harness: auto-skill general + memory spesifik + lossless compaction

Memory balik spesifik per-project (keep paths, THIS repo), general jadi
skill auto-create ke ~/.shiro-neko/skills/auto-*.md. Compaction lossless
dengan head preserved + droppedSpan normalized + retained note. Learner
throttled tiap 3 turn / delta 8, pakai subagentModel (learnerModel) dan
ceiling guard, hash dedup, JSON fallback, TTL tiap 6 turn.

.hermes/plans/harness-auto-skill.md

@@ -0,0 +1,20 @@
+# Plan: Harness auto-skill — memory spesifik, skill general
+
+## Koreksi desain
+- Sebelumnya: memory project + global layer (_global.json) untuk pattern universal. Salah.
+- Seharusnya: memory = spesifik per project (path, command, decision repo itu). Skill = general/universal, auto-create.
+
+## Goal
+Harness yang belajar terus: tiap sesi yang menghasilkan pattern reusable lintas-repo -> otomatis jadi file skill di ~/.shiro-neko/skills/auto-*.md, kebawa di sesi berikut via loadSkills tanpa manual.
+
+## Perubahan
+1. src/memory.ts — hapus global layer (globalFile/globalEntries/addGlobal/renderWithGlobal). Balikkan prompt summarize & suggestFromTranscript ke project-spesifik (keep paths/names, one-off file fix BOLEH disimpan kalau spesifik repo). MAX_TEXT 800, dedup normalize, TTL 90d tetap.
+2. src/skill-learner.ts — BARU. suggestSkillsFromTranscript(messages, model) -> SkillCandidate[] { name, description, body }, prompt: generalisasi ke peran (the auth layer, not src/auth.ts), condition->action->reason, skip repo-spesifik. writeAutoSkill(candidate): tulis ke ~/.shiro-neko/skills/auto-<slug>.md dengan frontmatter name/description, dedup by name+normalize(body), append section kalau file sudah ada, cap MAX_BODY 20k, best-effort.
+3. src/session.ts — import skill-learner, afterTurn hook: maybeLearnSkills() (best-effort, catch, hanya kalau messages.length>=6 dan model ada). Dipanggil di send() finally setelah plugins.afterTurn, atau setelah run selesai. Tanpa block turn.
+4. src/cli.tsx — tidak perlu wiring khusus; learner pakai SHIRO_HOME & Bun.write. Pastikan Memory ctor tidak lagi loadGlobal.
+5. Test — memory.test tetap hijau (project-only), tambah skill-learner.test untuk dedup & write.
+
+## Verifikasi
+- bun run typecheck
+- bun test (790 -> 79x)
+- manual: SHIRO_HOME=$(mktemp -d) bun run src/cli.tsx -p "hello" --yolo --json -> cek ~/.shiro-neko/skills/auto-*.md terbuat saat pattern kuat

