Taskagent

Taskagent - Agent Work Tracking

CRITICAL: Command Name

Always use taskagent, NEVER use task directly.

taskagent is a wrapper that uses a separate taskwarrior configuration for agent work. Using task directly will access the human's personal task list, not the agent task tracking system.

taskagent is globally available - no need to cd to any directory. It works from anywhere.

CRITICAL: Use Partial UUIDs, Not Short IDs

Always use partial UUIDs (first 8 characters) instead of short numeric IDs.

Short IDs (1, 2, 3...) renumber whenever tasks are completed, deleted, or modified. This makes them unreliable for persistent references. UUIDs are permanent.

# BAD - short ID may change
taskagent 3 info

# GOOD - partial UUID is stable
taskagent abc1234d info

All reports (list, ready, next) now display uuid.short instead of numeric IDs. Use these partial UUIDs in all commands.

When creating tasks or linking dependencies, always capture and use UUIDs:

taskagent add "New task" project:foo
# Note the UUID from output, use it for future references

Overview

taskagent is a taskwarrior-based tracker for persistent agent memory across sessions. Use for multi-session work with complex context; use TodoWrite for simple single-session tasks.

When to Use taskagent vs TodoWrite

Use taskagent when:

• Multi-session work - Tasks spanning multiple compaction cycles or days
• Complex context - Work needing detailed background or discovered subtasks
• Project work - Tasks tied to a specific project with a PLAN.md
• Side quests - Exploratory work that might pause the main task
• Handoff needed - Work that another session (or human) needs to pick up

Use TodoWrite when:

• Single-session tasks - Work that completes within current session
• Linear execution - Straightforward step-by-step tasks with no branching
• Immediate context - All information already in conversation
• Simple tracking - Just need a checklist to show progress

Key insight: If resuming work after 2 weeks would be difficult without taskagent, use taskagent.

Surviving Compaction Events

What survives compaction:

• All taskagent data (tasks, annotations, details, dependencies)
• PLAN.md in project root
• org/projects wiki files

What doesn't survive:

• Conversation history
• TodoWrite lists

Writing annotations for post-compaction recovery:

Write annotations as if explaining to a future agent with zero conversation context:

COMPLETED: Specific deliverables
IN PROGRESS: Current state + next immediate step
BLOCKERS: What's preventing progress
KEY DECISIONS: Important context or user guidance

Recording Learnings and Misalignments

During the session, annotate tasks with learnings and misalignments as they occur. These are critical for future sessions and handoff.

LEARNING = New information acquired (user preference, codebase pattern, correct approach discovered)

taskagent <uuid> annotate "LEARNING: User prefers fd over find for file searches."
taskagent <uuid> annotate "LEARNING: Deploy pipeline is in .github/workflows/deploy.yml, not CircleCI."

MISALIGNMENT = Agent went wrong direction, was corrected (implies wasted effort or wrong output)

taskagent <uuid> annotate "MISALIGNMENT: Assumed React class components, codebase uses functional only."
taskagent <uuid> annotate "MISALIGNMENT: Put config in src/, should have been lib/ per project convention."

When to record:

• LEARNING: When you discover something useful about the codebase, user preferences, or correct approach
• MISALIGNMENT: When the user corrects you or you realize you went the wrong direction

A misalignment often produces a learning - record both:

taskagent <uuid> annotate "MISALIGNMENT: Used grep, user corrected to use ugrep."
taskagent <uuid> annotate "LEARNING: User prefers ugrep (ug) over grep."

These annotations are retrieved during /handoff and written to the plan file's ## Learnings section for future sessions.

Session Start Protocol

At session start, always:

1. Check for ready work:

   taskagent ready
   

2. Check for active (in-progress) work:

   taskagent +ACTIVE list
   

3. If active work exists, read the task details and annotations:

   taskagent <uuid> agentinfo    # minimal info for agents (preferred)
   taskagent <uuid> info         # full details (shorthand for 'information')
   

4. Ask user which project/context if unclear, then check for org/projects file:

   • Look for ~/org/projects/<project>.md (vimwiki)
   • Read the ## Agent section for:
     • taskagent project: project name (defaults to filename)
     • plan file: path to PLAN.md (defaults to repo root)

5. Read the PLAN.md if it exists for additional context

6. Report to user: "X tasks ready, Y in progress. [Summary]. Should I continue with Z?"

Progress Checkpointing

Update taskagent annotations at these checkpoints:

Critical triggers:

• Context running low - User mentions compaction/token limit
• Token budget > 70% - Proactively checkpoint
• Major milestone reached - Completed significant work
• Hit a blocker - Can't proceed, capture what was tried
• Task transition - Switching tasks or closing one

Checkpoint command:

taskagent <uuid> annotate "COMPLETED: X. IN PROGRESS: Y. NEXT: Z."

