aiden-runtime 4.1.1 → 4.1.2

package/dist/core/v4/sessionDistiller.js

@@ -0,0 +1,405 @@
+"use strict";
+/**
+ * Copyright (c) 2026 Shiva Deore (Taracod).
+ * Licensed under AGPL-3.0. See LICENSE for details.
+ *
+ * Aiden — local-first agent.
+ */
+/**
+ * core/v4/sessionDistiller.ts — Phase v4.1.2-memory-AB.
+ *
+ * Replaces the lossy 5-bullet auxiliary summary with a structured
+ * SessionDistillation:
+ *
+ *   - bullets[]       (5 bullets, back-compat with MEMORY.md `## Recent sessions`)
+ *   - decisions[]     (higher-fidelity than bullets)
+ *   - open_items[]    (unfinished work, useful for next session)
+ *   - keywords[]      (for future retrieval ranking — Phase C)
+ *   - files_touched[] (DETERMINISTIC — derived from tool-call result payloads)
+ *   - tools_used[]    (DETERMINISTIC — counted from tool-call trace names)
+ *   - schema_version  (always 1; reserved for future migrations)
+ *   - exit_path       (which exit caused the distillation: quit/sigint/etc.)
+ *   - partial         (set true when LLM JSON parse falls back to bullets-only)
+ *
+ * Source-of-truth split:
+ *   - Programmatic fields (files_touched, tools_used) → trace inspection.
+ *   - Semantic fields (bullets, decisions, open_items, keywords) → single
+ *     auxiliary-LLM call with strict-then-lenient JSON parsing.
+ *
+ * Phase A's CLI ChatSession owns the per-session HonestyTraceEntry[]
+ * accumulator and passes it here. The auxiliary call sees the full
+ * message history (not the trace — the trace is purely for programmatic
+ * field derivation).
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.TOOL_RESULT_TRUNCATION = exports.SESSION_DISTILLATION_SCHEMA_VERSION = void 0;
+exports.deriveProgrammaticFields = deriveProgrammaticFields;
+exports.parseLLMDistillation = parseLLMDistillation;
+exports.filterMessagesForDistillation = filterMessagesForDistillation;
+exports.distillSession = distillSession;
+// ── Public surface ───────────────────────────────────────────────────────
+exports.SESSION_DISTILLATION_SCHEMA_VERSION = 1;
+// ── Programmatic field derivation ─────────────────────────────────────────
+/**
+ * Tools whose result payload SHOULD contain a `path` field naming the
+ * file they touched. Used to populate `files_touched`.
+ *
+ * Curated rather than "any tool with a path in its result" because
+ * read-only tools (`file_read`, `file_list`) shouldn't count as
+ * "touched" — only mutating ops do.
+ */
+const FILE_TOUCH_TOOLS = new Set([
+    'file_write',
+    'file_patch',
+    'file_create',
+    'file_delete',
+    'memory_add', // writes MEMORY.md / USER.md
+    'memory_remove',
+    'memory_replace',
+    'session_summary', // writes MEMORY.md
+]);
+/**
+ * Extract programmatic fields from the accumulated tool trace. Pure
+ * function — no I/O.
+ */
+function deriveProgrammaticFields(trace) {
+    // tools_used: count by name, sorted by count desc, name asc.
+    const counts = new Map();
+    for (const e of trace) {
+        counts.set(e.name, (counts.get(e.name) ?? 0) + 1);
+    }
+    const tools_used = Array.from(counts.entries())
+        .map(([name, count]) => ({ name, count }))
+        .sort((a, b) => b.count - a.count || a.name.localeCompare(b.name));
+    // files_touched: unique paths from mutating tool results.
+    // Each entry's `result` may be { success, path, ... } or { path: ... }
+    // depending on the tool. We accept either shape.
+    const paths = new Set();
+    for (const e of trace) {
+        if (e.error)
+            continue; // failed tool — don't credit
+        if (!FILE_TOUCH_TOOLS.has(e.name))
+            continue;
+        const candidate = extractPath(e.result);
+        if (candidate)
+            paths.add(candidate);
+    }
+    const files_touched = Array.from(paths).sort();
+    return { files_touched, tools_used };
+}
+function extractPath(result) {
+    if (!result || typeof result !== 'object')
+        return null;
+    // Top-level path field — most write tools.
