osborn 0.9.167 → 0.9.169

package/dist/claude-llm.d.ts

@@ -79,6 +79,14 @@ export declare const NAMED_AGENTS: {
         prompt: string;
     };
 };
+export declare const FAST_MODEL = "haiku";
+/**
+ * Apply turbo override to an agents roster.
+ * When turbo is true returns a DEEP COPY with every agent's model set to
+ * FAST_MODEL. When false returns the input object unchanged (zero allocation).
+ * NEVER mutates the input — NAMED_AGENTS and DB-sourced rows are untouched.
+ */
+export declare function applyTurbo(agents: Record<string, any>, turbo: boolean): Record<string, any>;
 /**
  * Claude LLM - Wraps Claude Agent SDK for LiveKit
  * Research mode: reads anything, writes only to session workspace
@@ -133,6 +141,13 @@ export declare class ClaudeLLM extends llm.LLM {
      * it was created with.
      */
     setAgents(agents: ClaudeLLMOptions['agents']): void;
+    /**
+     * Enable/disable Turbo mode. When on, the main model and every sub-agent
+     * spawn on FAST_MODEL. Takes effect at the next query cold start.
+     */
+    setTurbo(on: boolean): void;
+    /** Read-only accessor used by ClaudeLLMStream (private fields are class-scoped). */
+    get turbo(): boolean;
     /**
      * Set session ID to resume a specific conversation
      * Call this before sending the first message to resume from a previous session

