@probelabs/probe 0.6.0-rc277 → 0.6.0-rc279

package/build/agent/ProbeAgent.js

@@ -104,7 +104,6 @@ import {
   TaskManager,
   createTaskTool,
   taskSystemPrompt,
-  taskGuidancePrompt,
   createTaskCompletionBlockedMessage
 } from './tasks/index.js';
 import { z } from 'zod';
@@ -126,6 +125,27 @@ const MAX_HISTORY_MESSAGES = 100;
 // Maximum image file size (20MB) to prevent OOM attacks
 const MAX_IMAGE_FILE_SIZE = 20 * 1024 * 1024;
 
+/**
+ * Truncate a string for debug logging, showing first and last portion.
+ */
+export function debugTruncate(s, limit = 200) {
+  if (s.length <= limit) return s;
+  const half = Math.floor(limit / 2);
+  return s.substring(0, half) + ` ... [${s.length} chars] ... ` + s.substring(s.length - half);
+}
+
+/**
+ * Log tool results details for debug output.
+ */
+export function debugLogToolResults(toolResults) {
+  if (!toolResults || toolResults.length === 0) return;
+  for (const tr of toolResults) {
+    const argsStr = JSON.stringify(tr.args || {});
+    const resultStr = typeof tr.result === 'string' ? tr.result : JSON.stringify(tr.result || '');
+    console.log(`[DEBUG]   tool: ${tr.toolName} | args: ${debugTruncate(argsStr)} | result: ${debugTruncate(resultStr)}`);
+  }
+}
+
 /**
  * ProbeAgent class to handle AI interactions with code search capabilities
  */
@@ -194,6 +214,7 @@ export class ProbeAgent {
     this.enableExecutePlan = !!options.enableExecutePlan;
     this.debug = options.debug || process.env.DEBUG === '1';
     this.cancelled = false;
+    this._abortController = new AbortController();
     this.tracer = options.tracer || null;
     this.outline = !!options.outline;
     this.searchDelegate = options.searchDelegate !== undefined ? !!options.searchDelegate : true;
@@ -793,6 +814,7 @@ export class ProbeAgent {
       searchDelegateProvider: this.searchDelegateProvider,
       searchDelegateModel: this.searchDelegateModel,
       delegationManager: this.delegationManager,  // Per-instance delegation limits
+      parentAbortSignal: this._abortController.signal,  // Propagate cancellation to delegations
       outputBuffer: this._outputBuffer,
       concurrencyLimiter: this.concurrencyLimiter,  // Global AI concurrency limiter
       isToolAllowed,
@@ -1363,6 +1385,19 @@ export class ProbeAgent {
     const controller = new AbortController();
     const timeoutState = { timeoutId: null };
 
+    // Link agent-level abort to this operation's controller
+    // so that cancel() / cleanup() stops the current streamText call
+    if (this._abortController.signal.aborted) {
+      controller.abort();
+    } else {
+      const onAgentAbort = () => controller.abort();
+      this._abortController.signal.addEventListener('abort', onAgentAbort, { once: true });
+      // Clean up listener when this controller aborts (from any source)
+      controller.signal.addEventListener('abort', () => {
+        this._abortController.signal.removeEventListener('abort', onAgentAbort);
+      }, { once: true });
+    }
+
     // Set up overall operation timeout (default 5 minutes)
     if (this.maxOperationTimeout && this.maxOperationTimeout > 0) {
       timeoutState.timeoutId = setTimeout(() => {
@@ -1730,7 +1765,8 @@ export class ProbeAgent {
                 allowEdit: this.allowEdit,
                 allowedTools: allowedToolsForDelegate,
                 debug: this.debug,
-                tracer: this.tracer
+                tracer: this.tracer,
+                parentAbortSignal: this._abortController.signal
               };
 
               if (this.debug) {
@@ -3153,14 +3189,6 @@ Follow these instructions carefully:
       // Create user message with optional image support
       let userMessage = { role: 'user', content: message.trim() };
 