+    const top = result.path;
+    if (typeof top === 'string' && top.length > 0)
+        return top;
+    // Nested under .result (some adapters wrap output).
+    const inner = result.result;
+    if (inner && typeof inner === 'object') {
+        const innerPath = inner.path;
+        if (typeof innerPath === 'string' && innerPath.length > 0)
+            return innerPath;
+    }
+    return null;
+}
+// ── LLM extraction ────────────────────────────────────────────────────────
+/**
+ * Strict-then-lenient parser for the auxiliary LLM's distillation JSON.
+ *
+ * Strict path: parse as JSON, validate shape, return all four semantic
+ * fields. Lenient path (only when strict fails): try to extract a
+ * bullets array from a malformed body (codepath shared with slice2's
+ * parseSessionBulletsResponse fallback), set the other three fields to
+ * empty arrays, and signal `partial: true` to the caller.
+ *
+ * Pure function — no I/O. Caller decides what to do with `partial`.
+ */
+function parseLLMDistillation(raw) {
+    const trimmed = raw.trim();
+    if (!trimmed) {
+        return { bullets: [], decisions: [], open_items: [], keywords: [], partial: true };
+    }
+    // Strict path.
+    const strict = tryStrictParse(trimmed);
+    if (strict)
+        return { ...strict, partial: false };
+    // Lenient: scan for a JSON object embedded in prose (some models
+    // prefix "Here is the JSON:\n{...}"). Trim to the first '{' through
+    // the last '}' and retry.
+    const first = trimmed.indexOf('{');
+    const last = trimmed.lastIndexOf('}');
+    if (first >= 0 && last > first) {
+        const inner = trimmed.slice(first, last + 1);
+        const second = tryStrictParse(inner);
+        if (second)
+            return { ...second, partial: false };
+    }
+    // Bullets-only fallback — recover what we can. Tries a bare bullet
+    // list ("- ...", "* ...", numbered lines) or a JSON-array fragment.
+    const fallbackBullets = recoverBullets(trimmed);
+    return {
+        bullets: fallbackBullets,
+        decisions: [],
+        open_items: [],
+        keywords: [],
+        partial: true,
+    };
+}
+function tryStrictParse(s) {
+    try {
+        const obj = JSON.parse(s);
+        if (!obj || typeof obj !== 'object' || Array.isArray(obj))
+            return null;
+        const o = obj;
+        const bullets = toStringArray(o.bullets);
+        const decisions = toStringArray(o.decisions);
+        const open_items = toStringArray(o.open_items ?? o.openItems);
+        const keywords = toStringArray(o.keywords);
+        if (bullets.length === 0 && decisions.length === 0 && open_items.length === 0) {
+            return null; // nothing useful — let the lenient path try
+        }
+        return { bullets, decisions, open_items, keywords };
+    }
+    catch {
+        return null;
+    }
+}
+function toStringArray(v) {
+    if (!Array.isArray(v))
+        return [];
+    return v
+        .filter((x) => typeof x === 'string')
+        .map((x) => x.trim())
+        .filter((x) => x.length > 0);
+}
+function recoverBullets(raw) {
+    // Strategy 1: bullet-prefixed lines.
+    const lines = raw.split(/\r?\n/);
+    const bulleted = lines
+        .map((l) => l.replace(/^\s*(?:[-*•]|\d+\.)\s+/, '').trim())
+        .filter((l, i, arr) => l.length > 0 && /^\s*(?:[-*•]|\d+\.)\s+/.test(lines[i] ?? ''));
+    if (bulleted.length > 0)
+        return bulleted.slice(0, 5);
+    // Strategy 2: a JSON array of strings, with or without the object wrapper.
+    const arrMatch = raw.match(/\[\s*"[\s\S]*?"\s*\]/);
+    if (arrMatch) {
+        try {
+            const arr = JSON.parse(arrMatch[0]);
+            return toStringArray(arr).slice(0, 5);
+        }
+        catch { /* fall through */ }
+    }
+    return [];
+}
+const DEFAULT_TIMEOUT_MS = 4000;
+/**
+ * Phase v4.1.2-bug-Y: max chars of tool-result content surfaced to the
+ * auxiliary LLM. Covers typical error messages + JSON-payload heads
+ * without bloating the prompt with full tool-output dumps. User and
+ * assistant TEXT are never truncated — user intent must survive in
+ * full. Widen this only after eval shows truncation eating signal.
