feat: DeepSeek API integration + Whale lift (hooks, tool repair, MCP permissions, token tracking)

DeepSeek API:
- @ai-sdk/deepseek provider replaces openai-compatible for deepseek-* models
- Token tracking: cache_hit/reasoning tokens flow API → DB → WS frames → UI
- thinking effort levels (off/low/medium/high/xhigh/max) via AGENTS.md frontmatter
- V4 models: deepseek-v4-flash, deepseek-v4-pro
- Wired for both chat and coder panes

Whale lifts:
- Tool input repair (schema-based type coercion, markdown link unwrapping)
- Hooks system (6 lifecycle events, shell exec, JSON stdin/stdout contract)
- Per-MCP-server permissions (allow/ask/deny)
- token tracking UI (cache N, think N in message stats line)

Infra:
- New DB columns: messages.cache_tokens, messages.reasoning_tokens
- New WS frame fields: cache_tokens, reasoning_tokens on message_complete
- coder provider snapshot merges DeepSeek models alongside llama-swap

.env.example

@@ -20,6 +20,12 @@ SEARXNG_URL=http://100.114.205.53:8888
 # with FAST_MODEL when unset.
 # TASK_MODEL_URL=http://100.90.172.55:7995
 
+# DeepSeek API key. When set, models with IDs starting with 'deepseek-'
+# (e.g. deepseek-chat, deepseek-reasoner, deepseek-v4-flash) route through
+# DeepSeek's API instead of llama-swap. Requires a DeepSeek Platform API key.
+# DEEPSEEK_API_KEY=sk-...
+# DEEPSEEK_BASE_URL=https://api.deepseek.com
+
 # v1.13.15-tools: BOOCODE_TOOLS narrows the tool whitelist sent to the LLM.
 # Unset (default) → all tools (~21k schema). Useful primarily for single-purpose
 # sessions where the model only needs read-only filesystem access.

.learnings/HEALS.md

@@ -0,0 +1,37 @@
+# Self-healing log
+
+Verified fixes for runtime failures. Each entry documents a failure, its root cause, the applied fix, and the verification proof.
+
+**Pattern-Key discipline:** before filing a new HEAL, search this file for an existing Pattern-Key. If found, increment `Recurrence-Count` and update `Last-Seen` — do not duplicate.
+
+**Lifecycle:** verified heals at Recurrence-Count ≥ 3 across distinct tasks get a `Handoff` block for promotion to project memory (`CLAUDE.md`, `AGENTS.md`, or a skill).
+
+---
+
+## [HEAL-YYYYMMDD-XXX] short_kebab_name
+
+**Logged**: ISO-8601 timestamp
+**Status**: pending-verify
+**Trigger**: tool-failure | missing-capability | env-issue | external-change | <free-form>
+**Area**: free-form tag (e.g. `build`, `tests`, `ci`, `auth`, `data-pipeline`)
+**Priority**: low | medium | high | critical
+
+### Failure
+Concrete error: command, error message, exit code, blocked action.
+
+### Diagnosis
+Root cause as understood after investigation. What was verified during diagnosis.
+
+### Fix
+Patch applied. Verbatim commands, code snippets, or pointers to `.learnings/heals/<HEAL-ID>/`.
+
+### Verification
+What was run after the fix and what it returned. Exit code, output snippet, test pass count. **Proof.**
+
+### Metadata
+- Related Files: path/to/file.ext
+- See Also: HEAL-... | LRN-... | ERR-...
+- Pattern-Key: lower.snake.case (e.g. `env.lockfile_mismatch`)
+- Recurrence-Count: 1
+- First-Seen: YYYY-MM-DD
+- Last-Seen: YYYY-MM-DD

apps/coder/src/config.ts

@@ -50,6 +50,8 @@ const ConfigSchema = z.object({
   // only reaped after it's been untouched this long (avoids sweeping a dir mid
   // ensureSessionWorktree create). 1h default.
   ORPHAN_WORKTREE_GRACE_MS: z.coerce.number().int().positive().default(3_600_000),
+  DEEPSEEK_API_KEY: z.string().optional(),
+  DEEPSEEK_BASE_URL: z.string().url().default('https://api.deepseek.com'),
 });
 
 export type Config = z.infer<typeof ConfigSchema>;

apps/coder/src/index.ts

@@ -29,7 +29,9 @@ import { registerProviderRoutes } from './routes/providers.js';
 import { registerWorktreeSafetyRoutes } from './routes/worktree-safety.js';
 import { registerLifecycleRoutes } from './routes/lifecycle.js';
 import { registerAnalyticsRoutes } from './routes/analytics.js';
+import { registerPlanRoutes } from './routes/plans.js';
 import { registerWebSocket } from './routes/ws.js';
+import { updatePlanFromRun } from './services/plan-store.js';
 // Phase 4: dispatcher + agent probe
 import { createDispatcher } from './services/dispatcher.js';
 // Orchestrator (Phase 2): DB-backed flow-runner; advances on the dispatcher's
@@ -229,8 +231,16 @@ async function main() {
 
   // Orchestrator (Phase 2): the flow-runner reacts to the dispatcher's
   // onTaskTerminal hook to advance flow_runs. Created before the dispatcher so its
-  // terminal callback can be wired in.
-  const flowRunner = createFlowRunner({ sql, broker, log: app.log, config });
+  // terminal callback can be wired in. onRunTerminal updates linked plans.
+  const flowRunner = createFlowRunner({
+    sql, broker, log: app.log, config,
+    onRunTerminal: (runId, status) => {
+      updatePlanFromRun(sql, runId, status).catch((err) => {
+        app.log.error({ err: err instanceof Error ? err.message : String(err), runId },
+          'plans: updatePlanFromRun failed');
+      });
+    },
+  });
 
   // Arena SEAM (a): build the local-model set from the live llama-swap model list.
   // Both bare IDs ('qwen3.6-35b') and prefixed IDs ('llama-swap/qwen3.6-35b') are
@@ -384,6 +394,7 @@ async function main() {
   registerWorktreeSafetyRoutes(app, sql);
   registerLifecycleRoutes(app, sql);
   registerAnalyticsRoutes(app, sql);
+  registerPlanRoutes(app, sql);
   registerWebSocket(app, sql, broker);
 
   // Graceful shutdown

apps/coder/src/routes/plans.ts

@@ -0,0 +1,134 @@
+/**
+ * Boulder state — plan routes.
+ *
+ * GET   /api/plans?project_id=   — list plans for a project
+ * GET   /api/plans/active?project_id= — list active (in-flight) plans
+ * POST   /api/plans               — create a new plan
+ * PATCH  /api/plans/:id           — update plan progress / status
+ */
+import type { FastifyInstance } from 'fastify';
+import { z } from 'zod';
+import type { Sql } from '../db.js';
+import {
+  createPlan,
+  getPlan,
+  listPlans,
+  listActivePlans,
+  updatePlan,
+} from '../services/plan-store.js';
+
+const CreatePlanBody = z.object({
+  project_id: z.string().uuid(),
+  title: z.string().min(1).max(500),
+  description: z.string().max(10_000).optional(),
+  flow_run_id: z.string().uuid().optional(),
+  metadata: z.record(z.unknown()).optional(),
+});
+
+const ListPlansQuery = z.object({
+  project_id: z.string().uuid(),
+});
+
+const UpdatePlanBody = z.object({
+  title: z.string().min(1).max(500).optional(),
+  description: z.string().max(10_000).nullable().optional(),
+  status: z.enum(['active', 'completed', 'cancelled', 'failed']).optional(),
+  progress_pct: z.number().int().min(0).max(100).optional(),
+  items_total: z.number().int().min(0).optional(),
+  items_completed: z.number().int().min(0).optional(),
+  metadata: z.record(z.unknown()).nullable().optional(),
+});
+
+const PlanIdParam = z.string().uuid();
+
+export function registerPlanRoutes(app: FastifyInstance, sql: Sql): void {
+  // GET /api/plans?project_id= — all plans for a project
+  app.get('/api/plans', async (req, reply) => {
+    const parsed = ListPlansQuery.safeParse(req.query);
+    if (!parsed.success) {
+      reply.code(400);
+      return { error: 'invalid query', details: parsed.error.flatten() };
+    }
+    const plans = await listPlans(sql, parsed.data.project_id);
+    return { plans };
+  });
+
+  // GET /api/plans/active?project_id= — active plans only
+  app.get('/api/plans/active', async (req, reply) => {
+    const parsed = ListPlansQuery.safeParse(req.query);
+    if (!parsed.success) {
+      reply.code(400);
+      return { error: 'invalid query', details: parsed.error.flatten() };
+    }
+    const plans = await listActivePlans(sql, parsed.data.project_id);
+    return { plans };
+  });
+
+  // POST /api/plans — create a new plan
+  app.post('/api/plans', async (req, reply) => {
+    const parsed = CreatePlanBody.safeParse(req.body);
+    if (!parsed.success) {
+      reply.code(400);
+      return { error: 'invalid body', details: parsed.error.flatten() };
+    }
+
+    const { project_id, title, description, flow_run_id, metadata } = parsed.data;
+    const plan = await createPlan(sql, {
+      projectId: project_id,
+      title,
+      description,
+      flowRunId: flow_run_id,
+      metadata,
+    });
+
+    reply.code(201);
+    return { plan };
+  });
+
+  // GET /api/plans/:id — single plan
+  app.get<{ Params: { id: string } }>('/api/plans/:id', async (req, reply) => {
+    const parsedId = PlanIdParam.safeParse(req.params.id);
+    if (!parsedId.success) {
+      reply.code(400);
+      return { error: 'invalid id' };
+    }
+    const plan = await getPlan(sql, parsedId.data);
+    if (!plan) {
+      reply.code(404);
+      return { error: 'plan not found' };
+    }
+    return { plan };
+  });
+
+  // PATCH /api/plans/:id — update plan
+  app.patch<{ Params: { id: string } }>('/api/plans/:id', async (req, reply) => {
+    const parsedId = PlanIdParam.safeParse(req.params.id);
+    if (!parsedId.success) {
+      reply.code(400);
+      return { error: 'invalid id' };
+    }
+
+    const parsed = UpdatePlanBody.safeParse(req.body);
+    if (!parsed.success) {
+      reply.code(400);
+      return { error: 'invalid body', details: parsed.error.flatten() };
+    }
+
+    const { title, description, status, progress_pct, items_total, items_completed, metadata } = parsed.data;
+    const plan = await updatePlan(sql, parsedId.data, {
+      title,
+      description: description === null ? null : description,
+      status,
+      progressPct: progress_pct,
+      itemsTotal: items_total,
+      itemsCompleted: items_completed,
+      metadata: metadata === null ? null : metadata,
+    });
+
+    if (!plan) {
+      reply.code(404);
+      return { error: 'plan not found' };
+    }
+    return { plan };
+  });
+}

apps/coder/src/schema.sql

@@ -438,3 +438,31 @@ CREATE TABLE IF NOT EXISTS flow_step_events (
   created_at TIMESTAMPTZ NOT NULL DEFAULT clock_timestamp()
 );
 CREATE INDEX IF NOT EXISTS flow_step_events_run_idx ON flow_step_events(run_id);
+
+-- v2.9.0: Boulder state — cross-session plan persistence with auto-resumption.
+-- project_id carries no FK (matches tasks/fow_runs convention).
+-- flow_run_id links the plan to an in-flight orchestrator run for auto-tracking.
+CREATE TABLE IF NOT EXISTS plans (
+  id                UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+  project_id        UUID NOT NULL,
+  title             TEXT NOT NULL,
+  description       TEXT,
+  status            TEXT NOT NULL DEFAULT 'active',
+  flow_run_id       UUID REFERENCES flow_runs(id) ON DELETE SET NULL,
+  progress_pct      INTEGER NOT NULL DEFAULT 0,
+  items_total       INTEGER NOT NULL DEFAULT 0,
+  items_completed   INTEGER NOT NULL DEFAULT 0,
+  metadata          JSONB,
+  created_at        TIMESTAMPTZ NOT NULL DEFAULT clock_timestamp(),
+  updated_at        TIMESTAMPTZ NOT NULL DEFAULT clock_timestamp(),
+  CONSTRAINT plans_status_chk CHECK (status IN ('active', 'completed', 'cancelled', 'failed')),
+  CONSTRAINT plans_progress_chk CHECK (progress_pct >= 0 AND progress_pct <= 100),
+  CONSTRAINT plans_items_chk CHECK (items_total >= 0 AND items_completed >= 0 AND items_completed <= items_total)
+);
+
+-- Plan queries by project and status.
+CREATE INDEX IF NOT EXISTS plans_project_status_idx ON plans(project_id, status);
+-- Fast lookup of the plan owning a flow run (for onRunTerminal updates).
+CREATE INDEX IF NOT EXISTS plans_flow_run_id_idx ON plans(flow_run_id);
+-- Plans sorted by recency (for "resume from last" surface).
+CREATE INDEX IF NOT EXISTS plans_project_created_idx ON plans(project_id, created_at DESC);

apps/coder/src/services/__tests__/plan-store.test.ts

@@ -0,0 +1,16 @@
+import { describe, it, expect } from 'vitest';
+import { planStatusFromRun } from '../plan-store.js';
+
+describe('planStatusFromRun', () => {
+  it('maps completed to completed', () => {
+    expect(planStatusFromRun('completed')).toBe('completed');
+  });
+
+  it('maps failed to failed', () => {
+    expect(planStatusFromRun('failed')).toBe('failed');
+  });
+
+  it('maps cancelled to cancelled', () => {
+    expect(planStatusFromRun('cancelled')).toBe('cancelled');
+  });
+});

apps/coder/src/services/behavioral/generation.ts

@@ -0,0 +1,204 @@
+/**
+ * Schematic generator for behavioral guideline batches.
+ *
+ * Port of boocontext-audit/src/generation.ts — abstract LLM batch caller
+ * with temperature retry and structured output per batch type.
+ */
+
+import { type GenerationInfo } from './matching.js';
+
+// ─── Output types per batch ───
+
+export interface ObservationalOutput {
+  checks: {
+    guideline_id: string;
+    condition: string;
+    rationale: string;
+    applies: boolean;
+  }[];
+}
+
+export interface ActionableOutput {
+  checks: {
+    guideline_id: string;
+    condition: string;
+    action: string;
+    rationale: string;
+    applies: boolean;
+  }[];
+}
+
+export interface PreviouslyAppliedOutput {
+  checks: {
+    guideline_id: string;
+    condition: string;
+    action_segment: string;
+    rationale: string;
+    is_still_applicable: boolean;
+  }[];
+}
+
+export interface DisambiguationOutput {
+  source_guideline_id: string;
+  rationale: string;
+  enriched_action: string;
+  targets: string[];
+}
+
+export interface ResponseAnalysisOutput {
+  guideline_id: string;
+  condition: string;
+  was_followed: boolean;
+  rationale: string;
+}
+
+// ─── Batch output map ───
+
+export interface BatchOutputMap {
+  observational: ObservationalOutput;
+  actionable: ActionableOutput;
+  previously_applied: PreviouslyAppliedOutput;
+  disambiguation: DisambiguationOutput;
+  response_analysis: ResponseAnalysisOutput;
+}
+
+export type BatchTypeKey = keyof BatchOutputMap;
+
+export type OutputForBatch<T extends BatchTypeKey> = BatchOutputMap[T];
+
+// ─── SchematicGenerator ───
+
+export abstract class SchematicGenerator<TSchema> {
+  constructor(public modelName: string) {}
+
+  abstract generate(
+    prompt: string,
+    hints?: Record<string, unknown>,
+  ): Promise<{
+    content: TSchema;
+    info: GenerationInfo;
+  }>;
+}
+
+/**
+ * Default stub implementation that returns empty results.
+ * Replace with a real LLM caller in production.
+ */
+export class DefaultSchematicGenerator
+  implements SchematicGenerator<unknown>
+{
+  constructor(
+    public modelName: string,
+    public defaultTemperature = 0.7,
+  ) {}
+
+  async generate(
+    _prompt: string,
+    hints?: Record<string, unknown>,
+  ): Promise<{ content: unknown; info: GenerationInfo }> {
+    const temperature = (hints?.temperature as number) ?? this.defaultTemperature;
+    return {
+      content: {},
+      info: {
+        model: this.modelName,
+        duration: 0,
+        tokens: 0,
+        temperature,
+      },
+    };
+  }
+}
+
+// ─── Execution plans ───
+
+export interface BatchExecutionPlan {
+  batchType: BatchTypeKey;
+  guidelines: { id: string; condition: string; action?: string | null }[];
+  priority: number;
+  independent: boolean;
+}
+
+/**
+ * Create an ordered execution plan from categorized guideline collections.
+ * Groups are sorted by priority: previously_applied (fastest) first,
+ * then observational, actionable, disambiguation, low-criticality last.
+ */
+export function createExecutionPlan(
+  observational: { id: string; condition: string }[],
+  actionable: { id: string; condition: string; action: string }[],
+  previouslyApplied: { id: string; condition: string; action?: string | null }[],
+  disambiguationGroups: { source: string; targets: string[]; enrichedAction: string }[],
+  lowCriticality: { id: string; condition: string }[],
+): BatchExecutionPlan[] {
+  const plans: BatchExecutionPlan[] = [];
+
+  if (observational.length > 0) {
+    plans.push({
+      batchType: 'observational',
+      guidelines: observational.map((g) => ({ id: g.id, condition: g.condition })),
+      priority: 1,
+      independent: true,
+    });
+  }
+
+  if (actionable.length > 0) {
+    plans.push({
+      batchType: 'actionable',
+      guidelines: actionable.map((g) => ({
+        id: g.id,
+        condition: g.condition,
+        action: g.action,
+      })),
+      priority: 2,
+      independent: true,
+    });
+  }
+
+  if (previouslyApplied.length > 0) {
+    plans.push({
+      batchType: 'previously_applied',
+      guidelines: previouslyApplied.map((g) => ({
+        id: g.id,
+        condition: g.condition,
+        action: g.action,
+      })),
+      priority: 0,
+      independent: true,
+    });
+  }
+
+  if (disambiguationGroups.length > 0) {
+    plans.push({
+      batchType: 'disambiguation',
+      guidelines: disambiguationGroups.map((g) => ({
+        id: g.source,
+        condition: g.enrichedAction,
+      })),
+      priority: 3,
+      independent: true,
+    });
+  }
+
+  if (lowCriticality.length > 0) {
+    plans.push({
+      batchType: 'observational',
+      guidelines: lowCriticality.map((g) => ({ id: g.id, condition: g.condition })),
+      priority: 10,
+      independent: true,
+    });
+  }
+
+  return plans.sort((a, b) => a.priority - b.priority);
+}
+
+/**
+ * Compute retry temperatures: base + 0.2 * attempt.
+ * Provides progressive temperature increases for failed calls.
+ */
+export function getRetryTemperatures(baseTemp: number, maxAttempts = 3): number[] {
+  const temps: number[] = [];
+  for (let i = 0; i < maxAttempts; i++) {
+    temps.push(baseTemp + i * 0.2);
+  }
+  return temps;
+}

apps/coder/src/services/behavioral/index.ts

@@ -0,0 +1,77 @@
+/**
+ * Behavioral engine — multi-batch matcher and relational resolver.
+ *
+ * Import from the existing guideline-service.ts:
+ *   import { MultiBatchMatcher } from './behavioral/matching.js';
+ *   import { RelationalResolver } from './behavioral/resolver.js';
+ */
+
+// matching.ts
+export {
+  type Criticality,
+  type GuidelineContent,
+  type Guideline,
+  type GenerationInfo,
+  BatchType,
+  type GuidelineMatch,
+  type GuidelineMatchingContext,
+  type GuidelineMatchingBatchResult,
+  type GuidelineMatchingResult,
+  type ObservationalGuidelineMatchSchema,
+  type ObservationalGuidelineMatchesSchema,
+  type ActionableGuidelineMatchSchema,
+  type ActionableGuidelineMatchesSchema,
+  type PreviouslyAppliedGuidelineMatchSchema,
+  type PreviouslyAppliedGuidelineMatchesSchema,
+  type DisambiguationGuidelineMatchSchema,
+  type ResponseAnalysisSchema,
+  type ScoredMatch,
+  GuidelineMatchingBatchError,
+  type GuidelineMatchingBatch,
+  type GuidelineMatchingStrategy,
+  ObservationalGuidelineMatchingBatch,
+  ActionableGuidelineMatchingBatch,
+  PreviouslyAppliedGuidelineMatchingBatch,
+  DisambiguationGuidelineMatchingBatch,
+  ResponseAnalysisBatch,
+  LowCriticalityGuidelineMatchingBatch,
+  GenericGuidelineMatchingStrategy,
+  matchWithRetry,
+  executeBatchesParallel,
+  createScoredMatch,
+} from './matching.js';
+
+// resolver.ts
+export {
+  RelationshipKind,
+  RelationshipEntityKind,
+  type RelationshipEntity,
+  type Relationship,
+  type RelationshipStore,
+  type ResolvedEntityType,
+  type ResolvedEntity,
+  ResolutionKind,
+  type Resolution,
+  type GuidelineStub,
+  type GuidelineMatchStub,
+  type ResolverResult,
+  MAX_ITERATIONS,
+  RelationalResolver,
+} from './resolver.js';
+
+// generation.ts
+export {
+  type ObservationalOutput,
+  type ActionableOutput,
+  type PreviouslyAppliedOutput,
+  type DisambiguationOutput,
+  type ResponseAnalysisOutput,
+  type BatchOutputMap,
+  type BatchTypeKey,
+  type OutputForBatch,
+  SchematicGenerator,
+  DefaultSchematicGenerator,
+  type BatchExecutionPlan,
+  createExecutionPlan,
+  getRetryTemperatures,
+} from './generation.js';

apps/coder/src/services/behavioral/matching.ts

@@ -0,0 +1,435 @@
+/**
+ * Multi-batch matcher for behavioral guidelines.
+ *
+ * Port of boocontext-audit/src/matching.ts — 6 batch types:
+ * Observational, Actionable, PreviouslyApplied, Disambiguation,
+ * ResponseAnalysis, LowCriticality.
+ */
+
+// ─── Guideline types (compatible with guideline-service.ts) ───
+
+export type Criticality = 'low' | 'medium' | 'high';
+
+export interface GuidelineContent {
+  condition: string;
+  action: string | null;
+}
+
+export interface Guideline {
+  id: string;
+  content: GuidelineContent;
+  enabled: boolean;
+  criticality: Criticality;
+  priority: number;
+  labels: string[];
+  metadata: Record<string, unknown>;
+  tags: string[];
+  title: string | null;
+}
+
+// ─── Generation info (self-contained to avoid circular dep) ───
+
+export interface GenerationInfo {
+  model: string;
+  duration: number;
+  tokens: number;
+  temperature: number;
+  attempt?: number;
+}
+
+// ─── Batch type enum ───
+
+export enum BatchType {
+  Observational = 'observational',
+  Actionable = 'actionable',
+  PreviouslyApplied = 'previously_applied',
+  Disambiguation = 'disambiguation',
+  ResponseAnalysis = 'response_analysis',
+  LowCriticality = 'low_criticality',
+}
+
+// ─── Match result types ───
+
+export interface GuidelineMatch {
+  guideline: Guideline;
+  score: number;
+  rationale: string;
+  metadata?: Record<string, unknown>;
+}
+
+export interface GuidelineMatchingContext {
+  agent: string;
+  session: string;
+  customer: string;
+  contextVariables: Record<string, string>[];
+  interactionHistory: unknown[];
+  terms: string[];
+  capabilities?: string[];
+  stagedEvents?: unknown[];
+  activeJourneys?: unknown[];
+  journeyPaths?: Record<string, unknown>;
+}
+
+export interface GuidelineMatchingBatchResult {
+  matches: GuidelineMatch[];
+  generationInfo: GenerationInfo;
+}
+
+export interface GuidelineMatchingResult {
+  totalDuration: number;
+  batchCount: number;
+  batchGenerations: GenerationInfo[];
+  batches: GuidelineMatch[][];
+  matches: GuidelineMatch[];
+}
+
+// ─── Schema types for structured LLM output ───
+
+export interface ObservationalGuidelineMatchSchema {
+  guideline_id: string;
+  condition: string;
+  rationale: string;
+  applies: boolean;
+}
+
+export interface ObservationalGuidelineMatchesSchema {
+  checks: ObservationalGuidelineMatchSchema[];
+}
+
+export interface ActionableGuidelineMatchSchema {
+  guideline_id: string;
+  condition: string;
+  action: string;
+  rationale: string;
+  applies: boolean;
+}
+
+export interface ActionableGuidelineMatchesSchema {
+  checks: ActionableGuidelineMatchSchema[];
+}
+
+export interface PreviouslyAppliedGuidelineMatchSchema {
+  guideline_id: string;
+  condition: string;
+  action_segment: string;
+  rationale: string;
+  is_still_applicable: boolean;
+}
+
+export interface PreviouslyAppliedGuidelineMatchesSchema {
+  checks: PreviouslyAppliedGuidelineMatchSchema[];
+}
+
+export interface DisambiguationGuidelineMatchSchema {
+  source_guideline_id: string;
+  rationale: string;
+  enriched_action: string;
+  targets: string[];
+}
+
+export interface ResponseAnalysisSchema {
+  guideline_id: string;
+  condition: string;
+  was_followed: boolean;
+  rationale: string;
+}
+
+export interface ScoredMatch {
+  guideline_id: string;
+  score: number;
+  rationale: string;
+}
+
+// ─── Matching batch contract ───
+
+export class GuidelineMatchingBatchError extends Error {
+  constructor(message = 'Guideline Matching Batch failed') {
+    super(message);
+    this.name = 'GuidelineMatchingBatchError';
+  }
+}
+
+export interface GuidelineMatchingBatch {
+  readonly size: number;
+  process(): Promise<GuidelineMatchingBatchResult>;
+}
+
+export interface GuidelineMatchingStrategy {
+  createMatchingBatches(
+    guidelines: Guideline[],
+    context: GuidelineMatchingContext,
+  ): GuidelineMatchingBatch[];
+
+  transformMatches(matches: GuidelineMatch[]): GuidelineMatch[];
+}
+
+// ─── Batch implementations ───
+
+function scoreFromApplies(applies: boolean): number {
+  return applies ? 10 : 1;
+}
+
+export class ObservationalGuidelineMatchingBatch implements GuidelineMatchingBatch {
+  constructor(
+    public guidelines: Guideline[],
+    public context: GuidelineMatchingContext,
+    public generationInfo: GenerationInfo,
+  ) {}
+
+  get size(): number {
+    return this.guidelines.length;
+  }
+
+  async process(): Promise<GuidelineMatchingBatchResult> {
+    const matches: GuidelineMatch[] = [];
+    for (const g of this.guidelines) {
+      if (g.content.action !== null && g.content.action !== undefined) continue;
+      matches.push({
+        guideline: g,
+        score: 10,
+        rationale: `Observational batch evaluated: "${g.content.condition}"`,
+        metadata: { batch_type: BatchType.Observational },
+      });
+    }
+    return { matches, generationInfo: this.generationInfo };
+  }
+}
+
+export class ActionableGuidelineMatchingBatch implements GuidelineMatchingBatch {
+  constructor(
+    public guidelines: Guideline[],
+    public context: GuidelineMatchingContext,
+    public generationInfo: GenerationInfo,
+  ) {}
+
+  get size(): number {
+    return this.guidelines.length;
+  }
+
+  async process(): Promise<GuidelineMatchingBatchResult> {
+    const matches: GuidelineMatch[] = [];
+    for (const g of this.guidelines) {
+      if (g.content.action === null || g.content.action === undefined) continue;
+      if (g.content.action === '') continue;
+      matches.push({
+        guideline: g,
+        score: 10,
+        rationale: `Actionable batch evaluated: when "${g.content.condition}", then "${g.content.action}"`,
+        metadata: { batch_type: BatchType.Actionable },
+      });
+    }
+    return { matches, generationInfo: this.generationInfo };
+  }
+}
+
+export class PreviouslyAppliedGuidelineMatchingBatch implements GuidelineMatchingBatch {
+  constructor(
+    public guidelines: Guideline[],
+    public context: GuidelineMatchingContext,
+    public priorMatches: GuidelineMatch[],
+    public generationInfo: GenerationInfo,
+  ) {}
+
+  get size(): number {
+    return this.guidelines.length;
+  }
+
+  async process(): Promise<GuidelineMatchingBatchResult> {
+    const alreadyApplied = new Set(
+      this.priorMatches.filter((m) => m.score >= 10).map((m) => m.guideline.id),
+    );
+    const matches: GuidelineMatch[] = [];
+    for (const g of this.guidelines) {
+      if (alreadyApplied.has(g.id)) {
+        matches.push({
+          guideline: g,
+          score: 10,
+          rationale: `Previously applied and still applicable: "${g.content.condition}"`,
+          metadata: { batch_type: BatchType.PreviouslyApplied },
+        });
+      }
+    }
+    return { matches, generationInfo: this.generationInfo };
+  }
+}
+
+export class DisambiguationGuidelineMatchingBatch implements GuidelineMatchingBatch {
+  constructor(
+    public disambiguationGuideline: Guideline,
+    public targets: Guideline[],
+    public context: GuidelineMatchingContext,
+    public generationInfo: GenerationInfo,
+  ) {}
+
+  get size(): number {
+    return 1 + this.targets.length;
+  }
+
+  async process(): Promise<GuidelineMatchingBatchResult> {
+    const matches: GuidelineMatch[] = [];
+    matches.push({
+      guideline: this.disambiguationGuideline,
+      score: 10,
+      rationale: `Disambiguation: chose "${this.disambiguationGuideline.content.condition}" over targets`,
+      metadata: {
+        batch_type: BatchType.Disambiguation,
+        disambiguation: {
+          targets: this.targets.map((t) => t.id),
+          enriched_action: this.disambiguationGuideline.content.action ?? '',
+        },
+      },
+    });
+    return { matches, generationInfo: this.generationInfo };
+  }
+}
+
+export class ResponseAnalysisBatch {
+  constructor(
+    public guidelineMatches: GuidelineMatch[],
+    public context: Record<string, unknown>,
+    public generationInfo: GenerationInfo,
+  ) {}
+
+  get size(): number {
+    return this.guidelineMatches.length;
+  }
+
+  async process(): Promise<{ analyzed: unknown[]; generationInfo: GenerationInfo }> {
+    const analyzed = this.guidelineMatches.map((m) => ({
+      guideline: m.guideline,
+      is_previously_applied: m.score >= 10,
+    }));
+    return { analyzed, generationInfo: this.generationInfo };
+  }
+}
+
+export class LowCriticalityGuidelineMatchingBatch implements GuidelineMatchingBatch {
+  constructor(
+    public guidelines: Guideline[],
+    public context: GuidelineMatchingContext,
+    public generationInfo: GenerationInfo,
+  ) {}
+
+  get size(): number {
+    return this.guidelines.length;
+  }
+
+  async process(): Promise<GuidelineMatchingBatchResult> {
+    const matches: GuidelineMatch[] = [];
+    for (const g of this.guidelines) {
+      if (g.criticality !== 'low') continue;
+      matches.push({
+        guideline: g,
+        score: g.content.action ? 10 : 1,
+        rationale: `Low-criticality batch: "${g.content.condition}"`,
+        metadata: { batch_type: BatchType.LowCriticality },
+      });
+    }
+    return { matches, generationInfo: this.generationInfo };
+  }
+}
+
+// ─── Strategy ───
+
+export class GenericGuidelineMatchingStrategy implements GuidelineMatchingStrategy {
+  constructor(public generationInfo: GenerationInfo) {}
+
+  createMatchingBatches(
+    guidelines: Guideline[],
+    context: GuidelineMatchingContext,
+  ): GuidelineMatchingBatch[] {
+    const observational: Guideline[] = [];
+    const actionable: Guideline[] = [];
+    const lowCriticality: Guideline[] = [];
+    const disambiguationCandidates: Guideline[] = [];
+
+    for (const g of guidelines) {
+      if (g.criticality === 'low') {
+        lowCriticality.push(g);
+      } else if (!g.content.action) {
+        disambiguationCandidates.push(g);
+      } else if (g.content.action) {
+        actionable.push(g);
+      } else {
+        observational.push(g);
+      }
+    }
+
+    const batches: GuidelineMatchingBatch[] = [];
+
+    if (observational.length > 0) {
+      batches.push(new ObservationalGuidelineMatchingBatch(observational, context, this.generationInfo));
+    }
+
+    if (actionable.length > 0) {
+      batches.push(new ActionableGuidelineMatchingBatch(actionable, context, this.generationInfo));
+    }
+
+    if (lowCriticality.length > 0) {
+      batches.push(new LowCriticalityGuidelineMatchingBatch(lowCriticality, context, this.generationInfo));
+    }
+
+    return batches;
+  }
+
+  transformMatches(matches: GuidelineMatch[]): GuidelineMatch[] {
+    const seen = new Set<string>();
+    return matches.filter((m) => {
+      const key = m.guideline.id;
+      if (seen.has(key)) return false;
+      seen.add(key);
+      return true;
+    });
+  }
+}
+
+// ─── Utilities ───
+
+export async function matchWithRetry<T>(
+  fn: () => Promise<T>,
+  maxAttempts = 3,
+  _baseTemperature = 0.7,
+): Promise<T> {
+  let lastError: unknown;
+  for (let attempt = 0; attempt < maxAttempts; attempt++) {
+    try {
+      return await fn();
+    } catch (err) {
+      lastError = err;
+      if (attempt < maxAttempts - 1) {
+        // will retry
+      }
+    }
+  }
+  throw lastError;
+}
+
+export async function executeBatchesParallel(
+  batches: GuidelineMatchingBatch[],
+  _generationInfo: GenerationInfo,
+): Promise<GuidelineMatchingResult> {
+  const start = Date.now();
+  const results = await Promise.all(
+    batches.map((batch) => matchWithRetry(() => batch.process())),
+  );
+
+  const allBatches = results.map((r) => r.matches);
+  const allMatches = allBatches.flat();
+  const allGenInfos = results.map((r) => r.generationInfo);
+
+  return {
+    totalDuration: Date.now() - start,
+    batchCount: batches.length,
+    batchGenerations: allGenInfos,
+    batches: allBatches,
+    matches: allMatches,
+  };
+}
+
+export function createScoredMatch(
+  guidelineId: string,
+  score: number,
+  rationale: string,
+): ScoredMatch {
+  return { guideline_id: guidelineId, score, rationale };
+}