-      // START CHECKPOINT: Inject task guidance if tasks are enabled
-      if (this.enableTasks) {
-        userMessage.content = userMessage.content + '\n\n' + taskGuidancePrompt;
-        if (this.debug) {
-          console.log('[DEBUG] Task guidance injected into user message');
-        }
-      }
-
       // If schema is provided, prepend JSON format requirement to user message
       if (options.schema && !options._schemaFormatted) {
         const schemaInstructions = generateSchemaInstructions(options.schema, { debug: this.debug });
@@ -3378,6 +3406,11 @@ Follow these instructions carefully:
         completionAttempted = true;
       }, toolContext);
 
+      if (this.debug) {
+        const toolNames = Object.keys(tools);
+        console.log(`[DEBUG] Agent tools registered (${toolNames.length}): ${toolNames.join(', ')}`);
+      }
+
       let maxResponseTokens = this.maxResponseTokens;
       if (!maxResponseTokens) {
         maxResponseTokens = 4000;
@@ -3427,6 +3460,7 @@ Follow these instructions carefully:
 
               if (this.debug) {
                 console.log(`[DEBUG] Step ${currentIteration}/${maxIterations} finished (reason: ${finishReason}, tools: ${toolResults?.length || 0})`);
+                debugLogToolResults(toolResults);
               }
             }
           };
@@ -3638,6 +3672,7 @@ Double-check your response based on the criteria above. If everything looks good
               }
               if (this.debug) {
                 console.log(`[DEBUG] Completion prompt step finished (reason: ${finishReason}, tools: ${toolResults?.length || 0})`);
+                debugLogToolResults(toolResults);
               }
             }
           };
@@ -4554,6 +4589,11 @@ Convert your previous response content into actual JSON data that follows this s
    * Clean up resources (including MCP connections)
    */
   async cleanup() {
+    // Abort any in-flight operations (delegations, streaming, etc.)
+    if (!this._abortController.signal.aborted) {
+      this._abortController.abort();
+    }
+
     // Clean up MCP bridge
     if (this.mcpBridge) {
       try {
@@ -4583,12 +4623,24 @@ Convert your previous response content into actual JSON data that follows this s
   }
 
   /**
-   * Cancel the current request
+   * Cancel the current request and all in-flight delegations.
+   * Aborts the internal AbortController so streamText, subagents,
+   * and any code checking the signal will stop.
    */
   cancel() {
     this.cancelled = true;
+    this._abortController.abort();
     if (this.debug) {
       console.log(`[DEBUG] Agent cancelled for session ${this.sessionId}`);
     }
   }
+
+  /**
+   * Get the abort signal for this agent.
+   * Delegations and subagents should check this signal.
+   * @returns {AbortSignal}
+   */
+  get abortSignal() {
+    return this._abortController.signal;
+  }
 }

