestrator, and the split of labour inside it is\n * deliberate:\n *\n * - **Gathering** is deterministic. Finding session files, resolving which cwd\n *   they belong to, walking the active branch of a forked session — all exact,\n *   all cheap, all here.\n * - **Judgement** is the model's, in `mine.ts` and `coverage.ts`. What counts as\n *   a directive, what two phrasings have in common, whether a rule already\n *   covers something — none of that survives contact with a regex, and it used\n *   to be decided by one.\n * - **Counting** is deterministic again, in `reduce.ts`. The number is the\n *   product, and a model asked to count over a long context will be\n *   approximately right.\n *\n * The expensive step is memoized per session file (`cache.ts`), so a session is\n * read by the model exactly once in its life and the counts are still computed\n * over every session in the window on every run.\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 { getUserAgentsDir } from \"../../config.js\";\nimport { getSessionDirPath } from \"../session-manager.js\";\nimport { loadSkills } from \"../skills.js\";\nimport { hashSessionFile, pruneLearnCache, readCachedMining, writeCachedMining } from \"./cache.js\";\nimport type { CoverageIndex, CoverageJudge, CoverageQuery } from \"./coverage.js\";\nimport { noCoverageJudge } from \"./coverage.js\";\nimport type { MinableSession, Miner } from \"./mine.js\";\nimport type { DirectiveCluster, FixCandidate, MinedSession, Proposable, WorkflowCandidate } from \"./reduce.js\";\nimport { reduceDirectives, reduceFixes, reduceWorkflows } from \"./reduce.js\";\nimport { judge, type LearnState } from \"./state.js\";\n\nexport type { CoverageIndex, CoverageMatch } from \"./coverage.js\";\nexport { LEARN_DIGEST_MARKER } from \"./mine.js\";\nexport type { DirectiveCluster, DirectiveStatus, FixCandidate, WorkflowCandidate } from \"./reduce.js\";\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/** Occurrences a directive needs before it is proposed. */\nconst DEFAULT_MIN_DIRECTIVE_COUNT = 2;\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 * 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\n/** What the run cost, so the price of an LLM-read pipeline is visible rather than hidden. */\nexport interface MiningReport {\n\t/** Sessions whose candidates came from cache, free. */\n\tcached: number;\n\t/** Sessions sent to the model this run. */\n\tmined: number;\n\t/** Sessions the model failed on. Their signals are missing from the counts. */\n\tfailed: 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\t/** What was read by the model versus reused. */\n\tmining: MiningReport;\n\t/**\n\t * The run stopped before reading the whole window, so the counts below are\n\t * computed from part of it. Callers must not record these as surfaced: a\n\t * partial count can fall under the repeat threshold, and bookmarking it would\n\t * hide the item on the next run, when the evidence is complete.\n\t */\n\taborted: boolean;\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; text?: string }>;\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/** 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\n/** Everything the async pipeline needs beyond the window settings. */\nexport interface MineOptions extends ExtractOptions {\n\t/** Reads one session and reports what it saw. */\n\tminer: Miner;\n\t/** Decides which proposals are already written down. Defaults to \"none are\". */\n\tcoverageJudge?: CoverageJudge;\n\t/** Progress callback, so a cold-cache run is not a silent wait. */\n\tonProgress?: (progress: { done: number; total: number; cached: number }) => void;\n\tsignal?: AbortSignal;\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\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 miner from reading two turns that never happened in sequence as if\n * they did. Sessions written before entry ids existed are flat, and for those\n * 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 * mining anything. `/learn settings` and `/learn stats` report on the window\n * without paying for a model call.\n */\nexport function scanSessions(options: ExtractOptions): SessionScanReport {\n\treturn listSessions(options).scan;\n}\n\n/**\n * What a run would read, without reading it.\n *\n * Runs the real selection — the same age, cwd, cap and de-duplication rules\n * `mineLearnDigest` applies — and then asks the cache about each survivor. It\n * has to be the same selection: this number is what the confirmation prompt\n * quotes, and a prompt that says twelve before reading three is worse than no\n * prompt at all. Hashing the chosen files is cheap next to sending them to a\n * model.\n */\nexport function planMining(options: ExtractOptions): { total: number; cached: number; pending: number } {\n\tconst { sessions } = listSessions(options);\n\tlet cached = 0;\n\tlet pending = 0;\n\tfor (const session of sessions) {\n\t\tconst hash = hashSessionFile(session.file);\n\t\tif (hash && readCachedMining(options.agentDir, hash)) cached++;\n\t\telse pending++;\n\t}\n\treturn { total: sessions.length, cached, pending };\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\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/**\n * Run the miner over the window, reusing cached results wherever the file has\n * not changed.