apps/coder/src/services/behavioral/resolver.ts

@@ -0,0 +1,355 @@
+/**
+ * Relational resolver for behavioral guidelines.
+ *
+ * Port of boocontext-audit/src/resolver.ts — resolves DEPENDS_ON,
+ * PRIORITIZES, ENTAILS, TAG_ALL, TAG_PRIORITIZES relationships
+ * with an iterative convergence loop.
+ */
+
+// ─── Relationship types (self-contained) ───
+
+export enum RelationshipKind {
+  DEPENDS_ON = 'depends_on',
+  PRIORITIZES = 'prioritizes',
+  ENTAILS = 'entails',
+  TAG_ALL = 'tag_all',
+  TAG_PRIORITIZES = 'tag_prioritizes',
+}
+
+export enum RelationshipEntityKind {
+  GUIDELINE = 'guideline',
+  TAG = 'tag',
+}
+
+export interface RelationshipEntity {
+  id: string;
+  kind: RelationshipEntityKind;
+}
+
+export interface Relationship {
+  id: string;
+  creation_utc: string;
+  source: RelationshipEntity;
+  target: RelationshipEntity;
+  kind: RelationshipKind;
+  group_id?: string;
+}
+
+/**
+ * Minimal relationship store interface.
+ * The resolver only needs listRelationships. Implementations
+ * can back against files, postgres, or in-memory maps.
+ */
+export interface RelationshipStore {
+  listRelationships(
+    kind?: RelationshipKind,
+    sourceId?: string,
+    targetId?: string,
+  ): Promise<Relationship[]>;
+}
+
+// ─── Resolution types ───
+
+export type ResolvedEntityType = 'guideline' | 'journey' | 'tag';
+
+export interface ResolvedEntity {
+  entityType: ResolvedEntityType;
+  entityId: string;
+}
+
+export enum ResolutionKind {
+  NONE = 'none',
+  UNMET_DEPENDENCY = 'unmet_dependency',
+  DEPRIORITIZED = 'deprioritized',
+  ENTAILED = 'entailed',
+}
+
+export interface Resolution {
+  kind: ResolutionKind;
+  description: string;
+  relationshipId?: string;
+  counterparts?: ResolvedEntity[];
+}
+
+export interface GuidelineStub {
+  id: string;
+  priority: number;
+  tags: string[];
+}
+
+export interface GuidelineMatchStub {
+  guideline: GuidelineStub;
+}
+
+export interface ResolverResult {
+  matchedIds: Set<string>;
+  resolutions: Map<string, Resolution[]>;
+  converged: boolean;
+  iterations: number;
+}
+
+// ─── Constants ───
+
+export const MAX_ITERATIONS = 100;
+
+// ─── RelationalResolver ───
+
+export class RelationalResolver {
+  private store: RelationshipStore;
+
+  constructor(store: RelationshipStore) {
+    this.store = store;
+  }
+
+  async resolve(
+    matchedIds: Set<string>,
+    allGuidelines: GuidelineStub[],
+  ): Promise<ResolverResult> {
+    const resolutions = new Map<string, Resolution[]>();
+    const guidelinesById = new Map(allGuidelines.map((g) => [g.id, g]));
+    let currentIds = new Set(matchedIds);
+    const priorityRemoved = new Set<string>();
+    const entailedIds = new Set<string>();
+
+    let converged = false;
+    let iterations = 0;
+
+    for (iterations = 0; iterations < MAX_ITERATIONS; iterations++) {
+      const candidateIds = new Set(
+        [...currentIds].filter((id) => !priorityRemoved.has(id)),
+      );
+
+      const step1Ids = await this.applyDependencies(candidateIds, guidelinesById, resolutions);
+
+      const step2Ids = await this.applyPrioritization(
+        step1Ids,
+        guidelinesById,
+        resolutions,
+        priorityRemoved,
+      );
+
+      const step3Ids = this.applyNumericalPriority(
+        step2Ids,
+        guidelinesById,
+        resolutions,
+        priorityRemoved,
+        entailedIds,
+      );
+
+      const step4Ids = await this.applyEntailment(
+        step3Ids,
+        guidelinesById,
+        resolutions,
+        priorityRemoved,
+        entailedIds,
+      );
+
+      if (this.setsEqual(step4Ids, currentIds)) {
+        converged = true;
+        break;
+      }
+
+      currentIds = step4Ids;
+    }
+
+    for (const id of allGuidelines.map((g) => g.id)) {
+      if (!resolutions.has(id)) {
+        resolutions.set(id, [
+          { kind: ResolutionKind.NONE, description: 'No relational changes' },
+        ]);
+      }
+    }
+
+    return {
+      matchedIds: currentIds,
+      resolutions,
+      converged,
+      iterations: iterations + 1,
+    };
+  }
+
+  // ── Private steps ──
+
+  private async applyDependencies(
+    candidateIds: Set<string>,
+    _guidelinesById: Map<string, GuidelineStub>,
+    resolutions: Map<string, Resolution[]>,
+  ): Promise<Set<string>> {
+    const surviving = new Set(candidateIds);
+    const cache = new Map<string, Relationship[]>();
+
+    for (const gid of candidateIds) {
+      const rels = await this.getRelationshipsFromCache(cache, gid, RelationshipKind.DEPENDS_ON);
+
+      for (const rel of rels) {
+        const targetId = rel.target.id;
+        if (!candidateIds.has(targetId)) {
+          surviving.delete(gid);
+          this.addResolution(resolutions, gid, {
+            kind: ResolutionKind.UNMET_DEPENDENCY,
+            description: `Depends on ${targetId} which is not matched`,
+            relationshipId: rel.id,
+            counterparts: [{ entityType: 'guideline' as const, entityId: targetId }],
+          });
+          break;
+        }
+      }
+    }
+
+    return surviving;
+  }
+
+  private async applyPrioritization(
+    candidateIds: Set<string>,
+    guidelinesById: Map<string, GuidelineStub>,
+    resolutions: Map<string, Resolution[]>,
+    priorityRemoved: Set<string>,
+  ): Promise<Set<string>> {
+    const surviving = new Set(candidateIds);
+    const cache = new Map<string, Relationship[]>();
+
+    for (const gid of candidateIds) {
+      if (priorityRemoved.has(gid)) continue;
+
+      const allRels = await this.getAllRelationships(cache, gid);
+      const priorityRels = allRels.filter((r) => r.kind === RelationshipKind.PRIORITIZES);
+
+      for (const rel of priorityRels) {
+        const sourceId = rel.source.id;
+        if (sourceId !== gid) continue;
+        const targetId = rel.target.id;
+
+        if (candidateIds.has(targetId)) {
+          surviving.delete(targetId);
+          priorityRemoved.add(targetId);
+          this.addResolution(resolutions, targetId, {
+            kind: ResolutionKind.DEPRIORITIZED,
+            description: `Deprioritized by ${gid}`,
+            relationshipId: rel.id,
+            counterparts: [{ entityType: 'guideline' as const, entityId: gid }],
+          });
+        }
+      }
+    }
+
+    return surviving;
+  }
+
+  private applyNumericalPriority(
+    candidateIds: Set<string>,
+    guidelinesById: Map<string, GuidelineStub>,
+    resolutions: Map<string, Resolution[]>,
+    priorityRemoved: Set<string>,
+    entailedIds: Set<string>,
+  ): Set<string> {
+    if (candidateIds.size === 0) return candidateIds;
+
+    const nonEntailed = [...candidateIds].filter((id) => !entailedIds.has(id));
+    const entailed = [...candidateIds].filter((id) => entailedIds.has(id));
+
+    if (nonEntailed.length === 0) return new Set(entailed);
+
+    const priorities = nonEntailed.map((id) => guidelinesById.get(id)?.priority ?? 0);
+    const maxPriority = Math.max(...priorities);
+
+    const surviving = new Set<string>();
+
+    for (const id of nonEntailed) {
+      const priority = guidelinesById.get(id)?.priority ?? 0;
+      if (priority >= maxPriority) {
+        surviving.add(id);
+      } else {
+        priorityRemoved.add(id);
+        this.addResolution(resolutions, id, {
+          kind: ResolutionKind.DEPRIORITIZED,
+          description: `Lower priority (${priority} < ${maxPriority})`,
+        });
+      }
+    }
+
+    for (const id of entailed) {
+      surviving.add(id);
+    }
+
+    return surviving;
+  }
+
+  private async applyEntailment(
+    candidateIds: Set<string>,
+    guidelinesById: Map<string, GuidelineStub>,
+    resolutions: Map<string, Resolution[]>,
+    priorityRemoved: Set<string>,
+    entailedIds: Set<string>,
+  ): Promise<Set<string>> {
+    const result = new Set(candidateIds);
+    const cache = new Map<string, Relationship[]>();
+
+    for (const gid of candidateIds) {
+      if (priorityRemoved.has(gid)) continue;
+
+      const allRels = await this.getAllRelationships(cache, gid);
+      const entailRels = allRels.filter((r) => r.kind === RelationshipKind.ENTAILS);
+
+      for (const rel of entailRels) {
+        const targetId = rel.target.id;
+        if (!guidelinesById.has(targetId)) continue;
+        if (priorityRemoved.has(targetId)) continue;
+        if (entailedIds.has(targetId)) continue;
+
+        result.add(targetId);
+        entailedIds.add(targetId);
+        this.addResolution(resolutions, targetId, {
+          kind: ResolutionKind.ENTAILED,
+          description: `Entailed by ${gid}`,
+          relationshipId: rel.id,
+          counterparts: [{ entityType: 'guideline' as const, entityId: gid }],
+        });
+      }
+    }
+
+    return result;
+  }
+
+  // ── Cache helpers ──
+
+  private async getRelationshipsFromCache(
+    cache: Map<string, Relationship[]>,
+    gid: string,
+    kind: RelationshipKind,
+  ): Promise<Relationship[]> {
+    const key = `${kind}:${gid}`;
+    if (!cache.has(key)) {
+      cache.set(key, await this.store.listRelationships(kind, gid));
+    }
+    return cache.get(key)!;
+  }
+
+  private async getAllRelationships(
+    cache: Map<string, Relationship[]>,
+    gid: string,
+  ): Promise<Relationship[]> {
+    const result: Relationship[] = [];
+    const kinds = Object.values(RelationshipKind) as RelationshipKind[];
+    for (const kind of kinds) {
+      const rels = await this.getRelationshipsFromCache(cache, gid, kind);
+      const targetRels = await this.getRelationshipsFromCache(cache, `target:${gid}`, kind);
+      result.push(...rels, ...targetRels);
+    }
+    return result;
+  }
+
+  private addResolution(
+    resolutions: Map<string, Resolution[]>,
+    id: string,
+    resolution: Resolution,
+  ): void {
+    if (!resolutions.has(id)) resolutions.set(id, []);
+    resolutions.get(id)!.push(resolution);
+  }
+
+  private setsEqual(a: Set<string>, b: Set<string>): boolean {
+    if (a.size !== b.size) return false;
+    for (const item of a) if (!b.has(item)) return false;
+    return true;
+  }
+}

apps/coder/src/services/dispatcher.ts

@@ -30,6 +30,7 @@ import {
   type TerminalMessageStatus,
 } from './finalize-message.js';
 import { shouldFailOnMissingAgent } from './flow-runner-decisions.js';
+import { emitHook } from '../plugins/host.js';
 
 interface InferenceRunner {
   enqueue: (
@@ -123,6 +124,22 @@ export function createDispatcher(deps: Deps): {
     publishAgentStatus(broker.publishFrame, sessionId, chatId, agent, status, reason);
   }
 
+  // EmitHook: fire-and-forget turn.end notification. Best-effort — a hook throwing
+  // is silently swallowed so it never blocks the dispatch flow.
+  function emitTurnEnd(
+    sessionId: string,
+    taskId: string,
+    state: string,
+    agent?: string | null,
+    model?: string | null,
+    outputSummary?: string,
+  ): void {
+    void emitHook('turn.end', {
+      sessionId,
+      turnSummary: { taskId, state, agent, model: model ?? undefined, outputSummary },
+    });
+  }
+
   // F1 (OCE-001/OCE-002): finalize a streaming assistant message into a terminal
   // state and publish the matching message_complete frame. Best-effort + idempotent
   // (the helper's `WHERE status='streaming'` guard) — a failure here must never mask
@@ -318,6 +335,7 @@ export function createDispatcher(deps: Deps): {
 
     // Declared before try so the catch block can write it back on the task row.
     let chatId: string | null = null;
+    let sessionId: string | undefined;
 
     try {
       // Mark running
@@ -330,7 +348,6 @@ export function createDispatcher(deps: Deps): {
       // Session setup: reuse a pre-created session (e.g. Q&A arena contestants
       // whose persona is stamped on the session via agent_id) or create a fresh one.
       const model = task.model ?? config.DEFAULT_MODEL;
-      let sessionId: string;
       if (task.session_id) {
         sessionId = task.session_id;
       } else {
@@ -377,6 +394,7 @@ export function createDispatcher(deps: Deps): {
           SET state = 'cancelled', ended_at = clock_timestamp()
           WHERE id = ${taskId}
         `;
+        if (sessionId) emitTurnEnd(sessionId, taskId, 'cancelled', null, task.model);
         return;
       }
 
@@ -399,6 +417,7 @@ export function createDispatcher(deps: Deps): {
           WHERE id = ${taskId}
         `;
         log.info({ taskId, costTokens }, 'dispatcher: task completed (native)');
+        emitTurnEnd(sessionId, taskId, 'completed', null, task.model, summary);
       } else {
         const [msg] = await sql<{ content: string | null }[]>`
           SELECT content FROM messages WHERE id = ${assistantId}
@@ -410,6 +429,7 @@ export function createDispatcher(deps: Deps): {
           WHERE id = ${taskId}
         `;
         log.warn({ taskId, finalStatus }, 'dispatcher: task failed (native)');
+        emitTurnEnd(sessionId, taskId, 'failed', null, task.model, summary);
       }
     } catch (err) {
       const errMsg = err instanceof Error ? err.message : String(err);
@@ -419,6 +439,7 @@ export function createDispatcher(deps: Deps): {
         SET state = 'failed', ended_at = clock_timestamp(), output_summary = ${errMsg.slice(0, 500)}, chat_id = ${chatId}
         WHERE id = ${taskId}
       `.catch(() => {});
+      if (sessionId) emitTurnEnd(sessionId, taskId, 'failed', null, task.model, errMsg);
     }
   }
 
@@ -684,6 +705,7 @@ export function createDispatcher(deps: Deps): {
         await finalizeMessage(sessionId, chatId, assistantId, 'cancelled', task.model, assistantContent);
         await sql`UPDATE tasks SET state = 'cancelled', ended_at = clock_timestamp() WHERE id = ${taskId}`;
         emitAgentStatus(sessionId, chatId, agent, 'idle', stopping ? 'shutdown' : 'cancelled');
+        emitTurnEnd(sessionId, taskId, 'cancelled', agent, task.model);
         await cleanupWorktree(projectPath, taskId);
         clearTaskCommands(taskId);
         return;
@@ -738,6 +760,7 @@ export function createDispatcher(deps: Deps): {
       log.info({ taskId, agent, costTokens: extCostTokens }, 'dispatcher: task completed (external)');
       // #10: external-agent turn completed cleanly.
       emitAgentStatus(sessionId, chatId, agent, 'idle', 'turn_complete');
+      emitTurnEnd(sessionId, taskId, 'completed', agent, task.model, outputSummary);
       clearTaskCommands(taskId);
 
     } catch (err) {
@@ -762,6 +785,7 @@ export function createDispatcher(deps: Deps): {
       // preceded its assignment — guard so the status publish never masks the real
       // error.
       if (chatId) emitAgentStatus(sessionId, chatId, agent, status === 'cancelled' ? 'idle' : 'error', status === 'cancelled' ? 'cancelled' : 'failed');
+      if (sessionId) emitTurnEnd(sessionId, taskId, status, agent, task.model, errMsg);
 
       // Best-effort cleanup
       await cleanupWorktree(projectPath, taskId);
@@ -1030,6 +1054,7 @@ export function createDispatcher(deps: Deps): {
         await finalizeMessage(sessionId, chatId, assistantId, 'cancelled', task.model, assistantContent);
         await sql`UPDATE tasks SET state = 'cancelled', ended_at = clock_timestamp() WHERE id = ${taskId}`;
         emitAgentStatus(sessionId, chatId, agent, 'idle', stopping ? 'shutdown' : 'cancelled');
+        emitTurnEnd(sessionId, taskId, 'cancelled', agent, task.model);
         clearTaskCommands(taskId);
         return; // worktree persists (no cleanup); backend stays warm
       }
@@ -1090,6 +1115,7 @@ export function createDispatcher(deps: Deps): {
         result.ok ? 'idle' : 'error',
         result.ok ? 'turn_complete' : 'failed',
       );
+      emitTurnEnd(sessionId, taskId, finalState, agent, task.model, outputSummary);
       clearTaskCommands(taskId);
     } catch (err) {
       const errMsg = err instanceof Error ? err.message : String(err);
@@ -1104,6 +1130,7 @@ export function createDispatcher(deps: Deps): {
       await finalizeMessage(sessionId, chatId, assistantId, status, task.model);
       // #10: turn crashed.
       if (chatId) emitAgentStatus(sessionId, chatId, agent, status === 'cancelled' ? 'idle' : 'error', status === 'cancelled' ? 'cancelled' : 'crashed');
+      if (sessionId) emitTurnEnd(sessionId, taskId, status, agent, task.model, errMsg);
       clearTaskCommands(taskId);
       // No worktree cleanup (persistent); backend stays warm for the next turn.
     }
@@ -1308,6 +1335,7 @@ export function createDispatcher(deps: Deps): {
         await finalizeMessage(sessionId, chatId, assistantId, 'cancelled', task.model, assistantContent);
         await sql`UPDATE tasks SET state = 'cancelled', ended_at = clock_timestamp() WHERE id = ${taskId}`;
         emitAgentStatus(sessionId, chatId, agent, 'idle', stopping ? 'shutdown' : 'cancelled');
+        emitTurnEnd(sessionId, taskId, 'cancelled', agent, task.model);
         clearTaskCommands(taskId);
         return; // worktree persists (no cleanup); backend stays warm
       }
@@ -1367,6 +1395,7 @@ export function createDispatcher(deps: Deps): {
         result.ok ? 'idle' : 'error',
         result.ok ? 'turn_complete' : 'failed',
       );
+      emitTurnEnd(sessionId, taskId, finalState, agent, task.model, outputSummary);
       clearTaskCommands(taskId);
     } catch (err) {
       const errMsg = err instanceof Error ? err.message : String(err);
@@ -1381,6 +1410,7 @@ export function createDispatcher(deps: Deps): {
       await finalizeMessage(sessionId, chatId, assistantId, status, task.model);
       // #10: turn crashed.
       emitAgentStatus(sessionId, chatId, agent, status === 'cancelled' ? 'idle' : 'error', status === 'cancelled' ? 'cancelled' : 'crashed');
+      emitTurnEnd(sessionId, taskId, status, agent, task.model, errMsg);
       clearTaskCommands(taskId);
       // No worktree cleanup (persistent); backend stays warm for the next turn.
     }
@@ -1576,6 +1606,7 @@ export function createDispatcher(deps: Deps): {
         await finalizeMessage(sessionId, chatId, assistantId, 'cancelled', task.model, assistantContent);
         await sql`UPDATE tasks SET state = 'cancelled', ended_at = clock_timestamp() WHERE id = ${taskId}`;
         emitAgentStatus(sessionId, chatId, agent, 'idle', stopping ? 'shutdown' : 'cancelled');
+        emitTurnEnd(sessionId, taskId, 'cancelled', agent, task.model);
         clearTaskCommands(taskId);
         return; // worktree persists (no cleanup); backend stays warm
       }
@@ -1638,6 +1669,7 @@ export function createDispatcher(deps: Deps): {
         result.ok ? 'idle' : 'error',
         result.ok ? 'turn_complete' : 'failed',
       );
+      emitTurnEnd(sessionId, taskId, finalState, agent, task.model, outputSummary);
       clearTaskCommands(taskId);
     } catch (err) {
       const errMsg = err instanceof Error ? err.message : String(err);
@@ -1652,6 +1684,7 @@ export function createDispatcher(deps: Deps): {
       await finalizeMessage(sessionId, chatId, assistantId, status, task.model);
       // #10: turn crashed.
       emitAgentStatus(sessionId, chatId, agent, status === 'cancelled' ? 'idle' : 'error', status === 'cancelled' ? 'cancelled' : 'crashed');
+      emitTurnEnd(sessionId, taskId, status, agent, task.model, errMsg);
       clearTaskCommands(taskId);
       // No worktree cleanup (persistent); backend stays warm for the next turn.
     }

apps/coder/src/services/flow-runner.ts

@@ -89,6 +89,8 @@ interface Deps {
   broker: Broker;
   log: FastifyBaseLogger;
   config: Config;
+  /** Fired when a flow run reaches a terminal state (for plan-store integration). */
+  onRunTerminal?: (runId: string, status: 'completed' | 'failed' | 'cancelled') => void;
 }
 
 interface FlowStepRow {
@@ -479,6 +481,7 @@ export function createFlowRunner(deps: Deps): FlowRunner {
       WHERE id = ${runId} AND status = 'running'
     `;
     if (updated.count === 0) return; // already terminal (e.g. cancelled) — don't publish
+    deps.onRunTerminal?.(runId, 'completed');
     publishStep(runId, lastAgentStepId(flow, input, model), 'completed', {
       run_status: 'completed',
       report,
@@ -498,6 +501,7 @@ export function createFlowRunner(deps: Deps): FlowRunner {
       WHERE id = ${runId} AND status = 'running'
     `;
     if (updated.count === 0) return;
+    deps.onRunTerminal?.(runId, 'failed');
     const stepId = failedStepId ?? (flow ? lastAgentStepId(flow, input, model) : 'run');
     log.warn({ runId, error }, 'flow-runner: run failed');
     await appendStepEvent(sql, runId, stepId, 'failed', { error });
@@ -512,6 +516,7 @@ export function createFlowRunner(deps: Deps): FlowRunner {
       WHERE id = ${runId} AND status = 'running'
     `;
     if (updated.count === 0) return; // idempotent — already terminal
+    deps.onRunTerminal?.(runId, 'cancelled');
     // Any remaining pending steps are unreachable; mark + publish them so the
     // pane can show them as cancelled rather than stuck in pending.
     const pending = await sql<{ step_id: string; kind: string }[]>`
@@ -742,6 +747,7 @@ export function createFlowRunner(deps: Deps): FlowRunner {
       WHERE id = ${runId} AND status = 'running'
     `;
     if (updated.count === 0) return { cancelled: false, taskIds: [] };
+    deps.onRunTerminal?.(runId, 'cancelled');
 
     // Mark all non-terminal steps cancelled and collect in-flight task_ids.
     const steps = await sql<{ step_id: string; task_id: string | null; kind: string }[]>`

apps/coder/src/services/frame-emitter.ts

@@ -19,9 +19,10 @@
 import type { Broker } from '@boocode/server/broker';
 import type { WsFrame } from '@boocode/contracts/ws-frames';
 import type { AgentEvent } from './agent-backend.js';
-import { type AcpToolSnapshot, snapshotToWireToolCall } from './acp-tool-snapshot.js';
+import { type AcpToolSnapshot, snapshotToWireToolCall, mapToolLifecycleStatus } from './acp-tool-snapshot.js';
 import { mergeTaskCommands, getTaskCommands } from './agent-commands-cache.js';
 import type { DcpStreamStripper } from './dcp-strip.js';
+import { emitHook } from '../plugins/host.js';
 
 export interface FrameEmitterOpts {
   broker?: Broker;
@@ -91,8 +92,29 @@ export function makeFrameEmitter(opts: FrameEmitterOpts): FrameEmitter {
         }
         break;
       case 'tool_call':
+        toolSnapshots.set(e.toolCall.toolCallId, e.toolCall);
+        if (canStream()) {
+          broker!.publishFrame(sessionId!, {
+            type: 'tool_call',
+            message_id: assistantId!,
+            chat_id: chatId!,
+            tool_call: snapshotToWireToolCall(e.toolCall),
+          } as WsFrame);
+        }
+        break;
       case 'tool_update':
         toolSnapshots.set(e.toolCall.toolCallId, e.toolCall);
+        {
+          const lifecycle = mapToolLifecycleStatus(e.toolCall.status, e.toolCall.rawOutput);
+          if (lifecycle === 'completed' || lifecycle === 'failed') {
+            void emitHook('tool.execute.after', {
+              toolName: e.toolCall.title,
+              args: e.toolCall.rawInput,
+              result: e.toolCall.rawOutput,
+              duration: undefined,
+            });
+          }
+        }
         if (canStream()) {
           broker!.publishFrame(sessionId!, {
             type: 'tool_call',

apps/coder/src/services/hashline/constants.ts

@@ -0,0 +1,10 @@
+export const NIBBLE_STR = "ZPMQVRWSNKTXJBYH"
+
+export const HASHLINE_DICT = Array.from({ length: 256 }, (_, i) => {
+  const high = i >>> 4
+  const low = i & 0x0f
+  return `${NIBBLE_STR[high]}${NIBBLE_STR[low]}`
+})
+
+export const HASHLINE_REF_PATTERN = /^([0-9]+)#([ZPMQVRWSNKTXJBYH]{2})$/
+export const HASHLINE_OUTPUT_PATTERN = /^([0-9]+)#([ZPMQVRWSNKTXJBYH]{2})\|(.*)$/

apps/coder/src/services/hashline/hash-computation.ts

@@ -0,0 +1,31 @@
+import { HASHLINE_DICT } from "./constants.js"
+import { hashXxh32 } from "./xxhash32.js"
+
+const RE_SIGNIFICANT = /[\p{L}\p{N}]/u
+
+function computeNormalizedLineHash(lineNumber: number, normalizedContent: string): string {
+  const stripped = normalizedContent
+  const seed = RE_SIGNIFICANT.test(stripped) ? 0 : lineNumber
+  const hash = hashXxh32(stripped, seed)
+  const index = hash % 256
+  return HASHLINE_DICT[index]!
+}
+
+export function computeLineHash(lineNumber: number, content: string): string {
+  return computeNormalizedLineHash(lineNumber, content.replace(/\r/g, "").trimEnd())
+}
+
+export function computeLegacyLineHash(lineNumber: number, content: string): string {
+  return computeNormalizedLineHash(lineNumber, content.replace(/\r/g, "").replace(/\s+/g, ""))
+}
+
+export function formatHashLine(lineNumber: number, content: string): string {
+  const hash = computeLineHash(lineNumber, content)
+  return `${lineNumber}#${hash}|${content}`
+}
+
+export function formatHashLines(content: string): string {
+  if (!content) return ""
+  const lines = content.split("\n")
+  return lines.map((line, index) => formatHashLine(index + 1, line)).join("\n")
+}

apps/coder/src/services/hashline/index.ts

@@ -0,0 +1,11 @@
+/**
+ * Hashline editing core — content-hash anchors for edit_file stale-patch detection.
+ *
+ * Ported from oh-my-openagent/packages/hashline-core/.
+ * Bundles a runtime-aware xxHash32 (Bun fast-path, pure-JS fallback).
+ */
+export { computeLineHash, formatHashLines, formatHashLine, computeLegacyLineHash } from "./hash-computation.js"
+export { parseLineRef, validateLineRef, validateLineRefs, HashlineMismatchError, normalizeLineRef } from "./validation.js"
+export type { LineRef } from "./validation.js"
+export { NIBBLE_STR, HASHLINE_DICT, HASHLINE_REF_PATTERN, HASHLINE_OUTPUT_PATTERN } from "./constants.js"
+export type { ReplaceEdit, AppendEdit, PrependEdit, HashlineEdit } from "./types.js"

apps/coder/src/services/hashline/types.ts

@@ -0,0 +1,20 @@
+export interface ReplaceEdit {
+  op: "replace"
+  pos: string
+  end?: string
+  lines: string | string[]
+}
+
+export interface AppendEdit {
+  op: "append"
+  pos?: string
+  lines: string | string[]
+}
+
+export interface PrependEdit {
+  op: "prepend"
+  pos?: string
+  lines: string | string[]
+}
+
+export type HashlineEdit = ReplaceEdit | AppendEdit | PrependEdit

apps/coder/src/services/hashline/validation.ts

@@ -0,0 +1,192 @@
+import { computeLegacyLineHash, computeLineHash } from "./hash-computation.js"
+import { HASHLINE_REF_PATTERN } from "./constants.js"
+
+export interface LineRef {
+  line: number
+  hash: string
+}
+
+interface HashMismatch {
+  line: number
+  expected: string
+}
+
+const MISMATCH_CONTEXT = 2
+
+const LINE_REF_EXTRACT_PATTERN = /([0-9]+#[ZPMQVRWSNKTXJBYH]{2})/
+
+function isCompatibleLineHash(line: number, content: string, hash: string): boolean {
+  return computeLineHash(line, content) === hash || computeLegacyLineHash(line, content) === hash
+}
+
+export function normalizeLineRef(ref: string): string {
+  const originalTrimmed = ref.trim()
+  let trimmed = originalTrimmed
+  trimmed = trimmed.replace(/^(?:>>>|[+-])\s*/, "")
+  trimmed = trimmed.replace(/\s*#\s*/, "#")
+  trimmed = trimmed.replace(/\|.*$/, "")
+  trimmed = trimmed.trim()
+
+  if (HASHLINE_REF_PATTERN.test(trimmed)) {
+    return trimmed
+  }
+
+  const extracted = trimmed.match(LINE_REF_EXTRACT_PATTERN)
+  if (extracted) {
+    return extracted[1]!
+  }
+
+  return originalTrimmed
+}
+
+export function parseLineRef(ref: string): LineRef {
+  const normalized = normalizeLineRef(ref)
+  const match = normalized.match(HASHLINE_REF_PATTERN)
+  if (match) {
+    return {
+      line: Number.parseInt(match[1]!, 10),
+      hash: match[2]!,
+    }
+  }
+  const hashIdx = normalized.indexOf('#')
+  if (hashIdx > 0) {
+    const prefix = normalized.slice(0, hashIdx)
+    const suffix = normalized.slice(hashIdx + 1)
+    if (!/^\d+$/.test(prefix) && /^[ZPMQVRWSNKTXJBYH]{2}$/.test(suffix)) {
+      throw new Error(
+        `Invalid line reference: "${ref}". "${prefix}" is not a line number. ` +
+          `Use the actual line number from the read output.`
+      )
+    }
+  }
+  throw new Error(
+    `Invalid line reference format: "${ref}". Expected format: "{line_number}#{hash_id}"`
+  )
+}
+
+export function validateLineRef(lines: string[], ref: string): void {
+  const { line, hash } = parseLineRefWithHint(ref, lines)
+
+  if (line < 1 || line > lines.length) {
+    throw new Error(
+      `Line number ${line} out of bounds. File has ${lines.length} lines.`
+    )
+  }
+
+  const content = lines[line - 1]
+  if (content === undefined) {
+    throw new Error(
+      `Line number ${line} out of bounds. File has ${lines.length} lines.`
+    )
+  }
+  if (!isCompatibleLineHash(line, content, hash)) {
+    throw new HashlineMismatchError([{ line, expected: hash }], lines)
+  }
+}
+
+export class HashlineMismatchError extends Error {
+  readonly remaps: ReadonlyMap<string, string>
+
+  constructor(
+    private readonly mismatches: HashMismatch[],
+    private readonly fileLines: string[]
+  ) {
+    super(HashlineMismatchError.formatMessage(mismatches, fileLines))
+    this.name = "HashlineMismatchError"
+    const remaps = new Map<string, string>()
+    for (const mismatch of mismatches) {
+      const content = fileLines[mismatch.line - 1]
+      const actualLine = content ?? ""
+      const actual = computeLineHash(mismatch.line, actualLine)
+      remaps.set(`${mismatch.line}#${mismatch.expected}`, `${mismatch.line}#${actual}`)
+    }
+    this.remaps = remaps
+  }
+
+  static formatMessage(mismatches: HashMismatch[], fileLines: string[]): string {
+    const mismatchByLine = new Map<number, HashMismatch>()
+    for (const mismatch of mismatches) mismatchByLine.set(mismatch.line, mismatch)
+
+    const displayLines = new Set<number>()
+    for (const mismatch of mismatches) {
+      const low = Math.max(1, mismatch.line - MISMATCH_CONTEXT)
+      const high = Math.min(fileLines.length, mismatch.line + MISMATCH_CONTEXT)
+      for (let line = low; line <= high; line++) displayLines.add(line)
+    }
+
+    const sortedLines = [...displayLines].sort((a, b) => a - b)
+    const output: string[] = []
+    output.push(
+      `${mismatches.length} line${mismatches.length > 1 ? "s have" : " has"} changed since last read. ` +
+        "Use updated {line_number}#{hash_id} references below (>>> marks changed lines)."
+    )
+    output.push("")
+
+    let previousLine = -1
+    for (const line of sortedLines) {
+      if (previousLine !== -1 && line > previousLine + 1) {
+        output.push("    ...")
+      }
+      previousLine = line
+
+      const content = fileLines[line - 1] ?? ""
+      const hash = computeLineHash(line, content)
+      const prefix = `${line}#${hash}|${content}`
+      if (mismatchByLine.has(line)) {
+        output.push(`>>> ${prefix}`)
+      } else {
+        output.push(`    ${prefix}`)
+      }
+    }
+
+    return output.join("\n")
+  }
+}
+
+function suggestLineForHash(ref: string, lines: string[]): string | null {
+  const hashMatch = ref.trim().match(/#([ZPMQVRWSNKTXJBYH]{2})$/)
+  if (!hashMatch) return null
+  const hash = hashMatch[1]!
+  for (let i = 0; i < lines.length; i++) {
+    if (isCompatibleLineHash(i + 1, lines[i] ?? "", hash)) {
+      return `Did you mean "${i + 1}#${computeLineHash(i + 1, lines[i] ?? "")}"?`
+    }
+  }
+  return null
+}
+
+function parseLineRefWithHint(ref: string, lines: string[]): LineRef {
+  try {
+    return parseLineRef(ref)
+  } catch (parseError) {
+    const hint = suggestLineForHash(ref, lines)
+    if (hint && parseError instanceof Error) {
+      throw new Error(`${parseError.message} ${hint}`)
+    }
+    throw parseError
+  }
+}
+
+export function validateLineRefs(lines: string[], refs: string[]): void {
+  const mismatches: HashMismatch[] = []
+
+  for (const ref of refs) {
+    const { line, hash } = parseLineRefWithHint(ref, lines)
+
+    if (line < 1 || line > lines.length) {
+      throw new Error(`Line number ${line} out of bounds (file has ${lines.length} lines)`)
+    }
+
+    const content = lines[line - 1]
+    if (content === undefined) {
+      throw new Error(`Line number ${line} out of bounds (file has ${lines.length} lines)`)
+    }
+    if (!isCompatibleLineHash(line, content, hash)) {
+      mismatches.push({ line, expected: hash })
+    }
+  }
+
+  if (mismatches.length > 0) {
+    throw new HashlineMismatchError(mismatches, lines)
+  }
+}