package/dist/claude-llm.js

@@ -341,16 +341,32 @@ export const NAMED_AGENTS = {
             '## Your role',
             'Execute test suites, build commands, and linters. Interpret failures clearly.',
             'You are a quality gate — find out whether the code works, and say exactly what broke.',
+            'You protect the USER: the product must stay predictable across releases. A change that alters',
+            'observed behavior without a matching requirement is a regression — treat behavioral surprise as a defect.',
             '',
             '## Grounding — consult shared context before writing or running tests',
-            'Before deciding what to test, locate the session index (search-index.txt under .claude/projects/<slug>/osb/<session>/; newest if several) and Grep it for the changes/work under test. Also check any project docs and known-issues files (e.g. CONVENTIONS.md, docs/, a known-issues or gotchas doc if present) for: (a) KNOWN ISSUES and gotchas already recorded, and (b) what behavior is ALREADY covered by existing tests.',
+            'Before deciding what to test, locate the session index (search-index.txt under .claude/projects/<slug>/osb/<session>/; newest if several) and Grep it for the changes/work under test. Also check project docs and known-issues files. Key doc locations to consult: `/workspace/osborn/CLAUDE.md`, `/workspace/osborn/docs/critical-patterns.md`, the `docs/` directory, `README.md`, and `CHANGELOG.md`. Check these for: (a) KNOWN ISSUES and gotchas already recorded, and (b) what behavior is ALREADY covered by existing tests.',
             'Purpose: target regression coverage at real GAPS and known-risk areas rather than testing blind or duplicating coverage — and stay IN SYNC with the reviewer, which reads the same sources.',
             'Read only the relevant slice of the index, never the whole file. If you find nothing or no index/docs exist, proceed normally — this is an optimization, not a hard dependency.',
+            'If the change adds or renames a feature, flag any doc now out of date (see DOC STALENESS in "What to return").',
             '',
             '## Backward-compatibility / regression mandate (CRITICAL)',
             'Existing test suites MUST still pass — any pre-existing test that breaks is a BLOCKER; report it as such.',
             'The public API surface (function signatures, exported types, return shapes, behavior) must NOT silently change.',
             'Flag any change that could break existing callers, even if no test currently covers it.',
+            'Coverage targets — where applicable, explicitly include:',
+            '  - Backward-compatibility tests: verify the old calling contract still holds for any modified function or export.',
+            '  - Regression tests for existing behavior: confirm behaviors that existed before the change still work after.',
+            '  - End-to-end checks: verify that an existing route, HTTP endpoint, or exported function still works as documented.',
+            'Find existing tests first. If no tests exist for the affected area, create them FROM the documentation/requirements (not from the implementation — see "Write tests BLIND" below).',
+            '',
+            '## Write tests BLIND to the solution',
+            'When you WRITE or CREATE tests, derive them ONLY from the requirements/task, the documentation, and the',
+            'EXISTING code (pre-change behavior). Do NOT read, and do NOT shape your assertions around, the writer\'s',
+            'new diff, new code, or rationale. Tests must encode the behavior that is REQUIRED and DOCUMENTED, never',
+            'the behavior that happens to have been implemented — otherwise they are rigged to pass.',
+            '(Running/executing the tests against the code afterward is expected; this blindness applies to the',
+            'test-WRITING phase only.)',
             '',
             '## How to work',
             '0. **Get the diff first (MANDATORY):** Run `git diff HEAD~1 HEAD --name-only` to get the list of',
@@ -373,13 +389,19 @@ export const NAMED_AGENTS = {
             '- TEST FILES: path(s) to any test files written or modified',
             '- COVERAGE DELTA: what the change adds or leaves uncovered (before vs after where determinable); list notable uncovered lines/paths',
             '- REGRESSIONS / COMPAT BREAKS: explicit list of any pre-existing tests that now fail or API changes that could break existing callers — tag each as BLOCKER',
+            '- DOC STALENESS: list any doc file+section that no longer matches the change (path + what drifted), or "none". Tag each DOC-STALE.',
             '- What you checked but found to be unrelated',
             '',
             '## Backward-compatibility testing & building the test library',
             'GROW THE LIBRARY OVER TIME: where coverage is missing for the behavior being verified,',
-            'CREATE a targeted regression test so the suite accumulates over time. If NO test suite or',
-            'test infrastructure exists yet, establish a MINIMAL one — a single test file plus the',
-            'smallest runner wiring needed — do NOT stand up a heavy framework; keep it small and incremental.',
+            'CREATE a targeted regression test so the suite accumulates over time.',
+            '',
+            'AGENT TEST SUITE REALITY: as of now there is NO test suite for the agent (CLAUDE.md: "There is no test',
+            'suite"). You MAY write minimal `agent/tests/*.test.ts` files runnable via `npx tsx` (e.g.',
+            '`npx tsx agent/tests/my-feature.test.ts`). Do NOT attempt to wire a `package.json` "test" script —',
+            'that is a config write, is gate-denied, and is a WRITER task. Instead, FLAG that need as a follow-up.',
+            'If standing up even a minimal runner is out of scope for the change under test, report it as a',
+            'COVERAGE GAP rather than half-installing infra.',
             '',
             'HARD RESTRICTION: you may ONLY write TEST files — files whose names contain `.test.` or `.spec.`,',
             'or files located under a `__tests__/` or `tests/` directory.',
@@ -559,6 +581,24 @@ export const NAMED_AGENTS = {
         ].join('\n'),
     },
 };
+// Turbo mode — single fast model override. ONE place to update for future
+// Gemini/OpenAI fast model swap. Keep this immediately after NAMED_AGENTS.
+export const FAST_MODEL = 'haiku';
+/**
