aiden-runtime 4.6.1 → 4.8.0

package/dist/moat/honestyEnforcement.js

@@ -6,111 +6,92 @@
  * Aiden — local-first agent.
  */
 /**
- * moat/honestyEnforcement.ts — Aiden v4.0.0
+ * moat/honestyEnforcement.ts — Aiden v4.7.0 (Phase 2.3 — outcome-based verifier)
  *
- * Post-loop trace inspector. Runs after AidenAgent returns its final
- * response. Compares the response's stated actions to the actual tool
- * calls in the trace. If the model claims it did something but the trace
- * says no tool fired (or fired and failed verification), Honesty refuses
- * the claim and rewrites the response.
+ * The regex-based natural-language claim scanner (deleted in Phase 2.2)
+ * has been replaced with a deterministic outcome recorder that consumes
+ * `toolCallTrace` structurally. Two failure modes are recorded:
  *
- * The failure modes this catches:
- *   - "I saved your file to ~/notes/today.md"   → no file_write call
- *   - "I sent the email"                         → no email tool call
- *   - "I remembered that"                        → no memory_add OR memory_add returned verified=false
- *   - "I searched the web"                       → no web_search call
- *   - "I ran X"                                  → no shell_exec call
+ *   1. mutation_errored  — a tool tagged `mutates: true` (in the
+ *      registry, stamped onto trace entries at dispatch time via
+ *      `handlerMutates`) returned an `error` envelope. Path is
+ *      extracted from `result.path` when present.
  *
- * Three modes:
- *   off      — passes everything; no inspection.
- *   detect   — runs checks and populates findings, but does NOT modify the
- *              response. Useful for telemetry / canary measurement.
- *   enforce  — DEFAULT. Rewrites failed claims into honest text that lists
- *              the actual trace summary.
+ *   2. memory_unverified — a memory_* tool's result carries
+ *      `verified === false` (per Phase 9 MemoryGuard). This was
+ *      the v3 C20/C21 lying surface and remains the only memory-
+ *      specific check the verifier performs.
  *
- * Detection:
- *   1. Pattern-based (default, $0 cost) — past-tense action verbs matched
- *      against tool registry. This file owns the table.
- *   2. LLM-classified — auxiliary LLM call. Wired via the optional
- *      `llmAdapter`; defaulted off in Phase 12. Phase 13 turns it on.
+ * Modes:
+ *   off      — bypass entirely. No events recorded.
+ *   detect   — Record events; never user-visible. `findings` populated;
+ *              no `footer`.
+ *   enforce  — DEFAULT. Record events + append a short footer to the
+ *              assistant reply summarising the unverified outcomes.
+ *              The footer is APPEND-ONLY — the assistant's text is
+ *              never rewritten. (This is the key behaviour change vs
+ *              v4.6.x — append-only, never an in-place edit.)
  *
- * Critical invariant for memory:
- *   Every memory_add / memory_replace / memory_remove tool result carries
- *   a `verified` flag (per Phase 9 MemoryGuard). If the model claims
- *   "I remembered X" but `verified=false`, Honesty MUST flag this — even
- *   though a memory tool DID fire. This was the v3 C20/C21 lying surface.
- *
- * Status: PHASE 12.
+ * What the verifier intentionally does NOT do (delta vs the deleted
+ * scanner):
+ *   - It does not look at the assistant's natural-language text at all.
+ *     There's no regex matching of English verbs to tool names.
+ *   - It does not emit `no_tool_call` findings. The previous "model
+ *     claimed X but no tool fired" failure mode is gone — that was
+ *     the false-refusal class. We only record OUTCOMES that ran.
+ *   - It does not mutate `loopResult.messages`. The caller appends
+ *     the footer to its own `finalContent` string variable.