apps/coder/src/services/hashline/xxhash32.ts

@@ -0,0 +1,90 @@
+type BunHashRuntime = { hash: { xxHash32(data: string | Uint8Array, seed: number): number } }
+
+const runtime = globalThis as typeof globalThis & { Bun?: BunHashRuntime }
+const encoder = new TextEncoder()
+
+const PRIME32_1 = 0x9e3779b1
+const PRIME32_2 = 0x85ebca77
+const PRIME32_3 = 0xc2b2ae3d
+const PRIME32_4 = 0x27d4eb2f
+const PRIME32_5 = 0x165667b1
+
+function rotateLeft32(value: number, bits: number): number {
+  return ((value << bits) | (value >>> (32 - bits))) >>> 0
+}
+
+function readUint32LittleEndian(input: Uint8Array, offset: number): number {
+  return (
+    ((input[offset] ?? 0) |
+      ((input[offset + 1] ?? 0) << 8) |
+      ((input[offset + 2] ?? 0) << 16) |
+      ((input[offset + 3] ?? 0) << 24)) >>>
+    0
+  )
+}
+
+function round32(accumulator: number, value: number): number {
+  const added = (accumulator + Math.imul(value, PRIME32_2)) >>> 0
+  return Math.imul(rotateLeft32(added, 13), PRIME32_1) >>> 0
+}
+
+function xxHash32Js(input: Uint8Array, seed: number): number {
+  let offset = 0
+  const length = input.length
+  let hash: number
+
+  if (length >= 16) {
+    const limit = length - 16
+    let value1 = (seed + PRIME32_1 + PRIME32_2) >>> 0
+    let value2 = (seed + PRIME32_2) >>> 0
+    let value3 = seed >>> 0
+    let value4 = (seed - PRIME32_1) >>> 0
+
+    while (offset <= limit) {
+      value1 = round32(value1, readUint32LittleEndian(input, offset))
+      offset += 4
+      value2 = round32(value2, readUint32LittleEndian(input, offset))
+      offset += 4
+      value3 = round32(value3, readUint32LittleEndian(input, offset))
+      offset += 4
+      value4 = round32(value4, readUint32LittleEndian(input, offset))
+      offset += 4
+    }
+
+    hash = (rotateLeft32(value1, 1) + rotateLeft32(value2, 7)) >>> 0
+    hash = (hash + rotateLeft32(value3, 12)) >>> 0
+    hash = (hash + rotateLeft32(value4, 18)) >>> 0
+  } else {
+    hash = (seed + PRIME32_5) >>> 0
+  }
+
+  hash = (hash + length) >>> 0
+
+  while (offset + 4 <= length) {
+    hash = (hash + Math.imul(readUint32LittleEndian(input, offset), PRIME32_3)) >>> 0
+    hash = Math.imul(rotateLeft32(hash, 17), PRIME32_4) >>> 0
+    offset += 4
+  }
+
+  while (offset < length) {
+    hash = (hash + Math.imul(input[offset] ?? 0, PRIME32_5)) >>> 0
+    hash = Math.imul(rotateLeft32(hash, 11), PRIME32_1) >>> 0
+    offset += 1
+  }
+
+  hash = (hash ^ (hash >>> 15)) >>> 0
+  hash = Math.imul(hash, PRIME32_2) >>> 0
+  hash = (hash ^ (hash >>> 13)) >>> 0
+  hash = Math.imul(hash, PRIME32_3) >>> 0
+
+  return (hash ^ (hash >>> 16)) >>> 0
+}
+
+export function hashXxh32(input: string, seed: number): number {
+  const bun = runtime.Bun
+  if (bun !== undefined) {
+    return bun.hash.xxHash32(input, seed)
+  }
+
+  return xxHash32Js(encoder.encode(input), seed >>> 0)
+}

apps/coder/src/services/model-resolution/connected-providers-cache.ts

@@ -0,0 +1,34 @@
+import type { ModelMetadata } from "./provider-cache.js"
+
+export interface ProviderModelsCache {
+  readonly models: Record<string, readonly string[] | readonly ModelMetadata[]>
+  readonly connected: readonly string[]
+  readonly updatedAt: string
+}
+
+export interface ConnectedProvidersAdapter {
+  readConnectedProvidersCache(): string[] | null
+  findProviderModelMetadata(providerID: string, modelID: string): ModelMetadata | undefined
+  readProviderModelsCache(): ProviderModelsCache | null
+}
+
+export function readConnectedProvidersCache(): string[] | null {
+  return null
+}
+
+export function findProviderModelMetadata(
+  _providerID: string,
+  _modelID: string,
+): ModelMetadata | undefined {
+  return undefined
+}
+
+export function readProviderModelsCache(): ProviderModelsCache | null {
+  return null
+}
+
+export const connectedProvidersAdapter: ConnectedProvidersAdapter = {
+  readConnectedProvidersCache,
+  findProviderModelMetadata,
+  readProviderModelsCache,
+}

apps/coder/src/services/model-resolution/fallback-chain-from-models.ts

@@ -0,0 +1,128 @@
+import type { FallbackEntry } from "./model-requirement-types.js"
+import type { FallbackModelObject } from "./fallback-model-object.js"
+import { normalizeFallbackModels } from "./model-resolver.js"
+import { KNOWN_VARIANTS } from "./known-variants.js"
+
+function parseVariantFromModel(rawModel: string): { modelID: string; variant?: string } {
+  if (typeof rawModel !== "string") {
+    return { modelID: "" }
+  }
+  const trimmedModel = rawModel.trim()
+  if (!trimmedModel) {
+    return { modelID: "" }
+  }
+
+  const parenthesizedVariant = trimmedModel.match(/^(.*)\(([^()]+)\)\s*$/)
+  if (parenthesizedVariant) {
+    const modelID = parenthesizedVariant[1]?.trim() ?? ""
+    const variant = parenthesizedVariant[2]?.trim()
+    return variant ? { modelID, variant } : { modelID }
+  }
+
+  const spaceVariant = trimmedModel.match(/^(.*\S)\s+([a-z][a-z0-9_-]*)$/i)
+  if (spaceVariant) {
+    const modelID = spaceVariant[1]?.trim() ?? ""
+    const variant = spaceVariant[2]?.trim().toLowerCase()
+    if (variant && KNOWN_VARIANTS.has(variant)) {
+      return { modelID, variant }
+    }
+  }
+
+  return { modelID: trimmedModel }
+}
+
+export function parseFallbackModelEntry(
+  model: string,
+  contextProviderID: string | undefined,
+  defaultProviderID = "opencode",
+): FallbackEntry | undefined {
+  if (typeof model !== "string") return undefined
+  const trimmed = model.trim()
+  if (!trimmed) return undefined
+
+  const parts = trimmed.split("/")
+  const providerID =
+    parts.length >= 2 ? (parts[0]?.trim() ?? "") : (contextProviderID?.trim() || defaultProviderID)
+  const rawModelID = parts.length >= 2 ? parts.slice(1).join("/").trim() : trimmed
+  if (!providerID || !rawModelID) return undefined
+
+  const parsed = parseVariantFromModel(rawModelID)
+  if (!parsed.modelID) return undefined
+
+  return {
+    providers: [providerID],
+    model: parsed.modelID,
+    variant: parsed.variant,
+  }
+}
+
+export function parseFallbackModelObjectEntry(
+  obj: FallbackModelObject,
+  contextProviderID: string | undefined,
+  defaultProviderID = "opencode",
+): FallbackEntry | undefined {
+  const base = parseFallbackModelEntry(obj.model, contextProviderID, defaultProviderID)
+  if (!base) return undefined
+
+  return {
+    ...base,
+    variant: obj.variant ?? base.variant,
+    reasoningEffort: obj.reasoningEffort,
+    temperature: obj.temperature,
+    top_p: obj.top_p,
+    maxTokens: obj.maxTokens,
+    thinking: obj.thinking,
+  }
+}
+
+/**
+ * Find the most specific FallbackEntry whose `provider/model` is a prefix of
+ * the resolved `provider/modelID`.  Longest match wins so that e.g.
+ * `openai/gpt-5.4-preview` picks the entry for `openai/gpt-5.4-preview` over
+ * the shorter `openai/gpt-5.4`.
+ */
+export function findMostSpecificFallbackEntry(
+  providerID: string,
+  modelID: string,
+  chain: FallbackEntry[],
+): FallbackEntry | undefined {
+  const resolved = `${providerID}/${modelID}`.toLowerCase()
+
+  // Collect entries whose provider/model is a prefix of the resolved model,
+  // together with the length of the matching prefix (longest match wins).
+  const matches: { entry: FallbackEntry; matchLen: number }[] = []
+  for (const entry of chain) {
+    for (const p of entry.providers) {
+      const candidate = `${p}/${entry.model}`.toLowerCase()
+      if (resolved.startsWith(candidate)) {
+        matches.push({ entry, matchLen: candidate.length })
+        break // one match per entry is enough
+      }
+    }
+  }
+
+  if (matches.length === 0) return undefined
+  matches.sort((a, b) => b.matchLen - a.matchLen)
+  return matches[0]!.entry
+}
+
+export function buildFallbackChainFromModels(
+  fallbackModels: string | (string | FallbackModelObject)[] | undefined,
+  contextProviderID: string | undefined,
+  defaultProviderID = "opencode",
+): FallbackEntry[] | undefined {
+  const normalized = normalizeFallbackModels(fallbackModels)
+  if (!normalized || normalized.length === 0) return undefined
+
+  const parsed = normalized
+    .map((entry) => {
+      if (typeof entry === "string") {
+        return parseFallbackModelEntry(entry, contextProviderID, defaultProviderID)
+      }
+      return parseFallbackModelObjectEntry(entry, contextProviderID, defaultProviderID)
+    })
+    .filter((entry): entry is FallbackEntry => entry !== undefined)
+
+  if (parsed.length === 0) return undefined
+  return parsed
+}

apps/coder/src/services/model-resolution/fallback-model-object.ts

@@ -0,0 +1,9 @@
+export type FallbackModelObject = {
+  readonly model: string
+  readonly variant?: string
+  readonly reasoningEffort?: "none" | "minimal" | "low" | "medium" | "high" | "xhigh" | "max"
+  readonly temperature?: number
+  readonly top_p?: number
+  readonly maxTokens?: number
+  readonly thinking?: { readonly type: "enabled" | "disabled"; readonly budgetTokens?: number }
+}

apps/coder/src/services/model-resolution/index.ts

@@ -0,0 +1,80 @@
+export type {
+  FallbackEntry,
+  ModelRequirement,
+} from "./model-requirement-types.js"
+export type {
+  FallbackModelObject,
+} from "./fallback-model-object.js"
+export type {
+  DelegatedModelConfig,
+  ModelResolutionRequest,
+  ModelResolutionProvenance,
+  ModelResolutionResult,
+} from "./model-resolution-types.js"
+export type {
+  ModelResolutionInput,
+  ModelSource,
+  ExtendedModelResolutionInput,
+} from "./model-resolver.js"
+export {
+  resolveModel,
+  resolveModelWithFallback,
+  normalizeFallbackModels,
+  flattenToFallbackModelStrings,
+} from "./model-resolver.js"
+export {
+  normalizeModel,
+  normalizeModelID,
+} from "./model-normalization.js"
+export {
+  fuzzyMatchModel,
+  isModelAvailable,
+} from "./model-availability.js"
+export {
+  transformModelForProvider,
+  transformModelForProviderDisplay,
+} from "./provider-model-id-transform.js"
+export {
+  buildFallbackChainFromModels,
+  parseFallbackModelEntry,
+  parseFallbackModelObjectEntry,
+  findMostSpecificFallbackEntry,
+} from "./fallback-chain-from-models.js"
+export {
+  KNOWN_VARIANTS,
+} from "./known-variants.js"
+export {
+  _setModelResolutionLogImplementationForTesting,
+  resolveModelPipeline,
+} from "./model-resolution-pipeline.js"
+export type {
+  ModelResolutionRequest as PipelineModelResolutionRequest,
+  ModelResolutionProvenance as PipelineModelResolutionProvenance,
+  ModelResolutionResult as PipelineModelResolutionResult,
+  ModelResolutionDeps,
+} from "./model-resolution-pipeline.js"
+export {
+  isRetryableModelError,
+  shouldRetryError,
+  getNextFallback,
+  hasMoreFallbacks,
+  selectFallbackProvider,
+  selectFallbackProviderWithCache,
+} from "./model-error-classifier.js"
+export type {
+  ErrorInfo,
+} from "./model-error-classifier.js"
+export type {
+  ProviderCache,
+  ModelMetadata,
+} from "./provider-cache.js"
+export type {
+  ProviderModelsCache,
+  ConnectedProvidersAdapter,
+} from "./connected-providers-cache.js"
+export {
+  readConnectedProvidersCache,
+  findProviderModelMetadata,
+  readProviderModelsCache,
+  connectedProvidersAdapter,
+} from "./connected-providers-cache.js"

apps/coder/src/services/model-resolution/known-variants.ts

@@ -0,0 +1,16 @@
+/**
+ * Canonical set of recognised variant / effort tokens.
+ * Used by parseFallbackModelEntry (space-suffix detection) and
+ * flattenToFallbackModelStrings (inline-variant stripping).
+ */
+export const KNOWN_VARIANTS = new Set([
+  "low",
+  "medium",
+  "high",
+  "xhigh",
+  "max",
+  "minimal",
+  "none",
+  "auto",
+  "thinking",
+])

apps/coder/src/services/model-resolution/model-availability.ts

@@ -0,0 +1,64 @@
+function normalizeModelName(name: string): string {
+  return name
+    .toLowerCase()
+    .replace(/claude-(opus|sonnet|haiku)-(\d+)[.-](\d+)/g, "claude-$1-$2.$3")
+}
+
+export function fuzzyMatchModel(
+  target: string,
+  available: Set<string>,
+  providers?: string[],
+): string | null {
+  if (available.size === 0) {
+    return null
+  }
+
+  const targetNormalized = normalizeModelName(target)
+
+  let candidates = Array.from(available)
+  if (providers && providers.length > 0) {
+    const providerSet = new Set(providers)
+    candidates = candidates.filter((model) => {
+      const [provider] = model.split("/")
+      return providerSet.has(provider!)
+    })
+  }
+
+  if (candidates.length === 0) {
+    return null
+  }
+
+  const matches = candidates.filter((model) =>
+    normalizeModelName(model).includes(targetNormalized),
+  )
+
+  if (matches.length === 0) {
+    return null
+  }
+
+  const exactMatch = matches.find((model) => normalizeModelName(model) === targetNormalized)
+  if (exactMatch) {
+    return exactMatch
+  }
+
+  const exactModelIdMatches = matches.filter((model) => {
+    const modelId = model.split("/").slice(1).join("/")
+    return normalizeModelName(modelId) === targetNormalized
+  })
+  if (exactModelIdMatches.length > 0) {
+    return exactModelIdMatches.reduce((shortest, current) =>
+      current.length < shortest.length ? current : shortest,
+    )
+  }
+
+  return matches.reduce((shortest, current) =>
+    current.length < shortest.length ? current : shortest,
+  )
+}
+
+export function isModelAvailable(
+  targetModel: string,
+  availableModels: Set<string>,
+): boolean {
+  return fuzzyMatchModel(targetModel, availableModels) !== null
+}

apps/coder/src/services/model-resolution/model-error-classifier.ts

@@ -0,0 +1,261 @@
+import type { FallbackEntry } from "./model-requirement-types.js"
+import type { ProviderCache } from "./provider-cache.js"
+import * as connectedProvidersCache from "./connected-providers-cache.js"
+
+/**
+ * Error names that indicate a retryable model error.
+ * These errors halt execution and should trigger fallback retry.
+ */
+const RETRYABLE_ERROR_NAMES = new Set([
+  "providermodelnotfounderror",
+  "ratelimiterror",
+  "modelunavailableerror",
+  "providerconnectionerror",
+  "authenticationerror",
+])
+
+const STOP_ERROR_NAMES = new Set([
+  "quotaexceedederror",
+  "insufficientcreditserror",
+  "freeusagelimiterror",
+])
+
+/**
+ * Error names that should NOT trigger retry.
+ * These errors are typically user-induced or fixable without switching models.
+ */
+const NON_RETRYABLE_ERROR_NAMES = new Set([
+  "messageabortederror",
+  "permissiondeniederror",
+  "contextlengtherror",
+  "timeouterror",
+  "validationerror",
+  "syntaxerror",
+  "usererror",
+])
+
+/**
+ * Message patterns that indicate a retryable error even without a known error name.
+ */
+const RETRYABLE_MESSAGE_PATTERNS = [
+  "rate_limit",
+  "rate limit",
+  "usage_limit_reached",
+  "usage limit has been reached",
+  "quota",
+  "all credentials for model",
+  "cooling down",
+  "exhausted your capacity",
+  "not found",
+  "unavailable",
+  "insufficient",
+  "too many requests",
+  "over limit",
+  "overloaded",
+  "bad gateway",
+  "bad request",
+  "unknown provider",
+  "provider not found",
+  "model_not_supported",
+  "model not supported",
+  "model is not supported",
+  "connection error",
+  "network error",
+  "timeout",
+  "service unavailable",
+  "internal_server_error",
+  "free usage",
+  "usage exceeded",
+  "credit",
+  "balance",
+  "temporarily unavailable",
+  "try again",
+  "请稍后重试",
+  "503",
+  "502",
+  "504",
+  "429",
+  "529",
+  "selected provider is forbidden",
+  "provider is forbidden",
+  // Chinese retryable patterns (Zhipu, etc.)
+  "频率限制",           // "rate limit"
+  "请求过于频繁",       // "too many requests"
+  "暂时不可用",         // "temporarily unavailable"
+  "服务不可用",         // "service unavailable"
+  "server_error",
+  "an error occurred while processing",
+]
+
+/**
+ * Message patterns that indicate a non-retryable STOP error (quota/billing exhaustion).
+ * These take precedence over RETRYABLE_MESSAGE_PATTERNS.
+ */
+const STOP_MESSAGE_PATTERNS = [
+  "quota will reset after",
+  "quota exceeded",
+  "free usage limit",
+  "billing limit",
+  "billing hard limit",
+  "monthly limit",
+  "plan limit",
+  "subscription quota",
+  "subscription limit",
+  "payment required",
+  "out of credits",
+  "credits exhausted",
+  "insufficient credits",
+  "insufficient balance",
+  "credit balance",
+  "usage limit for this month",
+  "exhausted your capacity",
+  // GLM/Z.ai business error codes that indicate permanent quota/billing exhaustion
+  "daily call limit",
+  "daily limit",
+  "usage limit reached for",
+  "in arrears",
+  "fair use policy",
+  "recharge and try",
+  "使用上限",
+  "额度不足",
+  "余额不足",
+  "已耗尽",
+]
+
+const AUTO_RETRY_GATE_PATTERNS = [
+  "rate limit",
+  "cooling down",
+  "credentials for model",
+]
+
+function hasProviderAutoRetrySignal(message: string): boolean {
+  if (!message.includes("retrying in")) {
+    return false
+  }
+  return AUTO_RETRY_GATE_PATTERNS.some((pattern) => message.includes(pattern))
+}
+
+export interface ErrorInfo {
+  name?: string
+  message?: string
+  /** HTTP status code from the provider response (e.g., 429 for rate limit) */
+  statusCode?: number
+}
+
+/**
+ * Determines if an error is a retryable model error.
+ * Returns true if it's a known retryable type OR matches retryable message patterns.
+ */
+export function isRetryableModelError(error: ErrorInfo): boolean {
+  // If we have an error name, check against known lists
+  if (error.name) {
+    const errorNameLower = error.name.toLowerCase()
+    // Explicit non-retryable takes precedence
+    if (NON_RETRYABLE_ERROR_NAMES.has(errorNameLower)) {
+      return false
+    }
+    if (STOP_ERROR_NAMES.has(errorNameLower)) {
+      return false
+    }
+    // Check if it's a known retryable error
+    if (RETRYABLE_ERROR_NAMES.has(errorNameLower)) {
+      return true
+    }
+  }
+
+  // Check message patterns for unknown errors
+  const msg = error.message?.toLowerCase() ?? ""
+
+  // STOP patterns take precedence over retryable patterns
+  if (STOP_MESSAGE_PATTERNS.some((pattern) => msg.includes(pattern))) {
+    return false
+  }
+
+  if (hasProviderAutoRetrySignal(msg)) {
+    return true
+  }
+
+  // HTTP status code check: catches rate-limit errors regardless of message format/language.
+  // Uses the same codes as runtime-fallback config (400 excluded as it is a permanent client error).
+  if (
+    error.statusCode != null &&
+    (error.statusCode === 429 || error.statusCode === 503 || error.statusCode === 529)
+  ) {
+    return true
+  }
+
+  return RETRYABLE_MESSAGE_PATTERNS.some((pattern) => msg.includes(pattern))
+}
+
+/**
+ * Determines if an error should trigger a fallback retry.
+ * Returns true for errors that halt execution.
+ */
+export function shouldRetryError(error: ErrorInfo): boolean {
+  return isRetryableModelError(error)
+}
+
+/**
+ * Gets the next fallback model from the chain based on attempt count.
+ * Returns undefined if all fallbacks have been exhausted.
+ */
+export function getNextFallback(
+  fallbackChain: FallbackEntry[],
+  attemptCount: number,
+): FallbackEntry | undefined {
+  return fallbackChain[attemptCount]
+}
+
+/**
+ * Checks if there are more fallbacks available after the current attempt.
+ */
+export function hasMoreFallbacks(
+  fallbackChain: FallbackEntry[],
+  attemptCount: number,
+): boolean {
+  return attemptCount < fallbackChain.length
+}
+
+/**
+ * Selects the best provider for a fallback entry.
+ * Priority:
+ * 1) First connected provider in the entry's provider preference order
+ * 2) Preferred provider when connected (and entry providers are unavailable)
+ * 3) First provider listed in the fallback entry
+ */
+export function selectFallbackProvider(
+  providers: string[],
+  preferredProviderID?: string,
+): string {
+  return selectFallbackProviderWithCache(
+    providers,
+    connectedProvidersCache,
+    preferredProviderID,
+  )
+}
+
+export function selectFallbackProviderWithCache(
+  providers: string[],
+  providerCache: ProviderCache,
+  preferredProviderID?: string,
+): string {
+  const connectedProviders = providerCache.readConnectedProvidersCache()
+  if (connectedProviders) {
+    const connectedSet = new Set(connectedProviders.map(p => p.toLowerCase()))
+
+    for (const provider of providers) {
+      if (connectedSet.has(provider.toLowerCase())) {
+        return provider
+      }
+    }
+
+    if (
+      preferredProviderID &&
+      connectedSet.has(preferredProviderID.toLowerCase())
+    ) {
+      return preferredProviderID
+    }
+  }
+
+  return providers[0] ?? preferredProviderID ?? "opencode"
+}

apps/coder/src/services/model-resolution/model-normalization.ts

@@ -0,0 +1,8 @@
+export function normalizeModel(model?: string): string | undefined {
+  const trimmed = model?.trim()
+  return trimmed || undefined
+}
+
+export function normalizeModelID(modelID: string): string {
+  return modelID.replace(/\.(\d+)/g, "-$1")
+}

apps/coder/src/services/model-resolution/model-requirement-types.ts

@@ -0,0 +1,18 @@
+export type FallbackEntry = {
+  providers: string[];
+  model: string;
+  variant?: string; // Entry-specific variant (e.g., GPT->high, Opus->max)
+  reasoningEffort?: string;
+  temperature?: number;
+  top_p?: number;
+  maxTokens?: number;
+  thinking?: { type: "enabled" | "disabled"; budgetTokens?: number };
+};
+
+export type ModelRequirement = {
+  fallbackChain: FallbackEntry[];
+  variant?: string; // Default variant (used when entry doesn't specify one)
+  requiresModel?: string; // If set, only activates when this model is available (fuzzy match)
+  requiresAnyModel?: boolean; // If true, requires at least ONE model in fallbackChain to be available (or empty availability treated as unavailable)
+  requiresProvider?: string[]; // If set, only activates when any of these providers is connected
+};

apps/coder/src/services/model-resolution/model-resolution-pipeline.ts

@@ -0,0 +1,256 @@
+import { fuzzyMatchModel } from "./model-availability.js"
+import type { FallbackEntry } from "./model-requirement-types.js"
+import { transformModelForProvider } from "./provider-model-id-transform.js"
+import { normalizeModel } from "./model-normalization.js"
+import type { ProviderCache } from "./provider-cache.js"
+
+type LogImplementation = (message: string, data?: unknown) => void
+
+let logImplementationForTesting: LogImplementation | undefined
+
+function log(message: string, data?: unknown): void {
+  const logImpl = logImplementationForTesting
+  if (!logImpl) {
+    return
+  }
+  if (arguments.length === 1) {
+    logImpl(message)
+    return
+  }
+  logImpl(message, data)
+}
+
+export function _setModelResolutionLogImplementationForTesting(
+  logImplementation: LogImplementation | undefined,
+): void {
+  logImplementationForTesting = logImplementation
+}
+
+export type ModelResolutionRequest = {
+  intent?: {
+    uiSelectedModel?: string
+    userModel?: string
+    userFallbackModels?: string[]
+    categoryDefaultModel?: string
+  }
+  constraints: {
+    availableModels: Set<string>
+    connectedProviders?: string[] | null
+  }
+  policy?: {
+    fallbackChain?: FallbackEntry[]
+    systemDefaultModel?: string
+  }
+}
+
+export type ModelResolutionProvenance =
+  | "override"
+  | "category-default"
+  | "provider-fallback"
+  | "system-default"
+
+export type ModelResolutionResult = {
+  model: string
+  provenance: ModelResolutionProvenance
+  variant?: string
+  attempted?: string[]
+  reason?: string
+}
+
+export type ModelResolutionDeps = {
+  fuzzyMatchModel: (
+    target: string,
+    available: Set<string>,
+    providers?: string[],
+  ) => string | null
+  transformModelForProvider: (provider: string, model: string) => string
+}
+
+const DEFAULT_MODEL_RESOLUTION_DEPS: ModelResolutionDeps = {
+  fuzzyMatchModel,
+  transformModelForProvider,
+}
+
+
+export function resolveModelPipeline(
+  request: ModelResolutionRequest,
+  providerCache: ProviderCache = {
+    readConnectedProvidersCache: () => null,
+    findProviderModelMetadata: () => undefined,
+  },
+  deps: ModelResolutionDeps = DEFAULT_MODEL_RESOLUTION_DEPS,
+): ModelResolutionResult | undefined {
+  const attempted: string[] = []
+  const { intent, constraints, policy } = request
+  const availableModels = constraints.availableModels
+  const fallbackChain = policy?.fallbackChain
+  const systemDefaultModel = policy?.systemDefaultModel
+
+  const normalizedUiModel = normalizeModel(intent?.uiSelectedModel)
+  if (normalizedUiModel) {
+    log("Model resolved via UI selection", { model: normalizedUiModel })
+    return { model: normalizedUiModel, provenance: "override" }
+  }
+
+  const normalizedUserModel = normalizeModel(intent?.userModel)
+  if (normalizedUserModel) {
+    log("Model resolved via config override", { model: normalizedUserModel })
+    return { model: normalizedUserModel, provenance: "override" }
+  }
+
+  const normalizedCategoryDefault = normalizeModel(intent?.categoryDefaultModel)
+  if (normalizedCategoryDefault) {
+    attempted.push(normalizedCategoryDefault)
+    if (availableModels.size > 0) {
+      const parts = normalizedCategoryDefault.split("/")
+      const providerHint = parts.length >= 2 ? [parts[0]!] : undefined
+      const match = deps.fuzzyMatchModel(normalizedCategoryDefault, availableModels, providerHint)
+      if (match) {
+        log("Model resolved via category default (fuzzy matched)", {
+          original: normalizedCategoryDefault,
+          matched: match,
+        })
+        return { model: match, provenance: "category-default", attempted }
+      }
+    } else {
+      const connectedProviders = constraints.connectedProviders ?? providerCache.readConnectedProvidersCache()
+      if (connectedProviders === null) {
+        log("Model resolved via category default (no cache, first run)", {
+          model: normalizedCategoryDefault,
+        })
+        return { model: normalizedCategoryDefault, provenance: "category-default", attempted }
+      }
+      const parts = normalizedCategoryDefault.split("/")
+      if (parts.length >= 2) {
+        const provider = parts[0]!
+        if (connectedProviders.includes(provider)) {
+          const modelName = parts.slice(1).join("/")
+          const transformedModel = `${provider}/${deps.transformModelForProvider(provider, modelName)}`
+          log("Model resolved via category default (connected provider)", {
+            model: transformedModel,
+            original: normalizedCategoryDefault,
+          })
+          return { model: transformedModel, provenance: "category-default", attempted }
+        }
+      }
+    }
+    log("Category default model not available, falling through to fallback chain", {
+      model: normalizedCategoryDefault,
+    })
+  }
+
+  //#when - user configured fallback_models, try them before hardcoded fallback chain
+  const userFallbackModels = intent?.userFallbackModels
+  if (userFallbackModels && userFallbackModels.length > 0) {
+    if (availableModels.size === 0) {
+      const connectedProviders = constraints.connectedProviders ?? providerCache.readConnectedProvidersCache()
+      const connectedSet = connectedProviders ? new Set(connectedProviders) : null
+
+      if (connectedSet !== null) {
+        for (const model of userFallbackModels) {
+          attempted.push(model)
+          const parts = model.split("/")
+          if (parts.length >= 2) {
+            const provider = parts[0]!
+            if (connectedSet.has(provider)) {
+              const modelName = parts.slice(1).join("/")
+              const transformedModel = `${provider}/${deps.transformModelForProvider(provider, modelName)}`
+              log("Model resolved via user fallback_models (connected provider)", { model: transformedModel, original: model })
+              return { model: transformedModel, provenance: "provider-fallback", attempted }
+            }
+          }
+        }
+        log("No connected provider found in user fallback_models, falling through to hardcoded chain")
+      }
+    } else {
+      for (const model of userFallbackModels) {
+        attempted.push(model)
+        const parts = model.split("/")
+        const providerHint = parts.length >= 2 ? [parts[0]!] : undefined
+        const match = deps.fuzzyMatchModel(model, availableModels, providerHint)
+        if (match) {
+          log("Model resolved via user fallback_models (availability confirmed)", { model, match })
+          return { model: match, provenance: "provider-fallback", attempted }
+        }
+      }
+      log("No available model found in user fallback_models, falling through to hardcoded chain")
+    }
+  }
+
+  if (fallbackChain && fallbackChain.length > 0) {
+    if (availableModels.size === 0) {
+      const connectedProviders = constraints.connectedProviders ?? providerCache.readConnectedProvidersCache()
+      const connectedSet = connectedProviders ? new Set(connectedProviders) : null
+
+      if (connectedSet === null) {
+        log("Model fallback chain skipped (no connected providers cache) - falling through to system default")
+      } else {
+        for (const entry of fallbackChain) {
+          for (const provider of entry.providers) {
+            if (connectedSet.has(provider)) {
+              const transformedModelId = deps.transformModelForProvider(provider, entry.model)
+              const model = `${provider}/${transformedModelId}`
+              log("Model resolved via fallback chain (connected provider)", {
+                provider,
+                model: transformedModelId,
+                variant: entry.variant,
+              })
+              return {
+                model,
+                provenance: "provider-fallback",
+                variant: entry.variant,
+                attempted,
+              }
+            }
+          }
+        }
+        log("No connected provider found in fallback chain, falling through to system default")
+      }
+    } else {
+      for (const entry of fallbackChain) {
+        for (const provider of entry.providers) {
+          const fullModel = `${provider}/${entry.model}`
+          const match = deps.fuzzyMatchModel(fullModel, availableModels, [provider])
+          if (match) {
+            log("Model resolved via fallback chain (availability confirmed)", {
+              provider,
+              model: entry.model,
+              match,
+              variant: entry.variant,
+            })
+            return {
+              model: match,
+              provenance: "provider-fallback",
+              variant: entry.variant,
+              attempted,
+            }
+          }
+        }
+
+        const crossProviderMatch = deps.fuzzyMatchModel(entry.model, availableModels)
+        if (crossProviderMatch) {
+          log("Model resolved via fallback chain (cross-provider fuzzy match)", {
+            model: entry.model,
+            match: crossProviderMatch,
+            variant: entry.variant,
+          })
+          return {
+            model: crossProviderMatch,
+            provenance: "provider-fallback",
+            variant: entry.variant,
+            attempted,
+          }
+        }
+      }
+      log("No available model found in fallback chain, falling through to system default")
+    }
+  }
+
+  if (systemDefaultModel === undefined) {
+    log("No model resolved - systemDefaultModel not configured")
+    return undefined
+  }
+
+  log("Model resolved via system default", { model: systemDefaultModel })
+  return { model: systemDefaultModel, provenance: "system-default", attempted }
+}