.hermes/plans/memory-compaction-lifelong.md

@@ -0,0 +1,69 @@
+# Plan: Lifelong Learning — Memory + Compaction Fix
+
+## 1. Goal
+Bikin shiro-neko cukup untuk agent yang belajar terus (lifelong) tanpa ngulang kesalahan setelah context compact / sesi ganti.
+
+Status sekarang:
+- Harness (session + prune + notebook): 7.5/10 — ladder prune, orphan repair, repeat guard sudah benar.
+- Auto-memory: 4/10 — keyword literal, manual `remember`, boot block hits-only, MAX_TEXT 400, per-project silo, tanpa auto-consolidation.
+
+Target: harness 9/10, memory 8/10 dengan perubahan kecil, backward-compatible, semua tes hijau.
+
+## 2. Scope
+
+### A. Memory (`src/memory.ts`)
+- Search dari `terms.every(includes)` strict → scored hybrid (normalize + token + substring + hits + recency). Tetap require every term untuk kompatibilitas tes, tapi ranking jauh lebih baik dan tahan typo minor (underscore/hyphen/punct).
+- Render boot block dari hits-only → diverse: top-hits + most-recent, dedup. Mencegah 20 gotcha lama mengusir decision baru.
+- MAX_TEXT 400 → 800 (potong silent terlalu agresif untuk gotcha/decision berkonteks).
+- Dedup dari exact `text===clean` → normalized (`lower + collapse ws + strip punct`) biar near-duplicate tidak dobel.
+- TTL / decay: `pruneExpired(90d)` untuk entry hits==0 & umur >90d, dipanggil di load/persist. hits diberi recency bonus saat render, bukan decay destructive.
+- Global memory: `~/.shiro-neko/memory/_global.json` sebagai layer kedua; per-project tetap utama, global untuk pattern lintas-repo.
+- Auto-extract hook: `suggestFromTranscript(messages, model)` — 1x generateText yang meringkas sesi jadi 1-3 kandidat memory (dipakai session afterTurn, bukan auto-write diam-diam).
+
+### B. Compaction (`src/prune.ts` + `src/session.ts`)
+- estimateTokens: `len/4` → `len/3.6 + messages*8` (akun overhead role/tool envelope; terukur lebih dekat ke cl100k). Threshold tetap 120k.
+- pruneToFit: selalu preserve head (first user goal + first assistant) terlepas ladder, jadi goal tidak hilang di narrowest rung.
+- prepareStep di session: ladder tetap, tapi droppedSpan di-capture by identity dan di-ringkas via 1x generateText menjadi `Note (retained from compacted history)` user message (lossless compaction). Guarded try/catch — mock/dead model tidak break turn.
+- afterTurn auto-memory: jika turn menambah >=4 pesan dan memory ada, tawarkan 1-3 kandidat via suggestFromTranscript (tidak auto-persist tanpa konfirmasi model; tool `remember` tetap sumber write).
+- Repeat guard lintas-turn: `seen` sekarang Map dengan decay per 20 turn (opsional, ringan).
+- Persist: onChange debounce 400ms → flush juga di `beforeExit` + tiap done (sudah ada), tambah flush on compact.
+
+### C. Prompt / Wiring (`src/prompt.ts`, `src/cli.tsx`, `src/store.ts`)
+- Prompt: memory block sekarang labeled `Project memory (project + global)`; notebook tetap survive compaction.
+- cli.tsx: Memory ctor dapat global layer; `session.send` after done trigger suggest (off jika --no-memory).
+- store: tidak diubah format, hanya notebook persistance tetap.
+
+## 3. Files Touched
+- `src/memory.ts` — utama (search, render, MAX_TEXT, dedup, global, TTL, suggest)
+- `src/prune.ts` — preserveHead, estimate helper export, ladder keep
+- `src/session.ts` — estimateTokens baru, summarizeDiscarded, afterTurn hook, cross-turn seen
+- `src/prompt.ts` — minor label (opsional)
+- `src/cli.tsx` — wiring global memory + afterTurn suggestion
+- `test/memory.test.ts` — update expectation truncation 400→800, tambah tes diverse/ranking/TTL
+- `test/prune.test.ts` — tambah preserveHead test
+- `test/compact.test.ts` — tambah lossless-compaction test (mock summarize)
+
+## 4. Schema / Types
+- MemoryEntry tetap `{id,kind,text,createdAt,hits}` — tambah field opsional `lastHitAt?: string` untuk recency bonus (backward compat: fallback ke createdAt).
+- Config tidak berubah (SHIRO_HOME, maxSpendUsd, etc).
+- File: per-project `${hash(cwd)}.json`, global `_global.json` di root(). Format array JSON sama, corrupted → [].
+
+## 5. Verification
+- `bun run typecheck` — harus pass (verbatimModuleSyntax, noUncheckedIndexedAccess)
+- `bun test` — semua existing + baru hijau; khusus:
+  - memory.test: truncation 800, every-term strict tetap, hits+recency ranking, diverse render contains recent
+  - prune.test: head goal survive narrowest rung, estimate monotonic
+  - compact.test: droppedSpan summarized to retained note, tidak loop
+- `bun run build` — `dist/shiro` bisa di-compile
+- Manual: `SHIRO_HOME=$(mktemp -d) bun run src/cli.tsx -p "hello" --yolo --json` → compacted event + done, memory file terisi
+
+## 6. Non-Goals (defer)
+- Embedding vector search (butuh dep + infra) — trigram/BM25 cukup untuk sekarang, bisa follow-up.
+- Subagent parallel fan-out — butuh executor custom, defer (catat di ROADMAP).
+- MCP meta-tools — defer.
+
+## 7. Rollout
+- Implement memory.ts dulu → test
+- prune.ts preserveHead + estimate
+- session.ts summarizeDiscarded + wiring
+- Update tests incremental, typecheck tiap langkah, jangan bulk.

src/cli.tsx

@@ -232,7 +232,10 @@ const externalTools = await loadExternalTools(process.cwd(), async (command) =>
 );
 
 const memory = has('--no-memory') ? undefined : new Memory(process.cwd(), languageModel);
-if (memory) await memory.load();
+if (memory) {
+  await memory.load();
+  try { await memory.pruneExpired(); } catch {}
+}
 
 /** Placeholder until /provider supplies a key; it never gets called because the UI gates input. */
 const unconfiguredModel: LanguageModel = {
@@ -302,6 +305,8 @@ const session = new Session({
   model: languageModel ?? unconfiguredModel,
   modelId: cfg.model,
   ...(cfg.subagentModel ? { subagentModelId: cfg.subagentModel } : {}),
+  ...(subagentModel !== unconfiguredModel ? { learnerModel: subagentModel } : {}),
+  onNotice: (t) => notices.emit(t),
   askApproval: bridge.ask,
   yolo,
   instructions,

src/memory.ts

@@ -14,14 +14,19 @@ export type MemoryEntry = {
   createdAt: string;
   /** Bumped on each recall so summarisation can keep what gets used. */
   hits: number;
+  /** Last time this entry was recalled; falls back to createdAt when absent. */
+  lastHitAt?: string;
 };
 
 const MAX_ENTRIES = 300;
-const MAX_TEXT = 400;
+const MAX_TEXT = 800;
 const BOOT_ENTRIES = 20;
 const SEARCH_HITS = 15;
 /** Summarise once the store passes this, so the boot block stays small. */
 const SUMMARISE_AT = 60;
+/** Entries with 0 hits older than this are pruned (lifelong: drop stale cruft). */
+const TTL_DAYS = 90;
+const TTL_MS = TTL_DAYS * 24 * 60 * 60 * 1000;
 
 const root = () => join(process.env['SHIRO_HOME'] ?? homedir(), '.shiro-neko', 'memory');
 
@@ -35,6 +40,28 @@ const KIND_LABEL: Record<MemoryKind, string> = {
   command: 'command',
 };
 