+ */
+exports.TOOL_RESULT_TRUNCATION = 200;
+/**
+ * Pure: filter + format the conversation history into the transcript
+ * the auxiliary LLM sees. Phase v4.1.2-bug-Y root-cause fix:
+ *
+ *   The previous distiller dumped chatSession.history verbatim,
+ *   including the giant `role: 'system'` block PromptBuilder
+ *   constructs (SOUL.md identity, MEMORY.md, USER.md, Runtime slot,
+ *   Capabilities boilerplate, tool-catalog descriptions, personality
+ *   overlay, execution-discipline notes). Weak summarizer models
+ *   latched onto this longest-coherent-block in context as the
+ *   session topic, returning bullets like "I'm Aiden, a local-first
+ *   AI agent built by Taracod" regardless of what the user and
+ *   assistant actually discussed.
+ *
+ *   This filter drops ALL `role: 'system'` messages and emits the
+ *   remaining traffic as role-tagged lines:
+ *
+ *     [USER] full user message verbatim
+ *     [ASSISTANT] assistant text (if non-empty)
+ *     [TOOL:name] {args}
+ *     [TOOL:name] → result-payload, truncated to TOOL_RESULT_TRUNCATION
+ *
+ *   Tool results carry their tool name (resolved via toolCallId →
+ *   call-name map walked through preceding assistant turns) so the
+ *   model can correlate tool intent with output. Empty messages are
+ *   dropped entirely. Multi-line content within a message is
+ *   preserved.
+ */
+function filterMessagesForDistillation(messages) {
+    /** Per-toolCallId → toolName, populated as we walk assistant turns. */
+    const callNames = new Map();
+    const lines = [];
+    for (const m of messages) {
+        if (m.role === 'system')
+            continue; // entire boilerplate source — dropped
+        if (m.role === 'user') {
+            const text = m.content.trim();
+            if (text.length === 0)
+                continue;
+            lines.push(`[USER] ${text}`);
+            continue;
+        }
+        if (m.role === 'assistant') {
+            // Emit assistant text only if non-empty — avoid empty `[ASSISTANT]`
+            // placeholder for tool-only turns.
+            const text = (m.content ?? '').trim();
+            if (text.length > 0)
+                lines.push(`[ASSISTANT] ${text}`);
+            // Tool calls: cache the id → name pair so the matching tool
+            // result downstream can render with its tool name. Emit the
+            // call line in original order.
+            if (m.toolCalls && m.toolCalls.length > 0) {
+                for (const tc of m.toolCalls) {
+                    callNames.set(tc.id, tc.name);
+                    const argsStr = compactArgs(tc.arguments);
+                    lines.push(`[TOOL:${tc.name}] ${argsStr}`);
+                }
+            }
+            continue;
+        }
+        if (m.role === 'tool') {
+            const name = callNames.get(m.toolCallId) ?? 'unknown';
+            const truncated = truncateForTranscript(m.content);
+            lines.push(`[TOOL:${name}] → ${truncated}`);
+            continue;
+        }
+    }
+    return lines.join('\n');
+}
+/**
+ * Compact tool-call args into a one-line representation. JSON shape
+ * preserved; large strings get truncated alongside everything else
+ * to keep the transcript focused on intent, not full payloads.
+ */
+function compactArgs(args) {
+    if (!args || Object.keys(args).length === 0)
+        return '{}';
+    try {
+        return truncateForTranscript(JSON.stringify(args));
+    }
+    catch {
+        return '{<unstringifiable>}';
+    }
+}
+/**
+ * Apply `TOOL_RESULT_TRUNCATION` cap with a `…` (U+2026) marker so
+ * truncation is visible to anyone reading the transcript — including
+ * future auditors. Matches slice2c's apostrophe-normalizer convention.