apps/coder/src/services/model-resolution/model-resolution-types.ts

@@ -0,0 +1,41 @@
+import type { FallbackEntry } from "./model-requirement-types.js"
+
+export interface DelegatedModelConfig {
+  providerID: string
+  modelID: string
+  variant?: string
+  reasoningEffort?: string
+  temperature?: number
+  top_p?: number
+  maxTokens?: number
+  thinking?: { type: "enabled" | "disabled"; budgetTokens?: number }
+}
+
+export type ModelResolutionRequest = {
+  intent?: {
+    uiSelectedModel?: string
+    userModel?: string
+    categoryDefaultModel?: string
+  }
+  constraints: {
+    availableModels: Set<string>
+  }
+  policy?: {
+    fallbackChain?: FallbackEntry[]
+    systemDefaultModel?: string
+  }
+}
+
+export type ModelResolutionProvenance =
+  | "override"
+  | "category-default"
+  | "provider-fallback"
+  | "system-default"
+
+export type ModelResolutionResult = {
+  model: string
+  provenance: ModelResolutionProvenance
+  variant?: string
+  attempted?: string[]
+  reason?: string
+}

apps/coder/src/services/model-resolution/model-resolver.ts

@@ -0,0 +1,109 @@
+import type { FallbackEntry } from "./model-requirement-types.js"
+import type { FallbackModelObject } from "./fallback-model-object.js"
+import { normalizeModel } from "./model-normalization.js"
+import { resolveModelPipeline } from "./model-resolution-pipeline.js"
+import { KNOWN_VARIANTS } from "./known-variants.js"
+import type { ConnectedProvidersAdapter } from "./connected-providers-cache.js"
+import * as connectedProvidersCache from "./connected-providers-cache.js"
+
+export type ModelResolutionInput = {
+  userModel?: string
+  inheritedModel?: string
+  systemDefault?: string
+}
+
+export type ModelSource =
+  | "override"
+  | "category-default"
+  | "provider-fallback"
+  | "system-default"
+
+export type ModelResolutionResult = {
+  model: string
+  source: ModelSource
+  variant?: string
+}
+
+export type ExtendedModelResolutionInput = {
+  uiSelectedModel?: string
+  userModel?: string
+  userFallbackModels?: string[]
+  categoryDefaultModel?: string
+  fallbackChain?: FallbackEntry[]
+  availableModels: Set<string>
+  systemDefaultModel?: string
+}
+
+
+export function resolveModel(input: ModelResolutionInput): string | undefined {
+  return (
+    normalizeModel(input.userModel) ??
+    normalizeModel(input.inheritedModel) ??
+    input.systemDefault
+  )
+}
+
+export function resolveModelWithFallback(
+  input: ExtendedModelResolutionInput,
+  connectedProvidersAdapter: ConnectedProvidersAdapter = connectedProvidersCache,
+): ModelResolutionResult | undefined {
+  const { uiSelectedModel, userModel, userFallbackModels, categoryDefaultModel, fallbackChain, availableModels, systemDefaultModel } = input
+  const resolved = resolveModelPipeline({
+    intent: { uiSelectedModel, userModel, userFallbackModels, categoryDefaultModel },
+    constraints: { availableModels },
+    policy: { fallbackChain, systemDefaultModel },
+  }, connectedProvidersAdapter)
+
+  if (!resolved) {
+    return undefined
+  }
+
+  return {
+    model: resolved.model,
+    source: resolved.provenance,
+    variant: resolved.variant,
+  }
+}
+
+/**
+ * Normalizes fallback_models config to a mixed array.
+ * Accepts string, string[], or mixed arrays of strings and FallbackModelObject entries.
+ */
+export function normalizeFallbackModels(
+  models: string | (string | FallbackModelObject)[] | undefined,
+): (string | FallbackModelObject)[] | undefined {
+  if (!models) return undefined
+  if (typeof models === "string") return [models]
+  return models
+}
+
+/**
+ * Extracts plain model strings from a mixed fallback models array.
+ * Object entries are flattened to "model" or "model(variant)" strings.
+ * Use this when consumers need string[] (e.g., resolveModelForDelegateTask).
+ */
+export function flattenToFallbackModelStrings(
+  models: (string | FallbackModelObject)[] | undefined,
+): string[] | undefined {
+  if (!models) return undefined
+  return models.map((entry) => {
+    if (typeof entry === "string") return entry
+    const variant = entry.variant
+    if (variant) {
+      // Strip any supported inline variant syntax before appending explicit override.
+      // Supports both parenthesized and space-suffix forms so we don't emit
+      // invalid strings like "provider/model high(low)".
+      const model = entry.model
+        .replace(/\([^()]+\)\s*$/, "")
+        .replace(/\s+([a-z][a-z0-9_-]*)\s*$/i, (_match: string, suffix: string) => {
+          const normalized = String(suffix).toLowerCase()
+          return KNOWN_VARIANTS.has(normalized)
+            ? ""
+            : _match
+        })
+        .trim()
+      return `${model}(${variant})`
+    }
+    return entry.model
+  })
+}

apps/coder/src/services/model-resolution/provider-cache.ts

@@ -0,0 +1,27 @@
+export interface ModelMetadata {
+  readonly id: string
+  readonly provider?: string
+  readonly context?: number
+  readonly output?: number
+  readonly name?: string
+  readonly variants?: Record<string, unknown>
+  readonly limit?: {
+    readonly context?: number
+    readonly input?: number
+    readonly output?: number
+  }
+  readonly modalities?: {
+    readonly input?: string[]
+    readonly output?: string[]
+  }
+  readonly capabilities?: Record<string, unknown>
+  readonly reasoning?: boolean
+  readonly temperature?: boolean
+  readonly tool_call?: boolean
+  readonly [key: string]: unknown
+}
+
+export interface ProviderCache {
+  readConnectedProvidersCache(): string[] | null
+  findProviderModelMetadata(providerID: string, modelID: string): ModelMetadata | undefined
+}

apps/coder/src/services/model-resolution/provider-model-id-transform.ts

@@ -0,0 +1,69 @@
+function inferSubProvider(model: string): string | undefined {
+  if (model.startsWith("claude-")) return "anthropic"
+  if (model.startsWith("gpt-")) return "openai"
+  if (model.startsWith("gemini-")) return "google"
+  if (model.startsWith("grok-")) return "xai"
+  if (model.startsWith("minimax-")) return "minimax"
+  if (model.startsWith("kimi-")) return "moonshotai"
+  if (model.startsWith("glm-")) return "zai"
+  return undefined
+}
+
+const CLAUDE_VERSION_DOT = /claude-(\w+)-(\d+)-(\d+)/g
+const GEMINI_31_PRO_PREVIEW = /gemini-3\.1-pro(?!-)/g
+const GEMINI_3_FLASH_PREVIEW = /gemini-3-flash(?!-)/g
+
+function claudeVersionDot(model: string): string {
+  return model.replace(CLAUDE_VERSION_DOT, "claude-$1-$2.$3")
+}
+
+function applyGatewayTransforms(model: string): string {
+  return claudeVersionDot(model).replace(
+    GEMINI_31_PRO_PREVIEW,
+    "gemini-3.1-pro-preview",
+  )
+}
+
+function transformModelForProviderUsingAnthropicBehavior(
+  provider: string,
+  model: string,
+): string {
+  if (provider === "vercel") {
+    const slashIndex = model.indexOf("/")
+    if (slashIndex !== -1) {
+      const subProvider = model.substring(0, slashIndex)
+      const subModel = model.substring(slashIndex + 1)
+      return `${subProvider}/${applyGatewayTransforms(subModel)}`
+    }
+    const subProvider = inferSubProvider(model)
+    if (subProvider) {
+      return `${subProvider}/${applyGatewayTransforms(model)}`
+    }
+    return model
+  }
+  if (provider === "github-copilot") {
+    return claudeVersionDot(model)
+      .replace(GEMINI_31_PRO_PREVIEW, "gemini-3.1-pro-preview")
+      .replace(GEMINI_3_FLASH_PREVIEW, "gemini-3-flash-preview")
+  }
+  if (provider === "google") {
+    return model
+      .replace(GEMINI_31_PRO_PREVIEW, "gemini-3.1-pro-preview")
+      .replace(GEMINI_3_FLASH_PREVIEW, "gemini-3-flash-preview")
+  }
+  if (provider === "anthropic") {
+    return model
+  }
+  return model
+}
+
+export function transformModelForProvider(provider: string, model: string): string {
+  return transformModelForProviderUsingAnthropicBehavior(provider, model)
+}
+
+export function transformModelForProviderDisplay(
+  provider: string,
+  model: string,
+): string {
+  return transformModelForProviderUsingAnthropicBehavior(provider, model)
+}

apps/coder/src/services/plan-store.ts

@@ -0,0 +1,184 @@
+/**
+ * Boulder state — cross-session plan persistence for BooCode.
+ *
+ * Plans live above flow_runs: a plan tracks a user's work goal and can link to
+ * a flow run for automatic progress tracking. When the linked flow run reaches
+ * a terminal state (completed/failed/cancelled), the plan is auto-updated.
+ *
+ * Auto-resumption: on startup, plans with a linked in-flight flow_run are
+ * surfaced via the GET endpoint so the UI can show a resume prompt. The
+ * flow-runner's initResume() re-advances the actual run; this store surfaces
+ * the plan-level view.
+ */
+import type { Sql } from '../db.js';
+
+export interface Plan {
+  id: string;
+  project_id: string;
+  title: string;
+  description: string | null;
+  status: string;
+  flow_run_id: string | null;
+  progress_pct: number;
+  items_total: number;
+  items_completed: number;
+  metadata: Record<string, unknown> | null;
+  created_at: Date;
+  updated_at: Date;
+}
+
+export interface CreatePlanOpts {
+  projectId: string;
+  title: string;
+  description?: string;
+  flowRunId?: string;
+  metadata?: Record<string, unknown>;
+}
+
+export interface UpdatePlanOpts {
+  title?: string;
+  description?: string | null;
+  status?: 'active' | 'completed' | 'cancelled' | 'failed';
+  progressPct?: number;
+  itemsTotal?: number;
+  itemsCompleted?: number;
+  metadata?: Record<string, unknown> | null;
+}
+
+export function createPlan(sql: Sql, opts: CreatePlanOpts): Promise<Plan> {
+  return sql`
+    INSERT INTO plans (project_id, title, description, flow_run_id, metadata)
+    VALUES (
+      ${opts.projectId},
+      ${opts.title},
+      ${opts.description ?? null},
+      ${opts.flowRunId ?? null},
+      ${opts.metadata ? sql.json(opts.metadata as never) : null}
+    )
+    RETURNING *
+  `.then((rows) => rows[0] as unknown as Plan);
+}
+
+export function getPlan(sql: Sql, planId: string): Promise<Plan | null> {
+  return sql`
+    SELECT * FROM plans WHERE id = ${planId}
+  `.then((rows) => (rows[0] as unknown as Plan) ?? null);
+}
+
+export function listPlans(sql: Sql, projectId: string): Promise<Plan[]> {
+  return sql`
+    SELECT * FROM plans
+    WHERE project_id = ${projectId}
+    ORDER BY created_at DESC
+    LIMIT 100
+  ` as Promise<Plan[]>;
+}
+
+export function listActivePlans(sql: Sql, projectId: string): Promise<Plan[]> {
+  return sql`
+    SELECT * FROM plans
+    WHERE project_id = ${projectId} AND status = 'active'
+    ORDER BY created_at DESC
+  ` as Promise<Plan[]>;
+}
+
+export async function updatePlan(
+  sql: Sql,
+  planId: string,
+  opts: UpdatePlanOpts,
+): Promise<Plan | null> {
+  const sets: string[] = [];
+  const values: unknown[] = [];
+
+  if (opts.title !== undefined) {
+    sets.push(`title = $${values.length + 1}`);
+    values.push(opts.title);
+  }
+  if (opts.description !== undefined) {
+    sets.push(`description = $${values.length + 1}`);
+    values.push(opts.description);
+  }
+  if (opts.status !== undefined) {
+    sets.push(`status = $${values.length + 1}`);
+    values.push(opts.status);
+  }
+  if (opts.progressPct !== undefined) {
+    sets.push(`progress_pct = $${values.length + 1}`);
+    values.push(opts.progressPct);
+  }
+  if (opts.itemsTotal !== undefined) {
+    sets.push(`items_total = $${values.length + 1}`);
+    values.push(opts.itemsTotal);
+  }
+  if (opts.itemsCompleted !== undefined) {
+    sets.push(`items_completed = $${values.length + 1}`);
+    values.push(opts.itemsCompleted);
+  }
+  if (opts.metadata !== undefined) {
+    sets.push(`metadata = $${values.length + 1}::jsonb`);
+    values.push(opts.metadata !== null ? JSON.stringify(opts.metadata) : null);
+  }
+
+  if (sets.length === 0) return getPlan(sql, planId);
+
+  sets.push(`updated_at = clock_timestamp()`);
+
+  const query = `
+    UPDATE plans SET ${sets.join(', ')}
+    WHERE id = $${values.length + 1}
+    RETURNING *
+  `;
+  values.push(planId);
+
+  const result = await sql.unsafe(query, values as never[]);
+  return (result[0] as unknown as Plan) ?? null;
+}
+
+/**
+ * Called when a flow run reaches a terminal state. Updates the linked plan's
+ * status based on the run outcome:
+ *  - completed → plan completed
+ *  - failed    → plan failed
+ *  - cancelled → plan cancelled
+ * Returns true when a plan was updated, false when no plan is linked to the run.
+ */
+export async function updatePlanFromRun(
+  sql: Sql,
+  runId: string,
+  runStatus: 'completed' | 'failed' | 'cancelled',
+): Promise<boolean> {
+  const planStatus = planStatusFromRun(runStatus);
+  const updated = await sql`
+    UPDATE plans
+    SET status = ${planStatus}, progress_pct = 100,
+        items_completed = items_total, updated_at = clock_timestamp()
+    WHERE flow_run_id = ${runId} AND status = 'active'
+  `;
+  return updated.count > 0;
+}
+
+/** Map a flow-run terminal status to its corresponding plan status. Pure. */
+export function planStatusFromRun(runStatus: 'completed' | 'failed' | 'cancelled'): string {
+  return runStatus === 'completed' ? 'completed' : runStatus;
+}
+
+/**
+ * Find any active plan linked to a running flow run — used by the startup
+ * resume path to surface plans that have in-flight orchestrator runs.
+ */
+export async function findPlanWithRunningRun(
+  sql: Sql,
+  projectId: string,
+): Promise<(Plan & { run_status: string }) | null> {
+  const [row] = await sql`
+    SELECT p.*, fr.status AS run_status
+    FROM plans p
+    JOIN flow_runs fr ON fr.id = p.flow_run_id
+    WHERE p.project_id = ${projectId}
+      AND p.status = 'active'
+      AND fr.status = 'running'
+    ORDER BY p.created_at DESC
+    LIMIT 1
+  `;
+  return (row as unknown as Plan & { run_status: string }) ?? null;
+}

apps/coder/src/services/provider-snapshot.ts

@@ -29,6 +29,22 @@ interface AgentRow {
   last_probed_at: string | Date | null;
 }
 
+export async function fetchDeepSeekModels(config: Config): Promise<ProviderModel[]> {
+  if (!config.DEEPSEEK_API_KEY) return [];
+  try {
+    const baseURL = (config.DEEPSEEK_BASE_URL ?? 'https://api.deepseek.com').replace(/\/+$/, '');
+    const res = await fetch(`${baseURL}/v1/models`, {
+      headers: { Authorization: `Bearer ${config.DEEPSEEK_API_KEY}` },
+      signal: AbortSignal.timeout(5_000),
+    });
+    if (!res.ok) return [];
+    const parsed = (await res.json()) as { data?: Array<{ id: string }> };
+    return (parsed.data ?? []).map((m) => ({ id: m.id, label: m.id }));
+  } catch {
+    return [];
+  }
+}
+
 export async function fetchLlamaSwapModels(config: Config): Promise<ProviderModel[]> {
   try {
     const res = await fetch(`${config.LLAMA_SWAP_URL}/v1/models`);
@@ -256,7 +272,13 @@ export async function getProviderSnapshot(
   }
 
   const build = async (): Promise<ProviderSnapshotEntry[]> => {
-    const llamaModels = await fetchLlamaSwapModels(config);
+    const [llamaModels, deepseekModels] = await Promise.all([
+      fetchLlamaSwapModels(config),
+      fetchDeepSeekModels(config),
+    ]);
+    // Merge DeepSeek models into the llama-swap model pool so the boocode
+    // provider (which sources from llama-swap) also includes DeepSeek models.
+    const mergedModels = mergeModels(llamaModels, deepseekModels);
     const agents = await sql<AgentRow[]>`
       SELECT name, install_path, supports_acp, models, commands, label, transport, last_probed_at FROM available_agents
     `;
@@ -265,7 +287,7 @@ export async function getProviderSnapshot(
 
     const entries = await Promise.all(
       [...getResolvedRegistry().values()].map((resolved) =>
-        buildProviderEntry(resolved, agentMap.get(resolved.id), llamaModels, resolvedCwd, ttlMs, force),
+        buildProviderEntry(resolved, agentMap.get(resolved.id), mergedModels, resolvedCwd, ttlMs, force),
       ),
     );
 

apps/server/package.json

@@ -77,8 +77,9 @@
     "test": "vitest run"
   },
   "dependencies": {
-    "@boocode/contracts": "workspace:*",
+    "@ai-sdk/deepseek": "^2.0.35",
     "@ai-sdk/openai-compatible": "^2.0.47",
+    "@boocode/contracts": "workspace:*",
     "@fastify/static": "^7.0.4",
     "@fastify/websocket": "^10.0.1",
     "@modelcontextprotocol/sdk": "^1.29.0",

apps/server/src/config.ts

@@ -26,6 +26,14 @@ const ConfigSchema = z.object({
   FAST_MODEL: z.string().optional(),
   TASK_MODEL_URL: z.string().url().optional(),
   LLAMA_SIDECAR_URL: z.string().url().optional(),
+  // vDeepSeek: DeepSeek API key for direct API access. When set, models
+  // with IDs starting with 'deepseek-' route through DeepSeek's API instead
+  // of llama-swap. Defaults to empty (DeepSeek routing disabled).
+  DEEPSEEK_API_KEY: z.string().optional(),
+  // Optional base URL override for DeepSeek API. Defaults to api.deepseek.com.
+  DEEPSEEK_BASE_URL: z.string().url().default('https://api.deepseek.com'),
+  // vWhale hooks: path to hooks JSON config file. Missing file = no hooks.
+  HOOKS_CONFIG_PATH: z.string().default('/data/hooks.json'),
 });
 
 export type Config = z.infer<typeof ConfigSchema>;

apps/server/src/index.ts

@@ -31,6 +31,7 @@ import { loadMcpConfig } from './services/mcp-config.js';
 import { initialize as initMcp, getTools as getMcpTools, shutdown as shutdownMcp } from './services/mcp-client.js';
 import { appendMcpTools } from './services/tools.js';
 import { refreshToolNames, getAgentsForProject } from './services/agents.js';
+import { loadHooksConfig, createHookRunner } from './services/hooks.js';
 
 async function main() {
   const config = loadConfig();
@@ -136,11 +137,17 @@ async function main() {
     app.log.warn({ err }, 'skills boot walk failed');
   }
 
+  // vWhale hooks: load hook config and create runner. Missing file = no hooks.
+  loadHooksConfig(config.HOOKS_CONFIG_PATH);
+  const hookRunner = createHookRunner();
+  const hasHooks = Object.keys(loadHooksConfig(config.HOOKS_CONFIG_PATH).hooks).length > 0;
+
   const inference = createInferenceRunner(
     {
       sql,
       config,
       log: app.log,
+      hooks: hasHooks ? hookRunner : undefined,
       publish: (sessionId, frame) => {
         // v1.13.11-b: route through the typed publishFrame so the broker's
         // Zod gate validates every inference frame before delivery.
@@ -166,7 +173,7 @@ async function main() {
     // bubble up so the route can reply 500 — manual /compact failures
     // should be loud (the user just clicked a button).
     runCompaction: (chatId) =>
-      compaction.process({ sql, config, log: app.log, broker, chatId }),
+      compaction.process({ sql, config, log: app.log, broker, chatId, hooks: hasHooks ? hookRunner : undefined }),
     cancelInference: async (sessionId, chatId) => {
       return inference.cancel(sessionId, chatId);
     },

apps/server/src/routes/models.ts

@@ -2,26 +2,55 @@ import type { FastifyInstance } from 'fastify';
 import type { Config } from '../config.js';
 import type { ModelInfo } from '../types/api.js';
 
-interface LlamaSwapModelsResponse {
+interface ApiModelsResponse {
   data?: ModelInfo[];
 }
 
+const DEEPSEEK_STATIC_MODELS: ModelInfo[] = [
+  { id: 'deepseek-v4-flash', object: 'model', created: 0, owned_by: 'deepseek' },
+  { id: 'deepseek-v4-pro', object: 'model', created: 0, owned_by: 'deepseek' },
+];
+
 export function registerModelRoutes(app: FastifyInstance, config: Config): void {
   app.get('/api/models', async (_req, reply) => {
+    const models: ModelInfo[] = [];
+
+    // 1. Fetch llama-swap models
     try {
       const res = await fetch(`${config.LLAMA_SWAP_URL}/v1/models`);
-      if (!res.ok) {
-        reply.code(502);
-        return { error: `llama-swap returned ${res.status}` };
+      if (res.ok) {
+        const parsed = (await res.json()) as ApiModelsResponse;
+        if (parsed.data) models.push(...parsed.data);
       }
-      const parsed = (await res.json()) as LlamaSwapModelsResponse;
-      return parsed.data ?? [];
-    } catch (err) {
-      reply.code(502);
-      return {
-        error: 'failed to reach llama-swap',
-        details: err instanceof Error ? err.message : String(err),
-      };
+    } catch {
+      // llama-swap unreachable — proceed with whatever we have
     }
+
+    // 2. If DeepSeek is configured, fetch live models from their API
+    if (config.DEEPSEEK_API_KEY) {
+      try {
+        const baseURL = (config.DEEPSEEK_BASE_URL ?? 'https://api.deepseek.com').replace(/\/+$/, '');
+        const res = await fetch(`${baseURL}/v1/models`, {
+          headers: { Authorization: `Bearer ${config.DEEPSEEK_API_KEY}` },
+          signal: AbortSignal.timeout(5_000),
+        });
+        if (res.ok) {
+          const parsed = (await res.json()) as ApiModelsResponse;
+          if (parsed.data) models.push(...parsed.data);
+        } else {
+          // API call failed — fall back to static model list
+          models.push(...DEEPSEEK_STATIC_MODELS);
+        }
+      } catch {
+        // Network error — fall back to static model list
+        models.push(...DEEPSEEK_STATIC_MODELS);
+      }
+    }
+
+    if (models.length === 0) {
+      reply.code(502);
+      return { error: 'no models available from any provider' };
+    }
+    return models;
   });
 }

apps/server/src/schema.sql

@@ -126,8 +126,8 @@ SELECT
      FROM message_parts p
     WHERE p.message_id = m.id AND p.kind = 'reasoning' AND p.hidden_at IS NULL) AS reasoning_parts,
   -- NEW columns MUST be appended at the end: CREATE OR REPLACE VIEW can't
-  -- reorder/rename existing columns (42P16). m.model added last.
-  m.model
+  -- reorder/rename existing columns (42P16). cache_tokens and reasoning_tokens added last.
+  m.model, m.cache_tokens, m.reasoning_tokens
 FROM messages m;
 
 -- v1.13.20: drop legacy tool_calls/tool_results columns. Reads have routed
@@ -204,6 +204,8 @@ ALTER TABLE messages ADD COLUMN IF NOT EXISTS ctx_used INTEGER;
 ALTER TABLE messages ADD COLUMN IF NOT EXISTS ctx_max INTEGER;
 ALTER TABLE messages ADD COLUMN IF NOT EXISTS started_at TIMESTAMPTZ;
 ALTER TABLE messages ADD COLUMN IF NOT EXISTS finished_at TIMESTAMPTZ;
+ALTER TABLE messages ADD COLUMN IF NOT EXISTS cache_tokens INTEGER;
+ALTER TABLE messages ADD COLUMN IF NOT EXISTS reasoning_tokens INTEGER;
 
 ALTER TABLE sessions ADD COLUMN IF NOT EXISTS updated_at TIMESTAMPTZ NOT NULL DEFAULT clock_timestamp();
 

apps/server/src/services/agents.ts

@@ -106,6 +106,8 @@ interface ParsedFrontmatter {
   // allowed" — the model responds text-only.
   steps?: number;
   llama_extra_args?: string[];
+  // vDeepSeek: thinking effort for DeepSeek V4 models.
+  reasoning_effort?: string;
 }
 
 // P5: table-driven validation for the "soft-range" numeric frontmatter fields.
@@ -386,6 +388,7 @@ function parseAgentSection(section: RawSection): Omit<Agent, 'source'> {
     max_tool_calls: typeof fm.max_tool_calls === 'number' ? fm.max_tool_calls : null,
     steps: typeof fm.steps === 'number' ? fm.steps : null,
     llama_extra_args: Array.isArray(fm.llama_extra_args) ? fm.llama_extra_args : null,
+    reasoning_effort: typeof fm.reasoning_effort === 'string' ? (fm.reasoning_effort as Agent['reasoning_effort']) : null,
   };
 }
 

apps/server/src/services/boocontext_client.ts

@@ -0,0 +1,110 @@
+/**
+ * v2.7.18: shared MCP client wrapper for the boocontext sidecar.
+ *
+ * Calls into the existing multi-server MCP client infrastructure
+ * (services/mcp-client.ts) which connects to boocontext as a stdio
+ * MCP process defined in data/mcp.json (server name "boocontext",
+ * command: `node /opt/forks/boocontext/dist/standalone.js`).
+ *
+ * The boocontext MCP server is initialized once at app boot in
+ * index.ts via initMcp() and the actual MCP tool call routing is
+ * handled by mcp-client.ts:callTool() — this module is a thin
+ * convenience wrapper that prepends the "boocontext_" server prefix,
+ * normalises the response, and applies inline truncation matching
+ * the same pattern as codecontext_client.ts.
+ *
+ * Usage:
+ *   import { callBoocontext } from './services/boocontext_client.js';
+ *   const resp = await callBoocontext({
+ *     toolName: 'codesight_get_summary',
+ *     args: { directory: '/opt/boocode' },
+ *   });
+ */
+
+import { callTool } from './mcp-client.js';
+import { truncateIfNeeded } from './truncate.js';
+
+// ---- Exported types ----
+
+export interface BoocontextRequest {
+  /** Unprefixed tool name as defined on the boocontext MCP server
+   * (e.g. "codesight_scan", "boocontext_overview", "codesight_get_summary"). */
+  toolName: string;
+  /** Arguments to pass to the tool. */
+  args: Record<string, unknown>;
+}
+
+export interface BoocontextResponse {
+  /** The tool output text. */
+  result: string;
+  /** Whether the result was truncated to fit the inline limit. */
+  truncated: boolean;
+  /** Opaque id pointing at the full pre-slice content on tmpfs, set when
+   * truncated=true and storage succeeded. */
+  outputPath?: string;
+}
+
+// ---- Constants ----
+
+/** Must match the server name in data/mcp.json. */
+const BOOCONTEXT_SERVER_NAME = 'boocontext';
+
+/** Inline truncation limit, matching codecontext_client.ts. */
+const TRUNCATION_LIMIT = 32_000;
+
+// ---- Public API ----
+
+/**
+ * Call a boocontext MCP tool by its unprefixed name.
+ *
+ * Prepends the "boocontext_" server prefix, delegates to the
+ * multi-server MCP client's callTool(), and normalises the response
+ * into a BoocontextResponse with inline truncation.
+ *
+ * @param req   The tool name and arguments.
+ * @param log   Optional Fastify-compatible logger (for debug traces).
+ * @returns     The tool result, possibly truncated.
+ * @throws      If the boocontext server is not connected or the tool
+ *              returns an MCP-level error.
+ */
+export async function callBoocontext(
+  req: BoocontextRequest,
+  log?: { debug?: (obj: object, msg: string) => void; warn?: (obj: object, msg: string) => void },
+): Promise<BoocontextResponse> {
+  const prefixedName = `${BOOCONTEXT_SERVER_NAME}_${req.toolName}`;
+
+  log?.debug?.({ tool: prefixedName }, 'boocontext: calling tool');
+
+  const raw = await callTool(prefixedName, req.args);
+
+  // callTool returns { error: true, output: string } on failure (both
+  // for MCP-level isError and for network/protocol exceptions).
+  if (typeof raw === 'object' && raw !== null && (raw as Record<string, unknown>).error === true) {
+    const errOutput = (raw as Record<string, unknown>).output ?? 'Unknown MCP error';
+    throw new Error(`boocontext error: ${String(errOutput)}`);
+  }
+
+  const result = typeof raw === 'string' ? raw : JSON.stringify(raw);
+
+  // Inline truncation at 32 kB, matching codecontext_client.ts.
+  // The model gets a clear hint about how to narrow the next call
+  // rather than a silent cut.
+  if (result.length > TRUNCATION_LIMIT) {
+    const truncated = result.slice(0, TRUNCATION_LIMIT);
+    const omitted = result.length - TRUNCATION_LIMIT;
+    const slicedWithMarker =
+      `${truncated}\n\n[truncated, ${omitted} chars omitted; narrow with additional filters]`;
+    const wrapped = await truncateIfNeeded({
+      fullContent: result,
+      slicedContent: slicedWithMarker,
+      wasTruncated: true,
+    });
+    return {
+      result: wrapped.content,
+      truncated: wrapped.truncated,
+      ...(wrapped.outputPath ? { outputPath: wrapped.outputPath } : {}),
+    };
+  }
+
+  return { result, truncated: false };
+}

apps/server/src/services/compaction.ts

@@ -24,6 +24,8 @@ import { SUMMARY_TEMPLATE } from './compaction-prompt.js';
 import * as modelContextLookup from './model-context.js';
 import { SENTINEL_KINDS } from './inference/sentinels.js';
 import type { OpenAiMessage } from './inference/payload.js';
+import { resolveModelEndpoint } from './inference/provider.js';
+import type { HookRunner } from './hooks.js';
 
 // v1.13.9: ratio-only overflow trigger. Fires compaction at 85% of ctx_max
 // (opencode session/overflow.ts pattern). Replaces the v1.11.0-era
@@ -346,20 +348,22 @@ interface CompletionResult {
   completionTokens: number;
 }
 