+/** Normalize for dedup and tolerant matching: lower, collapse whitespace, strip leading/trailing punct. */
+function normalize(text: string): string {
+  return text.trim().toLowerCase().replace(/\s+/g, ' ').replace(/^[\p{P}\s]+|[\p{P}\s]+$/gu, '');
+}
+
+/** Tokenize query/entry for scoring: split on non-alnum, drop empties. Underscore/hyphen are separators. */
+function tokensOf(s: string): string[] {
+  return s
+    .toLowerCase()
+    .replace(/[_-]+/g, ' ')
+    .replace(/[^\p{L}\p{N}\s]/gu, ' ')
+    .split(/\s+/)
+    .filter(Boolean);
+}
+
+function isExpired(e: MemoryEntry, now: number): boolean {
+  if (e.hits > 0) return false;
+  const ts = Date.parse(e.createdAt);
+  if (Number.isNaN(ts)) return false;
+  return now - ts > TTL_MS;
+}
+
 /**
  * Durable per-project memory, separate from the session transcript.
  *
@@ -62,9 +89,15 @@ export class Memory {
         this.entries = [];
       }
     }
+    // TTL pruning on load (only drops stale 0-hit cruft)
+    const now = Date.now();
+    const beforeLen = this.entries.length;
+    this.entries = this.entries.filter((e) => !isExpired(e, now));
+    if (this.entries.length !== beforeLen) await this.persist();
     return this.entries;
   }
 
+
   all(): MemoryEntry[] {
     return [...this.entries];
   }
@@ -78,7 +111,8 @@ export class Memory {
     await this.load();
     const clean = text.trim().slice(0, MAX_TEXT);
     if (!clean) throw new Error('memory text is empty');
-    if (this.entries.some((e) => e.text === clean)) return undefined;
+    const norm = normalize(clean);
+    if (this.entries.some((e) => normalize(e.text) === norm)) return undefined;
 
     const entry: MemoryEntry = {
       id: Bun.randomUUIDv7(),
@@ -106,32 +140,99 @@ export class Memory {
     await this.persist();
   }
 
-  /** Every term must appear. Matching entries get a hit, which protects them from summarisation. */
-  async search(query: string): Promise<MemoryEntry[]> {
+  /** Prune stale 0-hit entries older than TTL. Returns count removed. */
+  async pruneExpired(): Promise<number> {
     await this.load();
-    const terms = query.toLowerCase().split(/\s+/).filter(Boolean);
-    if (terms.length === 0) throw new Error('query is empty');
-
-    const found = this.entries.filter((e) => {
-      const lower = e.text.toLowerCase();
-      return terms.every((t) => lower.includes(t));
-    });
-    for (const e of found) e.hits += 1;
-    if (found.length > 0) await this.persist();
-    return found.slice(-SEARCH_HITS).reverse();
+    const before = this.entries.length;
+    const now = Date.now();
+    this.entries = this.entries.filter((e) => !isExpired(e, now));
+    const removed = before - this.entries.length;
+    if (removed > 0) await this.persist();
+    return removed;
   }
 
-  /** The block injected at boot: most-used first, then most recent. */
+  /**
+   * Every term must appear (compat). Scoring adds hits + recency + exact-phrase bonus,
+   * so the best entries surface first instead of arbitrary file order.
+   */
+  async search(query: string): Promise<MemoryEntry[]> {
+    await this.load();
+    const terms = tokensOf(query);
+    if (terms.length === 0) throw new Error('query is empty');
+
+    const now = Date.now();
+    const scored: { e: MemoryEntry; score: number }[] = [];
+    for (const e of this.entries) {
+      const lower = e.text.toLowerCase();
+      const entryTokens = new Set(tokensOf(e.text));
+      // every query token must appear as substring or token (keeps compat with "snake_case" queries)
+      const allPresent = terms.every((t) => lower.includes(t) || entryTokens.has(t));
+      if (!allPresent) continue;
+      // scoring: hits weight + recency + phrase bonus
+      const lastHit = e.lastHitAt ? Date.parse(e.lastHitAt) : Date.parse(e.createdAt);
+      const ageDays = Number.isNaN(lastHit) ? 999 : (now - lastHit) / (24 * 60 * 60 * 1000);
+      const recency = Math.max(0, 10 - ageDays * 0.1); // ~10 pts fresh, decays over 100d
+      const phraseBonus = lower.includes(query.toLowerCase().trim()) ? 5 : 0;
+      const score = e.hits * 3 + recency + phraseBonus;
+      scored.push({ e, score });
+    }
+    scored.sort((a, b) => b.score - a.score || b.e.hits - a.e.hits || b.e.createdAt.localeCompare(a.e.createdAt));
+    const found = scored.map((s) => s.e);
+    const nowIso = new Date().toISOString();
+    for (const e of found) {
+      e.hits += 1;
+      e.lastHitAt = nowIso;
+    }
+    if (found.length > 0) await this.persist();
+    return found.slice(0, SEARCH_HITS);
+  }
+
+  /**
+   * The block injected at boot: diverse mix of most-used + most-recent, so a few
+   * stale high-hit entries cannot evict fresh decisions.
+   */
   render(limit = BOOT_ENTRIES): string {
     if (this.entries.length === 0) return '';
-    const ranked = [...this.entries]
-      .sort((a, b) => b.hits - a.hits || b.createdAt.localeCompare(a.createdAt))
-      .slice(0, limit);
+    const pool = this.entries;
+    const byHits = [...pool].sort((a, b) => b.hits - a.hits || b.createdAt.localeCompare(a.createdAt));
+    const byRecent = [...pool].sort((a, b) => b.createdAt.localeCompare(a.createdAt));
+
+    const seen = new Set<string>();
+    const ranked: MemoryEntry[] = [];
+    // interleave: take from hits, then recent, until limit
+    let hi = 0;
+    let ri = 0;
+    while (ranked.length < limit && (hi < byHits.length || ri < byRecent.length)) {
+      if (hi < byHits.length) {
+        const e = byHits[hi++]!;
+        if (!seen.has(e.id)) {
+          seen.add(e.id);
+          ranked.push(e);
+        }
+        if (ranked.length >= limit) break;
+      }
+      if (ri < byRecent.length) {
+        const e = byRecent[ri++]!;
+        if (!seen.has(e.id)) {
+          seen.add(e.id);
+          ranked.push(e);
+        }
+      }
+    }
+    // recency bonus visible in ordering already via interleave; keep hits bias slightly by stable sort
+    // final sort by composite score for determinism: hits*2 + recency
+    const now = Date.now();
+    ranked.sort((a, b) => {
+      const aRec = Math.max(0, 10 - (now - Date.parse(a.lastHitAt ?? a.createdAt)) / (24 * 60 * 60 * 1000) * 0.1);
+      const bRec = Math.max(0, 10 - (now - Date.parse(b.lastHitAt ?? b.createdAt)) / (24 * 60 * 60 * 1000) * 0.1);
+      return b.hits * 2 + bRec - (a.hits * 2 + aRec) || b.createdAt.localeCompare(a.createdAt);
+    });
+
     return [
       '',
       'What you learned about this project in earlier sessions. Trust it, but verify anything',
       'that contradicts what you can see in the code now:',
-      ...ranked.map((e) => `- (${KIND_LABEL[e.kind]}) ${e.text}`),
+      ...ranked.slice(0, limit).map((e) => `- (${KIND_LABEL[e.kind]}) ${e.text}`),
     ].join('\n');
   }
 