\n *\n * A session that fails to mine is counted and skipped rather than aborting the\n * run: one provider hiccup on one transcript should cost that transcript's\n * signals, not the whole digest. The failure count is reported so the reader\n * knows the numbers are short.\n *\n * Cancellation is different from failure and is reported separately. A run\n * stopped half way has counted only some of the window, so its numbers are not\n * merely short — they are wrong in a way that would poison the bookmark if the\n * digest were treated as a completed run.\n */\nasync function mineSessions(\n\tsessions: ParsedSession[],\n\toptions: MineOptions,\n): Promise<{ mined: MinedSession[]; report: MiningReport; aborted: boolean }> {\n\tconst mined: MinedSession[] = [];\n\tconst report: MiningReport = { cached: 0, mined: 0, failed: 0 };\n\n\tlet done = 0;\n\tfor (const session of sessions) {\n\t\tif (options.signal?.aborted) return { mined, report, aborted: true };\n\n\t\tconst hash = hashSessionFile(session.file);\n\t\tconst cached = hash ? readCachedMining(options.agentDir, hash) : undefined;\n\t\tif (cached) {\n\t\t\tmined.push({ sessionId: session.id, timestamp: session.timestamp, candidates: cached.candidates });\n\t\t\treport.cached++;\n\t\t\tdone++;\n\t\t\toptions.onProgress?.({ done, total: sessions.length, cached: report.cached });\n\t\t\tcontinue;\n\t\t}\n\n\t\tconst minable: MinableSession = { id: session.id, timestamp: session.timestamp, entries: session.entries };\n\t\ttry {\n\t\t\tconst candidates = await options.miner(minable, options.signal);\n\t\t\tmined.push({ sessionId: session.id, timestamp: session.timestamp, candidates });\n\t\t\treport.mined++;\n\t\t\tif (hash) {\n\t\t\t\twriteCachedMining(options.agentDir, hash, {\n\t\t\t\t\tsessionId: session.id,\n\t\t\t\t\ttimestamp: session.timestamp,\n\t\t\t\t\tcandidates,\n\t\t\t\t\tminedAt: new Date().toISOString(),\n\t\t\t\t});\n\t\t\t}\n\t\t} catch {\n\t\t\t// A cancelled request surfaces here as a rejection. That is not the\n\t\t\t// provider failing on this transcript, so it must not be counted as one.\n\t\t\tif (options.signal?.aborted) return { mined, report, aborted: true };\n\t\t\treport.failed++;\n\t\t}\n\t\tdone++;\n\t\toptions.onProgress?.({ done, total: sessions.length, cached: report.cached });\n\t}\n\n\treturn { mined, report, aborted: false };\n}\n\n/** Apply the coverage verdicts to the clusters they were asked about. */\nfunction applyCoverage(directives: DirectiveCluster[], verdicts: Map<string, { rule?: string; skill?: string }>): void {\n\tfor (const cluster of directives) {\n\t\tconst verdict = verdicts.get(cluster.label);\n\t\tif (!verdict) continue;\n\t\tif (verdict.rule) {\n\t\t\tcluster.status = \"restated\";\n\t\t\tcluster.existingRule = verdict.rule;\n\t\t} else if (verdict.skill) {\n\t\t\tcluster.status = \"has-skill\";\n\t\t\tcluster.existingSkill = verdict.skill;\n\t\t}\n\t}\n}\n\n/** Mine the recent sessions for this cwd and return the ranked digest. */\nexport async function mineLearnDigest(options: MineOptions): Promise<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\n\tconst { mined, report, aborted } = await mineSessions(sessions, options);\n\tpruneLearnCache(options.agentDir, options.now);\n\n\tconst minRepeats = options.minRepeats ?? DEFAULT_MIN_DIRECTIVE_COUNT;\n\tconst maxProposals = options.maxProposals ?? DEFAULT_MAX_PER_CATEGORY;\n\tconst directives = reduceDirectives(mined, minRepeats);\n\tconst fixes = reduceFixes(mined, minRepeats);\n\tconst workflows = reduceWorkflows(mined, options.minWorkflowRepeats ?? DEFAULT_MIN_WORKFLOW_COUNT);\n\n\t// Coverage is asked only about what survived the repeat threshold. Judging\n\t// everything would mean sending the context file alongside a long tail of\n\t// one-off observations that are never going to be proposed.\n\tconst coverage = buildCoverageIndex({ cwd: options.cwd, agentDir: options.agentDir, skills: options.skills });\n\tconst queries: CoverageQuery[] = directives.map((d) => ({ label: d.label, text: d.text }));\n\ttry {\n\t\tconst verdicts = await (options.coverageJudge ?? noCoverageJudge)(queries, coverage, options.signal);\n\t\tapplyCoverage(directives, verdicts);\n\t} catch {\n\t\t// A failed coverage call leaves everything `new`, which over-proposes\n\t\t// slightly. That is the right way to fail: the reader can reject a\n\t\t// duplicate, but cannot recover a proposal that was wrongly withheld.\n\t}\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 keptDirectives = applySuppression(\n\t\tdirectives,\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 keptFixes = applySuppression(fixes, state, maxProposals, () => false);\n\tconst keptWorkflows = applySuppression(workflows, state, maxProposals, () => false);\n\n\tconst surfaced = [\n\t\t// Directives carry their wording forward so a later `/learn stats` can ask\n\t\t// about coverage using the sentence rather than the slug that names it.\n\t\t...keptDirectives.kept.map((d) => ({\n\t\t\tkey: d.key,\n\t\t\tlastSeen: d.lastSeen,\n\t\t\tcovered: d.status !== \"new\",\n\t\t\ttext: d.text,\n\t\t})),\n\t\t...keptFixes.kept.map((f) => ({ key: f.key, lastSeen: f.lastSeen, covered: false })),\n\t\t...keptWorkflows.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\tmining: report,\n\t\taborted,\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: keptDirectives.kept,\n\t\tfixes: keptFixes.kept,\n\t\tworkflows: keptWorkflows.kept,\n\t\tsuppressed: keptDirectives.suppressed + keptFixes.suppressed + keptWorkflows.suppressed,\n\t\tsurfaced,\n\t};\n}\n"]}