-async function callLlamaSwap(
+async function callLlm(
   config: Config,
   model: string,
   messages: OpenAiMessage[],
   log: FastifyBaseLogger,
 ): Promise<CompletionResult> {
-  const res = await fetch(`${config.LLAMA_SWAP_URL}/v1/chat/completions`, {
+  const { url, headers, model: resolvedModel } = resolveModelEndpoint(config, model);
+  const res = await fetch(`${url}/v1/chat/completions`, {
     method: 'POST',
-    headers: { 'Content-Type': 'application/json' },
-    body: JSON.stringify({ model, messages, stream: false }),
+    headers,
+    body: JSON.stringify({ model: resolvedModel, messages, stream: false }),
   });
   if (!res.ok) {
     const text = await res.text().catch(() => '');
-    throw new Error(`llama-swap returned ${res.status}: ${text.slice(0, 200)}`);
+    const prefix = model.startsWith('deepseek-') ? 'deepseek' : 'llama-swap';
+    throw new Error(`${prefix} returned ${res.status}: ${text.slice(0, 200)}`);
   }
   const json = (await res.json()) as {
     choices?: Array<{ message?: { content?: string } }>;
@@ -383,6 +387,8 @@ export interface ProcessInput {
   log: FastifyBaseLogger;
   broker: Broker;
   chatId: string;
+  /** vWhale: lifecycle hooks runner. Undefined when no hooks configured. */
+  hooks?: HookRunner;
 }
 
 // Runs one round of anchored rolling compaction on `chatId`. No-ops cleanly
@@ -497,6 +503,17 @@ export async function process(input: ProcessInput): Promise<void> {
     at: new Date().toISOString(),
   });
 
+  // vWhale: PreCompact hook (best-effort, non-blocking).
+  const msgBefore = messages.length;
+  if (input.hooks) {
+    input.hooks.run('PreCompact', {
+      event: 'PreCompact',
+      session_id: sessionId,
+      chat_id: chatId,
+      messages_before: msgBefore,
+    }).catch(() => {});
+  }
+
   // try/finally so the dot ALWAYS drops back to idle, even if the LLM call
   // throws or a downstream DB write fails. The succeeded flag gates the
   // 'compacted' frame + final log: we only signal completion to the UI when
@@ -506,7 +523,7 @@ export async function process(input: ProcessInput): Promise<void> {
   let result: CompletionResult | undefined;
   try {
     // 7. Single completion (no tools). Throws on llama-swap failure.
-    result = await callLlamaSwap(config, session.model, payload, log);
+    result = await callLlm(config, session.model, payload, log);
 
     // 7b. v1.11.3: fetch the model's true context window from llama-swap's
     // /upstream/<model>/props (the streaming completion doesn't carry it).
@@ -558,6 +575,18 @@ export async function process(input: ProcessInput): Promise<void> {
     `;
 
     succeeded = true;
+
+    // vWhale: PostCompact hook (best-effort, non-blocking).
+    if (input.hooks) {
+      input.hooks.run('PostCompact', {
+        event: 'PostCompact',
+        session_id: sessionId,
+        chat_id: chatId,
+        messages_before: msgBefore,
+        messages_after: sel.head.length,
+        summary: (result?.content ?? '').slice(0, 500),
+      }).catch(() => {});
+    }
   } finally {
     // Always restore the dot. Status='idle' (not 'error') even on failure —
     // the caller logs/re-surfaces the error separately; the dot doesn't

apps/server/src/services/hooks.ts

@@ -0,0 +1,299 @@
+/**
+ * vWhale: lifecycle hook runner. Hooks are shell commands that fire at key
+ * points in the inference pipeline. Each hook receives a JSON payload on
+ * stdin and can return JSON on stdout to influence behavior.
+ *
+ * Inspired by Whale's hook system with 11 lifecycle events. BooCode
+ * implements the most relevant subset: PreToolUse, PostToolUse,
+ * UserPromptSubmit, Stop, PreCompact, PostCompact.
+ *
+ * Config: JSON file at HOOKS_CONFIG_PATH (default /data/hooks.json).
+ * Format:
+ * ```json
+ * {
+ *   "hooks": {
+ *     "PreToolUse": [
+ *       { "match": "shell_run", "command": "python3 /data/hooks/check_shell.py", "timeout": 30 }
+ *     ],
+ *     "Stop": [
+ *       { "command": "node /data/hooks/log_turn.mjs" }
+ *     ]
+ *   }
+ * }
+ * ```
+ */
+
+import { spawn } from 'node:child_process';
+import { readFileSync, existsSync } from 'node:fs';
+import type { FastifyBaseLogger } from 'fastify';
+
+// ─── Events ───────────────────────────────────────────────────────────────
+
+export type HookEvent =
+  | 'PreToolUse'
+  | 'PostToolUse'
+  | 'UserPromptSubmit'
+  | 'Stop'
+  | 'PreCompact'
+  | 'PostCompact';
+
+const ALL_EVENTS: HookEvent[] = [
+  'PreToolUse',
+  'PostToolUse',
+  'UserPromptSubmit',
+  'Stop',
+  'PreCompact',
+  'PostCompact',
+];
+
+// ─── Config ────────────────────────────────────────────────────────────────
+
+export interface HookConfig {
+  /** Glob or exact tool name to match (PreToolUse/PostToolUse only). Omit or '*' for all. */
+  match?: string;
+  /** Shell command to run. Receives JSON payload on stdin. */
+  command: string;
+  /** Timeout in seconds (default 30). */
+  timeout?: number;
+}
+
+export interface HooksConfig {
+  hooks: Partial<Record<HookEvent, HookConfig[]>>;
+}
+
+// ─── Payloads ──────────────────────────────────────────────────────────────
+
+export interface PreToolUsePayload {
+  event: 'PreToolUse';
+  session_id: string;
+  tool_name: string;
+  tool_args: Record<string, unknown>;
+}
+
+export interface PostToolUsePayload {
+  event: 'PostToolUse';
+  session_id: string;
+  tool_name: string;
+  tool_args: Record<string, unknown>;
+  tool_result: unknown;
+  tool_error?: string;
+}
+
+export interface UserPromptSubmitPayload {
+  event: 'UserPromptSubmit';
+  session_id: string;
+  chat_id: string;
+  prompt: string;
+}
+
+export interface StopPayload {
+  event: 'Stop';
+  session_id: string;
+  chat_id: string;
+  last_assistant_text: string;
+  turn: number;
+}
+
+export interface PreCompactPayload {
+  event: 'PreCompact';
+  session_id: string;
+  chat_id: string;
+  messages_before: number;
+}
+
+export interface PostCompactPayload {
+  event: 'PostCompact';
+  session_id: string;
+  chat_id: string;
+  messages_before: number;
+  messages_after: number;
+  summary: string;
+}
+
+export type HookPayload =
+  | PreToolUsePayload
+  | PostToolUsePayload
+  | UserPromptSubmitPayload
+  | StopPayload
+  | PreCompactPayload
+  | PostCompactPayload;
+
+// ─── Response ──────────────────────────────────────────────────────────────
+
+export type HookDecision = 'pass' | 'warn' | 'block';
+
+export interface HookResponse {
+  decision?: HookDecision;
+  reason?: string;
+  /** When present, replaces the original tool args / user prompt. */
+  updated_input?: Record<string, unknown> | string;
+  /** Injected into the model's context for the next turn. */
+  additional_context?: string;
+}
+
+// ─── Runner ────────────────────────────────────────────────────────────────
+
+export interface HookRunner {
+  /** Run all hooks for the given event. Returns the effective response. */
+  run(event: HookEvent, payload: HookPayload, log?: FastifyBaseLogger): Promise<HookResponse>;
+}
+
+let hooksConfig: HooksConfig | null = null;
+let hooksPath: string | null = null;
+
+/** Load hooks config from disk. Missing file = no hooks. Never throws. */
+export function loadHooksConfig(path: string): HooksConfig {
+  hooksPath = path;
+  if (!existsSync(path)) {
+    hooksConfig = { hooks: {} };
+    return hooksConfig;
+  }
+  try {
+    const raw = readFileSync(path, 'utf8');
+    const parsed = JSON.parse(raw) as HooksConfig;
+    hooksConfig = {
+      hooks: { ...parsed.hooks },
+    };
+    // Validate event names
+    for (const event of Object.keys(hooksConfig.hooks)) {
+      if (!ALL_EVENTS.includes(event as HookEvent)) {
+        console.warn(`hooks: unknown event '${event}' in ${path} — ignoring`);
+        delete hooksConfig.hooks[event as HookEvent];
+      }
+    }
+  } catch (err) {
+    console.error(`hooks: failed to load ${path}`, err);
+    hooksConfig = { hooks: {} };
+  }
+  return hooksConfig;
+}
+
+/** Reload the config file (call after a PATCH). */
+export function reloadHooksConfig(): HooksConfig {
+  if (hooksPath) return loadHooksConfig(hooksPath);
+  hooksConfig = { hooks: {} };
+  return hooksConfig;
+}
+
+function getConfig(): HooksConfig {
+  return hooksConfig ?? { hooks: {} };
+}
+
+/** Create a HookRunner for the current config. */
+export function createHookRunner(): HookRunner {
+  return {
+    async run(event, payload, log): Promise<HookResponse> {
+      const configs = getConfig().hooks[event];
+      if (!configs || configs.length === 0) return { decision: 'pass' };
+
+      // Pre-filter by match pattern for tool events
+      const toolName = 'tool_name' in payload ? (payload as PreToolUsePayload).tool_name : undefined;
+
+      let effective: HookResponse = { decision: 'pass' };
+
+      for (const cfg of configs) {
+        // Skip if match doesn't apply
+        if (toolName && cfg.match && cfg.match !== '*' && cfg.match !== toolName) continue;
+
+        const result = await runSingleHook(cfg, payload, log);
+        // Merge decisions: block > warn > pass
+        if (result.decision === 'block') {
+          effective = { ...result, decision: 'block' };
+          break; // block is terminal
+        }
+        if (result.decision === 'warn' && effective.decision !== 'block') {
+          effective = { ...result, decision: 'warn' };
+        }
+        // Merge additional_context and updated_input
+        if (result.additional_context) {
+          effective.additional_context = effective.additional_context
+            ? effective.additional_context + '\n' + result.additional_context
+            : result.additional_context;
+        }
+        if (result.updated_input && !effective.updated_input) {
+          effective.updated_input = result.updated_input;
+        }
+      }
+
+      return effective;
+    },
+  };
+}
+
+async function runSingleHook(
+  cfg: HookConfig,
+  payload: HookPayload,
+  log?: FastifyBaseLogger,
+): Promise<HookResponse> {
+  const timeoutMs = (cfg.timeout ?? 30) * 1000;
+
+  return new Promise((resolve) => {
+    const child = spawn('sh', ['-c', cfg.command], {
+      stdio: ['pipe', 'pipe', 'pipe'],
+      timeout: timeoutMs,
+      env: { ...process.env },
+    });
+
+    const stdout: Buffer[] = [];
+    const stderr: Buffer[] = [];
+
+    child.stdout.on('data', (chunk: Buffer) => stdout.push(chunk));
+    child.stderr.on('data', (chunk: Buffer) => stderr.push(chunk));
+
+    let settled = false;
+    const timer = setTimeout(() => {
+      if (!settled) {
+        settled = true;
+        child.kill('SIGTERM');
+        log?.warn({ event: payload.event, command: cfg.command }, 'hooks: timeout');
+        resolve({ decision: 'warn', reason: 'hook timed out' });
+      }
+    }, timeoutMs);
+
+    child.on('error', (err) => {
+      if (!settled) {
+        settled = true;
+        clearTimeout(timer);
+        log?.warn({ err, event: payload.event }, 'hooks: spawn error');
+        resolve({ decision: 'warn', reason: `hook failed: ${err.message}` });
+      }
+    });
+
+    child.on('close', (code) => {
+      if (settled) return;
+      settled = true;
+      clearTimeout(timer);
+
+      const out = Buffer.concat(stdout).toString('utf8').trim();
+      const errOut = Buffer.concat(stderr).toString('utf8').trim();
+
+      if (code !== 0 && !out) {
+        log?.warn({ event: payload.event, code, stderr: errOut.slice(0, 200) }, 'hooks: non-zero exit');
+        resolve({ decision: 'warn', reason: `hook exited ${code}` });
+        return;
+      }
+
+      // Parse stdout as JSON response
+      if (out) {
+        try {
+          const parsed = JSON.parse(out) as HookResponse;
+          resolve(parsed);
+          return;
+        } catch {
+          // Not JSON — treat as pass with stdout as context
+          if (out.length > 0) {
+            resolve({ decision: 'pass', additional_context: out });
+            return;
+          }
+        }
+      }
+
+      resolve({ decision: 'pass' });
+    });
+
+    // Write payload to stdin
+    const json = JSON.stringify(payload);
+    child.stdin.write(json);
+    child.stdin.end();
+  });
+}

apps/server/src/services/inference/error-handler.ts

@@ -122,6 +122,8 @@ export async function finalizeStreamedRow(
     completionTokens: number | null;
     promptTokens: number | null;
     startedAt: string | null;
+    cacheTokens?: number | null;
+    reasoningTokens?: number | null;
     beforeComplete?: () => Promise<void>;
   },
 ): Promise<void> {
@@ -137,6 +139,8 @@ export async function finalizeStreamedRow(
         tokens_used = ${opts.completionTokens},
         ctx_used = ${opts.promptTokens},
         ctx_max = ${nCtx},
+        cache_tokens = ${opts.cacheTokens ?? null},
+        reasoning_tokens = ${opts.reasoningTokens ?? null},
         finished_at = clock_timestamp()
     WHERE id = ${opts.messageId}
     RETURNING tokens_used, ctx_used, ctx_max, finished_at
@@ -149,6 +153,8 @@ export async function finalizeStreamedRow(
     tokens_used: updated?.tokens_used ?? null,
     ctx_used: updated?.ctx_used ?? null,
     ctx_max: updated?.ctx_max ?? null,
+    cache_tokens: opts.cacheTokens ?? null,
+    reasoning_tokens: opts.reasoningTokens ?? null,
     started_at: opts.startedAt,
     finished_at: updated?.finished_at ?? null,
     model: opts.model,
@@ -188,7 +194,7 @@ export async function finalizeCompletion(
 ): Promise<void> {
   const { sessionId, chatId, assistantMessageId } = args;
   const content = stripToolMarkup(result.content, { final: true });
-  const { finishReason, promptTokens, completionTokens } = result;
+  const { finishReason, promptTokens, completionTokens, cacheReadTokens, reasoningTokens } = result;
 
   // v1.11.3: see executeToolPhase for the rationale.
   const mctx = await modelContext.getModelContext(session.model);
@@ -203,6 +209,8 @@ export async function finalizeCompletion(
         tokens_used = ${completionTokens},
         ctx_used = ${promptTokens},
         ctx_max = ${nCtx},
+        cache_tokens = ${cacheReadTokens ?? null},
+        reasoning_tokens = ${reasoningTokens ?? null},
         model = ${session.model},
         finished_at = clock_timestamp()
     WHERE id = ${assistantMessageId}
@@ -268,6 +276,8 @@ export async function finalizeCompletion(
     tokens_used: updated?.tokens_used ?? null,
     ctx_used: updated?.ctx_used ?? null,
     ctx_max: updated?.ctx_max ?? null,
+    cache_tokens: cacheReadTokens ?? null,
+    reasoning_tokens: reasoningTokens ?? null,
     started_at: startedAt,
     finished_at: updated?.finished_at ?? null,
     model: session.model,

apps/server/src/services/inference/provider.ts

@@ -1,4 +1,5 @@
 import { createOpenAICompatible } from '@ai-sdk/openai-compatible';
+import { createDeepSeek } from '@ai-sdk/deepseek';
 import type { LanguageModel } from 'ai';
 
 // v1.13.1-A: AI SDK provider against llama-swap. baseURL is threaded from
@@ -11,6 +12,12 @@ import type { LanguageModel } from 'ai';
 // llama-sidecar instead. A fresh provider is created per call (not cached)
 // because the X-Agent-Flags header varies per agent. The llama-swap path
 // stays cached since it has no per-request headers.
+//
+// vDeepSeek: when the model ID starts with 'deepseek-' and DEEPSEEK_API_KEY
+// is set, route through the official @ai-sdk/deepseek provider (not
+// openai-compatible) so DeepSeek-specific features work: providerMetadata
+// with promptCacheHitTokens/promptCacheMissTokens, reasoning via
+// LanguageModelV4Usage.outputTokens.reasoning, and thinking-mode options.
 
 const swapCache = new Map<string, ReturnType<typeof createOpenAICompatible>>();
 
@@ -41,7 +48,28 @@ function sidecarProvider(
   });
 }
 
-export type InferenceRoute = 'swap' | 'sidecar';
+const DEEPSEEK_MODEL_PREFIX = 'deepseek-';
+
+export function isDeepSeekModel(modelId: string): boolean {
+  return modelId.startsWith(DEEPSEEK_MODEL_PREFIX);
+}
+
+let deepseekProviderCache: ReturnType<typeof createDeepSeek> | null = null;
+
+function getDeepSeekProvider(
+  apiKey: string,
+  baseURL: string,
+): ReturnType<typeof createDeepSeek> {
+  if (!deepseekProviderCache) {
+    deepseekProviderCache = createDeepSeek({
+      apiKey,
+      baseURL,
+    });
+  }
+  return deepseekProviderCache;
+}
+
+export type InferenceRoute = 'swap' | 'sidecar' | 'deepseek';
 
 export interface RoutingInfo {
   route: InferenceRoute;
@@ -55,12 +83,21 @@ interface AgentLike {
 interface ConfigLike {
   LLAMA_SWAP_URL: string;
   LLAMA_SIDECAR_URL?: string;
+  DEEPSEEK_API_KEY?: string;
+  DEEPSEEK_BASE_URL?: string;
 }
 
 export function resolveRoute(
   agent: AgentLike | null,
   config?: ConfigLike,
+  modelId?: string,
 ): RoutingInfo {
+  // vDeepSeek: if the model starts with deepseek- and DEEPSEEK_API_KEY is set,
+  // route through the DeepSeek provider. Checked first so DeepSeek models
+  // always bypass llama-swap/sidecar even when those are also configured.
+  if (modelId?.startsWith(DEEPSEEK_MODEL_PREFIX) && config?.DEEPSEEK_API_KEY) {
+    return { route: 'deepseek', flags: null };
+  }
   // When llama_extra_args are explicitly set, route through sidecar with them.
   const flags = agent?.llama_extra_args;
   if (flags && flags.length > 0) {
@@ -80,7 +117,13 @@ export function upstreamModel(
   modelId: string,
   agent?: AgentLike | null,
 ): LanguageModel {
-  const { route, flags } = resolveRoute(agent ?? null, config);
+  const { route, flags } = resolveRoute(agent ?? null, config, modelId);
+  if (route === 'deepseek') {
+    return getDeepSeekProvider(
+      config.DEEPSEEK_API_KEY!,
+      config.DEEPSEEK_BASE_URL ?? 'https://api.deepseek.com',
+    ).chat(modelId);
+  }
   if (route === 'sidecar') {
     const url = config.LLAMA_SIDECAR_URL;
     if (!url) {
@@ -90,3 +133,30 @@ export function upstreamModel(
   }
   return getSwapProvider(config.LLAMA_SWAP_URL).chatModel(modelId);
 }
+
+/** Resolve the API endpoint for non-streaming calls (compaction, task-model).
+ *  Returns the URL + model + optional auth header for direct fetch() usage. */
+export function resolveModelEndpoint(
+  config: ConfigLike,
+  modelId: string,
+): { url: string; model: string; headers: Record<string, string> } {
+  const baseHeaders: Record<string, string> = { 'Content-Type': 'application/json' };
+  if (modelId.startsWith(DEEPSEEK_MODEL_PREFIX) && config.DEEPSEEK_API_KEY) {
+    const baseURL = (config.DEEPSEEK_BASE_URL ?? 'https://api.deepseek.com').replace(/\/+$/, '');
+    return {
+      url: baseURL,
+      model: modelId,
+      headers: { ...baseHeaders, Authorization: `Bearer ${config.DEEPSEEK_API_KEY}` },
+    };
+  }
+  return {
+    url: config.LLAMA_SWAP_URL.replace(/\/+$/, ''),
+    model: modelId,
+    headers: baseHeaders,
+  };
+}
+
+/** Invalidate the cached DeepSeek provider (e.g. when env vars change at runtime). */
+export function resetDeepSeekProvider(): void {
+  deepseekProviderCache = null;
+}

apps/server/src/services/inference/stream-phase-adapter.ts

@@ -13,7 +13,7 @@ import type { OpenAiMessage } from './payload.js';
 import { extractToolCallBlocks } from './tool-call-parser.js';
 import { classifyStreamError } from './stream-error-classifier.js';
 import type { StreamResult } from './types.js';
-import { upstreamModel } from './provider.js';
+import { isDeepSeekModel, upstreamModel } from './provider.js';
 import {
   jsonSchema,
   streamText,
@@ -51,6 +51,9 @@ export interface StreamOptions {
   dry_base?: number | null;
   dry_allowed_length?: number | null;
   dry_penalty_last_n?: number | null;
+  // vDeepSeek: thinking/reasoning effort. Maps to DeepSeek's reasoning_effort
+  // API param for deepseek-v4-flash / deepseek-v4-pro models.
+  reasoning_effort?: 'off' | 'low' | 'medium' | 'high' | 'xhigh' | 'max';
 }
 
 // P5: the 10-field sampler-options literal that was copy-pasted at 4 sites
@@ -74,6 +77,7 @@ export function samplerOptsFromAgent(agent: Agent | null): SamplerOpts {
     dry_base: agent?.dry_base ?? undefined,
     dry_allowed_length: agent?.dry_allowed_length ?? undefined,
     dry_penalty_last_n: agent?.dry_penalty_last_n ?? undefined,
+    reasoning_effort: agent?.reasoning_effort ?? undefined,
   };
 }
 
@@ -272,6 +276,19 @@ export async function streamCompletion(
   // before this. They now go through the same extraBody path as the new params.
   const samplerBody = buildSamplerProviderOptions(opts);
 
+  // vDeepSeek: build providerOptions.deepseek for DeepSeek V4 models.
+  let deepseekProviderOptions:
+    | { thinking: { type: 'enabled' | 'disabled' }; reasoningEffort?: 'low' | 'medium' | 'high' | 'xhigh' | 'max' }
+    | undefined;
+  if (isDeepSeekModel(model)) {
+    const dsEffort = opts.reasoning_effort;
+    const thinkingEnabled = dsEffort && dsEffort !== 'off';
+    deepseekProviderOptions = {
+      thinking: { type: thinkingEnabled ? 'enabled' : 'disabled' },
+      ...(thinkingEnabled ? { reasoningEffort: dsEffort } : {}),
+    };
+  }
+
   // F6: per-chunk stall deadline. If the model stops emitting chunks for
   // STALL_TIMEOUT_MS the stallAc fires through AbortSignal.any; the post-loop
   // abort check below then throws AbortError → handleAbortOrError writes
@@ -297,7 +314,14 @@ export async function streamCompletion(
     ...(typeof opts.temperature === 'number' ? { temperature: opts.temperature } : {}),
     ...(typeof opts.top_p === 'number' ? { topP: opts.top_p } : {}),
     ...(typeof opts.presence_penalty === 'number' ? { presencePenalty: opts.presence_penalty } : {}),
-    ...(samplerBody ? { providerOptions: { openaiCompatible: samplerBody } } : {}),
+    ...(samplerBody || deepseekProviderOptions
+      ? {
+          providerOptions: {
+            ...(samplerBody ? { openaiCompatible: samplerBody } : {}),
+            ...(deepseekProviderOptions ? { deepseek: deepseekProviderOptions } : {}),
+          },
+        }
+      : {}),
     abortSignal: effectiveSignal,
   });
 
@@ -401,12 +425,26 @@ export async function streamCompletion(
 
   // Usage lands as a promise on the result; awaiting after fullStream is
   // drained is safe. AI SDK v6 names: `inputTokens` / `outputTokens`.
+  // Some providers (llama-swap via openai-compatible) return plain numbers;
+  // others (deepseek via @ai-sdk/deepseek) return {total, cacheRead, noCache, ...}.
   let promptTokens: number | null = null;
   let completionTokens: number | null = null;
+  let cacheReadTokens: number | null = null;
+  let reasoningTokens: number | null = null;
   try {
     const usage = await result.usage;
-    if (typeof usage.inputTokens === 'number') promptTokens = usage.inputTokens;
-    if (typeof usage.outputTokens === 'number') completionTokens = usage.outputTokens;
+    if (typeof usage.inputTokens === 'number') {
+      promptTokens = usage.inputTokens;
+    } else if (usage.inputTokens && typeof usage.inputTokens === 'object') {
+      promptTokens = (usage.inputTokens as Record<string, number | undefined>).total ?? null;
+      cacheReadTokens = (usage.inputTokens as Record<string, number | undefined>).cacheRead ?? null;
+    }
+    if (typeof usage.outputTokens === 'number') {
+      completionTokens = usage.outputTokens;
+    } else if (usage.outputTokens && typeof usage.outputTokens === 'object') {
+      completionTokens = (usage.outputTokens as Record<string, number | undefined>).total ?? null;
+      reasoningTokens = (usage.outputTokens as Record<string, number | undefined>).reasoning ?? null;
+    }
   } catch {
     // Some providers omit usage on partial streams; leave both null.
   }
@@ -422,6 +460,13 @@ export async function streamCompletion(
     );
   }
 
+  if (cacheReadTokens !== null || reasoningTokens !== null) {
+    ctx.log.debug(
+      { promptTokens, completionTokens, cacheReadTokens, reasoningTokens, model },
+      'streamCompletion: deepseek usage breakdown',
+    );
+  }
+
   return {
     finishReason,
     content,
@@ -429,6 +474,10 @@ export async function streamCompletion(
     promptTokens,
     completionTokens,
     reasoning: reasoningAccumulated,
+    // vDeepSeek: optional usage breakdown populated when the provider returns
+    // structured usage (cache hit tokens, reasoning tokens).
+    cacheReadTokens: cacheReadTokens ?? undefined,
+    reasoningTokens: reasoningTokens ?? undefined,
   };
   } finally {
     // Clear the stall timer whether the stream completes normally, throws, or

apps/server/src/services/inference/tool-input-repair.ts

@@ -0,0 +1,179 @@
+/**
+ * vWhale: schema-based tool input repair. When the model emits tool call args
+ * that don't match the expected types (common with weaker models), apply
+ * heuristic repairs before falling through to the Zod parse.
+ *
+ * Inspired by Whale's RepairToolInputForSpec:
+ *   - Coerce string "true"/"false" → boolean
+ *   - Unwrap markdown autolinks in string fields: <file:///path> → /path
+ *   - Wrap bare values in arrays when schema expects array
+ *   - Convert "42.0" decimal string → "42" for integer fields
+ *   - Recurse into objects to repair nested properties
+ */
+
+export interface ToolInputRepair {
+  field: string;
+  kind: string;
+  detail: string;
+}
+
+const MARKDOWN_AUTOLINK_RE = /^<(?:file|path):\/\/(.+?)>$/;
+
+/**
+ * Attempt to repair tool call args against the tool's JSON Schema.
+ * Returns the (possibly modified) args plus a list of repairs applied.
+ */
+export function repairToolInput(
+  schema: Record<string, unknown> | undefined,
+  args: Record<string, unknown>,
+): { repaired: Record<string, unknown>; repairs: ToolInputRepair[] } {
+  const repairs: ToolInputRepair[] = [];
+  if (!schema || typeof schema !== 'object') {
+    return { repaired: args, repairs };
+  }
+
+  const properties = (schema as Record<string, unknown>).properties as
+    Record<string, unknown> | undefined;
+  if (!properties) {
+    return { repaired: args, repairs };
+  }
+
+  const required = new Set<string>(
+    Array.isArray((schema as Record<string, unknown>).required)
+      ? (schema as Record<string, unknown>).required as string[]
+      : [],
+  );
+
+  const repaired: Record<string, unknown> = {};
+  for (const [key, value] of Object.entries(args)) {
+    const propSchema = properties[key] as Record<string, unknown> | undefined;
+    if (propSchema && value !== null && value !== undefined) {
+      repaired[key] = repairValue(key, propSchema, value, repairs, required.has(key));
+    } else {
+      repaired[key] = value;
+    }
+  }
+
+  // Drop keys not in the schema (only for required fields that are missing)
+  // to avoid polluting the model with hallucinated params.
+  for (const key of Object.keys(repaired)) {
+    if (!(key in properties)) {
+      repairs.push({ field: key, kind: 'removed_unknown', detail: `Removed unknown parameter '${key}'` });
+      delete repaired[key];
+    }
+  }
+
+  return { repaired, repairs };
+}
+
+function repairValue(
+  field: string,
+  schema: Record<string, unknown>,
+  value: unknown,
+  repairs: ToolInputRepair[],
+  required: boolean,
+): unknown {
+  const schemaType = schema.type;
+  const isArray = schemaType === 'array' || Array.isArray(schemaType)
+    ? schemaType === 'array' || (Array.isArray(schemaType) && schemaType.includes('array'))
+    : false;
+  const isObject = schemaType === 'object';
+  const isBoolean = schemaType === 'boolean';
+  const isInteger = schemaType === 'integer' || schemaType === 'number';
+  const isString = schemaType === 'string';
+
+  // --- Array repair: wrap bare value or empty object ---
+  if (isArray) {
+    if (!Array.isArray(value)) {
+      if (typeof value === 'string') {
+        // Try parsing as JSON array first
+        try {
+          const parsed = JSON.parse(value);
+          if (Array.isArray(parsed)) {
+            repairs.push({ field, kind: 'parsed_json_array', detail: `Parsed string as JSON array for '${field}'` });
+            return parsed;
+          }
+        } catch { /* not JSON */ }
+      }
+      if (typeof value === 'object' && value !== null && Object.keys(value).length === 0) {
+        if (required) {
+          repairs.push({ field, kind: 'empty_object_to_array', detail: `Converted empty object to empty array for '${field}'` });
+          return [];
+        }
+        repairs.push({ field, kind: 'empty_object_to_undefined', detail: `Removed empty object for optional array '${field}'` });
+        return undefined;
+      }
+      repairs.push({ field, kind: 'wrapped_in_array', detail: `Wrapped bare value in array for '${field}'` });
+      return [value];
+    }
+    // Recurse into array items
+    const itemsSchema = schema.items as Record<string, unknown> | undefined;
+    if (itemsSchema) {
+      return value.map((item, i) => repairValue(`${field}[${i}]`, itemsSchema, item, repairs, required));
+    }
+    return value;
+  }
+
+  // --- Object repair: recurse into properties ---
+  if (isObject && typeof value === 'object' && value !== null && !Array.isArray(value)) {
+    const props = (schema.properties as Record<string, unknown>) ?? {};
+    const repaired: Record<string, unknown> = {};
+    for (const [k, v] of Object.entries(value as Record<string, unknown>)) {
+      const propSchema = props[k] as Record<string, unknown> | undefined;
+      if (propSchema) {
+        repaired[k] = repairValue(`${field}.${k}`, propSchema, v, repairs, required);
+      } else {
+        repaired[k] = v;
+      }
+    }
+    return repaired;
+  }
+
+  // --- String repair: unwrap markdown autolinks ---
+  if (isString && typeof value === 'string') {
+    const match = value.match(MARKDOWN_AUTOLINK_RE);
+    if (match) {
+      repairs.push({ field, kind: 'unwrapped_markdown_link', detail: `Unwrapped markdown autolink for '${field}': ${value}` });
+      return match[1];
+    }
+    return value;
+  }
+
+  // --- Boolean coercion ---
+  if (isBoolean && typeof value === 'string') {
+    const lower = value.toLowerCase();
+    if (lower === 'true') {
+      repairs.push({ field, kind: 'coerced_to_boolean', detail: `Coerced string '${value}' → true for '${field}'` });
+      return true;
+    }
+    if (lower === 'false') {
+      repairs.push({ field, kind: 'coerced_to_boolean', detail: `Coerced string '${value}' → false for '${field}'` });
+      return false;
+    }
+    return value;
+  }
+
+  // --- Integer coercion: "42.0" → 42 ---
+  if (isInteger && typeof value === 'string') {
+    const num = Number(value);
+    if (!Number.isNaN(num)) {
+      repairs.push({ field, kind: 'coerced_to_number', detail: `Coerced string '${value}' → ${num} for '${field}'` });
+      return num;
+    }
+    return value;
+  }
+
+  // --- Integer coercion: boolean → 0/1 ---
+  if (isInteger && typeof value === 'boolean') {
+    repairs.push({ field, kind: 'coerced_boolean_to_integer', detail: `Coerced boolean ${value} → ${value ? 1 : 0} for '${field}'` });
+    return value ? 1 : 0;
+  }
+
+  // --- Empty string to null for optional fields ---
+  if (value === '' && !required) {
+    repairs.push({ field, kind: 'empty_string_to_undefined', detail: `Converted empty string for optional '${field}'` });
+    return undefined;
+  }
+
+  return value;
+}

apps/server/src/services/inference/tool-phase.ts

@@ -6,6 +6,7 @@ import type { ToolExecCtx } from '../tools.js';
 import { matchToolGlob } from '../agents.js';
 import { maybeFlagForCompaction } from './payload.js';
 import { insertParts, partsFromAssistantMessage, partsFromToolMessage } from './parts.js';
+import { getServerPermission } from '../mcp-client.js';
 // v1.13.16: richer unknown-tool error so the model can self-correct when it
 // drifts to a Claude Code tool name (e.g. read_file → suggest view_file).
 // Applies to all unknown tool names, not just <invoke>-derived ones — at the
@@ -17,6 +18,7 @@ import { formatUnknownToolError } from './tool-suggestions.js';
 // prompted about paths we couldn't grant anyway (e.g. /etc/passwd).
 import { resolveGrantRoot } from '../grant_resolver.js';
 import { stripToolMarkup } from './tool-call-parser.js';
+import { repairToolInput } from './tool-input-repair.js';
 import type { FailureKind } from './mistake-tracker.js';
 import type {
   InferenceContext,
@@ -34,6 +36,8 @@ async function executeToolCall(
   toolCall: ToolCall,
   extraRoots: readonly string[],
   toolCtx?: ToolExecCtx,
+  hooks?: import('../hooks.js').HookRunner,
+  sessionId?: string,
 ): Promise<{ output: unknown; truncated: boolean; error?: string; outcome: FailureKind | 'success' }> {
   // v#12 MistakeTracker: every return path carries an `outcome` so the turn
   // loop can detect a run of heterogeneous failures. The failure taxonomy
@@ -48,7 +52,61 @@ async function executeToolCall(
       outcome: 'tool_not_found',
     };
   }
-  const parsed = tool.inputSchema.safeParse(toolCall.args);
+  // MCP permission gate — block deny/ask before any Zod parsing or execution
+  const mcpPerm = getServerPermission(toolCall.name);
+  if (mcpPerm === 'deny') {
+    return { output: null, truncated: false, error: `blocked: MCP server denied tool '${toolCall.name}'`, outcome: 'permission_denied' };
+  }
+  if (mcpPerm === 'ask') {
+    return { output: null, truncated: false, error: `requires approval: tool '${toolCall.name}' needs user approval`, outcome: 'permission_denied' };
+  }
+  // vWhale: schema-based tool input repair. If the Zod parse fails, attempt
+  // heuristic repairs (type coercion, markdown-link unwrapping, array wrapping)
+  // and retry. Logs repairs for debugging.
+  let args = toolCall.args;
+  let parsed = tool.inputSchema.safeParse(args);
+  if (!parsed.success) {
+    const schema = tool.jsonSchema?.function?.parameters;
+    if (schema) {
+      const { repaired: repairedArgs, repairs } = repairToolInput(
+        schema as Record<string, unknown>,
+        args as Record<string, unknown>,
+      );
+      if (repairs.length > 0) {
+        const retry = tool.inputSchema.safeParse(repairedArgs);
+        if (retry.success) {
+          args = repairedArgs;
+          parsed = retry;
+        }
+      }
+    }
+  }
+  // vWhale: PreToolUse hook — can block execution.
+  if (hooks && sessionId) {
+    const hookResult = await hooks.run('PreToolUse', {
+      event: 'PreToolUse',
+      session_id: sessionId,
+      tool_name: toolCall.name,
+      tool_args: args as Record<string, unknown>,
+    });
+    if (hookResult.decision === 'block') {
+      return {
+        output: null,
+        truncated: false,
+        error: `blocked by hook: ${hookResult.reason ?? 'PreToolUse denied'}`,
+        outcome: 'permission_denied',
+      };
+    }
+    // Apply updated_input if the hook rewrote the args
+    if (hookResult.updated_input && typeof hookResult.updated_input === 'object') {
+      const reParsed = tool.inputSchema.safeParse(hookResult.updated_input);
+      if (reParsed.success) {
+        args = hookResult.updated_input as Record<string, unknown>;
+        parsed = reParsed;
+      }
+    }
+  }
+
   if (!parsed.success) {
     // v1.12 Track B.2: enrich the zod-reject path so the model sees a
     // one-line, tool-named hint ("tool 'search_symbols' rejected — query:
@@ -183,6 +241,8 @@ export async function executeToolPhase(
     tokens_used: updated?.tokens_used ?? null,
     ctx_used: updated?.ctx_used ?? null,
     ctx_max: updated?.ctx_max ?? null,
+    cache_tokens: result.cacheReadTokens ?? null,
+    reasoning_tokens: result.reasoningTokens ?? null,
     started_at: startedAt,
     finished_at: updated?.finished_at ?? null,
     model: session.model,
@@ -318,10 +378,22 @@ export async function executeToolPhase(
         });
         return;
       }
-      const tres = await executeToolCall(projectRoot, tc, session.allowed_read_paths, {
-        sql: ctx.sql,
-        sessionId,
-      });
+      const tres = await executeToolCall(
+        projectRoot, tc, session.allowed_read_paths,
+        { sql: ctx.sql, sessionId },
+        ctx.hooks, sessionId,
+      );
+      // vWhale: PostToolUse hook (best-effort, non-blocking).
+      if (ctx.hooks) {
+        ctx.hooks.run('PostToolUse', {
+          event: 'PostToolUse',
+          session_id: sessionId,
+          tool_name: tc.name,
+          tool_args: tc.args as Record<string, unknown>,
+          tool_result: tres.output,
+          tool_error: tres.error,
+        }).catch(() => {});
+      }
       // v#12 MistakeTracker: record the real execution outcome (success or a
       // FailureKind). This is the primary signal for heterogeneous-failure
       // detection.