@@ -184,6 +285,48 @@ export class Memory {
     return { before, after: this.entries.length };
   }
 
+  /**
+   * Suggest 1-3 memory candidates from a transcript. Used as afterTurn hook for
+   * lifelong learning; caller decides whether to persist (via add).
+   */
+  async suggestFromTranscript(
+    messages: { role: string; content: unknown }[],
+    modelOverride?: LanguageModel,
+  ): Promise<{ kind: MemoryKind; text: string }[]> {
+    const model = modelOverride ?? this.model;
+    if (!model) return [];
+    if ((model as unknown as { provider?: string }).provider === 'unconfigured') return [];
+    if (messages.length < 4) return [];
+    try {
+      const slice = messages.slice(-20);
+      const transcript = slice
+        .map((m) => {
+          const c = typeof m.content === 'string' ? m.content : JSON.stringify(m.content).slice(0, 2000);
+          return `${m.role}: ${c}`;
+        })
+        .join('\n')
+        .slice(0, 8000);
+      const { text } = await generateText({
+        model,
+        system:
+          'Extract 0-3 durable learnings from this coding session that will still be true next session in THIS repo. ' +
+          'Keep repo-specific paths/names/commands — they are the value. ' +
+          'Only decisions with reason, traps with cause+fix, or working commands — not narration. ' +
+          'Output one per line as [fact|decision|gotcha|command] text, or empty if nothing durable.',
+        prompt: transcript,
+        maxRetries: 1,
+      });
+      return text
+        .split('\n')
+        .map((line) => /^\s*\[(fact|decision|gotcha|command)\]\s*(.+?)\s*$/i.exec(line))
+        .filter((m): m is RegExpExecArray => m !== null)
+        .map((m) => ({ kind: m[1]!.toLowerCase() as MemoryKind, text: m[2]!.trim().slice(0, MAX_TEXT) }))
+        .slice(0, 3);
+    } catch {
+      return [];
+    }
+  }
+
   tools() {
     return {
       remember: tool({

src/prune.ts

@@ -173,8 +173,37 @@ export function prunePreservingItems(options: PruneOptions): ModelMessage[] {
  * One agent step is two messages — the assistant's tool call and the tool message
  * answering it — so 64 is about 32 steps of memory.
  */
+export function droppedSpan(before: ModelMessage[], after: ModelMessage[]): ModelMessage[] {
+  const norm = (m: ModelMessage) => {
+    const c = (m as { content?: unknown }).content;
+    if (typeof c === 'string') return `${m.role}:${c}`;
+    try {
+      // Ignore reasoning parts and all providerOptions: a kept-but-detached
+      // message (reasoning stripped, itemId removed) is not considered dropped.
+      const filtered = Array.isArray(c)
+        ? c.filter((p: unknown) => (p as { type?: string }).type !== 'reasoning')
+        : c;
+      const stripped = JSON.stringify(filtered, (k, v) => (k === 'providerOptions' ? undefined : k === 'itemId' ? undefined : v));
+      return `${m.role}:${stripped}`;
+    } catch {
+      return `${m.role}:${String(c)}`;
+    }
+  };
+  const keptNorm = new Set(after.map(norm));
+  return before.filter((m) => !keptNorm.has(norm(m)));
+}
+
 const KEEP_LADDER = [64, 32, 16, 8, 4] as const;
 