package/build/agent/index.js

@@ -3862,12 +3862,17 @@ async function delegate({
   mcpConfigPath = null,
   delegationManager = null,
   // Optional per-instance manager, falls back to default singleton
-  concurrencyLimiter = null
+  concurrencyLimiter = null,
   // Optional global AI concurrency limiter
+  parentAbortSignal = null
+  // Optional AbortSignal from parent to cancel this delegation
 }) {
   if (!task || typeof task !== "string") {
     throw new Error("Task parameter is required and must be a string");
   }
+  if (parentAbortSignal?.aborted) {
+    throw new Error("Delegation cancelled: parent operation was aborted");
+  }
   const hasExplicitTimeout = Object.prototype.hasOwnProperty.call(arguments?.[0] ?? {}, "timeout");
   if (!hasExplicitTimeout) {
     const envTimeoutMs = parseInt(process.env.DELEGATION_TIMEOUT_MS || "", 10);
@@ -3952,12 +3957,37 @@ async function delegate({
     }
     const timeoutPromise = new Promise((_, reject2) => {
       timeoutId = setTimeout(() => {
+        subagent.cancel();
         reject2(new Error(`Delegation timed out after ${timeout} seconds`));
       }, timeout * 1e3);
     });
+    let parentAbortHandler;
+    const parentAbortPromise = new Promise((_, reject2) => {
+      if (parentAbortSignal) {
+        if (parentAbortSignal.aborted) {
+          subagent.cancel();
+          reject2(new Error("Delegation cancelled: parent operation was aborted"));
+          return;
+        }
+        parentAbortHandler = () => {
+          subagent.cancel();
+          reject2(new Error("Delegation cancelled: parent operation was aborted"));
+        };
+        parentAbortSignal.addEventListener("abort", parentAbortHandler, { once: true });
+      }
+    });
     const answerOptions = schema ? { schema } : void 0;
     const answerPromise = answerOptions ? subagent.answer(task, [], answerOptions) : subagent.answer(task);
-    const response = await Promise.race([answerPromise, timeoutPromise]);
+    const racers = [answerPromise, timeoutPromise];
+    if (parentAbortSignal) racers.push(parentAbortPromise);
+    let response;
+    try {
+      response = await Promise.race(racers);
+    } finally {
+      if (parentAbortHandler && parentAbortSignal) {
+        parentAbortSignal.removeEventListener("abort", parentAbortHandler);
+      }
+    }
     if (timeoutId !== null) {
       clearTimeout(timeoutId);
       timeoutId = null;
@@ -4369,8 +4399,9 @@ Instructions:
       promptType: "code-researcher",
       allowedTools: ["extract"],
       maxIterations: 5,
-      delegationManager: options.delegationManager
+      delegationManager: options.delegationManager,
       // Per-instance delegation limits
+      parentAbortSignal: options.parentAbortSignal || null
       // timeout removed - inherit default from delegate (300s)
     });
     return { chunk, result };
@@ -4469,8 +4500,9 @@ Organize all findings into clear categories with items listed under each.${compl
       promptType: "code-researcher",
       allowedTools: [],
       maxIterations: 5,
-      delegationManager: options.delegationManager
+      delegationManager: options.delegationManager,
       // Per-instance delegation limits
+      parentAbortSignal: options.parentAbortSignal || null
       // timeout removed - inherit default from delegate (300s)
     });
     return result;
@@ -4534,8 +4566,9 @@ CRITICAL: Do NOT guess keywords. Actually run searches and see what returns resu
       promptType: "code-researcher",
       // Full tool access for exploration and experimentation
       maxIterations: 15,
-      delegationManager: options.delegationManager
+      delegationManager: options.delegationManager,
       // Per-instance delegation limits
+      parentAbortSignal: options.parentAbortSignal || null
       // timeout removed - inherit default from delegate (300s)
     });
     const plan = parsePlanningResult(stripResultTags(result));
@@ -4592,8 +4625,9 @@ When done, use the attempt_completion tool with your answer as the result.`;
       promptType: "code-researcher",
       allowedTools: [],
       maxIterations: 5,
-      delegationManager: options.delegationManager
+      delegationManager: options.delegationManager,
       // Per-instance delegation limits
+      parentAbortSignal: options.parentAbortSignal || null
       // timeout removed - inherit default from delegate (300s)
     });
     return stripResultTags(result);
@@ -9277,7 +9311,8 @@ var init_vercel = __esm({
               promptType: "code-searcher",
               allowedTools: ["search", "extract", "listFiles", "attempt_completion"],
               searchDelegate: false,
-              schema: CODE_SEARCH_SCHEMA
+              schema: CODE_SEARCH_SCHEMA,
+              parentAbortSignal: options.parentAbortSignal || null
             });
             const delegateResult = options.tracer?.withSpan ? await options.tracer.withSpan("search.delegate", runDelegation, {
               "search.query": searchQuery,
@@ -9491,7 +9526,7 @@ var init_vercel = __esm({
         name: "delegate",
         description: delegateDescription,
         inputSchema: delegateSchema,
-        execute: async ({ task, currentIteration, maxIterations, parentSessionId, path: path9, provider, model, tracer, searchDelegate }) => {
+        execute: async ({ task, currentIteration, maxIterations, parentSessionId, path: path9, provider, model, tracer, searchDelegate, parentAbortSignal }) => {
           if (!task || typeof task !== "string") {
             throw new Error("Task parameter is required and must be a non-empty string");
           }
@@ -9549,8 +9584,9 @@ var init_vercel = __esm({
             enableMcp,
             mcpConfig,
             mcpConfigPath,
-            delegationManager
+            delegationManager,
             // Per-instance delegation limits
+            parentAbortSignal
           });
           return result;
         }
@@ -9588,8 +9624,9 @@ var init_vercel = __esm({
               provider: options.provider,
               model: options.model,
               tracer: options.tracer,
-              delegationManager
+              delegationManager,
               // Per-instance delegation limits
+              parentAbortSignal: options.parentAbortSignal || null
             });
             return result;
           } catch (error) {
@@ -31932,7 +31969,7 @@ ${taskManager.formatTasksForPrompt()}`;
     }
   };
 }
-var taskItemSchema, taskSchema, taskSystemPrompt, taskGuidancePrompt;
+var taskItemSchema, taskSchema, taskSystemPrompt;
 var init_taskTool = __esm({
   "src/agent/tasks/taskTool.js"() {
     "use strict";
@@ -31993,11 +32030,6 @@ Tasks = logical units of work, not files or steps.
 - Circular dependencies are rejected
 - attempt_completion is blocked while tasks remain unresolved
 `;
-    taskGuidancePrompt = `Does this request have MULTIPLE DISTINCT GOALS?
-- "Do A AND B AND C" (multiple goals) \u2192 Create tasks for each goal
-- "Investigate/explain/find X" (single goal) \u2192 Skip tasks, just answer directly
-Multiple internal steps for ONE goal = NO tasks needed.
-If creating tasks, use the task tool with action="create" first.`;
   }
 });
 