  */
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.__test__ = exports.HonestyEnforcement = void 0;
-/** Allows optional adverbs/auxiliaries between "I" and the verb:
- *  "I have", "I also", "I just", "I successfully", "I have just", etc. */
-const I_PREFIX = String.raw `\bI\s+(?:have\s+|just\s+|also\s+|already\s+|successfully\s+|then\s+|now\s+){0,3}`;
-const PATTERNS = [
-    // ── File operations (past tense only) ─────────────────────────
-    {
-        pattern: new RegExp(`${I_PREFIX}(?:saved|wrote|created|modified|patched|updated)\\b[^.]*\\b(?:file|to|at|in)\\b`, 'i'),
-        tools: ['file_write', 'file_patch', 'skill_manage'],
-        label: 'file_write',
-    },
-    {
-        pattern: new RegExp(`${I_PREFIX}(?:deleted|removed)\\s+(?:the\\s+)?(?:file|directory|folder)\\b`, 'i'),
-        tools: ['file_delete'],
-        label: 'file_delete',
-    },
-    // ── Web ────────────────────────────────────────────────────────
-    {
-        pattern: new RegExp(`${I_PREFIX}(?:searched|looked\\s+up|found(?:\\s+online)?|googled)\\b`, 'i'),
-        tools: ['web_search', 'deep_research'],
-        label: 'web_search',
-    },
-    {
-        pattern: new RegExp(`${I_PREFIX}(?:fetched|downloaded|retrieved)\\b`, 'i'),
-        tools: ['web_fetch', 'fetch_url'],
-        label: 'web_fetch',
-    },
-    // ── Shell / execution ──────────────────────────────────────────
-    {
-        pattern: new RegExp(`${I_PREFIX}(?:ran|executed|called)\\b`, 'i'),
-        tools: ['shell_exec', 'execute_code', 'run_python', 'run_node'],
-        label: 'shell_exec',
-    },
-    // ── Browser ────────────────────────────────────────────────────
-    {
-        pattern: new RegExp(`${I_PREFIX}(?:navigated|clicked|typed|scrolled)\\b`, 'i'),
-        tools: [
-            'browser_navigate',
-            'open_browser',
-            'browser_click',
-            'browser_type',
-            'browser_scroll',
-        ],
-        label: 'browser_action',
-    },
-    // ── Memory (verified=true required) ────────────────────────────
-    {
-        pattern: new RegExp(`${I_PREFIX}(?:remembered|memori[sz]ed|noted\\s+that|saved\\s+that\\s+to\\s+memory)\\b`, 'i'),
-        tools: ['memory_add', 'memory_upsert'],
-        label: 'memory_add',
-        kind: 'memory',
-    },
-    {
-        pattern: new RegExp(`${I_PREFIX}(?:forgot(?:ten)?|removed)\\b[^.]*\\bmemory\\b`, 'i'),
-        tools: ['memory_remove', 'memory_forget'],
-        label: 'memory_remove',
-        kind: 'memory',
-    },
-    // ── Model switch ───────────────────────────────────────────────
-    {
-        pattern: new RegExp(`${I_PREFIX}(?:switched\\s+to|changed\\s+(?:to|model\\s+to)|am\\s+now\\s+using)\\s+\\S+`, 'i'),
-        tools: ['model_switch'],
-        label: 'model_switch',
-    },
-];
-/** Negation patterns. If matched at the start of a sentence containing
- *  the claim, the claim is NOT flagged. */
-const NEGATION_RE = /\b(?:couldn'?t|cannot|can'?t|wasn'?t\s+able|unable\s+to|failed\s+to|did\s+not|didn'?t|won'?t|will\s+not)\b/i;
+exports.HonestyEnforcement = void 0;
+/**
+ * Memory tools whose results carry the `verified` flag set by
+ * MemoryGuard. The list is closed — adding a new memory_* tool
+ * means extending this set.
+ */
+const MEMORY_TOOLS = new Set([
+    'memory_add',
+    'memory_replace',
+    'memory_remove',
+]);
+/**
+ * Read `result.path` when present (file_* tools' result envelopes
+ * carry it). Returns undefined otherwise. Used only for cosmetic
+ * footer detail — never affects pass/fail outcome.
+ */
+function extractPath(result) {
+    if (result && typeof result === 'object' && 'path' in result) {
+        const p = result.path;
+        if (typeof p === 'string')
+            return p;
+    }
+    return undefined;
+}
+/**
+ * Translate a `HonestyEvent` to the legacy `HonestyFinding` shape so
+ * existing downstream consumers (chatSession, telemetry) keep working.
+ * The fine-grained kind is preserved via `reason`.