+/**
+ * Token estimate used by the session harness. `len/4` undercounts tool envelopes
+ * (role + toolCallId + providerOptions); `len/3.6 + 8*msgs` tracks cl100k closer
+ * without pulling a tokenizer. Exported so session and tests share it.
+ */
+export function estimateTokens(messages: ModelMessage[]): number {
+  return Math.round(JSON.stringify(messages).length / 3.6 + messages.length * 8);
+}
+
 export type FitOptions = {
   messages: ModelMessage[];
   /** Estimated tokens the wire history must come in under. */
@@ -202,21 +231,46 @@ export type FitOptions = {
  * not, the narrowest is returned, because sending something is better than sending a
  * request that will be rejected for size.
  */
+/** Head that must never be pruned: the initial user goal and first assistant ack. */
+function headOf(messages: ModelMessage[]): ModelMessage[] {
+  if (messages.length === 0) return [];
+  const firstUser = messages.find((m) => m.role === 'user');
+  if (!firstUser) return [];
+  // Keep first user message; if an assistant immediately follows, keep it too (goal ack).
+  const idx = messages.indexOf(firstUser);
+  const next = messages[idx + 1];
+  if (next && next.role === 'assistant' && idx === 0) return messages.slice(0, 2);
+  return [firstUser];
+}
+
+function withHeadPreserved(all: ModelMessage[], pruned: ModelMessage[]): ModelMessage[] {
+  const head = headOf(all);
+  if (head.length === 0) return pruned;
+  // identity + full content equality (not 80-char slice which collides)
+  const headNorm = JSON.stringify(head[0]!.content);
+  if (pruned.some((m) => m === head[0]! || JSON.stringify(m.content) === headNorm)) return pruned;
+  return [...head, ...pruned];
+}
+
 export function pruneToFit({ messages, threshold, estimate }: FitOptions): ModelMessage[] {
-  const withoutReasoning = detachProviderItems(
-    prunePreservingItems({ messages, reasoning: 'all', emptyMessages: 'remove' }),
+  const withoutReasoning = withHeadPreserved(
+    messages,
+    detachProviderItems(prunePreservingItems({ messages, reasoning: 'all', emptyMessages: 'remove' })),
   );
   if (estimate(withoutReasoning) <= threshold) return withoutReasoning;
 
   let narrowest = withoutReasoning;
   for (const keep of KEEP_LADDER) {
-    narrowest = detachProviderItems(
-      prunePreservingItems({
-        messages,
-        reasoning: 'all',
-        toolCalls: `before-last-${keep}-messages`,
-        emptyMessages: 'remove',
-      }),
+    narrowest = withHeadPreserved(
+      messages,
+      detachProviderItems(
+        prunePreservingItems({
+          messages,
+          reasoning: 'all',
+          toolCalls: `before-last-${keep}-messages`,
+          emptyMessages: 'remove',
+        }),
+      ),
     );
     if (estimate(narrowest) <= threshold) return narrowest;
   }

src/session.ts

@@ -17,8 +17,9 @@ import { Permissions, type PermissionConfig } from './permission';
 import type { PluginHost } from './plugins';
 import { costOf, formatUsd } from './pricing';
 import { systemPrompt } from './prompt';
-import { detachProviderItems, pruneToFit } from './prune';
+import { detachProviderItems, droppedSpan, estimateTokens as pruneEstimateTokens, pruneToFit } from './prune';
 import { createSkillTool, renderSkills, type Skill } from './skills';
+import { suggestSkillsFromTranscript, writeAutoSkill } from './skill-learner';
 import { disabledToolNames, onBashOutput, tools as builtinTools, type ToolSetName } from './tools';
 
 export type ApprovalRequest = {
@@ -92,9 +93,15 @@ export type SessionOptions = {
   /** Live stdout/stderr from bash, for a UI that wants progress. */
   onToolOutput?: (id: string, chunk: string) => void;
   onNotebookChange?: (state: NotebookState) => void;
+  /** Cheaper model for background learning; falls back to main model. */
+  learnerModel?: LanguageModel;
+  /** Emit learner notices to the UI. */
+  onNotice?: (text: string) => void;
+  /** Disable background auto-learn (tests). */
+  disableAutoLearn?: boolean;
 };
 
-const estimateTokens = (messages: ModelMessage[]) => Math.round(JSON.stringify(messages).length / 4);
+const estimateTokens = pruneEstimateTokens;
 
 /** Estimated tokens at which the wire history is pruned. */
 const DEFAULT_COMPACT_THRESHOLD = 120_000;
@@ -115,8 +122,65 @@ const isStaleItemError = (error: unknown): boolean =>
 const STALE_ITEM_NOTICE =
   'The provider no longer had part of this session stored. Re-sent the history inline and carried on.';
 
+async function summarizeDiscarded(span: ModelMessage[], model: LanguageModel): Promise<string | undefined> {
+  if (span.length === 0) return undefined;
+  const excerpt = span
+    .map((m) => {
+      const c = typeof m.content === 'string' ? m.content : JSON.stringify(m.content).slice(0, 2000);
+      return `${m.role}: ${c}`;
+    })
+    .join('\n')
+    .slice(0, 6000);
+  if (!excerpt.trim()) return undefined;
+  try {
+    const { text } = await generateText({
+      model,
+      system: 'Summarize the dropped part of a long coding session so nothing important is lost. Keep: user goals, files touched with paths, decisions and why, tool results that matter, and what remains. One short handover note, 3-6 lines, no preamble.',
+      prompt: excerpt,
+      maxRetries: 1,
+    });
+    const t = text.trim();
+    return t.length > 0 ? t : undefined;
+  } catch {
+    return undefined;
+  }
+}
+
 type ApprovalContext = Pick<ApprovalRequest, 'matchedPattern' | 'suggestedPattern' | 'repeated'>;
 
