work-manager

work-manager

Work lifecycle management plugin for Claude Code. Start, track, recall, and complete development tasks with automatic progress logging and knowledge capture.

Features

• Three-phase workflow: research → plan → implement, with allowed back-transitions
• Structural mode enforcement: each phase has a dedicated agent with restricted tools — no prompt-based honor system
• Notes as primary deliverable: phase agents write findings/decisions/results to _notes/ as their main output, not a side-effect
• Auto-progress logging: Session end hook captures what was accomplished and checks off criteria
• Auto work notes capture: Hooks detect architecture decisions, debugging findings, and research results
• Requirement detection: User prompt hook automatically adds new requirements mentioned mid-session
• Multi-repo PR creation: Discovers all repos with unpushed commits and creates PRs
• Task workspace management (optional): mise tasks for git worktree-based task isolation

Workflow

1. Create workspace     mise run task           # interactive: pick repos, create worktrees
                        cd tasks/MILAB-1234-fix-auth/

2. Start tracking       /work start             # creates _notes/, detects branch, gathers scope

3. Research phase       (ask questions, explore code)
                        → auto-saves findings to _notes/research-*.md

4. Plan phase           /work update move to plan
                        → build task list, acceptance criteria in _notes/

5. Implement phase      /work update move to implement
                        → write code, run tests, results in _notes/impl-*.md

6. Iterate              /work recall            # re-orient: what's done, what's next
                        /work update <message>  # log progress, capture knowledge

7. Finish               /work done              # verify criteria, mark complete
                        /work pr                # create PRs across all affected repos

8. Cleanup              mise run task-remove    # detach worktrees, delete task folder

Mise tasks are optional — the core plugin (/work start, /work recall, etc.) works in any git repo.

Capability	With mise tasks	Without mise tasks
Branch management	Automatic worktrees per repo	Manual git checkout -b
Repo isolation	Each task gets its own working tree	Shared working tree, stash to switch
Task creation	mise run task (interactive fzf)	Create branch + cd + /work start
Task cleanup	mise run task-remove (auto worktree detach)	Manual branch deletion
MCP config	Auto-copied .mcp.json per worktree	Manual setup
Multi-repo	Select multiple repos in one step	Set up each repo individually

Commands

Command	Description
/work install	Guided setup — plugin, QMD, mise, task scripts
/work start	Initialize a work session — creates _notes/ with _summary.md and work notes
/work status	Show current work summary
/work recall [topic]	Re-orient to current work with dynamic knowledge loading
/work recall --deep	Load everything for comprehensive context
/work update <message>	Log progress, capture knowledge, or transition phase
/work done	Mark work as complete
/work pr	Create PRs for all repos with unpushed commits
/work help	Show usage guide

Architecture

Router + Phase Agents + Phase Skills

User → work-manager (router) → reads phase from _notes/_summary.md
                              → skill commands: executes directly
                              → phase work: delegates to phase agent
                              → transitions: updates _summary.md

Phase agents:
  work-researcher  (green)  — Read, Glob, Grep, Bash (r/o), Explore agents, Write (_notes/ only)
  work-planner     (yellow) — Read, Glob, Grep, Bash (r/o), Write (_notes/ only)
  work-implementer (red)    — All tools, spawns code subagents

Phase skills (loaded by agents on demand):
  work-research   — scope breakdown, exploration workflow, findings template
  work-plan       — acceptance criteria, task lists, work split, decision template
  work-implement  — task execution, test runs, results template, blocker handling

Why three agents? The tools field in agent frontmatter is hard enforcement — the agent physically cannot call tools not listed. A single agent with prompt-based restrictions fails because models ignore behavioral constraints when focused on the user's request.

Why phase skills? Skills describe what to do (workflow steps, file templates, writing rules, completion signals). Agents define how to do it (tool access, delegation strategy). This separation keeps agents lean and allows progressive disclosure — skill content loads only when the phase activates.