@@ -70753,8 +70785,20 @@ If the solution is clear, you can jump to implementation right away. If not, ask
 - Check imports and existing utilities before creating new helpers \u2014 the project may already have what you need.
 
 # Task Planning
-- If the task tool is available, use it to break complex work into milestones before starting implementation.
-- Stay flexible \u2014 if your understanding changes mid-task, add, remove, or reorganize tasks as needed. The plan should serve you, not constrain you.
+When the request has **multiple distinct goals** (e.g. "Fix bug A AND add feature B"), use the task tool to track them:
+- Call the task tool with action="create" and a tasks array. Each task must have an "id" field.
+- Update task status to "in_progress" when starting and "completed" when done.
+- All tasks must be completed or cancelled before calling attempt_completion.
+- Stay flexible \u2014 add, remove, or reorganize tasks as your understanding changes.
+
+Do NOT create tasks for single-goal requests, even complex ones. Multiple internal steps for one goal (search, read, analyze, implement) do not need tasks.
+
+# Discovering Project Commands
+Before building or testing, determine the project's toolchain:
+- Check for Makefile, package.json (scripts), Cargo.toml, go.mod, pyproject.toml, or similar
+- Look for CI config (.github/workflows/, .gitlab-ci.yml) to see what commands CI runs
+- Read README for build/test instructions if the above are unclear
+- Common patterns: \`make build\`/\`make test\`, \`npm run build\`/\`npm test\`, \`cargo build\`/\`cargo test\`, \`go build ./...\`/\`go test ./...\`, \`python -m pytest\`
 
 # During Implementation
 - Always create a new branch before making changes to the codebase.
@@ -70765,12 +70809,22 @@ If the solution is clear, you can jump to implementation right away. If not, ask
 - When editing files, keep edits focused and minimal. For changes spanning more than a few lines, prefer line-targeted editing (start_line/end_line) over text replacement (old_string) \u2014 it constrains scope and prevents accidental removal of adjacent content. Never include unrelated sections in an edit operation.
 - After every significant change, verify the project still builds and passes linting. Do not wait until the end to discover breakage.
 
-# After Implementation
-- Verify the project builds successfully. If it doesn't, fix the build before moving on.
-- Run lint and typecheck commands if known for the project. Fix any new warnings or errors you introduced.
-- Add tests for any new or changed functionality. Tests must cover the main path and important edge cases.
-- Run the project's full test suite. If any tests fail (including pre-existing ones you may have broken), fix them before finishing.
-- When the task is done, respond to the user with a concise summary of what was implemented, what files were changed, and any relevant details. Include links (e.g. pull request URL) so the user has everything they need.
+# Writing Tests
+Every change must include tests. Before writing them:
+- Find existing test files for the module you changed \u2014 look in \`tests/\`, \`__tests__/\`, \`*_test.go\`, \`*.test.js\`, \`*.spec.ts\`, or co-located test modules (\`#[cfg(test)]\` in Rust).
+- Read those tests to understand the project's testing patterns: framework, assertion style, mocking approach, file naming, test organization.
+- Prefer extending an existing test file over creating a new one when your change is in the same module.
+- Write tests that cover the main path and important edge cases. Include a failing-input test when relevant.
+- When fixing a bug, write a failing test first that reproduces the bug, then fix the code to make it pass.
+
+# Verify Changes
+Before committing or creating a PR, run through this checklist:
+1. **Build** \u2014 run the project-appropriate build command (go build, npm run build, cargo build, make, etc.). Fix any compilation errors.
+2. **Lint & typecheck** \u2014 run linter/formatter if the project has one (eslint, clippy, golangci-lint, etc.). Fix any new warnings.
+3. **Test** \u2014 run the full test suite (go test ./..., npm test, cargo test, make test, pytest, etc.). Fix any failures, including pre-existing tests you may have broken.
+4. **Review** \u2014 re-read your diff. Ensure no debug code, no unrelated changes, no secrets, no missing files.
+
+Do NOT skip verification. Do NOT proceed to PR creation with a broken build or failing tests.
 