apps/server/src/services/inference/turn.ts

@@ -144,6 +144,7 @@ export async function runAssistantTurn(
           log: ctx.log,
           broker: ctx.broker,
           chatId,
+          hooks: ctx.hooks,
         });
       } catch (err) {
         ctx.log.warn({ err, chatId }, 'auto-compaction failed; clearing flag and proceeding');
@@ -214,6 +215,16 @@ export async function runAssistantTurn(
 
     // ---- non-tool finish → finalize and exit ----
     if (result.toolCalls.length === 0) {
+      // vWhale: Stop hook (best-effort, non-blocking).
+      if (ctx.hooks) {
+        ctx.hooks.run('Stop', {
+          event: 'Stop',
+          session_id: sessionId,
+          chat_id: chatId,
+          last_assistant_text: result.content.slice(0, 500),
+          turn: stepNumber,
+        }).catch(() => {});
+      }
       await finalizeCompletion(ctx, iterArgs, result, state.startedAt, iterSession);
       break;
     }
@@ -309,6 +320,22 @@ export async function runAssistantTurn(
     assistantMessageId = toolPhaseResult.nextAssistantId!;
   }
 
+  // vWhale: Stop hook at post-loop exit (best-effort, non-blocking).
+  if (ctx.hooks) {
+    const loaded = await loadContext(ctx.sql, sessionId, chatId);
+    const lastAssistant = loaded?.history?.slice().reverse().find(
+      (m: import('../../types/api.js').Message) => m.role === 'assistant',
+    );
+    const content = lastAssistant?.content ?? '';
+    ctx.hooks.run('Stop', {
+      event: 'Stop',
+      session_id: sessionId,
+      chat_id: chatId,
+      last_assistant_text: content.slice(0, 500),
+      turn: stepNumber,
+    }).catch(() => {});
+  }
+
   // ---- post-loop: step-cap sentinel ----
   // When the loop exits because stepNumber reached effectiveCap, the last
   // iteration's tool phase returned 'continue' with a nextAssistantId that

apps/server/src/services/inference/types.ts

@@ -19,6 +19,7 @@ import type {
   UserStreamFrame,
 } from '../../types/api.js';
 import type { Broker } from '../broker.js';
+import type { HookRunner } from '../hooks.js';
 import type { MistakeState } from './mistake-tracker.js';
 
 export interface StreamPhaseState {
@@ -77,6 +78,8 @@ export interface InferenceFrame {
   started_at?: string | null;
   finished_at?: string | null;
   model?: string;
+  cache_tokens?: number | null;
+  reasoning_tokens?: number | null;
   session_id?: string;
   name?: string;
   // orchestrator frames ([D-6])
@@ -117,6 +120,9 @@ export interface InferenceContext {
   // inference goes through `publish`); keeping a separate field avoids
   // tempting other code paths into bypassing the session-id binding.
   broker: Broker;
+  // vWhale: lifecycle hooks runner. Undefined when no hooks configured.
+  // Hook calls are best-effort — a failing hook never blocks inference.
+  hooks?: HookRunner;
 }
 
 export interface StreamResult {
@@ -128,6 +134,12 @@ export interface StreamResult {
   // v1.13.1-C: reasoning text accumulated across reasoning-delta parts.
   // Empty string when the model doesn't emit reasoning (most cases).
   reasoning: string;
+  // vDeepSeek: optional cache-hit token count from DeepSeek's API.
+  // Only populated when using @ai-sdk/deepseek provider (not llama-swap).
+  cacheReadTokens?: number;
+  // vDeepSeek: optional reasoning token count from DeepSeek's API.
+  // Only populated when using @ai-sdk/deepseek provider (not llama-swap).
+  reasoningTokens?: number;
 }
 
 export interface TurnArgs {

apps/server/src/services/mcp-client.ts

@@ -31,11 +31,14 @@ interface McpToolDef {
   annotations?: McpToolAnnotations;
 }
 
+export type McpPermission = 'allow' | 'ask' | 'deny';
+
 interface ServerState {
   client: Client;
   transport: StreamableHTTPClientTransport | StdioClientTransport;
   tools: ToolDef<Record<string, unknown>>[];
   type: 'streamableHttp' | 'stdio';
+  permission: McpPermission;
 }
 
 // ---- Module-level state ----
@@ -137,6 +140,14 @@ export async function callTool(
   }
 }
 
+/** Return the permission level for a given MCP server. Defaults to 'allow' if unknown. */
+export function getServerPermission(prefixedToolName: string): McpPermission {
+  const serverName = toolToServer.get(prefixedToolName);
+  if (!serverName) return 'allow';
+  const state = servers.get(serverName);
+  return state?.permission ?? 'allow';
+}
+
 /** Return all wrapped ToolDefs from all connected servers, flattened. */
 export function getTools(): ToolDef<Record<string, unknown>>[] {
   const all: ToolDef<Record<string, unknown>>[] = [];
@@ -214,7 +225,8 @@ async function connectServer(entry: McpServerEntry): Promise<void> {
     toolToServer.set(wrapped.name, name);
   }
 
-  servers.set(name, { client, transport, tools, type: config.type });
+  const permission = (config as { permission?: McpPermission }).permission ?? 'allow';
+  servers.set(name, { client, transport, tools, type: config.type, permission });
 
   log!.info(
     { server: name, type: config.type, count: tools.length, names: tools.map((t) => t.name) },

apps/server/src/services/mcp-config.ts

@@ -17,12 +17,15 @@ import type { FastifyBaseLogger } from 'fastify';
 
 // ---- Zod schema ----
 
+const McpPermissionSchema = z.enum(['allow', 'ask', 'deny']).default('allow');
+
 const McpServerConfigSchema = z.discriminatedUnion('type', [
   z.object({
     type: z.literal('streamableHttp'),
     url: z.string().url(),
     headers: z.record(z.string()).optional(),
     enabled: z.boolean().default(true),
+    permission: McpPermissionSchema,
   }),
   z.object({
     type: z.literal('stdio'),
@@ -30,6 +33,7 @@ const McpServerConfigSchema = z.discriminatedUnion('type', [
     args: z.array(z.string()).default([]),
     env: z.record(z.string()).optional(),
     enabled: z.boolean().default(true),
+    permission: McpPermissionSchema,
   }),
 ]);
 

apps/server/src/services/message-columns.ts

@@ -7,10 +7,12 @@
 
 export const MESSAGE_COLUMNS =
   'id, session_id, chat_id, role, content, kind, tool_calls, tool_results, status, last_seq, ' +
-  'tokens_used, ctx_used, ctx_max, started_at, finished_at, created_at, metadata, ' +
+  'tokens_used, ctx_used, ctx_max, cache_tokens, reasoning_tokens, ' +
+  'started_at, finished_at, created_at, metadata, ' +
   'summary, tail_start_id, compacted_at, model';
 
 export const INFERENCE_MESSAGE_COLUMNS =
   'id, session_id, chat_id, role, content, kind, tool_calls, tool_results, status, last_seq, ' +
-  'tokens_used, ctx_used, ctx_max, started_at, finished_at, created_at, metadata, ' +
+  'tokens_used, ctx_used, ctx_max, cache_tokens, reasoning_tokens, ' +
+  'started_at, finished_at, created_at, metadata, ' +
   'reasoning_parts, model';

apps/server/src/services/model-context.ts

@@ -37,7 +37,18 @@ export function configureModelContext(opts: { llamaSwapUrl: string }): void {
   llamaSwapUrl = opts.llamaSwapUrl;
 }
 
+// vDeepSeek: DeepSeek models don't have a /upstream/<model>/props endpoint.
+// Return a reasonable default context so compaction estimates work.
+const DEEPSEEK_DEFAULT_N_CTX = 131_072;
+const DEEPSEEK_MODEL_PREFIX = 'deepseek-';
+
 export async function getModelContext(model: string): Promise<ModelContext | null> {
+  // vDeepSeek: DeepSeek models have no /upstream/<model>/props. Use a static
+  // default so compaction doesn't fall to the buffer-only path with tiny limits.
+  if (model.startsWith(DEEPSEEK_MODEL_PREFIX)) {
+    return { n_ctx: DEEPSEEK_DEFAULT_N_CTX };
+  }
+
   // 1. Positive cache hit — no TTL check, model n_ctx is invariant.
   const pos = positiveCache.get(model);
   if (pos) return pos;

apps/server/src/services/system-prompt.ts

@@ -101,7 +101,7 @@ export interface PrefixFingerprint {
   has_agent_system_prompt: boolean;
   has_session_override: boolean;
   has_project_override: boolean;
-  route: 'swap' | 'sidecar';
+  route: 'swap' | 'sidecar' | 'deepseek';
 }
 
 export interface PrefixDrift {
@@ -129,7 +129,7 @@ interface ObservedInputs {
   has_agent_system_prompt: boolean;
   has_session_override: boolean;
   has_project_override: boolean;
-  route: 'swap' | 'sidecar';
+  route: 'swap' | 'sidecar' | 'deepseek';
 }
 
 interface ObserverEntry {

apps/server/src/services/tools/codecontext/get_code_health.ts

@@ -0,0 +1,62 @@
+import { z } from 'zod';
+import type { ToolDef } from '../types.js';
+import { callBoocontext } from '../../boocontext_client.js';
+
+export const GetCodeHealthInput = z.object({
+  directory: z.string().optional().describe('Directory to analyze (defaults to project root)'),
+  file: z.string().optional().describe('Optional: specific file to analyze'),
+});
+export type GetCodeHealthInputT = z.infer<typeof GetCodeHealthInput>;
+
+const DESCRIPTION =
+  'Code health analysis. Returns A–F grades per file across 7 dimensions ' +
+  '(cohesion, coupling, complexity, documentation, duplication, unit size, test coverage). ' +
+  'Includes project health summary and refactoring candidates.';
+
+/**
+ * Standalone execute function — calls the boocontext MCP server's
+ * boocontext_health tool and returns the raw report text.
+ *
+ * Structured for direct test access: accepts input + projectPath,
+ * no side effects beyond the MCP call.
+ */
+export async function executeGetCodeHealth(
+  input: GetCodeHealthInputT,
+  projectPath: string,
+): Promise<string> {
+  const args: Record<string, unknown> = {};
+  if (input.directory) args['directory'] = input.directory;
+  if (input.file) args['file'] = input.file;
+  const resp = await callBoocontext({ toolName: 'boocontext_health', args });
+  return resp.result;
+}
+
+export const getCodeHealth: ToolDef<GetCodeHealthInputT> = {
+  name: 'get_code_health',
+  description: DESCRIPTION,
+  inputSchema: GetCodeHealthInput,
+  jsonSchema: {
+    type: 'function',
+    function: {
+      name: 'get_code_health',
+      description: DESCRIPTION,
+      parameters: {
+        type: 'object',
+        properties: {
+          directory: {
+            type: 'string',
+            description: 'Directory to analyze (defaults to project root)',
+          },
+          file: {
+            type: 'string',
+            description: 'Optional: specific file to analyze',
+          },
+        },
+        additionalProperties: false,
+      },
+    },
+  },
+  async execute(input, projectRoot) {
+    return executeGetCodeHealth(input, projectRoot);
+  },
+};

apps/server/src/services/tools/codecontext/get_code_impact.ts

@@ -0,0 +1,228 @@
+import { spawn } from 'node:child_process';
+import { resolve } from 'node:path';
+import { z } from 'zod';
+import type { ToolDef } from '../types.js';
+import type { CodecontextResponse } from '../../codecontext_client.js';
+
+// ======================= MCP Client =======================
+
+const BOOCONTEXT_PATH = resolve('/opt/forks/boocontext/dist/standalone.js');
+const TOOL_CALL_TIMEOUT_MS = 60_000;
+
+interface JsonRpcMessage {
+  jsonrpc: '2.0';
+  id?: number | string;
+  result?: {
+    content?: Array<{ type: string; text: string }>;
+  };
+  error?: { code?: number; message: string };
+}
+
+/**
+ * Single-shot MCP JSON-RPC client for boocontext.
+ * Spawns the process, sends initialize + tools/call over NDJSON, returns the
+ * text result from the content array.  The boocontext MCP server auto-detects
+ * newline-delimited JSON transport when the first input lacks Content-Length
+ * headers, which is exactly what we send.
+ */
+async function callBoocontext(
+  toolName: string,
+  args: Record<string, unknown>,
+): Promise<string> {
+  return new Promise<string>((resolvePromise, reject) => {
+    const child = spawn(process.execPath, [BOOCONTEXT_PATH], {
+      stdio: ['pipe', 'pipe', 'pipe'],
+      timeout: TOOL_CALL_TIMEOUT_MS,
+    });
+
+    let stdout = '';
+    let stderr = '';
+    let resolved = false;
+
+    function finalize(err?: Error, result?: string): void {
+      if (resolved) return;
+      resolved = true;
+      if (err) reject(err);
+      else resolvePromise(result!);
+      child.kill();
+    }
+
+    child.stdout!.on('data', (chunk: Buffer) => {
+      stdout += chunk.toString();
+    });
+
+    child.stderr!.on('data', (chunk: Buffer) => {
+      stderr += chunk.toString();
+    });
+
+    child.on('error', (err: Error) => {
+      finalize(new Error(`boocontext spawn error: ${err.message}`));
+    });
+
+    child.on('close', (code: number | null) => {
+      if (resolved) return;
+
+      // Parse newline-delimited JSON responses from stdout
+      const lines = stdout.split('\n').filter((l) => l.trim().length > 0);
+      let toolText: string | undefined;
+      let toolError: string | undefined;
+
+      for (const line of lines) {
+        try {
+          const msg = JSON.parse(line) as JsonRpcMessage;
+          if (msg.id === 2) {
+            if (msg.error) {
+              toolError = msg.error.message ?? 'boocontext tool call failed';
+            } else if (msg.result?.content?.[0]?.text !== undefined) {
+              toolText = msg.result.content[0].text;
+            }
+          }
+        } catch {
+          // skip malformed JSON lines
+        }
+      }
+
+      if (toolError) {
+        finalize(new Error(toolError));
+      } else if (toolText !== undefined) {
+        finalize(undefined, toolText);
+      } else {
+        const errSuffix =
+          stderr.length > 0 ? ` stderr: ${stderr.slice(0, 500)}` : '';
+        finalize(
+          new Error(`boocontext MCP call failed (exit ${code})${errSuffix}`),
+        );
+      }
+    });
+
+    // Step 1: initialize — establishes MCP protocol version + capabilities
+    child.stdin!.write(
+      JSON.stringify({
+        jsonrpc: '2.0',
+        id: 1,
+        method: 'initialize',
+        params: {
+          protocolVersion: '2024-11-05',
+          capabilities: {},
+          clientInfo: { name: 'boocode-server', version: '1.0.0' },
+        },
+      }) + '\n',
+    );
+
+    // Step 2: tools/call — invoke the named boocontext tool
+    child.stdin!.write(
+      JSON.stringify({
+        jsonrpc: '2.0',
+        id: 2,
+        method: 'tools/call',
+        params: { name: toolName, arguments: args },
+      }) + '\n',
+    );
+
+    child.stdin!.end();
+
+    // Safety timeout — prevent hung processes
+    setTimeout(() => {
+      finalize(
+        new Error(
+          `boocontext call timed out after ${TOOL_CALL_TIMEOUT_MS}ms`,
+        ),
+      );
+    }, TOOL_CALL_TIMEOUT_MS);
+  });
+}
+
+// ======================= Tool Definition =======================
+
+const TRUNCATION_LIMIT = 32_000;
+
+export const GetCodeImpactInput = z.object({
+  symbol: z.string().min(1).describe('Symbol name for TSA trace_impact'),
+  file: z.string().optional().describe('File path for codesight blast_radius'),
+  directory: z
+    .string()
+    .optional()
+    .describe('Directory (defaults to project root)'),
+  depth: z
+    .number()
+    .int()
+    .min(1)
+    .max(5)
+    .optional()
+    .describe('Max blast-radius traversal depth (default 1)'),
+});
+export type GetCodeImpactInputT = z.infer<typeof GetCodeImpactInput>;
+
+const DESCRIPTION =
+  'Impact analysis. Merges symbol-level call trace with file-level blast radius. ' +
+  'Use before making changes to understand change propagation. ' +
+  'Single call replaces separate get_symbol_info + get_blast_radius steps.';
+
+/**
+ * Standalone execute function — calls the boocontext MCP `boocontext_impact`
+ * tool via a short-lived child process, then wraps the result in the standard
+ * CodecontextResponse shape with inline truncation at 32 KB.
+ */
+export async function executeGetCodeImpact(
+  input: GetCodeImpactInputT,
+  projectPath: string,
+): Promise<CodecontextResponse> {
+  const args: Record<string, unknown> = {
+    symbol: input.symbol,
+    directory: input.directory ?? projectPath,
+  };
+  if (input.file) args['file'] = input.file;
+
+  const text = await callBoocontext('boocontext_impact', args);
+
+  // Inline truncation matching codecontext_client.ts patterns (32 KB ceiling).
+  if (text.length > TRUNCATION_LIMIT) {
+    const sliced = text.slice(0, TRUNCATION_LIMIT);
+    const omitted = text.length - TRUNCATION_LIMIT;
+    return {
+      result: `${sliced}\n\n[truncated, ${omitted} chars omitted; narrow with symbol or file parameters]`,
+      truncated: true,
+    };
+  }
+
+  return { result: text, truncated: false };
+}
+
+export const getCodeImpact: ToolDef<GetCodeImpactInputT> = {
+  name: 'get_code_impact',
+  description: DESCRIPTION,
+  inputSchema: GetCodeImpactInput,
+  jsonSchema: {
+    type: 'function',
+    function: {
+      name: 'get_code_impact',
+      description: DESCRIPTION,
+      parameters: {
+        type: 'object',
+        properties: {
+          symbol: {
+            type: 'string',
+            description: 'Symbol name for TSA trace_impact',
+          },
+          file: {
+            type: 'string',
+            description: 'File path for codesight blast_radius',
+          },
+          directory: {
+            type: 'string',
+            description: 'Directory (defaults to project root)',
+          },
+          depth: {
+            type: 'number',
+            description: 'Max blast-radius traversal depth (default 1)',
+          },
+        },
+        required: ['symbol'],
+        additionalProperties: false,
+      },
+    },
+  },
+  execute(input, projectRoot) {
+    return executeGetCodeImpact(input, projectRoot);
+  },
+};

apps/server/src/services/tools/codecontext/get_code_map.ts

@@ -0,0 +1,192 @@
+import { spawn } from 'node:child_process';
+import { z } from 'zod';
+import type { ToolDef } from '../types.js';
+
+export const GetCodeMapInput = z.object({
+  directory: z.string().optional().describe('Directory to scan (defaults to project root)'),
+  compress: z.boolean().optional().describe('Apply DCP compression if payload exceeds threshold (default: true)'),
+});
+export type GetCodeMapInputT = z.infer<typeof GetCodeMapInput>;
+
+const DESCRIPTION =
+  'DCP-compressed codebase context map. Returns filenames, sizes, import relationships in a compressed format. ' +
+  'Use compress=false for full detail, compress=true (default) for token-efficient overview.';
+
+const BOOCONTEXT_PATH = '/opt/forks/boocontext/dist/standalone.js';
+const TOOL_TIMEOUT_MS = 30_000;
+const MAX_RESULT_BYTES = 32_768;
+
+export interface CodeMapResponse {
+  result: string;
+  truncated: boolean;
+}
+
+/**
+ * Calls the boocontext MCP server over stdio JSON-RPC to invoke
+ * the boocontext_map tool. Spawns the standalone binary, sends
+ * initialize + tools/call, collects NDJSON responses, and kills
+ * the child process.
+ */
+function callBoocontextMap(args: Record<string, unknown>): Promise<CodeMapResponse> {
+  return new Promise((resolve, reject) => {
+    const child = spawn('node', [BOOCONTEXT_PATH], {
+      stdio: ['pipe', 'pipe', 'pipe'],
+    });
+
+    let stdoutBuf = '';
+    const lines: string[] = [];
+    let timedOut = false;
+    let resolved = false;
+
+    const timer = setTimeout(() => {
+      timedOut = true;
+      child.kill('SIGKILL');
+      reject(new Error(`boocontext MCP call timed out after ${TOOL_TIMEOUT_MS}ms`));
+    }, TOOL_TIMEOUT_MS);
+
+    function tryParse(): void {
+      if (resolved || timedOut) return;
+
+      // Accumulate complete NDJSON lines
+      const parts = stdoutBuf.split('\n');
+      stdoutBuf = parts.pop()! ?? '';
+      for (const p of parts) {
+        const t = p.trim();
+        if (t) lines.push(t);
+      }
+
+      // Need at least 2 responses: initialize + tools/call
+      if (lines.length < 2) return;
+
+      resolved = true;
+      clearTimeout(timer);
+      child.kill();
+
+      try {
+        const callResponse = JSON.parse(lines[1]!);
+        if (callResponse.error) {
+          reject(new Error(`MCP error: ${callResponse.error.message}`));
+          return;
+        }
+
+        const content = callResponse.result?.content;
+        if (!content?.[0]?.text) {
+          reject(new Error('Unexpected MCP response shape — missing content[0].text'));
+          return;
+        }
+
+        // content[0].text is JSON-stringified VerdictEnvelope from boocontext
+        const envelope = JSON.parse(content[0].text as string);
+        const details = envelope.details;
+
+        let result: string;
+        if (details && typeof details === 'object' && 'data' in details) {
+          // DcpEnvelope shape: { compressed, originalLength, compressedLength, data }
+          if (details.compressed) {
+            // Return the full DcpEnvelope as JSON so the LLM can pass it
+            // transparently to a decompression step
+            result = JSON.stringify(details);
+          } else {
+            // Uncompressed — data is the raw output
+            result = details.data;
+          }
+        } else {
+          result = JSON.stringify(details ?? envelope);
+        }
+
+        const truncated = Buffer.byteLength(result, 'utf-8') > MAX_RESULT_BYTES;
+        if (truncated) {
+          result = result.substring(0, MAX_RESULT_BYTES);
+        }
+
+        resolve({ result, truncated });
+      } catch (e: any) {
+        reject(new Error(`Failed to parse boocontext response: ${e.message}`));
+      }
+    }
+
+    child.stdout!.on('data', (chunk: Buffer) => {
+      if (timedOut) return;
+      stdoutBuf += chunk.toString('utf-8');
+      tryParse();
+    });
+
+    child.stderr!.on('data', (_chunk: Buffer) => {
+      // Captured but not surfaced — logged only on parse failure
+    });
+
+    child.on('error', (err: Error) => {
+      clearTimeout(timer);
+      if (!resolved) {
+        resolved = true;
+        reject(new Error(`boocontext spawn failed: ${err.message}`));
+      }
+    });
+
+    child.on('close', () => {
+      clearTimeout(timer);
+      if (!resolved && !timedOut) {
+        tryParse();
+        if (!resolved) {
+          resolved = true;
+          reject(new Error('boocontext process closed without producing a valid response'));
+        }
+      }
+    });
+
+    // Step 1: initialize
+    child.stdin!.write(
+      JSON.stringify({ jsonrpc: '2.0', id: 1, method: 'initialize' }) + '\n',
+    );
+
+    // Step 2: tools/call for boocontext_map
+    child.stdin!.write(
+      JSON.stringify({
+        jsonrpc: '2.0',
+        id: 2,
+        method: 'tools/call',
+        params: { name: 'boocontext_map', arguments: args },
+      }) + '\n',
+    );
+  });
+}
+
+export const getCodeMap: ToolDef<GetCodeMapInputT> = {
+  name: 'get_code_map',
+  description: DESCRIPTION,
+  inputSchema: GetCodeMapInput,
+  jsonSchema: {
+    type: 'function',
+    function: {
+      name: 'get_code_map',
+      description: DESCRIPTION,
+      parameters: {
+        type: 'object',
+        properties: {
+          directory: { type: 'string', description: 'Directory to scan (defaults to project root)' },
+          compress: {
+            type: 'boolean',
+            description: 'Apply DCP compression if payload exceeds threshold (default: true)',
+          },
+        },
+        additionalProperties: false,
+      },
+    },
+  },
+  async execute(input, projectRoot): Promise<CodeMapResponse> {
+    return callBoocontextMap({
+      directory: input.directory ?? projectRoot,
+      compress: input.compress ?? true,
+    });
+  },
+};
+
+export async function executeGetCodeMap(
+  input: GetCodeMapInputT,
+  projectRoot: string,
+): Promise<CodeMapResponse> {
+  return callBoocontextMap({
+    directory: input.directory ?? projectRoot,
+    compress: input.compress ?? true,
+  });
+}