+async function maybeLearn(
+  messages: import('ai').ModelMessage[],
+  mainModel: import('ai').LanguageModel,
+  learnerModel: import('ai').LanguageModel | undefined,
+  memory: import('./memory').Memory | undefined,
+  spend: () => { overWarn: boolean },
+  onNotice?: (t: string) => void,
+): Promise<void> {
+  if (spend().overWarn) return;
+  const model = learnerModel ?? mainModel;
+  if ((model as unknown as { provider?: string }).provider === 'unconfigured') return;
+  // project memory: spesifik repo, keep paths — use cheaper learner model when available
+  if (memory) {
+    try {
+      const cands = await memory.suggestFromTranscript(messages as { role: string; content: unknown }[], model);
+      // at most 1 per turn to avoid spam; best-effort
+      if (cands.length > 0) {
+        const c = cands[0]!;
+        const added = await memory.add(c.kind, c.text);
+        if (added && onNotice) onNotice(`auto-memory: remembered (${c.kind}) ${c.text.slice(0, 80)}`);
+      }
+    } catch (e) { if (onNotice) onNotice(`auto-memory skipped: ${(e as Error).message?.slice(0, 120)}`); }
+  }
+  // general skill: universal pattern
+  try {
+    const skills = await suggestSkillsFromTranscript(messages as { role: string; content: unknown }[], model);
+    for (const c of skills) {
+      const p = await writeAutoSkill(c);
+      if (p && onNotice) onNotice(`auto-skill: ${c.name} → ${p}`);
+    }
+  } catch (e) { if (onNotice) onNotice(`auto-skill skipped: ${(e as Error).message?.slice(0, 120)}`); }
+}
+
 export class Session {
   readonly messages: ModelMessage[];
   readonly tools: ToolSet;
@@ -136,6 +200,8 @@ export class Session {
   /** The 80% spend warning is shown once, not on every turn past the line. */
   private warnedSpend = false;
   private controller: AbortController | undefined;
+  private learnTurns = 0;
+  private lastLearnLen = 0;
 
   constructor(private readonly opts: SessionOptions) {
     this.messages = opts.messages ?? [];
@@ -279,11 +345,13 @@ export class Session {
   }
 
   private systemFor(): string {
+    const mem = this.opts.memory;
+    const memoryBlock = mem ? mem.render() : '';
     return systemPrompt({
       cwd: this.opts.cwd ?? process.cwd(),
       instructions: this.opts.instructions ?? [],
       notebook: this.notebook.render(),
-      memory: this.opts.memory?.render() ?? '',
+      memory: memoryBlock,
       skills: renderSkills(this.opts.skills ?? []),
       agent: renderAgent(this.variant),
       plugins: this.opts.plugins?.appendix ?? '',
@@ -415,6 +483,19 @@ export class Session {
     } finally {
       onBashOutput(undefined);
       await this.opts.plugins?.afterTurn();
+      if (!this.opts.disableAutoLearn && this.messages.length >= 6) {
+        this.learnTurns += 1;
+        const delta = this.messages.length - this.lastLearnLen;
+        const throttled = this.learnTurns % 3 !== 0 && delta < 8;
+        if (!throttled) {
+          this.lastLearnLen = this.messages.length;
+          try {
+            await maybeLearn(this.messages, this.model, this.opts.learnerModel, this.opts.memory, () => this.spend(), this.opts.onNotice);
+          } catch {}
+          // periodic TTL prune so long session doesn't bloat
+          if (this.learnTurns % 6 === 0 && this.opts.memory) { try { await this.opts.memory.pruneExpired(); } catch {} }
+        }
+      }
     }
   }
 
@@ -439,6 +520,7 @@ export class Session {
     // Each iteration is one model run. A run ends either finished, or suspended
     // on tool approvals, in which case we collect decisions and run again.
     let compactionReported = false;
+    let compactionSpan: ModelMessage[] | undefined;
     while (true) {
       const pending: ApprovalRequest[] = [];
       const compactions: Extract<AgentEvent, { type: 'compacted' }>[] = [];
@@ -465,6 +547,10 @@ export class Session {
           const instructions = this.systemFor();
           if (estimateTokens(messages) <= threshold) return { instructions };
           const pruned = pruneToFit({ messages, threshold, estimate: estimateTokens });
+          // Capture what was dropped by reference identity — the lossless note is built after the stream.
+          if (!compactionReported && pruned.length < messages.length) {
+            compactionSpan = droppedSpan(messages, pruned);
+          }
           // prepareStep cannot yield, so queue the notice and drain it in the loop.
           if (!compactionReported) {
             compactions.push({ type: 'compacted', before: messages.length, after: pruned.length });
@@ -581,6 +667,16 @@ export class Session {
       this.messages.push(...(await result.responseMessages));
       this.opts.onChange?.(this.messages);
 
+      // Lossless compaction: summarize what the wire pruned so future turns keep it.
+      if (compactionSpan && compactionSpan.length > 0) {
+        const retained = await summarizeDiscarded(compactionSpan, this.model);
+        if (retained) {
+          this.messages.push({ role: 'user', content: `Note (retained from compacted history):\n${retained}` });
+          this.opts.onChange?.(this.messages);
+        }
+        compactionSpan = undefined;
+      }
+
       if (pending.length === 0) {
         const usage = await result.usage;
         this.inputTokens += usage.inputTokens ?? 0;