 # GitHub Integration
 - Use the \`gh\` CLI for all GitHub operations: issues, pull requests, checks, releases.
@@ -81751,7 +81805,9 @@ __export(ProbeAgent_exports, {
   ENGINE_ACTIVITY_TIMEOUT_DEFAULT: () => ENGINE_ACTIVITY_TIMEOUT_DEFAULT,
   ENGINE_ACTIVITY_TIMEOUT_MAX: () => ENGINE_ACTIVITY_TIMEOUT_MAX,
   ENGINE_ACTIVITY_TIMEOUT_MIN: () => ENGINE_ACTIVITY_TIMEOUT_MIN,
-  ProbeAgent: () => ProbeAgent
+  ProbeAgent: () => ProbeAgent,
+  debugLogToolResults: () => debugLogToolResults,
+  debugTruncate: () => debugTruncate
 });
 import dotenv2 from "dotenv";
 import { createAnthropic as createAnthropic2 } from "@ai-sdk/anthropic";
@@ -81764,6 +81820,19 @@ import { EventEmitter as EventEmitter5 } from "events";
 import { existsSync as existsSync7 } from "fs";
 import { readFile as readFile3, stat, readdir as readdir3 } from "fs/promises";
 import { resolve as resolve7, isAbsolute as isAbsolute6, dirname as dirname5, basename, normalize as normalize2, sep as sep5 } from "path";
+function debugTruncate(s, limit = 200) {
+  if (s.length <= limit) return s;
+  const half = Math.floor(limit / 2);
+  return s.substring(0, half) + ` ... [${s.length} chars] ... ` + s.substring(s.length - half);
+}
+function debugLogToolResults(toolResults) {
+  if (!toolResults || toolResults.length === 0) return;
+  for (const tr of toolResults) {
+    const argsStr = JSON.stringify(tr.args || {});
+    const resultStr = typeof tr.result === "string" ? tr.result : JSON.stringify(tr.result || "");
+    console.log(`[DEBUG]   tool: ${tr.toolName} | args: ${debugTruncate(argsStr)} | result: ${debugTruncate(resultStr)}`);
+  }
+}
 var ENGINE_ACTIVITY_TIMEOUT_DEFAULT, ENGINE_ACTIVITY_TIMEOUT_MIN, ENGINE_ACTIVITY_TIMEOUT_MAX, MAX_TOOL_ITERATIONS, MAX_HISTORY_MESSAGES, MAX_IMAGE_FILE_SIZE, ProbeAgent;
 var init_ProbeAgent = __esm({
   "src/agent/ProbeAgent.js"() {
@@ -81872,6 +81941,7 @@ var init_ProbeAgent = __esm({
         this.enableExecutePlan = !!options.enableExecutePlan;
         this.debug = options.debug || process.env.DEBUG === "1";
         this.cancelled = false;
+        this._abortController = new AbortController();
         this.tracer = options.tracer || null;
         this.outline = !!options.outline;
         this.searchDelegate = options.searchDelegate !== void 0 ? !!options.searchDelegate : true;
@@ -82317,6 +82387,8 @@ var init_ProbeAgent = __esm({
           searchDelegateModel: this.searchDelegateModel,
           delegationManager: this.delegationManager,
           // Per-instance delegation limits
+          parentAbortSignal: this._abortController.signal,
+          // Propagate cancellation to delegations
           outputBuffer: this._outputBuffer,
           concurrencyLimiter: this.concurrencyLimiter,
           // Global AI concurrency limiter
@@ -82764,6 +82836,15 @@ var init_ProbeAgent = __esm({
         }
         const controller = new AbortController();
         const timeoutState = { timeoutId: null };
+        if (this._abortController.signal.aborted) {
+          controller.abort();
+        } else {
+          const onAgentAbort = () => controller.abort();
+          this._abortController.signal.addEventListener("abort", onAgentAbort, { once: true });
+          controller.signal.addEventListener("abort", () => {
+            this._abortController.signal.removeEventListener("abort", onAgentAbort);
+          }, { once: true });
+        }
         if (this.maxOperationTimeout && this.maxOperationTimeout > 0) {
           timeoutState.timeoutId = setTimeout(() => {
             controller.abort();
@@ -83067,7 +83148,8 @@ var init_ProbeAgent = __esm({
                     allowEdit: this.allowEdit,
                     allowedTools: allowedToolsForDelegate,