+ */
+function toFinding(event) {
+    switch (event.kind) {
+        case 'mutation_errored':
+            return {
+                claim: event.tool,
+                expectedTool: event.tool,
+                found: false,
+                confidence: 1,
+                reason: 'tool_errored',
+            };
+        case 'memory_unverified':
+            return {
+                claim: event.tool,
+                expectedTool: event.tool,
+                found: false,
+                confidence: 1,
+                reason: 'memory_verified_false',
+            };
+    }
+}
 class HonestyEnforcement {
     constructor(mode = 'enforce', llmAdapter, logger) {
         this.llmAdapter = llmAdapter;
@@ -124,20 +105,63 @@ class HonestyEnforcement {
         return this.mode;
     }
     /**
-     * Inspect a finished response against the actual tool-call trace.
-     * Returns a structured result. Caller (AidenAgent) decides whether to
-     * use `correctedResponse` or `originalResponse` based on `passed`.
+     * v4.7.0 Phase 2.3 — record deterministic unverified outcomes from
+     * the per-turn tool trace. Pure function; no I/O, no side effects.
      */
-    async check(response, _messages, toolCallTrace) {
-        if (this.mode === 'off') {
-            return {
-                passed: true,
-                findings: [],
-                confidence: 1,
-                originalResponse: response,
-            };
+    recordOutcomes(trace) {
+        const events = [];
+        for (const t of trace) {
+            if (t.error && t.handlerMutates === true) {
+                events.push({
+                    kind: 'mutation_errored',
+                    tool: t.name,
+                    reason: t.error,
+                    path: extractPath(t.result),
+                });
+                continue;
+            }
+            if (MEMORY_TOOLS.has(t.name) && t.verified === false) {
+                events.push({
+                    kind: 'memory_unverified',
+                    tool: t.name,
+                    reason: 'verification failed',
+                });
+            }
         }
-        if (!response || !response.trim()) {
+        return events;
+    }
+    /**
+     * v4.7.0 Phase 2.3 — render the append-only footer used in enforce
+     * mode. Caller concatenates with a blank line; we own the lines
+     * inside. Format: one summary line + one row per event.
+     */
+    buildFooter(events) {
+        const lines = [];
+        lines.push(`⚠️ Verifier: ${events.length} tool outcome(s) not verified this turn.`);
+        for (const e of events) {
+            if (e.kind === 'mutation_errored') {
+                const where = e.path ? ` (path: ${e.path})` : '';
+                lines.push(`- ${e.tool}${where}: errored — ${e.reason}`);
+            }
+            else {
+                lines.push(`- ${e.tool}: not verified`);
+            }
+        }
+        return lines.join('\n');
+    }
+    /**
+     * v4.7.0 Phase 2.3 — entry point. Records outcome events from the
+     * trace, converts to legacy `HonestyFinding[]` for downstream
+     * consumers, and renders an append-only footer in enforce mode.
+     *
+     * NEVER rewrites `response`. The returned `footer` is what the
+     * caller appends; the original text is preserved verbatim.
+     *
+     * Off mode short-circuits without touching the trace — minimal cost
+     * for users who opt out.
+     */
+    async check(response, _messages, trace) {
+        if (this.mode === 'off') {
             return {
                 passed: true,
                 findings: [],
@@ -145,145 +169,23 @@ class HonestyEnforcement {
                 originalResponse: response,
             };
         }
-        const findings = this.detectClaimsPattern(response, toolCallTrace);
-        const failed = findings.filter((f) => !f.found);
-        const passed = failed.length === 0;
-        const confidence = findings.length === 0
-            ? 1
-            : findings.reduce((s, f) => s + f.confidence, 0) /
-                findings.length;
-        if (this.mode === 'detect') {
-            this.logger?.('info', `[HonestyEnforcement] detect mode: ${findings.length} findings (${failed.length} failed)`);
-            return {
-                passed,
-                findings,
-                confidence,
-                originalResponse: response,
-            };
+        const events = this.recordOutcomes(trace);
+        const findings = events.map(toFinding);
+        const passed = findings.length === 0;
+        let footer;
+        if (this.mode === 'enforce' && !passed) {
+            footer = this.buildFooter(events);
         }
-        // enforce mode
-        let correctedResponse;
         if (!passed) {
-            correctedResponse = this.buildCorrection(response, failed, toolCallTrace);
-            this.logger?.('warn', `[HonestyEnforcement] enforce: rewrote response (${failed.length} failed claims)`);
+            this.logger?.('info', `honesty: ${events.length} unverified outcome(s) this turn`);
         }
         return {
             passed,
             findings,
-            confidence,
+            confidence: 1,
             originalResponse: response,
-            correctedResponse,
+            footer,
         };
     }
-    // ─────────────────────────────────────────────────────────────────────
-    // pattern detection
-    // ─────────────────────────────────────────────────────────────────────
-    detectClaimsPattern(response, trace) {
-        const findings = [];
-        const sentences = splitSentences(response);
-        for (const sentence of sentences) {
-            // Skip negated sentences entirely.