Core Operations

Note: All commands below use taskagent, not task. The taskagent command is specifically configured for agent work tracking and is separate from the human's personal taskwarrior.

Direct execution: Basic taskagent operations (add, annotate, modify, start, stop, list, info) should be run directly - no subagent needed. Only invoke subagents for reviews (@task-reviewer, @code-reviewer) and complex queries (@taskagent-reader).

Check ready work (built-in report)

taskagent ready                    # built-in taskwarrior report, shows unblocked tasks
taskagent project:<name> ready
taskagent blocked                  # show tasks blocked by dependencies
taskagent blocking                 # show tasks that are blocking others

Create new task

taskagent add "Task title" project:<name>
taskagent add "Task title" project:<name> details:"Longer explanation of context and why"
taskagent add "Task title" project:<name> +side_quest

Link discovered work

# Use the parent's UUID for discovered_from
taskagent add "Found issue" project:<name> discovered_from:<parent_uuid>

Update task status

taskagent <uuid> start        # marks in progress
taskagent <uuid> stop         # pauses work
taskagent <uuid> done         # REQUIRES @task-reviewer first! See "Task Completion Review"

⚠️ NEVER mark a task done without invoking @task-reviewer first.

Add session notes (annotations)

taskagent <uuid> annotate "COMPLETED: auth endpoint. IN PROGRESS: tests. NEXT: integration."

View task details

taskagent <uuid> agentinfo    # minimal info for agents (preferred)
taskagent <uuid> info         # full details (shorthand for 'information')
taskagent project:<name> list

Dependencies

taskagent <uuid> modify depends:<other_uuid>

Task Fields Reference

PLAN.md Convention

Optional: Projects can have a PLAN.md in the repo root. This can be used independently of org/projects files, or together with them.

# Project Plan

## Current Focus
What the agent should prioritize right now.

## Context
Background information helpful for the agent.

## Notes
Human-agent collaborative notes, session summaries.

The agent should:

• Read PLAN.md at session start for context
• Update the ## Notes section with significant progress/decisions (ask first)
• Never modify ## Current Focus without asking

Task Completion Review

MANDATORY: Before marking ANY task done, you MUST invoke @task-reviewer. There are NO exceptions.

This applies to ALL task types:

• Code tasks
• Verification tasks (reviewer confirms the verification was actually performed)
• Research tasks (reviewer confirms findings were documented)
• Documentation tasks
• ANY task type

External Ticket Integration

For tasks linked to external tickets (Jira, GitHub Issues):

1. Before PR/merge: Invoke @code-reviewer with the ticket ID
2. Before taskagent done: Invoke @task-reviewer as usual

Code-reviewer checks ticket alignment; task-reviewer checks task completion. Both are required for externally-tracked work.

Then invoke the reviewer with the UUID (from the report output):

@task-reviewer Review task <uuid>
- Project: <project-name>
- PLAN.md: <path or "none">
- org/projects file: <path or "none">
- Summary: <what you did in 1-2 sentences>
- Key files changed: <list main files touched, or "none" for non-code tasks>

Always pass the UUID, not the short ID or description. The reviewer needs the UUID to look up the task.

Include whichever other fields you have - the reviewer can work with partial context.

After receiving the review verdict:

• PASS: You may mark the task done with taskagent <uuid> done
• PASS WITH RESERVATIONS: Do NOT mark done automatically. Ask the human to review and decide.
• NEEDS WORK: Address the feedback, then request another review before closing.

Integration with TodoWrite

Temporal layering:

• TodoWrite (this hour): Tactical execution, immediate steps
• taskagent (this week/month): Strategic objectives, persistent context

Handoff pattern:

1. Session start: Read taskagent task -> Create TodoWrite for immediate actions
2. During work: Mark TodoWrite items completed
3. Reach milestone: Update taskagent annotations
4. Session end: TodoWrite disappears, taskagent survives

org/projects File Format

Optional: The org/projects vimwiki file (~/org/projects/<project>.md) can be used independently of PLAN.md, or together with it. Format:

# Project Name

## Overview
High-level description.

## Resources
- Links, repos, docs

## Tasks
- [ ] Human tasks

## Agent
- **taskagent project**: `project-name`
- **plan file**: `~/path/to/repo/PLAN.md`

If taskagent project is omitted, defaults to the filename. If plan file is omitted, defaults to PLAN.md in repo root.