+ * Apply turbo override to an agents roster.
+ * When turbo is true returns a DEEP COPY with every agent's model set to
+ * FAST_MODEL. When false returns the input object unchanged (zero allocation).
+ * NEVER mutates the input — NAMED_AGENTS and DB-sourced rows are untouched.
+ */
+export function applyTurbo(agents, turbo) {
+    if (!turbo)
+        return agents;
+    const out = {};
+    for (const [name, agent] of Object.entries(agents)) {
+        out[name] = { ...agent, model: FAST_MODEL };
+    }
+    return out;
+}
 const RESEARCH_TOOLS = [
     'Read', 'Write', 'Edit', 'Glob', 'Grep',
     'Bash', 'WebSearch', 'WebFetch',
@@ -642,6 +682,9 @@ export class ClaudeLLM extends llm.LLM {
     // Dedup guard — prevents double-firing reviewer/gate if SubagentStop fires
     // more than once for the same agent_id (e.g. retry edge cases).
     #dispatchedFor = new Set();
+    // Turbo mode — when true, every spawned agent (main + sub-agents) runs on
+    // FAST_MODEL regardless of individual model config. Default off = no-op.
+    #turbo = false;
     constructor(opts = {}) {
         super();
         // Session resume/continue options
@@ -801,6 +844,16 @@ export class ClaudeLLM extends llm.LLM {
         this.#opts.agents = agents;
         console.log(`🤖 Named agents ${agents ? `set (${Object.keys(agents).join(', ')})` : 'reset to built-ins'} — applies at next query cold start`);
     }
+    /**
+     * Enable/disable Turbo mode. When on, the main model and every sub-agent
+     * spawn on FAST_MODEL. Takes effect at the next query cold start.
+     */
+    setTurbo(on) {
+        this.#turbo = on;
+        console.log(`⚡ Turbo mode ${on ? 'ON' : 'OFF'} — main model → ${on ? FAST_MODEL : (this.#opts.model || 'claude-opus-4-8[1m]')}; applies at next query cold start`);
+    }
+    /** Read-only accessor used by ClaudeLLMStream (private fields are class-scoped). */
+    get turbo() { return this.#turbo; }
     /**
      * Set session ID to resume a specific conversation
      * Call this before sending the first message to resume from a previous session
@@ -1248,15 +1301,13 @@ export class ClaudeLLM extends llm.LLM {
      * if the verdict is REJECT. A reviewer failure must never crash the consumer.
      * Public so ClaudeLLMStream can call it via this.#llmRef.spawnReviewer().
      */
+    // writerOutput intentionally NOT fed to the reviewer (neutrality); kept for signature stability
     async spawnReviewer(agentId, writerOutput, emitter) {
         // Dedup guard — SubagentStop may fire more than once for the same agent_id.
         if (this.#dispatchedFor.has(agentId))
             return;
         this.#dispatchedFor.add(agentId);
         try {
-            const idxPathReviewer = (this.#sessionId && this.#opts.workingDirectory)
-                ? getIndexPath(this.#sessionId, this.#opts.workingDirectory)
-                : null;
             // Get the actual diff to give reviewer concrete evidence instead of just the narrative
             let gitDiff = '';
             try {
@@ -1269,15 +1320,12 @@ export class ClaudeLLM extends llm.LLM {
                 // non-fatal: if git fails, proceed without diff
             }
             const prompt = [
-                'Use the reviewer sub-agent to review this writer output for correctness/spec-adherence/obvious bugs.',
-                'The git diff is provided below — use it as the authoritative source of what changed.',
+                'Review the change below for correctness, spec/requirement adherence, and obvious bugs.',
+                'The git diff is the AUTHORITATIVE source of what changed. Judge it on its own merits —',
+                'you are deliberately NOT given the writer\'s rationale or the session index; form an independent verdict from the code and the project\'s own documented standards.',
+                'Read the actual modified files and their callers as needed (you are not limited to the diff).',
                 'End your reply with exactly `VERDICT: ACCEPT` or `VERDICT: REJECT`.',
-                ...(idxPathReviewer ? [`The session index is at ${idxPathReviewer} — you MUST read it before reviewing.`] : []),
                 gitDiff,
-                '',
-                '<writer_output>',
-                writerOutput.slice(0, 6000),
-                '</writer_output>',
             ].join('\n');
             // Do NOT pass agents here — the reviewer must be review-only and must not
             // be able to spawn writer/researcher/reasoner sub-agents. Passing an empty
@@ -1546,7 +1594,8 @@ class ClaudeLLMStream extends llm.LLMStream {
                 permissionMode: this.#opts.permissionMode,
                 allowedTools,
                 // model: this.#opts.model || 'haiku', // haiku for speed with limited tools, sonnet for full research capabilities (including tool use trace in response)
-                model: this.#opts.model || 'claude-opus-4-8[1m]', // Opus 4.8 + [1m] → 1M context (see get model() note); prevents ~153k early compaction
+                // Turbo: when on, override main model to FAST_MODEL regardless of config.
+                model: this.#llmRef.turbo ? FAST_MODEL : (this.#opts.model || 'claude-opus-4-8[1m]'), // Opus 4.8 + [1m] → 1M context (see get model() note); prevents ~153k early compaction
                 enableFileCheckpointing: true,
                 settingSources: ['project', 'user'],
                 extraArgs: { 'replay-user-messages': null },
@@ -1977,7 +2026,13 @@ class ClaudeLLMStream extends llm.LLMStream {
                 // Named sub-agents — the orchestrator delegates to these specialists.
                 // Built-in definitions live at module level (NAMED_AGENTS); per-user
                 // DB-backed definitions (opts.agents, via set_agents) take precedence.
-                agents: this.#opts.agents ?? NAMED_AGENTS,
+                // TURBO BYPASS: when turbo is on and no custom agents were provided,
+                // opts.agents is undefined — the ?? NAMED_AGENTS fallback would skip
+                // the override entirely. Explicitly apply applyTurbo(NAMED_AGENTS)
+                // so built-in agents always get FAST_MODEL when turbo is on.
+                agents: this.#llmRef.turbo
+                    ? applyTurbo(this.#opts.agents ?? NAMED_AGENTS, true)
+                    : (this.#opts.agents ?? NAMED_AGENTS),
             };
             // Run Claude Agent SDK query() and stream results
             let hasOutput = false;