apps/server/src/services/tools/codecontext/get_type_info.ts

@@ -0,0 +1,262 @@
+import { z } from 'zod';
+import { spawn } from 'node:child_process';
+import type { ToolDef } from '../types.js';
+import type { CodecontextResponse } from '../../codecontext_client.js';
+
+const BOOCONTEXT_PATH = '/opt/forks/boocontext/dist/standalone.js';
+const TRUNCATION_LIMIT = 32_000;
+
+export const GetTypeInfoInput = z.object({
+  file: z.string().min(1).describe('File path to resolve types in'),
+  symbol: z.string().optional().describe('Symbol name to resolve (supports regex)'),
+  directory: z.string().optional().describe('Project directory for type resolution context'),
+});
+export type GetTypeInfoInputT = z.infer<typeof GetTypeInfoInput>;
+
+const DESCRIPTION =
+  'TypeScript type recovery. Returns type signatures, interface definitions, ' +
+  'generic constraints, and JSDoc for symbols in a file. Uses type-inject MCP server.';
+
+// ---- JSON-RPC-over-stdio MCP caller for boocontext --------------------------
+
+async function callBoocontext(
+  toolName: string,
+  args: Record<string, unknown>,
+): Promise<CodecontextResponse> {
+  const child = spawn(process.execPath, [BOOCONTEXT_PATH], {
+    stdio: ['pipe', 'pipe', 'pipe'],
+    timeout: 60_000,
+  });
+
+  let stderrBuf = '';
+  child.stderr!.on('data', (chunk: Buffer) => {
+    stderrBuf += chunk.toString('utf-8');
+  });
+
+  let killed = false;
+  const killChild = () => {
+    if (killed) return;
+    killed = true;
+    child.kill();
+  };
+
+  try {
+    // Read one complete JSON-RPC response from stdout (handles both
+    // Content-Length framed and newline-delimited transport).
+    async function readResponse(timeoutMs = 30_000): Promise<unknown> {
+      return new Promise((resolve, reject) => {
+        const timer = setTimeout(() => {
+          cleanup();
+          reject(new Error('Timeout reading boocontext response'));
+        }, timeoutMs);
+
+        let buf = '';
+
+        const cleanup = () => {
+          clearTimeout(timer);
+          child.stdout!.removeListener('data', onData);
+          child.stdout!.removeListener('end', onEnd);
+          child.stdout!.removeListener('error', onError);
+        };
+
+        const onData = (chunk: Buffer) => {
+          buf += chunk.toString('utf-8');
+
+          const msg = tryExtractMessage(buf);
+          if (msg !== null) {
+            cleanup();
+            resolve(msg);
+            return;
+          }
+
+          if (buf.length > 1_024 * 1_024) {
+            cleanup();
+            reject(new Error('Boocontext response exceeded 1 MB'));
+          }
+        };
+
+        const onEnd = () => {
+          cleanup();
+          if (buf.trim()) {
+            try {
+              resolve(JSON.parse(buf.trim()));
+            } catch {
+              reject(new Error('Boocontext stream ended with incomplete data'));
+            }
+          } else {
+            reject(new Error('Boocontext stream ended unexpectedly'));
+          }
+        };
+
+        const onError = (err: Error) => {
+          cleanup();
+          reject(err);
+        };
+
+        child.stdout!.on('data', onData);
+        child.stdout!.on('end', onEnd);
+        child.stdout!.on('error', onError);
+      });
+    }
+
+    // Wait for the process to be fully spawned.
+    await new Promise<void>((resolve, reject) => {
+      child.on('error', reject);
+      child.on('spawn', () => resolve());
+    });
+
+    // Step 1 — MCP initialize
+    let reqId = 0;
+    reqId++;
+    child.stdin!.write(
+      JSON.stringify({ jsonrpc: '2.0', id: reqId, method: 'initialize' }) + '\n',
+    );
+
+    const initResp = await readResponse() as { error?: { message: string } };
+    if (initResp.error) {
+      throw new Error(`Boocontext init failed: ${initResp.error.message}`);
+    }
+
+    // Step 2 — tools/call
+    reqId++;
+    child.stdin!.write(
+      JSON.stringify({
+        jsonrpc: '2.0',
+        id: reqId,
+        method: 'tools/call',
+        params: { name: toolName, arguments: args },
+      }) + '\n',
+    );
+
+    const callResp = await readResponse() as {
+      error?: { message: string };
+      result?: { content?: Array<{ type: string; text: string }> };
+    };
+    if (callResp.error) {
+      throw new Error(`Boocontext tool call failed: ${callResp.error.message}`);
+    }
+
+    // Extract text from the MCP tool result shape:
+    // { content: [{ type: "text", text: "…" }] }
+    const content = callResp.result?.content;
+    let text: string;
+    if (Array.isArray(content) && content.length > 0 && content[0]!.type === 'text') {
+      text = content[0]!.text;
+    } else {
+      text = JSON.stringify(callResp.result);
+    }
+
+    // Inline truncation at 32 KB.
+    if (text.length > TRUNCATION_LIMIT) {
+      const omitted = text.length - TRUNCATION_LIMIT;
+      return {
+        result:
+          text.slice(0, TRUNCATION_LIMIT) +
+          `\n\n[truncated, ${omitted} chars omitted; narrow with file or symbol filter]`,
+        truncated: true,
+      };
+    }
+
+    return { result: text, truncated: false };
+  } finally {
+    killChild();
+    // Give the process a moment to release resources.
+    await new Promise<void>((resolve) => {
+      const timer = setTimeout(resolve, 2_000);
+      child.on('exit', () => {
+        clearTimeout(timer);
+        resolve();
+      });
+    });
+  }
+}
+
+/**
+ * Attempt to extract one complete JSON-RPC message from the head of a
+ * buffer.  Handles both Content-Length framed and newline-delimited
+ * formats.  Returns `null` when more data is needed.
+ */
+function tryExtractMessage(buf: string): unknown | null {
+  // --- Content-Length framed ---
+  const headerEnd = buf.indexOf('\r\n\r\n');
+  if (headerEnd !== -1) {
+    const header = buf.substring(0, headerEnd);
+    const lengthMatch = header.match(/Content-Length:\s*(\d+)/i);
+    if (lengthMatch) {
+      const contentLength = parseInt(lengthMatch[1]!, 10);
+      const bodyStart = headerEnd + 4;
+      if (buf.length >= bodyStart + contentLength) {
+        const jsonStr = buf.substring(bodyStart, bodyStart + contentLength);
+        return JSON.parse(jsonStr);
+      }
+      return null; // need more data
+    }
+    // Has \r\n\r\n but no Content-Length — junk segment; skip and retry.
+    return tryExtractMessage(buf.substring(headerEnd + 4));
+  }
+
+  // --- Newline-delimited ---
+  const nlIndex = buf.indexOf('\n');
+  if (nlIndex !== -1) {
+    const line = buf.substring(0, nlIndex).trim();
+    if (line && line.startsWith('{')) {
+      return JSON.parse(line);
+    }
+    // Non-JSON line (e.g. stderr echo), skip and continue.
+    return tryExtractMessage(buf.substring(nlIndex + 1));
+  }
+
+  return null; // need more data
+}
+
+// ---- ToolDef ----------------------------------------------------------------
+
+export const getTypeInfo: ToolDef<GetTypeInfoInputT> = {
+  name: 'get_type_info',
+  description: DESCRIPTION,
+  inputSchema: GetTypeInfoInput,
+  jsonSchema: {
+    type: 'function',
+    function: {
+      name: 'get_type_info',
+      description: DESCRIPTION,
+      parameters: {
+        type: 'object',
+        properties: {
+          file: { type: 'string', description: 'File path to resolve types in' },
+          symbol: {
+            type: 'string',
+            description: 'Symbol name to resolve (supports regex)',
+          },
+          directory: {
+            type: 'string',
+            description: 'Project directory for type resolution context',
+          },
+        },
+        required: ['file'],
+        additionalProperties: false,
+      },
+    },
+  },
+  async execute(input): Promise<CodecontextResponse> {
+    const args: Record<string, unknown> = { file: input.file };
+    if (input.symbol) args['symbol'] = input.symbol;
+    return callBoocontext('boocontext_types', args);
+  },
+};
+
+/**
+ * Standalone execute function matching the `execute` shape returned by
+ * `makeCodecontextTool` — useful for direct callers and tests.
+ *
+ * Note: unlike the HTTP-backed codecontext tools this does NOT accept a
+ * `fetcher` override because it communicates over stdio rather than HTTP.
+ */
+export async function executeGetTypeInfo(
+  input: GetTypeInfoInputT,
+  _projectPath?: string,
+): Promise<CodecontextResponse> {
+  const args: Record<string, unknown> = { file: input.file };
+  if (input.symbol) args['symbol'] = input.symbol;
+  return callBoocontext('boocontext_types', args);
+}

apps/server/src/services/tools/codecontext/index.ts

@@ -13,3 +13,8 @@ export { getBlastRadius } from './get_blast_radius.js';
 export { getHotFiles } from './get_hot_files.js';
 export { getRoutes } from './get_routes.js';
 export { getMiddleware } from './get_middleware.js';
+// v2.8.14-domain2-phase1: boocontext-backed tools.
+export { getCodeHealth } from './get_code_health.js';
+export { getCodeImpact } from './get_code_impact.js';
+export { getTypeInfo } from './get_type_info.js';
+export { getCodeMap } from './get_code_map.js';

apps/server/src/services/tools/registry.ts

@@ -19,6 +19,10 @@ import {
   getHotFiles,
   getRoutes,
   getMiddleware,
+  getCodeHealth,
+  getCodeImpact,
+  getTypeInfo,
+  getCodeMap,
 } from './codecontext/index.js';
 // v1.13.17-cross-repo-reads: cross-repo read grant request tool. Paired
 // with the pause-on-pending-grant branch in inference/tool-phase.ts and the
@@ -75,6 +79,12 @@ export let ALL_TOOLS: ToolDef<unknown>[] = [
   // v2.6.x: read a tab's transcript by its session-scoped tab number.
   // Read-only; uses the ToolExecCtx 4th arg for DB/session access.
   readTabByNumber as ToolDef<unknown>,
+  // v2.8.14-domain2-phase1: boocontext-backed tools. Backed by the boocontext
+  // MCP server. All read-only. Health, impact, types, map analysis.
+  getCodeHealth as ToolDef<unknown>,
+  getCodeImpact as ToolDef<unknown>,
+  getTypeInfo as ToolDef<unknown>,
+  getCodeMap as ToolDef<unknown>,
 ].sort((a, b) => a.name.localeCompare(b.name));
 
 export let TOOLS_BY_NAME: Record<string, ToolDef<unknown>> = Object.fromEntries(

apps/server/src/types/api.ts

@@ -127,6 +127,9 @@ export interface Agent {
   // bounded only by MAX_STEPS (200). 0 means "no tool calls allowed."
   steps: number | null;
   llama_extra_args: string[] | null;
+  // vDeepSeek: thinking/reasoning effort for DeepSeek V4 models.
+  // Maps to DeepSeek's reasoning_effort API param.
+  reasoning_effort: 'off' | 'low' | 'medium' | 'high' | 'xhigh' | 'max' | null;
 }
 
 // One entry per malformed `## Name` block. Per-block errors don't fail the
@@ -206,6 +209,8 @@ export interface Message {
   tokens_used: number | null;
   ctx_used: number | null;
   ctx_max: number | null;
+  cache_tokens: number | null;
+  reasoning_tokens: number | null;
   started_at: string | null;
   finished_at: string | null;
   created_at: string;

apps/web/src/api/types.ts

@@ -152,6 +152,8 @@ export interface Message {
   tokens_used: number | null;
   ctx_used: number | null;
   ctx_max: number | null;
+  cache_tokens: number | null;
+  reasoning_tokens: number | null;
   // model-attribution: which model produced this assistant message (null for
   // user/system rows + pre-attribution messages). Rendered as a chip.
   model: string | null;

apps/web/src/components/MessageBubble.tsx

@@ -156,9 +156,16 @@ function StatsLine({ message }: { message: Message }) {
         : `${ctxUsed} ctx`
       : null;
 
+  const cacheHit = message.cache_tokens;
+  const reasoning = message.reasoning_tokens;
+  const cachePart = typeof cacheHit === 'number' && cacheHit > 0 ? `cache ${cacheHit}` : null;
+  const reasoningPart = typeof reasoning === 'number' && reasoning > 0 ? `think ${reasoning}` : null;
+
   const parts: string[] = [`${tokens} tokens`];
   if (tps !== null) parts.push(`${tps.toFixed(1)} tok/s`);
   if (ctxPart) parts.push(ctxPart);
+  if (cachePart) parts.push(cachePart);
+  if (reasoningPart) parts.push(reasoningPart);
 
   return (
     <div className="text-[10px] font-mono text-muted-foreground">

apps/web/src/hooks/useSessionStream.ts

@@ -123,6 +123,8 @@ function applyFrame(state: State, frame: WsFrame): State {
               ...(frame.tokens_used !== undefined ? { tokens_used: frame.tokens_used } : {}),
               ...(frame.ctx_used !== undefined ? { ctx_used: frame.ctx_used } : {}),
               ...(frame.ctx_max !== undefined ? { ctx_max: frame.ctx_max } : {}),
+              ...(frame.cache_tokens !== undefined ? { cache_tokens: frame.cache_tokens } : {}),
+              ...(frame.reasoning_tokens !== undefined ? { reasoning_tokens: frame.reasoning_tokens } : {}),
               ...(frame.started_at !== undefined ? { started_at: frame.started_at } : {}),
               ...(frame.finished_at !== undefined ? { finished_at: frame.finished_at } : {}),
               ...(frame.model !== undefined ? { model: frame.model } : {}),

data/AGENTS.md

@@ -6,7 +6,7 @@ Operating rules for every agent in this registry. Full procedures live in the `c
 
 **Worktrees** — Isolate work in a worktree when it is parallel to in-progress work, risky/experimental, a hotfix interrupting other work, or splits into independent units — just create when clear, propose in one line when ambiguous, skip quick/small single-stream work. Branch from a stable base (default branch); worktrees persist (never auto-remove or auto-merge); they isolate code state, not runtime (ports/DBs/services still collide). Full heuristic: invoke `using-worktrees`.
 
-**Sampling knobs** — Each `## Name` frontmatter block accepts these per-agent sampler fields, threaded into the llama-swap chat-completion request: `temperature`, `top_p`, `top_k`, `min_p`, `presence_penalty`, and (v2.6) `top_n_sigma`, `dry_multiplier`, `dry_base`, `dry_allowed_length`, `dry_penalty_last_n`. The `top_n_sigma` + `dry_*` repetition family curb the doom-loop-prone local model. Omit a field to leave it at the server default. Example: `top_n_sigma: 1.0`, `dry_multiplier: 0.8`, `dry_base: 1.75`, `dry_allowed_length: 2`, `dry_penalty_last_n: -1` (-1 = whole context).
+**Sampling knobs** — Each `## Name` frontmatter block accepts these per-agent sampler fields, threaded into the llama-swap chat-completion request: `temperature`, `top_p`, `top_k`, `min_p`, `presence_penalty`, and (v2.6) `top_n_sigma`, `dry_multiplier`, `dry_base`, `dry_allowed_length`, `dry_penalty_last_n`. The `top_n_sigma` + `dry_*` repetition family curb the doom-loop-prone local model. Omit a field to leave it at the server default. Example: `top_n_sigma: 1.0`, `dry_multiplier: 0.8`, `dry_base: 1.75`, `dry_allowed_length: 2`, `dry_penalty_last_n: -1` (-1 = whole context). DeepSeek V4 models also accept `reasoning_effort` (low/medium/high/xhigh/max); omit to disable thinking mode. Example: `reasoning_effort: 'high'`.
 
 **Reasoning budget** — To cap a reasoning model's thinking tokens, pass `--reasoning-budget` through `llama_extra_args` (already permitted by the deny-list validator; routes the agent to llama-sidecar). Example frontmatter line: `llama_extra_args: ["--reasoning-budget", "2048"]`. This is a sidecar process flag, not a chat-completion body param — distinct from the sampling knobs above.
 
@@ -17,7 +17,7 @@ top_p: 0.95
 top_k: 20
 min_p: 0.0
 presence_penalty: 0.0
-tools: [find_files, get_codebase_overview, get_dependencies, get_file_analysis, get_framework_analysis, get_semantic_neighborhoods, get_symbol_info, grep, list_dir, search_symbols, view_file, watch_changes, request_read_access, view_truncated_output, ask_user_input, git_status, get_blast_radius, get_hot_files, get_middleware, get_routes]
+tools: [find_files, get_blast_radius, get_codebase_overview, get_dependencies, get_file_analysis, get_framework_analysis, get_hot_files, get_middleware, get_routes, get_semantic_neighborhoods, get_symbol_info, git_status, grep, list_dir, request_read_access, search_symbols, view_file, view_truncated_output]
 description: Reviews code for bugs, security issues, and maintainability. Read-only.
 ---
 You review code. Find real problems, not style nits.
@@ -56,7 +56,7 @@ top_p: 0.95
 top_k: 20
 min_p: 0.0
 presence_penalty: 0.0
-tools: [find_files, get_codebase_overview, get_dependencies, get_file_analysis, get_framework_analysis, get_semantic_neighborhoods, get_symbol_info, grep, list_dir, search_symbols, view_file, watch_changes, request_read_access, view_truncated_output, ask_user_input, git_status, get_blast_radius, get_hot_files, get_middleware, get_routes]
+tools: [ask_user_input, find_files, get_blast_radius, get_codebase_overview, get_dependencies, get_file_analysis, get_framework_analysis, get_hot_files, get_semantic_neighborhoods, get_symbol_info, git_status, grep, list_dir, request_read_access, search_symbols, view_file, view_truncated_output, watch_changes]
 description: Diagnoses bugs from error messages, logs, or described symptoms.
 ---
 You diagnose bugs. Form a hypothesis, prove it with evidence from the code.
@@ -82,7 +82,7 @@ top_k: 20
 min_p: 0.0
 presence_penalty: 0.0
 steps: 5
-tools: [find_files, get_codebase_overview, get_dependencies, get_file_analysis, get_framework_analysis, get_semantic_neighborhoods, get_symbol_info, grep, list_dir, search_symbols, view_file, watch_changes, request_read_access, view_truncated_output, ask_user_input, git_status, get_blast_radius, get_hot_files, get_middleware, get_routes]
+tools: [find_files, get_blast_radius, get_codebase_overview, get_dependencies, get_file_analysis, get_framework_analysis, get_hot_files, get_middleware, get_routes, get_semantic_neighborhoods, get_symbol_info, git_status, grep, list_dir, request_read_access, search_symbols, view_file, view_truncated_output, watch_changes]
 description: Proposes refactors for clarity, deduplication, or decoupling. Read-only — outputs plans, not edits.
 ---
 You propose refactors. You do not apply them. The user applies via OpenCode or Claude Code.
@@ -125,7 +125,7 @@ top_k: 20
 min_p: 0.0
 presence_penalty: 1.5
 steps: 20
-tools: [find_files, get_codebase_overview, get_dependencies, get_file_analysis, get_framework_analysis, get_semantic_neighborhoods, get_symbol_info, grep, list_dir, search_symbols, view_file, watch_changes, request_read_access, view_truncated_output, ask_user_input, git_status, get_blast_radius, get_hot_files, get_middleware, get_routes]
+tools: [find_files, get_blast_radius, get_codebase_overview, get_dependencies, get_file_analysis, get_framework_analysis, get_hot_files, get_middleware, get_routes, get_semantic_neighborhoods, get_symbol_info, git_status, grep, list_dir, request_read_access, search_symbols, view_file, view_truncated_output, watch_changes, web_fetch, web_search]
 description: Designs new features, modules, or architectural changes. Outputs a build plan.
 ---
 You design. You produce build plans, not code.
@@ -167,7 +167,7 @@ top_p: 0.95
 top_k: 20
 min_p: 0.0
 presence_penalty: 0.0
-tools: [find_files, get_codebase_overview, get_dependencies, get_file_analysis, get_framework_analysis, get_semantic_neighborhoods, get_symbol_info, grep, list_dir, search_symbols, view_file, watch_changes, request_read_access, view_truncated_output, ask_user_input, git_status, get_blast_radius, get_hot_files, get_middleware, get_routes]
+tools: [find_files, get_codebase_overview, get_dependencies, get_file_analysis, grep, list_dir, request_read_access, search_symbols, view_file, view_truncated_output]
 description: Audits code for security vulnerabilities. Read-only.
 ---
 You audit for security issues. Concrete findings only, no generic warnings.
@@ -212,7 +212,7 @@ top_p: 0.95
 top_k: 20
 min_p: 0.0
 presence_penalty: 0.0
-tools: [view_file, list_dir, grep, find_files]
+tools: [find_files, get_codebase_overview, grep, list_dir, view_file]
 description: Builds prompts for OpenCode, Claude Code, or BooCode dispatch.
 ---
 You write prompts that another coding agent will execute. Your output is the prompt, not the work.
@@ -250,7 +250,7 @@ top_p: 0.95
 top_k: 20
 min_p: 0.0
 presence_penalty: 0.0
-tools: [find_files, get_codebase_overview, get_dependencies, get_file_analysis, get_framework_analysis, get_semantic_neighborhoods, get_symbol_info, grep, list_dir, search_symbols, view_file, watch_changes, request_read_access, view_truncated_output, ask_user_input, git_status, get_blast_radius, get_hot_files, get_middleware, get_routes]
+tools: [find_files, get_blast_radius, get_codebase_overview, get_dependencies, get_file_analysis, get_framework_analysis, get_hot_files, get_middleware, get_routes, get_semantic_neighborhoods, get_symbol_info, grep, list_dir, request_read_access, search_symbols, view_file, view_truncated_output, watch_changes]
 description: Discovers and maps unfamiliar codebases. Reads architecture, traces data flow, identifies key symbols.
 ---
 You map codebases. Start broad, then drill into specifics.
@@ -278,7 +278,7 @@ top_k: 20
 min_p: 0.0
 presence_penalty: 0.0
 steps: 10
-tools: [find_files, get_codebase_overview, get_dependencies, get_file_analysis, get_framework_analysis, get_semantic_neighborhoods, get_symbol_info, grep, list_dir, search_symbols, view_file, watch_changes, request_read_access, view_truncated_output, ask_user_input, git_status, get_blast_radius, get_hot_files, get_middleware, get_routes]
+tools: [ask_user_input, find_files, get_blast_radius, get_codebase_overview, get_dependencies, get_file_analysis, get_framework_analysis, get_hot_files, get_middleware, get_routes, get_semantic_neighborhoods, get_symbol_info, git_status, grep, list_dir, request_read_access, search_symbols, view_file, watch_changes]
 description: Produces actionable step plans from requirements. Read-only — never modifies files.
 ---
 You produce actionable step plans. You do not modify files.

data/skills/boocode/self-healing/SKILL.md

@@ -0,0 +1,287 @@
+---
+name: self-healing
+description: "Active runtime recovery for coding agents: when something breaks mid-task, diagnose the root cause, write a fix, VERIFY by re-running the broken thing, then file a `HEAL-` entry to `.learnings/HEALS.md` with proof. Use whenever a command, test, build, or lint fails or exits non-zero; on missing tooling, dependency/lockfile mismatch, wrong runtime version, venv or permission errors, port conflicts, dirty git state, or a missing `.env`; when the agent needs a helper or one-off script that doesn't exist yet; when an external API, tool, or MCP errors or rate-limits; or when a test flakes. Search `HEALS.md` by `Pattern-Key` first — most heals are recurrences, so increment `Recurrence-Count` instead of duplicating. Verify is mandatory: mark `pending-verify` honestly if sandboxed, `abandoned` if the fix can't be made to work. Pairs with `self-improvement` (which promotes recurring heals to durable memory) but owns the verify-before-persist discipline self-improvement doesn't."
+---
+
+# Self-Healing
+
+Active runtime recovery for coding agents. When something breaks, run the loop: **diagnose → patch → verify → file**. Leave behind a reusable, verified artifact instead of a swept-under-the-rug failure.
+
+The premise mirrors [browser-use/browser-harness](https://github.com/browser-use/browser-harness): *the harness improves itself every run*. An agent that hits a gap doesn't fail — it writes the fix during execution, verifies it works, and files the durable artifact for future runs. Coding tasks deserve the same loop.
+
+## What this skill is for
+
+When a coding agent hits a wall mid-task, the default failure modes are:
+
+1. **Paper over it** — "let me try a different approach" — and lose the recovery
+2. **Pretend the fix worked** — without re-running the broken thing
+3. **Symptom-fix** — skip the test, swallow the error, retry until green
+
+All three turn a one-time failure into a recurrence. The next agent on the same project hits the same wall.
+
+This skill enforces one discipline: **verify before persist**. A patch isn't real until you've re-run the failing operation and watched it succeed. When it does, file the verified fix so the next run benefits.
+
+## Relationship to self-improvement
+
+These two skills are deliberately split. Run both — they feed each other but don't overlap.
+
+| Aspect      | `self-healing` (this skill)                                          | `self-improvement`                                            |
+| ----------- | -------------------------------------------------------------------- | ------------------------------------------------------------- |
+| **When**    | During execution, failure is live                                    | After the fact, at natural breakpoints                        |
+| **Verb**    | Heal now — restore working state                                     | Remember for later — accumulate knowledge                     |
+| **Outcome** | Verified patch + (optional) reusable artifact                        | Logged learning, correction, request                          |
+| **Verify**  | **Mandatory** — no persist without proof                             | Not required                                                  |
+| **Files**   | `.learnings/HEALS.md` + `.learnings/heals/<HEAL-ID>/` (lazy)         | `.learnings/ERRORS.md`, `LEARNINGS.md`, `FEATURE_REQUESTS.md` |
+| **Trigger** | Failure observed mid-task                                            | Correction, knowledge gap, feature request, recurrence        |
+
+**Boundary rule:** if you're capturing a fact, a correction, or a wish — that's `self-improvement`. If you're applying and verifying a fix to a live failure — that's `self-healing`.
+
+## The Heal Loop
+
+```
+  ● failure observed
+  │
+  ● 1. DIAGNOSE  capture context — command, error, env, what was attempted
+  │              search HEALS.md for the same Pattern-Key first
+  │              (most heals are recurrences; don't reinvent)
+  │
+  ● 2. PATCH     write the fix — script, helper, env tweak, alt command
+  │              artifacts → .learnings/heals/<HEAL-ID>/  (only if needed)
+  │
+  ● 3. VERIFY    re-run the failing op — must succeed
+  │              ↻ if still failing: refine and retry, cap at 3 attempts
+  │              ✗ if uncrackable: file Status: abandoned with notes
+  │
+  ● 4. FILE      write HEAL-YYYYMMDD-XXX to .learnings/HEALS.md
+  │              with Pattern-Key, status, verification proof
+  │
+  ✓ working state restored, heal persisted
+
+  (conditional) PROMOTE  if Pattern-Key recurrence ≥ 3 across distinct tasks,
+                           append a Handoff block → self-improvement promotes to memory
+```
+
+If you abandon a heal mid-loop, don't pretend it succeeded. File a `HEAL-` entry with `Status: abandoned` and notes on what didn't work. The next agent learns from the dead end too.
+
+## When to trigger
+
+Self-healing fires on **active failures during execution** — the agent has just observed something not working and needs to make it work to continue. Five shapes:
+
+### 1. Tool failure (command / test / build / lint)
+Any invocation exits non-zero or produces wrong output. Don't acknowledge and retry verbatim — diagnose, patch, verify.
+
+*Examples:* `npm install` errors when a `pnpm-lock.yaml` is present (switch tool); `pytest` fails with `ModuleNotFoundError` (activate the venv); `tsc` flags a stale type (regenerate the client); `eslint` reports a config error (install the missing parser).
+
+### 2. Missing capability / tool gap
+The agent needs something that doesn't exist yet — a script, a helper, a wrapper, a glue function. Write it in the moment. This is the closest analog to browser-harness's `agent_helpers.py`.
+
+*Examples:* dedupe a CSV by custom key (write a small Python helper); bootstrap 12 microservices the same way (write `scripts/bootstrap-all.sh`); bulk-rename branches matching a pattern (write a `gh`-based shell helper).
+
+### 3. Environment issue
+The local environment isn't what the project expects. Detect, patch, verify.
+
+*Examples:* runtime version mismatch (`nvm use`, `pyenv local`, `rustup override`); stale dependency cache after a branch switch; dirty git state blocking a checkout; missing `.env` (copy from `.env.example` and surface gaps).
+
+### 4. External service / API change
+A service the agent depends on returns something unexpected. Find a workaround and capture it.
+
+*Examples:* an MCP tool returns `InputValidationError` because the schema changed (patch the call shape); a public API hits a rate limit (back off, switch endpoint, batch); an upstream lib bumped a default and broke a script (pin the version).
+
+### 5. About-to-retry-the-same-broken-approach
+The agent catches itself about to redo the failing step. That self-recognition is a heal forming — capture the alternate approach as the patch.
+
+### Detection signals to watch for
+
+- Non-zero exit codes
+- Stack traces in tool output
+- The same operation failing twice with the same error
+- "I'll try a different approach" — capture it as a heal
+- `command not found` / `module not found` / `permission denied`
+- Stale assertions, snapshot mismatches, type errors that weren't there before
+- "Weird" output that suggests environmental rather than logical bugs
+
+## HEAL Entry Format
+
+Append to `.learnings/HEALS.md` (create if missing):
+
+```markdown
+## [HEAL-YYYYMMDD-XXX] short_kebab_name
+
+**Logged**: ISO-8601 timestamp
+**Status**: verified | pending-verify | abandoned
+**Trigger**: tool-failure | missing-capability | env-issue | external-change | <free-form>
+**Active-Context**: (optional) — current skill, task phase, or workflow stage; omit if not applicable
+**Area**: free-form tag — what part of the system (`build`, `tests`, `ci`, `auth`, `data-pipeline`, `mobile`, ...)
+**Priority**: low | medium | high | critical
+
+### Failure
+What broke — concrete: the command, the error message, the action that was blocked. Include exit codes and verbatim error lines.
+
+### Diagnosis
+The root cause as understood after investigation. Why the obvious approach didn't work. Not a guess — what was actually verified during the heal.
+
+### Fix
+The patch that was applied. Verbatim commands, code snippets, or pointers to files under `.learnings/heals/<HEAL-ID>/`. Keep it minimal — just enough to reproduce.
+
+### Verification
+What was run after the fix and what it returned. Exit code, output snippet, test pass count. **This is the proof.** Without it, the entry is `pending-verify` or `abandoned`.
+
+### Artifacts
+(omit this section if no files were generated; otherwise list relative paths under `.learnings/heals/<HEAL-ID>/`)
+
+### Metadata
+- Related Files: path/to/file.ext
+- See Also: HEAL-... | LRN-... | ERR-... (related entries)
+- Pattern-Key: lower.snake.case key for recurrence detection (e.g. `env.lockfile_mismatch`)
+- Recurrence-Count: 1
+- First-Seen / Last-Seen: YYYY-MM-DD
+
+---
+```
+
+### Field guidance
+
+- **Status** — `verified` = the verify step passed. `pending-verify` = patch applied but couldn't be fully proven (sandboxed/offline/CI-only) — surface to the user. `abandoned` = patch didn't work or diagnosis was wrong — document what was tried.
+- **Trigger** — free-form is fine. The listed values are common shapes; what matters is that the failure shape is described enough for future agents to match against.
+- **Active-Context** — optional. Use it if your environment has a meaningful "what was I doing" tag (an active skill, a current task phase, a build stage, an agent role). Skip if not applicable. The browser-harness analog is the per-domain scoping of `domain-skills/<site>/`.
+- **Area** — free-form. Pick whatever helps future agents find this. `frontend`, `data-pipeline`, `ci`, `auth`, `terraform`, `mobile`, `embedded` — anything that fits your project shape.
+- **Pattern-Key** — lower.snake.case, stable, reusable across projects. Two heals with the same key are recurrences. `env.lockfile_mismatch` is good; `fixed_thing_tuesday` isn't.
+
+## ID generation
+
+Format: `HEAL-YYYYMMDD-XXX`. `XXX` is sequential 3-digit or 3-char random alphanumeric. Examples: `HEAL-20260524-001`, `HEAL-20260524-A7B`.
+
+## Artifacts directory (lazy)
+
+Only create `.learnings/heals/<HEAL-ID>/` when the heal generated something worth preserving. One-line fixes don't need a folder; the HEAL entry text is enough. Abandoned heals with no applied patch also skip the folder.
+
+```
+.learnings/
+├── HEALS.md
+├── ERRORS.md / LEARNINGS.md / FEATURE_REQUESTS.md  (self-improvement)
+└── heals/
+    └── HEAL-20260524-001/
+        ├── helper.sh
+        ├── patch.diff
+        └── notes.md
+```
+
+**Put here:** generated scripts/helpers, patch files, supplementary notes, output captures that document the diagnosis.
+**Don't put here:** project source changes (those go in the project tree, referenced via Related Files); secrets; output already captured in the HEAL text.
+
+## Verification rules
+
+Verify is the load-bearing wall. The whole point of self-healing over self-improvement is that the fix is *proven*, not theorized.
+
+### What counts as proof
+
+| Failure shape                         | Verification                                                       |
+| ------------------------------------- | ------------------------------------------------------------------ |
+| Tool / command / test / build / lint  | Re-run the original invocation; expect exit 0 / pass               |
+| Missing capability                    | Invoke the helper end-to-end on a real input; expect the intent    |
+| Environment drift                     | Re-run the operation that triggered the diagnosis                  |
+| External service workaround           | Re-run the failed call with the patch; expect a usable response    |
+
+### Sandboxed / offline / CI-only failures
+
+When you genuinely can't run the verify step (no network, no real remote, sandboxed shell, CI-only reproduction), file `Status: pending-verify` with:
+
+- The exact command the user / CI should run
+- The acceptance criteria — what counts as proof
+- A simulated proof if you can construct one (e.g. a dry-run mode, a stub of the failing call, a sandbox script)
+
+`pending-verify` is honest. Faking `verified` is the failure mode this skill exists to prevent.
+
+### When to invest in a proof script
+
+Most heals don't need a separate proof script — the verify step is just re-running the failing thing. Build a proper proof script when:
+
+- The heal generates a reusable helper that needs to be exercised across cases
+- The failure can't be reproduced live but can be reproduced in a sandbox (clean git repo, mock service, fake input)
+- You expect the heal to be re-applied across projects — the proof script then doubles as a regression check
+
+### If verification fails
+
+1. **Once** — refine the patch and retry. First diagnosis is often wrong.
+2. **Twice** — step back and reconsider the diagnosis. Maybe the root cause is elsewhere.
+3. **Three times** — stop. File `Status: abandoned` with notes on what you tried. Surface to the user. Don't flail.
+
+### What does NOT count as verification
+
+- "It looks right" / "I think this should work"
+- Re-running a *different* command than the one that originally failed
+- Suppressing the failure (`|| true`, `--ignore-errors`) — that's hiding
+- Skipping or deleting the failing test — that's regression
+- Passing because the cache was warm from before the fix
+
+### Reversibility
+
+Prefer reversible patches. If your heal modifies project files, capture the diff in `patch.diff`. If the heal is destructive (deletes generated files, rewrites locks), note it explicitly — a future agent reading the HEAL needs to know what was destroyed.
+
+## Recurrence and promotion
+
+Most heals are recurrences. Before filing a new HEAL, search:
+
+```bash
+grep -n "Pattern-Key: <your-pattern-key>" .learnings/HEALS.md
+```
+
+If found:
+
+- Increment `Recurrence-Count`
+- Update `Last-Seen`
+- Add the current occurrence as a See Also link
+- **Do not** create a duplicate entry
+
+### Promotion threshold
+
+Add a `Handoff` block to an existing entry when **all** are true:
+
+- `Recurrence-Count >= 3`
+- Seen across at least 2 distinct tasks
+- The fix is generalizable (not project-specific in a way that's already in a memory file)
+
+```markdown
+### Handoff
+- **Promoted To**: self-improvement at YYYY-MM-DD
+- **Promotion Target**: CLAUDE.md | AGENTS.md | .github/copilot-instructions.md | new-skill
+- **Distilled Rule**: One-line prevention guidance derived from the heal
+```
+
+Then `self-improvement` (or a learning aggregator) takes over: distills the rule, writes it into the right context file, or extracts a reusable skill. The HEAL stays for traceability.
+
+## Anti-patterns
+
+1. **Logging without verifying.** A HEAL filed before the fix is proven turns this into noisier self-improvement. If verify hasn't passed, the entry is `pending-verify` or `abandoned`.
+2. **Healing the symptom, not the cause.** A failing test isn't healed by skipping it (`pytest.skip`, `it.skip`, `xit`). A flaky CI isn't healed by `--retry`. Find the root cause; if you can't, abandon honestly.
+3. **Generating a new fix without trying existing ones first.** Search `HEALS.md` by Pattern-Key. Most heals are recurrences.
+4. **Inventing helpers when the project already has them.** Look in `scripts/`, `Makefile`, `justfile`, `package.json`, `pyproject.toml` first. Heal = write what's missing, not what's there.
+5. **Scope creep.** A heal is scoped to one failure. Cleanup belongs in a quality pass; refactors are features. Scope creep makes heals unreviewable.
+6. **Empty artifact folders.** Don't create `.learnings/heals/<HEAL-ID>/` if nothing goes in it.
+
+## Best practices
+
+1. **Heal eagerly, file always.** Even abandoned heals teach the next agent what doesn't work.
+2. **Verify before persist.** The non-negotiable rule.
+3. **Minimal and reversible patches.** A 3-line fix is a heal; a 300-line refactor is a feature.
+4. **Stable Pattern-Keys.** `env.node_version_mismatch` is reusable; `fixed_the_thing_on_tuesday` isn't.
+5. **Reference, don't duplicate.** Cross-link related HEAL/LRN/ERR via See Also.
+6. **Hand off recurrences.** A heal seen 3 times deserves to be in the project's permanent memory.
+7. **Don't gate the main tree on heal artifacts.** Files under `.learnings/heals/` are reference material; if a script becomes load-bearing, promote it to `scripts/`.
+
+## Setup
+
+```bash
+mkdir -p .learnings        # heals/ is lazy — created only when artifacts exist
+touch .learnings/HEALS.md
+```
+
+Gitignore choices match `self-improvement`. Keep heals local (`.learnings/` in `.gitignore`) or share them as team knowledge (don't gitignore — they become reviewable durable context).
+
+## Multi-agent use
+
+The skill is agent-agnostic. The `.learnings/HEALS.md` format is plain markdown — any agent (Claude Code, BooCode agents, OpenCode, Copilot, Cursor, Aider, ...) can read and write it.
+
+## See also
+
+- [`references/examples.md`](references/examples.md) — canonical HEAL entry shapes (command failure, missing capability, env drift, external API workaround, abandoned heal)