                     debug: this.debug,
-                    tracer: this.tracer
+                    tracer: this.tracer,
+                    parentAbortSignal: this._abortController.signal
                   };
                   if (this.debug) {
                     console.log(`[DEBUG] Executing delegate tool`);
@@ -84229,12 +84311,6 @@ You are working with a workspace. Available paths: ${workspaceDesc}
           });
           const systemMessage = await this.getSystemMessage();
           let userMessage = { role: "user", content: message.trim() };
-          if (this.enableTasks) {
-            userMessage.content = userMessage.content + "\n\n" + taskGuidancePrompt;
-            if (this.debug) {
-              console.log("[DEBUG] Task guidance injected into user message");
-            }
-          }
           if (options.schema && !options._schemaFormatted) {
             const schemaInstructions = generateSchemaInstructions(options.schema, { debug: this.debug });
             userMessage.content = message.trim() + schemaInstructions;
@@ -84390,6 +84466,10 @@ You are working with a workspace. Available paths: ${workspaceDesc}
             completionResult = result;
             completionAttempted = true;
           }, toolContext);
+          if (this.debug) {
+            const toolNames = Object.keys(tools2);
+            console.log(`[DEBUG] Agent tools registered (${toolNames.length}): ${toolNames.join(", ")}`);
+          }
           let maxResponseTokens = this.maxResponseTokens;
           if (!maxResponseTokens) {
             maxResponseTokens = 4e3;
@@ -84429,6 +84509,7 @@ You are working with a workspace. Available paths: ${workspaceDesc}
                   }
                   if (this.debug) {
                     console.log(`[DEBUG] Step ${currentIteration}/${maxIterations} finished (reason: ${finishReason}, tools: ${toolResults?.length || 0})`);
+                    debugLogToolResults(toolResults);
                   }
                 }
               };
@@ -84585,6 +84666,7 @@ Double-check your response based on the criteria above. If everything looks good
                   }
                   if (this.debug) {
                     console.log(`[DEBUG] Completion prompt step finished (reason: ${finishReason}, tools: ${toolResults?.length || 0})`);
+                    debugLogToolResults(toolResults);
                   }
                 }
               };
@@ -85295,6 +85377,9 @@ Convert your previous response content into actual JSON data that follows this s
        * Clean up resources (including MCP connections)
        */
       async cleanup() {
+        if (!this._abortController.signal.aborted) {
+          this._abortController.abort();
+        }
         if (this.mcpBridge) {
           try {
             await this.mcpBridge.cleanup();
@@ -85318,14 +85403,25 @@ Convert your previous response content into actual JSON data that follows this s
         this.clearHistory();
       }
       /**
-       * Cancel the current request
+       * Cancel the current request and all in-flight delegations.
+       * Aborts the internal AbortController so streamText, subagents,
+       * and any code checking the signal will stop.
        */
       cancel() {
         this.cancelled = true;
+        this._abortController.abort();
         if (this.debug) {
           console.log(`[DEBUG] Agent cancelled for session ${this.sessionId}`);
         }
       }
+      /**
+       * Get the abort signal for this agent.
+       * Delegations and subagents should check this signal.
+       * @returns {AbortSignal}
+       */
+      get abortSignal() {
+        return this._abortController.signal;
+      }
     };
   }
 });