-            if (NEGATION_RE.test(sentence))
-                continue;
-            for (const pat of PATTERNS) {
-                if (!pat.pattern.test(sentence))
-                    continue;
-                const matched = sentence.match(pat.pattern);
-                const claimText = matched?.[0] ?? sentence.trim();
-                const found = this.traceSatisfies(pat, trace);
-                let reason;
-                if (!found) {
-                    if (pat.kind === 'memory' && memoryFiredButUnverified(pat, trace)) {
-                        reason = 'memory_verified_false';
-                    }
-                    else if (toolFiredButErrored(pat, trace)) {
-                        reason = 'tool_errored';
-                    }
-                    else {
-                        reason = 'no_tool_call';
-                    }
-                }
-                findings.push({
-                    claim: claimText.trim(),
-                    expectedTool: pat.tools.length === 1 ? pat.tools[0] : pat.tools,
-                    found,
-                    confidence: 0.8,
-                    reason,
-                });
-            }
-        }
-        return findings;
-    }
-    traceSatisfies(pat, trace) {
-        const matching = trace.filter((t) => pat.tools.includes(t.name) && !t.error);
-        if (matching.length === 0)
-            return false;
-        if (pat.kind === 'memory') {
-            // verified must be explicitly true
-            return matching.some((m) => m.verified === true);
-        }
-        return true;
-    }
-    // ─────────────────────────────────────────────────────────────────────
-    // correction builder
-    // ─────────────────────────────────────────────────────────────────────
-    buildCorrection(_original, failed, trace) {
-        const lines = [];
-        lines.push("I shouldn't claim actions I didn't take. Honest summary of what I actually did:");
-        lines.push('');
-        if (trace.length === 0) {
-            lines.push('- No tools were called this turn.');
-        }
-        else {
-            for (const entry of trace) {
-                const status = entry.error ? `errored (${entry.error})` : 'succeeded';
-                const verified = entry.verified === false
-                    ? ' (NOT VERIFIED)'
-                    : entry.verified === true
-                        ? ' (verified)'
-                        : '';
-                lines.push(`- ${entry.name}: ${status}${verified}`);
-            }
-        }
-        lines.push('');
-        lines.push('Refused claims:');
-        for (const f of failed) {
-            const tool = Array.isArray(f.expectedTool)
-                ? f.expectedTool.join('/')
-                : f.expectedTool;
-            const why = f.reason === 'memory_verified_false'
-                ? `(memory write returned verified=false — fact was not stored)`
-                : f.reason === 'tool_errored'
-                    ? `(tool errored)`
-                    : `(no ${tool} call in trace)`;
-            lines.push(`- "${f.claim}" ${why}`);
-        }
-        return lines.join('\n');
-    }
 }
 exports.HonestyEnforcement = HonestyEnforcement;
-// ─────────────────────────────────────────────────────────────────────
-// helpers (exported for tests)
-// ─────────────────────────────────────────────────────────────────────
-function splitSentences(text) {
-    // Split on sentence terminators while keeping reasonable bounds.
-    // Don't try to be clever about abbreviations — false positives are
-    // benign (we just inspect more granular slices).