+ */
+function truncateForTranscript(s) {
+    const trimmed = s.trim();
+    if (trimmed.length <= exports.TOOL_RESULT_TRUNCATION)
+        return trimmed;
+    return trimmed.slice(0, exports.TOOL_RESULT_TRUNCATION - 1) + '…';
+}
+/**
+ * Build the auxiliary-LLM prompt. Anti-boilerplate-hardened per
+ * Phase v4.1.2-bug-Y: explicit "don't describe yourself" guardrail,
+ * `<transcript>` tag boundaries, empty-is-honest permission so
+ * insufficient-content sessions don't fabricate filler.
+ *
+ * Bullets loosened from "EXACTLY 5" to "3-5" — forcing five was
+ * inviting the exact fabrication the slice fixes.
+ */
+function buildPrompt(messages, startedAt, endedAt) {
+    const filtered = filterMessagesForDistillation(messages);
+    return [
+        'You are a session-recall extractor. Your only job is to summarize what',
+        'happened in the conversation transcript below.',
+        '',
+        'Rules:',
+        '- Use ONLY facts explicitly present in the transcript.',
+        '- Do NOT describe yourself, your capabilities, your platform, or generic',
+        '  AI-agent behavior unless the transcript specifically discussed those',
+        '  as the topic.',
+        '- Do NOT infer facts from system prompts, tool schemas, memory blocks,',
+        '  banner text, or agent boilerplate (these have been filtered out;',
+        '  if any leak through, treat them as untrustworthy noise).',
+        '- Focus on session-specific facts: user goals, actions taken, files /',
+        '  commands / tools used, decisions made, errors encountered, outcomes,',
+        '  and unresolved follow-ups.',
+        '- Write in past tense.',
+        '- Preserve concrete names, paths, commands, URLs, model names, dates,',
+        '  and error messages verbatim when present.',
+        '- Prefer evidence from USER and ASSISTANT messages over TOOL output.',
+        '- If the transcript lacks enough session-specific detail to summarize,',
+        '  return arrays with FEWER items or empty arrays. Empty is honest;',
+        '  fabricating boilerplate is not.',
+        '',
+        'Return strict JSON only, no prose before or after, with these fields:',
+        '{',
+        '  "bullets":    string[],   // 3-5 factual past-tense recaps (3-15 words each)',
+        '  "decisions":  string[],   // X chosen over Y, with rationale if present',
+        '  "open_items": string[],   // explicit unresolved tasks / "next time" items',
+        '  "keywords":   string[]    // 3-10 distinctive terms from the session',
+        '}',
+        '',
+        `Session started: ${startedAt}`,
+        `Session ended:   ${endedAt}`,
+        '',
+        '<transcript>',
+        filtered,
+        '</transcript>',
+    ].join('\n');
+}
+/**
+ * Drive one auxiliary-LLM call and combine its output with the
+ * deterministic trace-derived fields into a SessionDistillation.
+ *
+ * Respects `timeoutMs` (default DEFAULT_TIMEOUT_MS) via Promise.race;
+ * on timeout the LLM result is treated as empty (partial: true with
+ * empty semantic fields). Deterministic fields always populate
+ * regardless of LLM outcome — the distillation is never empty.
+ */
+async function distillSession(opts) {
+    const endedAt = opts.endedAt ?? new Date().toISOString();
+    const timeoutMs = opts.timeoutMs ?? DEFAULT_TIMEOUT_MS;
+    const programmatic = deriveProgrammaticFields(opts.toolTrace);
+    // Run the auxiliary call under a hard timeout. The race resolves
+    // with `{timedOut: true}` if the LLM doesn't return in time — we
+    // record that as a partial distillation.