package/build/agent/shared/prompts.js

@@ -81,8 +81,20 @@ If the solution is clear, you can jump to implementation right away. If not, ask
 - Check imports and existing utilities before creating new helpers — the project may already have what you need.
 
 # Task Planning
-- If the task tool is available, use it to break complex work into milestones before starting implementation.
-- Stay flexible — if your understanding changes mid-task, add, remove, or reorganize tasks as needed. The plan should serve you, not constrain you.
+When the request has **multiple distinct goals** (e.g. "Fix bug A AND add feature B"), use the task tool to track them:
+- Call the task tool with action="create" and a tasks array. Each task must have an "id" field.
+- Update task status to "in_progress" when starting and "completed" when done.
+- All tasks must be completed or cancelled before calling attempt_completion.
+- Stay flexible — add, remove, or reorganize tasks as your understanding changes.
+
+Do NOT create tasks for single-goal requests, even complex ones. Multiple internal steps for one goal (search, read, analyze, implement) do not need tasks.
+
+# Discovering Project Commands
+Before building or testing, determine the project's toolchain:
+- Check for Makefile, package.json (scripts), Cargo.toml, go.mod, pyproject.toml, or similar
+- Look for CI config (.github/workflows/, .gitlab-ci.yml) to see what commands CI runs
+- Read README for build/test instructions if the above are unclear
+- Common patterns: \`make build\`/\`make test\`, \`npm run build\`/\`npm test\`, \`cargo build\`/\`cargo test\`, \`go build ./...\`/\`go test ./...\`, \`python -m pytest\`
 
 # During Implementation
 - Always create a new branch before making changes to the codebase.
@@ -93,12 +105,22 @@ If the solution is clear, you can jump to implementation right away. If not, ask
 - When editing files, keep edits focused and minimal. For changes spanning more than a few lines, prefer line-targeted editing (start_line/end_line) over text replacement (old_string) — it constrains scope and prevents accidental removal of adjacent content. Never include unrelated sections in an edit operation.
 - After every significant change, verify the project still builds and passes linting. Do not wait until the end to discover breakage.
 
-# After Implementation
-- Verify the project builds successfully. If it doesn't, fix the build before moving on.
-- Run lint and typecheck commands if known for the project. Fix any new warnings or errors you introduced.
-- Add tests for any new or changed functionality. Tests must cover the main path and important edge cases.
-- Run the project's full test suite. If any tests fail (including pre-existing ones you may have broken), fix them before finishing.
-- When the task is done, respond to the user with a concise summary of what was implemented, what files were changed, and any relevant details. Include links (e.g. pull request URL) so the user has everything they need.
+# Writing Tests
+Every change must include tests. Before writing them:
+- Find existing test files for the module you changed — look in \`tests/\`, \`__tests__/\`, \`*_test.go\`, \`*.test.js\`, \`*.spec.ts\`, or co-located test modules (\`#[cfg(test)]\` in Rust).
+- Read those tests to understand the project's testing patterns: framework, assertion style, mocking approach, file naming, test organization.
+- Prefer extending an existing test file over creating a new one when your change is in the same module.
+- Write tests that cover the main path and important edge cases. Include a failing-input test when relevant.
+- When fixing a bug, write a failing test first that reproduces the bug, then fix the code to make it pass.
+
+# Verify Changes
+Before committing or creating a PR, run through this checklist:
+1. **Build** — run the project-appropriate build command (go build, npm run build, cargo build, make, etc.). Fix any compilation errors.
+2. **Lint & typecheck** — run linter/formatter if the project has one (eslint, clippy, golangci-lint, etc.). Fix any new warnings.
+3. **Test** — run the full test suite (go test ./..., npm test, cargo test, make test, pytest, etc.). Fix any failures, including pre-existing tests you may have broken.
+4. **Review** — re-read your diff. Ensure no debug code, no unrelated changes, no secrets, no missing files.
+
+Do NOT skip verification. Do NOT proceed to PR creation with a broken build or failing tests.
 
 # GitHub Integration
 - Use the \`gh\` CLI for all GitHub operations: issues, pull requests, checks, releases.