-    return text
-        .split(/(?<=[.!?])\s+|\n+/)
-        .map((s) => s.trim())
-        .filter((s) => s.length > 0);
-}
-function memoryFiredButUnverified(pat, trace) {
-    if (pat.kind !== 'memory')
-        return false;
-    return trace.some((t) => pat.tools.includes(t.name) && !t.error && t.verified === false);
-}
-function toolFiredButErrored(pat, trace) {
-    return trace.some((t) => pat.tools.includes(t.name) && !!t.error);
-}
-exports.__test__ = {
-    splitSentences,
-    PATTERNS,
-    NEGATION_RE,
-};

package/dist/tools/v4/index.js

@@ -246,6 +246,60 @@ function registerWriteTools(registry) {
 function registerAllTools(registry) {
     registerReadOnlyTools(registry);
     registerWriteTools(registry);
+    // v4.8.0 Phase 2.2 — register the 7 semantic ui_* event tools.
+    // All uiOnly: true → the dispatch loop in core/v4/aidenAgent.ts
+    // bypasses execute and fires onUiEvent on the caller. execute()
+    // throws as a safety guard: if the uiOnly branch ever misfires
+    // and an executor is reached, that's a wiring bug, not a render.
+    // Renderer is a no-op stub in this phase; Phase 2.3 lands chrome.
+    const ui = (name, description, properties, required) => ({
+        schema: { name, description, inputSchema: { type: 'object', properties, required } },
+        execute: async () => { throw new Error(`${name} is uiOnly — dispatch branch should bypass execute`); },
+        category: 'read', mutates: false, uiOnly: true,
+    });
+    const str = { type: 'string' };
+    const num = { type: 'number' };
+    registry.register(ui('ui_task_update', 'Signal current task state for the live task panel. Append-only stream.', { task_id: str, label: { type: 'string', description: '≤80 chars' },
+        status: { type: 'string', enum: ['running', 'blocked', 'paused'] },
+        kind: { type: 'string', enum: ['task', 'subagent'] }, depth: num, parent_id: str }, ['task_id', 'label', 'status']));
+    registry.register(ui('ui_task_done', 'Signal a task is complete. Pairs with a prior ui_task_update.', { task_id: str, status: { type: 'string', enum: ['success', 'failure', 'blocked'] },
+        summary: { type: 'string', description: 'Optional, ≤120 chars' } }, ['task_id', 'status']));
+    registry.register(ui('ui_command_result', 'Surface shell output as a formatted block.', { command: str, stdout: str, stderr: str, exit_code: num }, ['command']));
+    registry.register(ui('ui_test_result', 'Pass/fail count after a test run.', { framework: { type: 'string', description: 'e.g. "vitest", "pytest"' },
+        passed: num, failed: num, skipped: num, duration_ms: num }, ['framework', 'passed', 'failed']));
+    registry.register(ui('ui_approval_request', 'Structured approval prompt before a privileged action.', { prompt: { type: 'string', description: '≤160 chars' },
+        risk_tier: { type: 'string', enum: ['low', 'medium', 'high', 'critical'] },
+        reason: { type: 'string', description: 'Optional, ≤200 chars' } }, ['prompt', 'risk_tier']));
+    registry.register(ui('ui_toast', 'Transient notice to surface without interrupting flow.', { message: { type: 'string', description: '≤120 chars' },
+        kind: { type: 'string', enum: ['info', 'success', 'warning', 'error'] } }, ['message', 'kind']));
+    registry.register(ui('ui_artifact_created', 'Surface a file or skill created/modified this turn.', { path: str, kind: { type: 'string', enum: ['file', 'skill', 'directory'] },
+        preview: { type: 'string', description: 'Optional, ≤200 chars' } }, ['path', 'kind']));
+    // v4.8.0 Phase 2.1 — env-gated uiOnly smoke stub. Never registers
+    // in production. Set AIDEN_TEST_UI_STUB=1 to enable for the
+    // dispatch-branch smoke harness. The execute() throws on purpose:
+    // if the uiOnly branch is wired correctly the model can never
+    // reach the executor.