package/dist/index.js

@@ -30,7 +30,7 @@ const __dirname = dirname(__filename);
 import { createPatch } from 'diff';
 import { loadConfig, getMcpServers, getEnabledMcpServerNames, getVoiceMode, getRealtimeConfig, getDirectConfig, listAllClaudeSessions, getMostRecentSessionId, sessionExists, getSessionSummary, getConversationHistory, ensureSessionWorkspace, getSessionWorkspace, getMcpServerStatusList, buildMcpServersForKeys, listWorkspaceArtifacts } from './config.js';
 import { createSTT, createTTS, createRealtimeModelFromConfig, DIRECT_MODE_STT, DIRECT_MODE_TTS } from './voice-io.js';
-import { createClaudeLLM, NAMED_AGENTS } from './claude-llm.js';
+import { createClaudeLLM, NAMED_AGENTS, applyTurbo } from './claude-llm.js';
 import { clearPipelineFastBrainSession, prewarmBM25Index } from './pipeline-fastbrain.js';
 import { ensureClaudeAuth } from './claude-auth.js';
 import { createSmitheryProxy, destroySmitheryProxy, parseSmitheryUrl, isSmitheryUrl, SmitheryAuthorizationError } from './smithery-proxy.js';
@@ -2239,6 +2239,9 @@ async function main() {
     // switch/resume) reuse them without waiting for a frontend resend. Merged
     // OVER the built-in NAMED_AGENTS (same-name user rows shadow built-ins).
     let userNamedAgents = null;
+    // Turbo mode — read from set_agents handshake (data.turbo). When true every
+    // agent (main + sub-agents) runs on FAST_MODEL. Default false = perfect no-op.
+    let turboMode = false;
     // Built-ins the user has removed (tombstoned in user_agents with enabled=false).
     // A removed built-in disappears from the effective set but stays re-addable
     // from the frontend's "Available to add" list — default agents are never
@@ -3266,14 +3269,21 @@ async function main() {
             voiceMode: 'direct',
             skipTTSQueue: true,
             onCompactionEvent: buildOnCompactionEvent(),
-            // Per-user named agents survive LLM recreations (session switch/resume)
+            // Per-user named agents survive LLM recreations (session switch/resume).
+            // applyTurbo runs post-merge so custom models are also overridden when turbo is on.
+            // When neither custom agents nor turbo are active, pass undefined so the
+            // ClaudeLLM cold-start uses NAMED_AGENTS directly (the ?? NAMED_AGENTS path).
+            // When turbo IS on with no custom agents, undefined would skip the override —
+            // ClaudeLLM handles that internally via the turbo bypass in ClaudeLLMStream.
             agents: (userNamedAgents || userRemovedAgents.length) ? (() => {
                 const base = { ...NAMED_AGENTS };
                 for (const r of userRemovedAgents)
                     delete base[r];
-                return { ...base, ...(userNamedAgents || {}) };
+                return applyTurbo({ ...base, ...(userNamedAgents || {}) }, turboMode);
             })() : undefined,
         });
+        // Apply turbo to the freshly created LLM so the main model also picks it up.
+        directLLM.setTurbo(turboMode);
         currentLLM = directLLM;
         // Reset the session always-allow list for each new direct session
         sessionAlwaysAllowPaths = new Set();
@@ -6008,11 +6018,15 @@ async function main() {
                 userNamedAgents = Object.keys(validated).length ? validated : null;
                 // Removed built-ins (tombstones) — excluded from the effective set.
                 userRemovedAgents = Array.isArray(data.removed) ? data.removed.map(String) : [];
+                // Turbo preference from the frontend (sourced from user_settings DB row).
+                turboMode = !!data.turbo;
                 const base = { ...NAMED_AGENTS };
                 for (const r of userRemovedAgents)
                     delete base[r];
-                const merged = { ...base, ...(userNamedAgents || {}) };
-                const customized = !!userNamedAgents || userRemovedAgents.length > 0;
+                // applyTurbo wraps AFTER the base+custom merge so custom models also get overridden.
+                const merged = applyTurbo({ ...base, ...(userNamedAgents || {}) }, turboMode);
+                const customized = !!userNamedAgents || userRemovedAgents.length > 0 || turboMode;
+                currentLLM?.setTurbo?.(turboMode);
                 currentLLM?.setAgents?.(customized ? merged : undefined);
                 console.log(`🤖 set_agents: ${Object.keys(validated).length} user agent(s), ${userRemovedAgents.length} removed → effective [${Object.keys(merged).join(', ')}]`);
                 await sendToFrontend({

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "osborn",
-  "version": "0.9.167",
+  "version": "0.9.169",
   "description": "Voice AI coding assistant - local agent that connects to Osborn frontend",
   "type": "module",
   "bin": {