src/skill-learner.ts

@@ -0,0 +1,108 @@
+import { generateText, type LanguageModel } from 'ai';
+import { createHash } from 'node:crypto';
+import { homedir } from 'node:os';
+import { join } from 'node:path';
+
+export type SkillCandidate = {
+  name: string;
+  description: string;
+  body: string;
+};
+
+const MAX_BODY = 12_000;
+
+function slugOf(s: string): string {
+  return s.toLowerCase().replace(/[^a-z0-9]+/g, '-').replace(/^-|-$/g, '').slice(0, 48) || 'auto-skill';
+}
+function normalizeBody(s: string): string {
+  return s.trim().toLowerCase().replace(/\s+/g, ' ').replace(/^[\p{P}\s]+|[\p{P}\s]+$/gu, '');
+}
+function hashBody(s: string): string {
+  return createHash('sha256').update(normalizeBody(s)).digest('hex').slice(0, 16);
+}
+function skillDir(): string {
+  return join(process.env['SHIRO_HOME'] ?? homedir(), '.shiro-neko', 'skills');
+}
+
+/** Parse candidates: primary ---SKILL block, fallback JSON array */
+function parseCandidates(rawText: string): SkillCandidate[] {
+  const out: SkillCandidate[] = [];
+  const re = /---SKILL\s+name:\s*([a-z0-9-]{2,48})\s*---\s*description:\s*(.+?)\s*\nbody:\s*\n([\s\S]*?)\s*---END---/gi;
+  let m: RegExpExecArray | null;
+  while ((m = re.exec(rawText)) !== null) {
+    const name = slugOf(m[1]!);
+    const description = m[2]!.trim().slice(0, 120);
+    const body = m[3]!.trim().slice(0, MAX_BODY);
+    if (body.length < 80) continue;
+    if (!/^use when\b/i.test(description)) continue;
+    out.push({ name: `auto-${name}`, description, body });
+    if (out.length >= 2) break;
+  }
+  if (out.length > 0) return out;
+  // JSON fallback: [{name, description, body}]
+  try {
+    const j = JSON.parse(rawText.trim());
+    const arr = Array.isArray(j) ? j : [j];
+    for (const e of arr) {
+      if (!e || typeof e.name !== 'string' || typeof e.description !== 'string' || typeof e.body !== 'string') continue;
+      const name = slugOf(e.name);
+      const description = String(e.description).trim().slice(0, 120);
+      const body = String(e.body).trim().slice(0, MAX_BODY);
+      if (body.length < 80) continue;
+      if (!/^use when\b/i.test(description)) continue;
+      out.push({ name: `auto-${name}`, description, body });
+      if (out.length >= 2) break;
+    }
+  } catch {}
+  return out;
+}
+
+export async function suggestSkillsFromTranscript(
+  messages: { role: string; content: unknown }[],
+  model: LanguageModel,
+): Promise<SkillCandidate[]> {
+  if ((model as unknown as { provider?: string }).provider === 'unconfigured') return [];
+  if (messages.length < 6) return [];
+  try {
+    const transcript = messages.slice(-24).map(m => {
+      const c = typeof m.content === 'string' ? m.content : JSON.stringify(m.content).slice(0, 2000);
+      return `${m.role}: ${c}`;
+    }).join('\n').slice(0, 9000);
+    if (!transcript.trim()) return [];
+    const { text } = await generateText({
+      model,
+      system: 'You distill UNIVERSAL, cross-repo skills from a coding transcript. Abstract away repo-specific paths/names/IDs into roles (e.g. "the ORM migration" not "src/db/migrate.ts"). Each skill must be a reusable pattern: when to use it, steps, why. Worth keeping for months across projects. Output 0-2 skills, each as:\n---SKILL name: kebab-case---\ndescription: one-line trigger, <80 chars, starts with "Use when ..."\nbody:\nmarkdown body (headings, steps, pitfalls) — no frontmatter, no Line truncated
+      prompt: transcript,
+      maxRetries: 1,
+    });
+    return parseCandidates(text);
+  } catch { return []; }
+}
+
+export async function writeAutoSkill(candidate: SkillCandidate): Promise<string | undefined> {
+  const dir = skillDir();
+  const { mkdirSync } = await import('node:fs');
+  try { mkdirSync(dir, { recursive: true }); } catch {}
+  const path = join(dir, `${candidate.name}.md`);
+  const safeDesc = candidate.description.replace(/[\n\r]+/g, ' ').replace(/"/g, "'").trim().slice(0, 120);
+  const frontmatter = `---\nname: ${candidate.name}\ndescription: ${safeDesc}\n---\n\n`;
+  const content = `${frontmatter}${candidate.body.trim()}\n`;
+  const candHash = hashBody(candidate.body);
+  try {
+    const existing = await Bun.file(path).text();
+    // exact hash dedup: body already present in any section
+    const existingHash = hashBody(existing);
+    if (existingHash === candHash) return undefined;
+    // also check if candidate body hash appears as substring section hash
+    // split existing by ## Update markers and hash each chunk
+    const chunks = existing.split(/## Update/);
+    for (const ch of chunks) if (hashBody(ch) === candHash) return undefined;
+    const appended = `${existing.trimEnd()}\n\n---\n\n## Update \u2014 ${new Date().toISOString().slice(0, 10)}\n\n${candidate.body.trim()}\n`;
+    if (appended.length > 20_000) return undefined;
+    await Bun.write(path, appended);
+    return path;
+  } catch {
+    await Bun.write(path, content);
+    return path;
+  }
+}

