osborn 0.9.220 → 0.9.221

package/dist/claude-llm.js

@@ -86,13 +86,13 @@ async function buildRecallInjection(sessionId, workingDir, prompt) {
         return ''; // recall is best-effort — never break the turn
     }
 }
-// ≤3 direct tool call budget per turn. Reset on every UserPromptSubmit (new user message).
+// ≤5 direct tool call budget per turn. Reset on every UserPromptSubmit (new user message).
 // Enforced mechanically in PreToolUse — the model CANNOT exceed this regardless of JSONL history.
 // Task/Agent delegations are exempt (delegation is what we WANT). Sub-agent tool calls
 // (agent_type !== null) are exempt (they're inside a delegation). Only the main orchestrator
 // agent's direct tool calls count against the budget.
 let turnToolCallCount = 0;
-const TOOL_CALL_BUDGET = 3;
+const TOOL_CALL_BUDGET = 5;
 /**
  * Strip markdown formatting for TTS (text-to-speech)
  * Removes **bold**, ##headers, ```code```, etc. so TTS doesn't read them literally
@@ -1704,6 +1704,17 @@ export class ClaudeLLM extends llm.LLM {
                                     return {};
                                 }],
                         }],
+                    // Self-review gate — block ONCE (stop_hook_active-guarded) so the reviewer
+                    // re-checks its own verdict before returning, instead of stopping on a
+                    // first-pass judgment.
+                    Stop: [{
+                            matcher: '.*',
+                            hooks: [async (input) => {
+                                    if (input?.stop_hook_active)
+                                        return {};
+                                    return { decision: 'block', reason: 'Before you finalize: re-read your findings against the diff once. Confirm each BLOCKER/MAJOR is real and reproducible (not a style nit or a false positive), and that your verdict matches the severity of what you actually found. Then end with exactly `VERDICT: ACCEPT` or `VERDICT: REJECT`.' };
+                                }],
+                        }],
                 },
             };
             console.log(`[DISPATCH] spawning reviewer for agentId=${agentId.slice(0, 8)}`);
@@ -1793,6 +1804,16 @@ export class ClaudeLLM extends llm.LLM {
                                     return {};
                                 }],
                         }],
+                    // Self-review gate — block ONCE (stop_hook_active-guarded) so the tester
+                    // re-checks its own conclusion before returning.
+                    Stop: [{
+                            matcher: '.*',
+                            hooks: [async (input) => {
+                                    if (input?.stop_hook_active)
+                                        return {};
+                                    return { decision: 'block', reason: 'Before you finalize: re-check your conclusion once. Did you actually RUN the commands (not assume the outcome)? Does your result match the real output, and did you distinguish a genuine regression from a flaky/environment failure? Then end with exactly `RESULT: PASS` or `RESULT: FAIL` followed by a brief summary.' };
+                                }],
+                        }],
                 },
             };
             console.log(`[DISPATCH] spawning tester for agentId=${agentId.slice(0, 8)}`);
@@ -1879,6 +1900,16 @@ export class ClaudeLLM extends llm.LLM {
                                     return {};
                                 }],
                         }],
+                    // Self-review gate — block ONCE (stop_hook_active-guarded) so the gate
+                    // re-checks its own PASS/NEEDS-MORE call before returning.
+                    Stop: [{
+                            matcher: '.*',
+                            hooks: [async (input) => {
+                                    if (input?.stop_hook_active)
+                                        return {};
+                                    return { decision: 'block', reason: 'Before you finalize: re-check your gate decision once. Is the research genuinely complete and well-sourced for the original question — not passing thin work, nor failing solid work? Then end with exactly `GATE: PASS` or `GATE: NEEDS-MORE`.' };
+                                }],
+                        }],
                 },
             };
             console.log(`[DISPATCH] spawning research-gate for agentId=${agentId.slice(0, 8)}`);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "osborn",
-  "version": "0.9.220",
+  "version": "0.9.221",
   "description": "Voice AI coding assistant - local agent that connects to Osborn frontend",
   "type": "module",
   "bin": {