task-extractor

# Task Extractor Parse multi-task messages → numbered queue → execute sequentially → verify each → report. ## Activation Trigger on ANY user message containing 3+ distinct actionable items. Signs: - Multiple sentences starting with verbs ("build", "fix", "send", "check", "add", "research") - Comma-separated or period-separated instructions - Mixed topics in one paragraph - "also", "and then", "plus", "oh and", "one more thing" If uncertain whether to activate: activate. False positives (structuring a 2-item request) cost nothing. False negatives (dropping task #7 of 12) cost trust. ## Step 1: EXTRACT (before any execution) Parse the message into individual tasks. Each task gets: - **Number** (sequential) - **Summary** (one line, imperative verb) - **Type**: BUILD | FIX | RESEARCH | SEND | DEPLOY | CONFIG | OTHER - **Estimated effort**: QUICK (< 5 min) | MEDIUM (5-30 min) | HEAVY (30+ min / sub-agent) Write to `workspace/TASK_QUEUE.md`: ```markdown # TASK_QUEUE — [date] [time] # Source: [channel] # Total: [N] # Status: IN PROGRESS | # | Task | Type | Effort | Status | Artifact | |---|------|------|--------|--------|----------| | 1 | [summary] | BUILD | HEAVY | ⏳ | | | 2 | [summary] | FIX | QUICK | ⏳ | | | 3 | [summary] | SEND | QUICK | ⏳ | | ... ``` ## Step 2: RECEIPT Reply to the user with the extracted checklist BEFORE starting work: ``` 📋 Extracted [N] tasks from your message: 1. ⏳ [task summary] 2. ⏳ [task summary] 3. ⏳ [task summary] ... Starting now. I'll check each off as I go. ``` Do NOT ask "is this right?" unless genuinely ambiguous. Convert ambiguity into a task and execute. Ryan's rule: don't interrogate, just do. ## Step 3: EXECUTE Work through tasks in dependency order (not necessarily numerical order): - Independent QUICK tasks first (batch them in parallel tool calls) - MEDIUM tasks next - HEAVY tasks: spawn sub-agents with clear scope For each completed task, update TASK_QUEUE.md: - `⏳` → `✅` (done) or `❌` (failed) or `⚠️` (partial) or `🔄` (spawned sub-agent) - Fill in the Artifact column (file path, URL, commit hash, or "n/a") ## Step 4: RECONCILE After all tasks are attempted (or sub-agents spawned), reply with the final checklist: ``` 📋 Task Report ([completed]/[total]): 1. ✅ [task] → [artifact] 2. ✅ [task] → [artifact] 3. 🔄 [task] → sub-agent running, will announce when done 4. ❌ [task] → [reason for failure] 5. ⚠️ [task] → [what was done, what's left] ``` ## Step 5: VERIFY (on sub-agent completion) When a sub-agent announces completion: - Update TASK_QUEUE.md - If ALL tasks are now ✅/❌, send final summary - If tasks remain ⏳, continue working ## Rules 1. **NEVER skip the extract step.** Even if the answer seems obvious. The extract step IS the safety net. 2. **NEVER mark a task ✅ without evidence.** Evidence = file exists, command succeeded, API returned 200, deploy URL works. 3. **If a task fails, say WHY.** Not just ❌ — include the error, the blocker, or what's needed. 4. **Sub-agent tasks get 🔄 until completion event arrives.** Don't mark them ✅ optimistically. 5. **TASK_QUEUE.md is the source of truth.** If context overflows, re-read it. If a session restarts, re-read it. 6. **One TASK_QUEUE at a time.** If a new multi-task message arrives while one is active, append to the existing queue (renumber). 7. **Checkpoint at 5 completed tasks.** Update the user with progress so far. ## Edge Cases ### "Do X and also Y but wait on Z" - X and Y get ⏳, Z gets 🕐 (blocked — note the dependency) ### Sub-agent timeout - Mark as ⚠️, note what was attempted, offer to retry or take over manually ### User changes mind mid-execution - Update TASK_QUEUE.md, cross out cancelled tasks with ~~strikethrough~~, continue with remaining ### Overlapping tasks - If task 3 and task 7 are really the same thing, merge them. Note in Artifact: "merged with #3" ## Anti-patterns - ❌ Reading the message, immediately jumping into task #1 without extracting all tasks - ❌ Marking a sub-agent task ✅ before the completion event - ❌ Saying "I'll get to that" and then forgetting - ❌ Only reporting on the tasks you completed, silently dropping the ones you didn't - ❌ Asking "which one should I do first?" — just prioritize by dependency + effort and go