data/skills/boocode/self-healing/eval.yaml

@@ -0,0 +1,35 @@
+skill: self-healing
+tasks:
+  - prompt: "I'm in a project root that has pnpm-lock.yaml present but no package-lock.json. I just tried to run `npm install` and it failed. Get me to a working state so I can keep working — I have other things to do, just unblock me. After fixing it, make sure future agents in this project know what happened."
+    grader:
+      - the response invokes the self-healing skill
+      - the response diagnoses pnpm vs npm mismatch as the root cause
+      - the response runs pnpm install successfully
+      - the response files a HEAL entry to .learnings/HEALS.md with Status: verified
+      - the HEAL entry has Trigger: tool-failure
+      - the HEAL entry has a Pattern-Key resembling env.lockfile_mismatch
+      - the HEAL entry includes the verification output
+  - prompt: "I need to bulk-rename 8 git branches in this repo from `feat-XXX-name` to `feat/XXX-name`. There's no existing script for this and `gh` doesn't have a bulk-rename. Write what's needed, prove it works on a dry run, and capture the work so it's not lost if I need it again."
+    grader:
+      - the response invokes the self-healing skill
+      - the response recognizes this as a missing-capability heal
+      - the response writes a helper script under .learnings/heals/HEAL-<date>-<seq>/
+      - the response runs a dry-run verification
+      - the response files a HEAL entry with Status: verified and Trigger: missing-capability
+      - the HEAL entry references the helper script in the Artifacts section
+  - prompt: "I just ran `pytest` and got `ModuleNotFoundError: No module named 'pydantic'`. There's already a `.learnings/HEALS.md` in this project with a prior heal for a similar venv-not-activated issue. Fix this, and do the right thing with the heal records."
+    grader:
+      - the response invokes the self-healing skill
+      - the response searches HEALS.md first (using find-similar-heals.sh or grep) before writing a new fix
+      - the response finds the existing HEAL entry and applies its fix (activate venv)
+      - the response increments Recurrence-Count on the existing entry
+      - the response updates Last-Seen on the existing entry
+      - the response does NOT create a duplicate HEAL entry
+  - prompt: "A test in this repo is failing intermittently — the snapshot for `Card.test.tsx` flakes. I've already tried fixing it once by stubbing the date; it passes twice then flakes again because there's a UUID that's also non-deterministic. I don't have time to refactor the Card component to inject dependencies. Just do the right thing — get me to a state that's honest about what's known and not known, and don't pretend the heal worked."
+    grader:
+      - the response invokes the self-healing skill
+      - the response diagnoses that the initial patch attempt was incomplete
+      - the response files a HEAL entry with Status: abandoned
+      - the HEAL entry documents what was tried and why it failed
+      - the response does NOT mark anything as verified
+      - the response surfaces the situation honestly to the user

data/skills/boocode/self-healing/references/examples.md

@@ -0,0 +1,248 @@
+# Self-Healing Examples
+
+Concrete HEAL entries showing the format applied to real failure shapes. Use these as templates when filing your own heals. All examples use the iteration-2 schema (free-form `Trigger` / `Area`, optional `Active-Context`, no `Source` field, lazy artifact folders).
+
+---
+
+## Example 1 — Tool failure (lockfile mismatch)
+
+```markdown
+## [HEAL-20260524-001] npm_install_pnpm_lockfile
+
+**Logged**: 2026-05-24T14:22:01Z
+**Status**: verified
+**Trigger**: tool-failure
+**Area**: build
+**Priority**: medium
+
+### Failure
+`npm install` exited 1 with `npm ERR! code EUSAGE` and a notice that `pnpm-lock.yaml` is present but `package-lock.json` is missing. The project uses pnpm workspaces; npm refuses to install against a pnpm lockfile.
+
+### Diagnosis
+Project root contains `pnpm-lock.yaml`. The README and CI both invoke `pnpm`. `npm` was a habit from previous projects, not the actual project's package manager.
+
+### Fix
+Use pnpm instead:
+```bash
+pnpm install
+```
+
+### Verification
+```
+$ pnpm install
+Lockfile is up to date, resolution step is skipped
+Already up to date
+✓ Done in 1.4s
+```
+Exit 0.
+
+### Metadata
+- Related Files: package.json, pnpm-lock.yaml
+- See Also: (none yet)
+- Pattern-Key: env.lockfile_mismatch
+- Recurrence-Count: 1
+- First-Seen: 2026-05-24
+- Last-Seen: 2026-05-24
+
+---
+```
+
+Pattern-Key `env.lockfile_mismatch` is reusable across projects (yarn.lock, bun.lockb, etc.). At Recurrence ≥ 3, this should be promoted to `CLAUDE.md` or `AGENTS.md` as a verification step.
+
+No Artifacts section — the fix is a tool swap, no files generated. Lazy folder pattern: nothing to put in `.learnings/heals/HEAL-20260524-001/`, so the folder isn't created.
+
+---
+
+## Example 2 — Missing capability (helper written on the fly)
+
+```markdown
+## [HEAL-20260524-002] bulk_rename_branches_helper
+
+**Logged**: 2026-05-24T15:10:44Z
+**Status**: verified
+**Trigger**: missing-capability
+**Area**: ci
+**Priority**: low
+
+### Failure
+Need to rename 12 feature branches from `feat-XXX-name` to `feat/XXX-name`. No existing project script handles this; `gh` doesn't have a bulk-rename primitive.
+
+### Diagnosis
+This is glue work, not a project bug. A small shell helper using `gh api` per branch is the right level — not worth a top-level script, but worth keeping the file for the next time someone asks.
+
+### Fix
+Wrote `.learnings/heals/HEAL-20260524-002/rename-branches.sh`:
+
+```bash
+#!/usr/bin/env bash
+set -euo pipefail
+git fetch --all
+for branch in $(git branch -r | grep 'origin/feat-' | sed 's|origin/||'); do
+  new="${branch/feat-/feat/}"
+  echo "$branch → $new"
+  gh api -X POST "repos/{owner}/{repo}/git/refs" \
+    -f "ref=refs/heads/$new" \
+    -f "sha=$(git rev-parse "origin/$branch")"
+  gh api -X DELETE "repos/{owner}/{repo}/git/refs/heads/$branch"
+done
+```
+
+### Verification
+Dry-run (commented out the API calls) printed the 12 expected mappings.
+Live run renamed all 12; `git branch -r | grep 'feat-' | wc -l` returns 0.
+
+### Artifacts
+- `.learnings/heals/HEAL-20260524-002/rename-branches.sh`
+
+### Metadata
+- Related Files: (none — operates on git refs)
+- See Also: (none)
+- Pattern-Key: tool.gh.bulk_branch_rename
+- Recurrence-Count: 1
+- First-Seen: 2026-05-24
+- Last-Seen: 2026-05-24
+
+---
+```
+
+Helper script lives under `.learnings/heals/<HEAL-ID>/` — referenceable, but not assumed to be load-bearing. If it gets reused frequently, promote to `scripts/`.
+
+---
+
+## Example 3 — Environment issue (runtime version)
+
+```markdown
+## [HEAL-20260524-003] nvm_use_project_node
+
+**Logged**: 2026-05-24T16:01:12Z
+**Status**: verified
+**Trigger**: env-issue
+**Active-Context**: verify-gate
+**Area**: tests
+**Priority**: medium
+
+### Failure
+`pnpm test` exited 1 with `engine "node" is incompatible with this module. Expected version "^20.10.0". Got "18.19.0"`.
+
+### Diagnosis
+`.nvmrc` requests node 20.10.0; current shell has 18.19.0 from a previous project context. The shell's nvm wasn't switched after `cd`-ing into the repo.
+
+### Fix
+```bash
+nvm use   # reads .nvmrc
+```
+
+### Verification
+```
+$ node --version
+v20.10.0
+$ pnpm test
+✓ 47 tests passed
+```
+
+### Metadata
+- Related Files: .nvmrc, package.json
+- See Also: (none)
+- Pattern-Key: env.node_version_mismatch
+- Recurrence-Count: 1
+- First-Seen: 2026-05-24
+- Last-Seen: 2026-05-24
+
+---
+```
+
+`Active-Context: verify-gate` because that's the workflow phase the agent was in when the test step blew up. An upstream context loader could surface this entry next time `verify-gate` runs in a node project. If you don't have an analogous concept in your pipeline, omit the field.
+
+---
+
+## Example 4 — External service workaround
+
+```markdown
+## [HEAL-20260524-004] gh_api_rate_limit_backoff
+
+**Logged**: 2026-05-24T17:33:08Z
+**Status**: verified
+**Trigger**: external-change
+**Area**: ci
+**Priority**: high
+
+### Failure
+Looping `gh api repos/.../issues` over 200 issues started returning `403 rate limit exceeded` after ~60 calls. Unauthenticated burst limit (abuse detection on rapid successive calls).
+
+### Diagnosis
+Script was using `gh api` REST without batching. `gh` is authenticated but the secondary rate limit fires on rapid successive calls — not the primary 5000/hour limit. Switching to a single paginated GraphQL query bypasses the secondary limit entirely.
+
+### Fix
+```bash
+gh api graphql -f query='
+  query($owner:String!,$repo:String!,$cursor:String) {
+    repository(owner:$owner,name:$repo) {
+      issues(first:100,after:$cursor) { ... }
+    }
+  }' -F owner=... -F repo=...
+```
+Took ~3 calls total instead of 200.
+
+### Verification
+Full run completed in 4.8s, no 403s, all 200 issues retrieved. Compared output against a sample of the original per-issue calls — fields match.
+
+### Artifacts
+- `.learnings/heals/HEAL-20260524-004/fetch-issues.sh`
+
+### Metadata
+- Related Files: (none — ad-hoc query)
+- See Also: (none)
+- Pattern-Key: api.gh.rate_limit
+- Recurrence-Count: 1
+- First-Seen: 2026-05-24
+- Last-Seen: 2026-05-24
+
+---
+```
+
+---
+
+## Example 5 — Abandoned heal (diagnosis was wrong)
+
+```markdown
+## [HEAL-20260524-005] vitest_flaky_snapshot
+
+**Logged**: 2026-05-24T18:14:22Z
+**Status**: abandoned
+**Trigger**: tool-failure
+**Active-Context**: verify-gate
+**Area**: tests
+**Priority**: medium
+
+### Failure
+`vitest` snapshot test `Card > renders default` flaked twice in three runs. Diff showed a timestamp string differing by ~3 seconds.
+
+### Diagnosis (initial — wrong)
+Assumed flake was timezone drift in the snapshot fixture. Patched the fixture to use a fixed `Date.now()` stub.
+
+### Diagnosis (current — correct)
+The snapshot depends on multiple non-deterministic values: timestamp AND a `crypto.randomUUID()`. The clock stub addressed only one of them. The UUID is still random per render, so the snapshot keeps drifting on subsequent runs.
+
+### Fix (attempted)
+Added `vi.useFakeTimers({ now: 1700000000000 })` to the test setup.
+
+### Verification
+Test passed twice, then flaked again on the third run — same `Card > renders default`, different diff (this time the UUID changed). Original diagnosis was incomplete.
+
+### Abandonment notes
+The right fix is to make the component deterministic via dependency injection (pass a `clock` and `idGen` prop), not to stub globally. That's a real change to the component contract — out of scope for a heal. Filed `FEAT-20260524-001` via self-improvement; surfaced to the user.
+
+### Metadata
+- Related Files: src/components/Card.tsx, src/components/Card.test.tsx
+- See Also: FEAT-20260524-001
+- Pattern-Key: tests.flaky_snapshot_multi_nondeterminism
+- Recurrence-Count: 1
+- First-Seen: 2026-05-24
+- Last-Seen: 2026-05-24
+
+---
+```
+
+Abandoned heals are first-class. They document a dead end so the next agent doesn't re-walk it. The handoff to a `FEAT-` entry via self-improvement is the right next step when the real fix is a feature, not a heal.
+
+No Artifacts section — the attempted patch was reverted; nothing reusable was generated.

data/skills/boocode/self-healing/scripts/detect-failure.sh

@@ -0,0 +1,54 @@
+#!/usr/bin/env bash
+# detect-failure.sh — PostToolUse hook for Bash invocations.
+# Reads the tool result JSON on stdin (per Claude Code hook spec); if exit_code != 0,
+# emits a system reminder pointing the agent at self-healing.
+#
+# Wire up in .claude/settings.json:
+#   "hooks": {
+#     "PostToolUse": [{ "matcher": "Bash",
+#       "hooks": [{ "type": "command",
+#         "command": "./data/skills/boocode/self-healing/scripts/detect-failure.sh" }] }]
+#   }
+
+set -euo pipefail
+
+# Hook payload arrives on stdin. We tolerate either jq-style JSON or raw text.
+PAYLOAD="$(cat || true)"
+
+# Try to parse exit_code; fall through silently on parse failure.
+EXIT_CODE=$(printf '%s' "$PAYLOAD" | python3 -c '
+import json, sys
+try:
+    data = json.loads(sys.stdin.read() or "{}")
+    # Common shapes: {"tool_result": {"exit_code": N}}, {"exit_code": N}, {"output": "...", "exit_code": N}
+    for path in (("tool_result","exit_code"), ("exit_code",), ("result","exit_code")):
+        d = data
+        ok = True
+        for k in path:
+            if isinstance(d, dict) and k in d:
+                d = d[k]
+            else:
+                ok = False
+                break
+        if ok and isinstance(d, int):
+            print(d)
+            sys.exit(0)
+except Exception:
+    pass
+print(0)
+' 2>/dev/null || echo 0)
+
+if [[ "$EXIT_CODE" != "0" ]]; then
+  cat <<'EOF'
+<self-healing-trigger>
+A Bash command just exited non-zero. This is a heal opportunity.
+
+Before retrying the same command verbatim:
+  1. DIAGNOSE — read the error; identify the root cause (env? missing dep? wrong tool?)
+  2. Search .learnings/HEALS.md for a matching Pattern-Key (don't re-solve a solved problem)
+  3. PATCH — write the fix (or apply a known one)
+  4. VERIFY — re-run the command; require exit 0
+  5. FILE — append a HEAL entry to .learnings/HEALS.md via data/skills/boocode/self-healing/scripts/new-heal.sh
+</self-healing-trigger>
+EOF
+fi

data/skills/boocode/self-healing/scripts/find-similar-heals.sh

@@ -0,0 +1,52 @@
+#!/usr/bin/env bash
+# find-similar-heals.sh — Search existing heals before generating a new fix.
+# Usage: ./find-similar-heals.sh <pattern-key-or-keyword>
+#
+# Prints matching HEAL entries with their Pattern-Key, Status, and Recurrence-Count
+# so the agent can decide whether to re-apply an existing fix or write a new one.
+
+set -euo pipefail
+
+QUERY="${1:-}"
+HEALS_FILE="$(pwd)/.learnings/HEALS.md"
+
+if [[ -z "$QUERY" ]]; then
+  echo "usage: $0 <pattern-key-or-keyword>" >&2
+  exit 2
+fi
+
+if [[ ! -f "$HEALS_FILE" ]]; then
+  echo "(no .learnings/HEALS.md yet — no prior heals to consult)"
+  exit 0
+fi
+
+# Find HEAL section headers that contain the query in their body (Pattern-Key, name, or text).
+python3 - <<PY "$QUERY" "$HEALS_FILE"
+import sys, re
+query, path = sys.argv[1].lower(), sys.argv[2]
+with open(path) as f:
+    text = f.read()
+# Split into entries by ^## [HEAL-...]
+entries = re.split(r"(?m)^## \[HEAL-", text)[1:]
+hits = []
+for body in entries:
+    if query in body.lower():
+        head = body.splitlines()[0]
+        pk = re.search(r"Pattern-Key:\s*(\S+)", body)
+        status = re.search(r"Status\*\*:\s*(\S+)", body) or re.search(r"Status:\s*(\S+)", body)
+        rc = re.search(r"Recurrence-Count:\s*(\d+)", body)
+        hits.append({
+            "id": "HEAL-" + head.split("]")[0],
+            "name": head.split("]", 1)[1].strip() if "]" in head else head,
+            "pattern_key": pk.group(1) if pk else "?",
+            "status": status.group(1) if status else "?",
+            "recurrence": rc.group(1) if rc else "1",
+        })
+if not hits:
+    print(f"(no heals match '{query}')")
+else:
+    print(f"Found {len(hits)} matching heal(s):\n")
+    for h in hits:
+        print(f"  {h['id']} {h['name']}")
+        print(f"    pattern={h['pattern_key']}  status={h['status']}  recurrence={h['recurrence']}")
+PY

data/skills/boocode/self-healing/scripts/new-heal.sh

@@ -0,0 +1,74 @@
+#!/usr/bin/env bash
+# new-heal.sh — Initialize a new HEAL-<date>-<seq> entry skeleton.
+# Usage: ./new-heal.sh <short_kebab_name> [trigger]
+#   trigger: tool-failure | missing-capability | env-issue | external-change | <free-form>
+#
+# Appends a templated HEAL entry to .learnings/HEALS.md and prints the HEAL-ID.
+# Does NOT create .learnings/heals/<HEAL-ID>/ — that folder is lazy, created
+# only when artifacts are written.
+
+set -euo pipefail
+
+NAME="${1:-}"
+TRIGGER="${2:-tool-failure}"
+
+if [[ -z "$NAME" ]]; then
+  echo "usage: $0 <short_kebab_name> [trigger]" >&2
+  exit 2
+fi
+
+LEARNINGS_DIR="$(pwd)/.learnings"
+HEALS_FILE="$LEARNINGS_DIR/HEALS.md"
+mkdir -p "$LEARNINGS_DIR"
+
+DATE="$(date +%Y%m%d)"
+SEQ=$(grep -c "^## \[HEAL-${DATE}-" "$HEALS_FILE" 2>/dev/null || echo 0)
+NEXT=$(printf "%03d" $((SEQ + 1)))
+HEAL_ID="HEAL-${DATE}-${NEXT}"
+
+# Active-Context is optional. The agent / harness can set ACTIVE_CONTEXT in env.
+ACTIVE_CONTEXT="${ACTIVE_CONTEXT:-}"
+ACTIVE_LINE=""
+if [[ -n "$ACTIVE_CONTEXT" ]]; then
+  ACTIVE_LINE="**Active-Context**: $ACTIVE_CONTEXT
+"
+fi
+
+cat >> "$HEALS_FILE" <<EOF
+
+## [$HEAL_ID] $NAME
+
+**Logged**: $(date -u +%Y-%m-%dT%H:%M:%SZ)
+**Status**: pending-verify
+**Trigger**: $TRIGGER
+${ACTIVE_LINE}**Area**: TODO
+**Priority**: medium
+
+### Failure
+TODO — concrete error, command, exit code
+
+### Diagnosis
+TODO — root cause after investigation
+
+### Fix
+TODO — patch applied (commands, snippets, or pointers to .learnings/heals/$HEAL_ID/ if files were generated)
+
+### Verification
+TODO — what was run after the fix, what it returned. **Update Status to "verified" only after this passes.**
+
+### Metadata
+- Related Files: TODO
+- See Also: TODO
+- Pattern-Key: TODO
+- Recurrence-Count: 1
+- First-Seen: $(date +%Y-%m-%d)
+- Last-Seen: $(date +%Y-%m-%d)
+
+---
+EOF
+
+# stdout = the HEAL-ID alone, so `ID=$(new-heal.sh ...)` captures it cleanly.
+# Human guidance goes to stderr.
+echo "$HEAL_ID"
+echo "$HEALS_FILE" >&2
+echo "(create .learnings/heals/$HEAL_ID/ only if you generate artifacts to put there)" >&2

data/skills/boocode/verify-gate/SKILL.md

@@ -0,0 +1,178 @@
+---
+name: verify-gate
+description: "Runs project compile, test, and lint commands between implementation and quality review. Gates simplify-and-harden behind machine verification. If checks fail, enters a fix loop with diagnostics. If checks pass, signals ready for quality pass. Use after any implementation work completes and before signaling done. Essential for the inner loop's verify step."
+---
+
+# Verify Gate
+
+Machine verification gate between implementation and quality review. Runs the project's compile, test, and lint commands. If any fail, enters a fix loop. If all pass, unblocks the quality pass.
+
+This is the inner loop's **verify** step. Without it, the agent hands off code with zero machine signal about whether it actually works.
+
+## When to Use
+
+- After any implementation work completes, before signaling "done"
+- Before running simplify-and-harden or quality review
+- After fixing audit findings from code review
+- Any time you want a machine-verified green signal
+
+## Pipeline Position
+
+```
+[implementation] → verify-gate → [quality review / simplify-and-harden]
+                   ↻ fix loop — on failure, diagnose and retry
+```
+
+## Step 1: Discover Project Commands
+
+Read the project's configuration to find verification commands. Check these sources in order:
+
+1. **Project instruction files** (`CLAUDE.md`, `data/AGENTS.md`) — look for a `## Verification` or `## Test Commands` section
+2. **package.json** — `scripts.test`, `scripts.lint`, `scripts.typecheck`, `scripts.build`. BooCode uses pnpm, so prefer `pnpm run <script>` when `pnpm-lock.yaml` is present.
+3. **Makefile** / **Justfile** — `test`, `lint`, `check`, `build` targets
+4. **Cargo.toml** — `cargo build`, `cargo test`, `cargo clippy`
+5. **pyproject.toml** / **setup.cfg** — `pytest`, `mypy`, `ruff`
+6. **go.mod** — `go build ./...`, `go test ./...`, `go vet ./...`
+7. **deno.json** / **deno.jsonc** — `deno task <name>` for any defined tasks
+
+If no commands are discoverable, ask the user once and suggest they add a `## Verification` section to `CLAUDE.md` for future sessions:
+
+```markdown
+## Verification
+
+- Build: `pnpm run build`
+- Test: `pnpm test`
+- Lint: `pnpm run lint`
+- Type check: `npx tsc -p apps/server/tsconfig.json --noEmit`
+```
+
+## Step 2: Run Verification
+
+Run discovered commands in this order. Stop at the first failure category.
+
+### Phase 1: Compile / Type Check
+Run the build or type-check command. These catch structural errors before wasting time on tests.
+
+```
+Exit 0 → proceed to Phase 2
+Exit non-zero → enter fix loop with compiler output
+```
+
+### Phase 2: Tests
+Run the test command. Scope to changed files if the test runner supports it.
+
+```
+Exit 0 → proceed to Phase 3
+Exit non-zero → enter fix loop with test output
+```
+
+### Phase 3: Lint (optional, skippable with --skip-lint)
+Run the lint command. Lint failures are lower severity but still worth catching.
+
+```
+Exit 0 → all phases green, gate passes
+Exit non-zero → enter fix loop with lint output
+```
+
+## Step 3: Fix Loop
+
+When a phase fails:
+
+1. **Read the output.** Parse the error output for actionable diagnostics — file paths, line numbers, error messages.
+2. **Scope the fix.** Only fix what the verification caught. Do not refactor, improve, or touch unrelated code.
+3. **Apply the fix.** Make the minimal change to resolve the failure.
+4. **Re-run the failed phase.** Not all phases — just the one that failed.
+5. **If it passes**, continue to the next phase.
+6. **If it fails again**, increment the attempt counter.
+
+### Fix Loop Limits
+
+- **Default max attempts:** 3 per phase (configurable via `--fix-limit N`)
+- **Counter increments on every attempt**, even if the error changes. Fixing Error A and uncovering Error B counts as attempt 2, not attempt 1. The counter tracks fix attempts, not unique errors.
+- **If limit reached:** Stop. Report what failed, what was tried, and the remaining error output. Do not guess further — signal to the user that manual intervention is needed.
+- **Total budget:** The fix loop should not exceed 20% of the original implementation effort. If fixes are snowballing, stop and report.
+
+## Step 4: Gate Signal
+
+When all phases pass:
+
+```markdown
+## Verify Gate: PASSED
+
+- Build: passed
+- Tests: passed (N tests, M suites)
+- Lint: passed (or skipped)
+
+Ready for quality review.
+```
+
+When the fix loop is exhausted:
+
+```markdown
+## Verify Gate: BLOCKED
+
+- Build: passed
+- Tests: FAILED (attempt 3/3)
+  - [file:line] error description
+  - [file:line] error description
+- Lint: not reached
+
+Fix loop exhausted. Manual intervention needed before quality review.
+```
+
+## Integration with Other Skills
+
+### boocode simplify-and-harden / quality review
+verify-gate should gate any quality pass. Run verify-gate first; only proceed to review if the gate passes.
+
+### self-healing (if available)
+On any failure during the verify run, consider handing the diagnostics to a self-healing loop (diagnose → patch → verify → persist). Verify-gate then re-runs the checks. Up to 3 heal attempts per phase before abandoning.
+
+### self-improvement
+If a recurring error pattern emerges across verify runs, capture it in `CLAUDE.md` or as a new skill under `data/skills/boocode/` so future verify-gate runs don't rediscover the same fix.
+
+## What This Skill Does NOT Do
+
+- Does not review code quality (that's a separate review pass)
+- Does not check security
+- Does not verify spec compliance
+- Does not modify test files or add new tests
+- Does not run tests for code it didn't change (unless the test runner doesn't support scoping)
+
+## Configuration
+
+If the project has a `verify-gate` section in `CLAUDE.md` or `data/AGENTS.md`:
+
+```yaml
+## Verify Gate Config
+
+build: pnpm run build
+test: pnpm test
+lint: pnpm run lint
+type_check: npx tsc -p apps/server/tsconfig.json --noEmit
+fix_limit: 3
+skip_lint: false
+test_scope: changed  # changed | all
+```
+
+If no configuration exists, discover commands automatically (Step 1) and suggest persisting them.
+
+### Custom Verification Steps
+
+Projects with custom invariants can define additional verification phases. These run as extra phases after the standard compile/test/lint checks.
+
+Example — a project that needs API schema validation:
+
+```yaml
+## Verify Gate Config
+
+custom_checks:
+  - name: validate-schema
+    command: python scripts/validate_schema.py --strict
+  - name: check-no-legacy-imports
+    command: grep -r "from legacy" src/ --include="*.py" && exit 1 || exit 0
+```
+
+When custom checks are defined, verify-gate runs them as **Phase 4** after lint. Each check's exit code determines pass/fail. Failed checks enter the same fix loop as standard phases.
+
+This moves project-specific invariants from "knowledge in your head" to "knowledge in the harness" — exactly where the agent can reach it.

packages/contracts/src/ws-frames.ts

@@ -116,6 +116,8 @@ export const MessageCompleteFrame = z.object({
   tokens_used: z.number().int().nonnegative().nullable().optional(),
   ctx_used: z.number().int().nonnegative().nullable().optional(),
   ctx_max: z.number().int().positive().nullable().optional(),
+  cache_tokens: z.number().int().nonnegative().nullable().optional(),
+  reasoning_tokens: z.number().int().nonnegative().nullable().optional(),
   started_at: IsoTimestamp.nullable().optional(),
   finished_at: IsoTimestamp.nullable().optional(),
   // nullable: external-coder turns carry task.model, which is null when no

scripts/omo-paseo-bridge.sh

@@ -0,0 +1,160 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# =============================================================================
+# omo-paseo-bridge.sh — Import OMO task() child sessions as Paseo agents
+#
+# Automates calling `paseo import` on child session IDs so OMO subagents
+# appear in `paseo ls` alongside native Paseo agents.
+#
+# Usage:
+#   omo-paseo-bridge.sh import [--type <category>] <session-id>...
+#       Import session(s) as Paseo agents with omo=true labels
+#
+#   omo-paseo-bridge.sh archive <agent-id>...
+#       Archive (soft-delete) agent(s) imported by this bridge
+#
+#   omo-paseo-bridge.sh ls [--all]
+#       List agents tagged omo=true via paseo ls
+#
+#   omo-paseo-bridge.sh --dry-run <command> ...
+#       Print what would be done without executing
+#
+# Examples:
+#   omo-paseo-bridge.sh import ses_abc123 ses_def456
+#   omo-paseo-bridge.sh import --type research ses_abc123
+#   omo-paseo-bridge.sh archive agt_789
+#   omo-paseo-bridge.sh ls
+#   omo-paseo-bridge.sh --dry-run import ses_abc123
+# =============================================================================
+
+SCRIPT_NAME="$(basename "$0")"
+PASEO="$(which paseo 2>/dev/null || echo "paseo")"
+DRY_RUN=false
+
+# ── helpers ──────────────────────────────────────────────────────────────────
+
+log()  { printf "[%s] %s\n" "$SCRIPT_NAME" "$*"; }
+warn() { printf "[%s] WARNING: %s\n" "$SCRIPT_NAME" "$*" >&2; }
+err()  { printf "[%s] ERROR: %s\n" "$SCRIPT_NAME" "$*" >&2; exit 1; }
+
+paseo_cmd() {
+  if $DRY_RUN; then
+    log "[DRY-RUN] would run: $PASEO $*"
+    return 0
+  fi
+  "$PASEO" "$@" 2>&1 || warn "'paseo $*' exited with code $?"
+}
+
+paseo_import() {
+  local session_id="$1"
+  shift
+  local type_label="${1:-}"
+  local labels=("--label" "omo=true")
+
+  # Add parent session label if OMO_SESSION_ID is set (injected by agent)
+  if [[ -n "${OMO_SESSION_ID:-}" ]]; then
+    labels+=("--label" "parent=${OMO_SESSION_ID}")
+  fi
+
+  if [[ -n "$type_label" ]]; then
+    labels+=("--label" "type=${type_label}")
+  fi
+
+  log "Importing session ${session_id} as Paseo agent ..."
+  paseo_cmd import "$session_id" --provider opencode "${labels[@]}"
+}
+
+paseo_archive() {
+  local agent_id="$1"
+  log "Archiving agent ${agent_id} ..."
+  paseo_cmd archive "$agent_id" --force
+}
+
+paseo_list() {
+  local all_flag="${1:-}"
+  if [[ "$all_flag" == "--all" ]]; then
+    paseo_cmd ls --label "omo=true" --all
+  else
+    paseo_cmd ls --label "omo=true"
+  fi
+}
+
+# ── usage ────────────────────────────────────────────────────────────────────
+
+usage() {
+  cat <<EOF
+Usage: $SCRIPT_NAME [--dry-run] <command> [options] [args...]
+
+Commands:
+  import [--type <category>] <session-id>...
+      Import OMO child session(s) as Paseo agents
+  archive <agent-id>...
+      Archive Paseo agent(s) (soft-delete)
+  ls [--all]
+      List agents tagged omo=true
+
+Options:
+  --dry-run   Print actions without executing them
+  -h, --help  Show this help
+
+Examples:
+  $SCRIPT_NAME import --type research ses_abc123
+  $SCRIPT_NAME archive agt_789
+  $SCRIPT_NAME ls --all
+EOF
+  exit 0
+}
+
+# ── main ─────────────────────────────────────────────────────────────────────
+
+# Peel off global flags
+while [[ $# -gt 0 ]]; do
+  case "$1" in
+    --dry-run) DRY_RUN=true; shift ;;
+    -h|--help) usage ;;
+    *) break ;;
+  esac
+done
+
+[[ $# -eq 0 ]] && usage
+
+COMMAND="$1"
+shift
+
+case "$COMMAND" in
+  import)
+    TYPE_LABEL=""
+    SESSION_IDS=()
+
+    while [[ $# -gt 0 ]]; do
+      case "$1" in
+        --type) TYPE_LABEL="$2"; shift 2 ;;
+        --type=*) TYPE_LABEL="${1#*=}"; shift ;;
+        -*) err "Unknown option for import: $1" ;;
+        *) SESSION_IDS+=("$1"); shift ;;
+      esac
+    done
+
+    [[ ${#SESSION_IDS[@]} -eq 0 ]] && err "import requires at least one session-id"
+
+    for sid in "${SESSION_IDS[@]}"; do
+      paseo_import "$sid" "$TYPE_LABEL"
+    done
+    ;;
+
+  archive)
+    [[ $# -eq 0 ]] && err "archive requires at least one agent-id"
+    for aid in "$@"; do
+      paseo_archive "$aid"
+    done
+    ;;
+
+  ls)
+    paseo_list "${1:-}"
+    ;;
+
+  *)
+    err "Unknown command: $COMMAND\n$(usage)"
+    ;;
+esac