test/compact.test.ts

@@ -358,6 +358,69 @@ test('a compacted turn inlines a plain assistant item instead of referencing rem
  * rejected with 404 "Item with id 'msg_...' not found", and every retry of the same
  * history is rejected the same way, so resuming a session could never get going.
  */
+
+test('lossless compaction appends a retained note when tool content was dropped', async () => {
+  const messages = [...bulkyExchange(0), ...bulkyExchange(1), ...bulkyExchange(2), ...bulkyExchange(3), ...bulkyExchange(4)];
+  let generateCalls = 0;
+  const session = new Session({
+    messages: [...messages],
+    compactThreshold: 1500,
+    disableAutoLearn: true,
+    model: new MockLanguageModelV4({
+      doStream: async () => stream(text('ok')),
+      doGenerate: async () => {
+        generateCalls++;
+        return { content: [{ type: 'text', text: 'Retained: files f0..f2' }], finishReason: { unified: 'stop', raw: 'stop' }, usage, warnings: [] } as any;
+      },
+    }),
+    askApproval: async () => 'deny',
+  });
+  const events: AgentEvent[] = [];
+  for await (const ev of session.send('next')) events.push(ev);
+  expect(generateCalls).toBe(1);
+  expect(events.some((e) => e.type === 'compacted')).toBe(true);
+  expect(session.messages.some((m) => String(m.content).includes('retained from compacted'))).toBe(true);
+});
+
+test('no retained note when history fits', async () => {
+  let generateCalls = 0;
+  const session = new Session({
+    messages: [...bulkyExchange(0)],
+    compactThreshold: 1_000_000,
+    disableAutoLearn: true,
+    model: new MockLanguageModelV4({
+      doStream: async () => stream(text('ok')),
+      doGenerate: async () => {
+        generateCalls++;
+        return { content: [{ type: 'text', text: 'should not be called' }], finishReason: { unified: 'stop', raw: 'stop' }, usage, warnings: [] } as any;
+      },
+    }),
+    askApproval: async () => 'deny',
+  });
+  for await (const _ of session.send('next')) void _;
+  expect(generateCalls).toBe(0);
+  expect(session.messages.some((m) => String(m.content).includes('retained from compacted'))).toBe(false);
+});
+
+test('a failing retained-note model does not break the turn', async () => {
+  const messages = [...bulkyExchange(0), ...bulkyExchange(1), ...bulkyExchange(2), ...bulkyExchange(3)];
+  const session = new Session({
+    messages: [...messages],
+    compactThreshold: 1000,
+    disableAutoLearn: true,
+    model: new MockLanguageModelV4({
+      doStream: async () => stream(text('ok')),
+      doGenerate: async () => { throw new Error('down'); },
+    }),
+    askApproval: async () => 'deny',
+  });
+  const events: AgentEvent[] = [];
+  for await (const ev of session.send('next')) events.push(ev);
+  expect(events.map((e) => e.type)).toContain('done');
+  expect(events.map((e) => e.type)).not.toContain('error');
+  expect(session.messages.some((m) => String(m.content).includes('retained from compacted'))).toBe(false);
+});
+
 const staleItem = (id: string) =>
   new APICallError({
     message: `Item with id '${id}' not found.`,

test/memory.test.ts

@@ -82,7 +82,7 @@ test('an empty note is refused', async () => {
 test('long text is truncated', async () => {
   const m = new Memory('/repo');
   const entry = await m.add('fact', 'x'.repeat(2000));
-  expect(entry?.text.length).toBe(400);
+  expect(entry?.text.length).toBe(800);
 });
 
 test('search requires every term and records a hit', async () => {

test/prune.test.ts

@@ -1,6 +1,6 @@
 import { expect, test } from 'bun:test';
 import type { ModelMessage } from 'ai';
-import { detachOrphanedItems, dropOrphanedResults, pruneToFit, prunePreservingItems } from '../src/prune';
+import { detachOrphanedItems, droppedSpan, dropOrphanedResults, estimateTokens, pruneToFit, prunePreservingItems } from '../src/prune';
 
 const kinds = (messages: ModelMessage[]) =>
   messages.map((m) => (Array.isArray(m.content) ? `${m.role}:${m.content.map((p) => p.type).join('+')}` : m.role));
@@ -438,6 +438,34 @@ test('compaction removes a plain assistant item reference without inline reasoni
   expect(JSON.stringify(fitted)).toContain('answer');
 });
 
+test('droppedSpan is empty when only reasoning was removed', () => {
+  const messages: ModelMessage[] = [
+    { role: 'user', content: 'q' },
+    {
+      role: 'assistant',
+      content: [
+        { type: 'reasoning', text: 't', providerOptions: { openai: { itemId: 'rs1' } } },
+        { type: 'text', text: 'ans', providerOptions: { openai: { itemId: 'msg1' } } },
+      ],
+    },
+  ];
+  // pruneToFit with high threshold only strips reasoning, keeping ans
+  const fitted = pruneToFit({ messages, threshold: 20000, estimate: estimateTokens });
+  expect(droppedSpan(messages, fitted)).toEqual([]);
+});
+
+test('droppedSpan captures pruned tool content', () => {
+  const msgs: ModelMessage[] = [{ role: 'user', content: 'do thing' }];
+  for (let i = 0; i < 20; i++) {
+    msgs.push({ role: 'assistant', content: [{ type: 'tool-call', toolCallId: `t${i}`, toolName: 'grep', input: { pattern: 'x' } }] });
+    msgs.push({ role: 'tool', content: [{ type: 'tool-result', toolCallId: `t${i}`, toolName: 'grep', output: { type: 'text', value: 'x'.repeat(3000) } }] });
+  }
+  const fitted = pruneToFit({ messages: msgs, threshold: 6000, estimate: estimateTokens });
+  const span = droppedSpan(msgs, fitted);
+  expect(span.length).toBeGreaterThan(0);
+  expect(span.length + fitted.length).toBe(msgs.length);
+});
+
 test('the user prompt survives even the narrowest rung', () => {
   const messages = transcript(200, 4000);
   const fitted = pruneToFit({ messages, threshold: 100, estimate });