+    if (process.env.AIDEN_TEST_UI_STUB === '1') {
+        registry.register({
+            schema: {
+                name: '_test_ui_stub',
+                description: 'Test-only uiOnly stub for v4.8.0 Phase 2.1 smoke. Set AIDEN_TEST_UI_STUB=1 to enable.',
+                inputSchema: {
+                    type: 'object',
+                    properties: {
+                        message: { type: 'string', description: 'Arbitrary message to echo via onUiEvent' },
+                    },
+                    required: ['message'],
+                },
+            },
+            execute: async () => {
+                throw new Error('_test_ui_stub should never execute — uiOnly branch should bypass it');
+            },
+            category: 'read',
+            mutates: false,
+            uiOnly: true,
+        });
+    }
 }
 var subagentFanout_2 = require("./subagent/subagentFanout");
 Object.defineProperty(exports, "makeSubagentFanoutTool", { enumerable: true, get: function () { return subagentFanout_2.makeSubagentFanoutTool; } });

package/dist/tools/v4/subagent/spawnSubAgentTool.js

@@ -289,6 +289,19 @@ function makeSpawnSubAgentTool(factory) {
             // ── 3. Resolve optional parent run / session identifiers ─────────────
             const parentRunId = factory.resolveParentRunId?.();
             const parentSessionId = factory.resolveParentSessionId?.();
+            // v4.8.0 Phase 2.5 — emit ui_task_update for the subagent start.
+            // Stable task_id correlates with the matching ui_task_done emit
+            // after the spawnSubAgent call returns. depth:1 hardcoded today
+            // — childBuilder caps recursion at 1 (see SUBAGENT_BLOCKED_TOOL_NAMES
+            // 'spawn_sub_agent'). TODO: thread real depth when nested spawns ship.
+            const subTaskId = `subagent-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`;
+            factory.onUiEvent?.('ui_task_update', {
+                task_id: subTaskId,
+                label: goalPreview,
+                status: 'running',
+                kind: 'subagent',
+                depth: 1,
+            });
             // ── 4. Invoke the primitive. NEVER throws — always envelope. ─────────
             const result = await (0, spawnSubAgent_1.spawnSubAgent)(spec, {
                 // ChildBuilderDeps fields:
@@ -310,6 +323,16 @@ function makeSpawnSubAgentTool(factory) {
                 parentRunId,
                 parentSessionId,
             });
+            // v4.8.0 Phase 2.5 — emit ui_task_done with the same subTaskId
+            // so the display layer can finalize the in-flight row.
+            const doneStatus = result.ok ? 'success' :
+                result.status === 'interrupted' ? 'blocked' :
+                    result.status === 'timeout' ? 'blocked' : 'failure';
+            factory.onUiEvent?.('ui_task_done', {
+                task_id: subTaskId,
+                status: doneStatus,
+                summary: `${result.metrics.apiCalls} calls · ${result.exitReason}`,
+            });
             // Completion log — pairs with "spawn_sub_agent invoked" so a
             // grep on parentSessionId surfaces invoke → complete in order.
             logger.info('spawn_sub_agent completed', {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "aiden-runtime",
-  "version": "4.6.1",
+  "version": "4.8.0",
   "publishConfig": {
     "access": "public"
   },
@@ -254,7 +254,7 @@
     "epub2": "^3.0.2",
     "execa": "^8.0.1",
     "express": "^4.18.2",
-    "form-data": "^4.0.0",
+    "form-data": "^4.0.4",
     "imap-simple": "^5.1.0",
     "js-tiktoken": "^1.0.21",
     "js-yaml": "^4.1.1",
@@ -284,7 +284,7 @@
     "uuid": "^9.0.0",
     "whatsapp-web.js": "^1.26.0",
     "wrap-ansi": "^9.0.2",
-    "ws": "^8.20.0"
+    "ws": "^8.20.1"
   },
   "optionalDependencies": {
     "decibri": "*",
@@ -296,7 +296,13 @@
     "semver": "^7.5.2",
     "postcss": "^8.5.10",
     "hono": "^4.12.16",
-    "minimatch": "^9.0.9"
+    "minimatch": "^9.0.9",
+    "qs": ">=6.14.1",
+    "tough-cookie": ">=4.1.3",
+    "protobufjs": ">=7.5.8",
+    "request": {
+      "form-data": "^2.5.5"
+    }
   },
   "devDependencies": {
     "@types/better-sqlite3": "^7.6.13",