+    const prompt = buildPrompt(opts.messages, opts.startedAt, endedAt);
+    const llmRaw = await Promise.race([
+        opts.auxiliaryClient
+            .call({ purpose: 'session_summary', prompt, maxTokens: 800 })
+            .then((r) => ({ ok: true, content: r.content ?? '' }))
+            .catch((e) => ({ ok: false, error: e })),
+        new Promise((resolve) => {
+            setTimeout(() => resolve({ ok: false, error: new Error(`auxiliary call timed out after ${timeoutMs}ms`), timedOut: true }), timeoutMs);
+        }),
+    ]);
+    let semantic;
+    if (llmRaw.ok) {
+        semantic = parseLLMDistillation(llmRaw.content);
+    }
+    else {
+        semantic = {
+            bullets: [],
+            decisions: [],
+            open_items: [],
+            keywords: [],
+            partial: true,
+        };
+    }
+    const dist = {
+        schema_version: exports.SESSION_DISTILLATION_SCHEMA_VERSION,
+        session_id: opts.sessionId,
+        started_at: opts.startedAt,
+        ended_at: endedAt,
+        exit_path: opts.exitPath,
+        user_turns: opts.userTurns,
+        bullets: semantic.bullets,
+        decisions: semantic.decisions,
+        open_items: semantic.open_items,
+        keywords: semantic.keywords,
+        files_touched: programmatic.files_touched,
+        tools_used: programmatic.tools_used,
+    };
+    if (semantic.partial)
+        dist.partial = true;
+    return dist;
+}

package/dist/core/v4/skillMining/skillMiner.js

@@ -101,6 +101,7 @@ class SkillMiner {
         this.auxiliaryClient = opts.auxiliaryClient;
         this.sessionCap = opts.sessionCap ?? exports.SESSION_SKILL_LIMIT;
         this.skeletonOnly = opts.skeletonOnly ?? false;
+        this.healthTracker = opts.healthTracker;
     }
     /** Test/reset hook. */
     _resetForTests() {
@@ -133,14 +134,32 @@ class SkillMiner {
                 return { status: 'opt-out' };
             }
         }
-        // Fingerprint + dedup.
+        // Fingerprint + dedup. Phase v4.1.2-slice3: wrap the store reads
+        // so disk-level failures (read of pending list, read of rejected
+        // list) surface to the health tracker. We re-throw to preserve
+        // existing crash-on-disk-error semantics — the surface is purely
+        // observational.
         const fpEntries = obs.trace.map((e) => ({ name: e.name, args: e.args }));
         const fingerprint = (0, traceFingerprint_1.traceFingerprint)(fpEntries);
-        const pending = await this.store.list();
+        let pending;
+        try {
+            pending = await this.store.list();
+        }
+        catch (e) {
+            this.healthTracker?.recordFailure(e);
+            throw e;
+        }
         if (pending.some((c) => c.fingerprint === fingerprint)) {
             return { status: 'dedup-pending' };
         }
-        const rejected = await this.store.loadRejected();
+        let rejected;
+        try {
+            rejected = await this.store.loadRejected();
+        }
+        catch (e) {
+            this.healthTracker?.recordFailure(e);
+            throw e;
+        }
         if (rejected.has(fingerprint)) {
             return { status: 'dedup-rejected' };
         }
@@ -163,7 +182,17 @@ class SkillMiner {
         };
         let skill = (0, proposalBuilder_1.draft)(obs.trace, ctx);
         if (!this.skeletonOnly) {
-            skill = await (0, extractorPrompt_1.refine)(skill, { client: this.auxiliaryClient });
+            // Phase v4.1.2-slice3: refine is an LLM call that historically
+            // crashed the whole turn when it threw. Wrap and surface; on
+            // failure fall back to the skeleton so mining still produces
+            // something rather than nothing.
+            try {
+                skill = await (0, extractorPrompt_1.refine)(skill, { client: this.auxiliaryClient });
+            }
+            catch (e) {
+                this.healthTracker?.recordFailure(e);
+                // Skeleton retained; carry on.
+            }
         }
         // Final validation — must round-trip through parseSkillContent
         // (the canonical loader parser). If it doesn't, drop the
@@ -171,7 +200,8 @@ class SkillMiner {
         try {
             (0, skillSpec_1.parseSkillContent)(skill);
         }
-        catch {
+        catch (e) {
+            this.healthTracker?.recordFailure(e);
             return { status: 'invalid-skill', confidence };
         }
         const candidate = {
@@ -183,8 +213,15 @@ class SkillMiner {
             candidateConfidence: confidence,
             skillContent: skill,
         };
-        await this.store.append(candidate);
+        try {
+            await this.store.append(candidate);
+        }
+        catch (e) {
+            this.healthTracker?.recordFailure(e);
+            throw e;
+        }
         this.perSessionCount.set(obs.sessionId, this.countForSession(obs.sessionId) + 1);
+        this.healthTracker?.recordSuccess();
         return { status: 'queued', candidate, confidence };
     }
 }