pm

A project manager for AI coding agents. Not for you — for Claude.

pm gives your AI agent structure so it doesn't just dive into code and lose track of what it's doing. It enforces planning before editing, tracks scope so changes don't sprawl across dozens of files, and logs decisions so the next session doesn't redo work the last one already figured out.

You install it, run pm in your project, and the setup wizard handles the rest. From that point on, your AI agent manages itself — creating tasks, logging progress, recording why it chose approach A over approach B. You just tell it what to build.

pm supports multiple AI agents (Claude Code, OpenCode, etc.) working on the same project concurrently. Each agent is identified automatically by hooks, and pm tracks which agent and model is working on each task and issue.

The problem

AI agents are productive but forgetful. Each session starts from zero. Without structure, they'll re-investigate decisions already made, let a "small fix" balloon into a 15-file refactor, or lose track of what was done halfway through a feature. pm solves this by giving the agent a persistent, enforced workflow — the same way a project manager keeps a team on track, except the team is an LLM.

What carries over between sessions

Every session starts with a briefing. pm tells Claude what's in progress, what was done recently, and what decisions were made. Nothing gets lost.

• pm recap — the agent runs this automatically on session start. It sees active tasks, recent progress, and open issues.
• pm log — full history of everything done: which tasks were started, completed, failed, and what notes the agent left.
• pm why "search term" — searches all recorded decisions. Before the agent re-decides something, it checks if a previous session already settled it.
• pm show <id> — deep view of a feature with all phases, tasks, decisions, and timestamps.

This is the core value. A five-session feature doesn't lose context between sessions — the agent picks up where it left off.

Install

npm install -g @piotrjura/pm

Or run without installing:

npx @piotrjura/pm

pm works with Claude Code (default) or OpenCode. During pm init you can choose which agent to set up, or use flags for non-interactive setup:

pm init              # Set up Claude Code (default)
pm init --opencode   # Set up OpenCode
pm init --opencode --claude-code  # Set up both

Getting started

Navigate to your project and run:

pm

If this is a fresh project, pm detects that and launches the setup wizard. It walks through three steps:

1. Whitelist pm commands — adds pm * to Claude Code's allowed commands so it can call pm without asking permission every time
2. Set up hooks — installs Claude Code hooks that enforce the workflow (more on this below)
3. Create data store — initializes .pm/data.json in your project

Once setup is done, you're dropped straight into the TUI. Next time you run pm, it opens directly.

How it works

pm enforces a simple rule: log work before writing code.

You don't need to run these commands yourself — Claude does. When it starts a session, pm briefs it on what's in progress. When it tries to edit a file without an active task, pm blocks it. When scope creeps beyond 3 files, pm warns about it. The agent learns the workflow on its first session and follows it from there.

There are two ways to track work:

Features — for structured work (multiple files, multiple steps):

pm add-feature "Add user preferences" --description "Store and apply user prefs"
# Use --type fix for bug/regression tracking
pm add-feature "Fix login bug" --fix --description "Session token expiry issue"
pm add-phase <featureId> "Implementation"
pm add-task <featureId> <phaseId> "Add preferences table" --files "schema.ts,migrations"
pm start <taskId>
# agent does the work
pm done <taskId> --note "Added prefs table with defaults"

Issues — for quick one-off fixes (1-2 files, one logical change):

pm add-issue "Fix pagination off-by-one" --type bug --priority high
# Issue status flows: triage → backlog → todo → in-progress → done
# agent does the work
pm done <id> --note "Fixed fence post error in offset calc"

Each task is a focused unit — one logical change across 1-3 files. If the agent touches more, pm warns it to break the work down further.

Decision logging

When the agent makes a design choice, it records it:

pm decide <id> "Use JSON file storage instead of SQLite" \
  --reasoning "Single file, no native deps, portable across machines"

Decisions attach to features, tasks, or issues and persist across sessions. Search them later:

pm why "storage"
# Returns all decisions matching "storage" with context

This is what makes pm useful across sessions. Claude doesn't re-discover or re-decide things that were already settled — it checks first.

Claude Code hooks

This is the enforcement layer. pm installs four hooks that run automatically — no action needed from you:

Hook	Trigger	What it does
PreToolUse	Before any file edit	Blocks edits without an active task or issue
PostToolUse	After any file edit	Tracks which files were edited, warns if scope exceeds 3 files
UserPromptSubmit	Every prompt	Injects active task context so Claude knows what it's working on
SessionStart	New session	Captures agent/model identity, resets stuck tasks, briefs Claude on current state

The hooks are non-destructive — they merge with any existing Claude Code hooks in your project.

OpenCode plugin

pm also supports OpenCode. Run pm init --opencode to install the plugin at .opencode/plugins/pm.ts. It provides equivalent enforcement:

• tool.execute.before — blocks edits without active work
• tool.execute.after — tracks file edits, warns on scope creep
• tui.prompt.append — injects task context into prompts

Multi-agent support

pm supports multiple AI agents working on the same project concurrently. Identity is tracked automatically — you don't need to configure anything beyond pm init.

Agent — which tool is running (e.g., claude-code, opencode). Set automatically by hooks via --agent flag.

Instance — distinguishes concurrent sessions of the same agent. Set automatically via --instance $PPID in hooks, so two Claude Code windows don't collide.

Model — which LLM model is running (e.g., claude-opus-4-6, claude-sonnet-4-6). For Claude Code, captured automatically from the SessionStart hook input and persisted to .pm/identity.json. The prompt-context hook then reminds the agent to pass --agent and --model flags on every pm command, so all tasks, issues, and log entries are tagged. For OpenCode, identity comes from the plugin. The --model flag on any command works as a manual override.

All three are recorded on tasks, issues, and log entries. The TUI displays agent and model together (e.g., claude-code/opus-4-6).

TUI

Run pm with no arguments to open the terminal UI. This is the human interface — while the agent uses CLI commands, you use the TUI to see what's going on.

Navigation: arrow keys to move, Enter to open, Esc to go back, q to quit.

List view shows all features and issues sorted by date, with progress bars, type badges, priority levels, and decision counts. Use / to search, [/] to paginate, d to delete.

Detail view shows phases, tasks, decisions, and status for a feature or issue. Feature descriptions with numbered items are split into separate lines for readability.

Decisions view — press w from the list view to browse all project decisions. Shows decisions sorted newest-first with reasoning and source context. Use / to search and arrow keys to navigate.

CLI reference

Status

pm              Open TUI
pm next         Show the next pending task (priority-aware)
pm list         List all features and issues with progress
pm log [N]      Show last N log entries (default: 20)
pm recap        Session briefing: active work, next steps, recent decisions

Track work

pm add-feature <title> [--description "..."] [--fix]
pm add-phase <featureId> <title>
pm add-task <featureId> <phaseId> <title> [--description "..."] [--files "a,b"] [--priority 1-5]
pm add-issue <title> [--type bug|change] [--priority urgent|high|medium|low] [--description "..."]

Task lifecycle

pm start <id>              Mark task as in-progress
pm done <id>               Mark as done [--note "..."] [--review]
pm error <id>              Mark as failed [--note "reason"]
pm retry <id>              Retry a failed task [--note "context"]
pm review <id>             Request review [--approve|--reject] [--note "..."]

Agent, instance, and model are auto-detected (see Multi-agent support). All commands also accept --agent, --model as manual overrides.

Tasks can require review by adding --review when marking done. Failed tasks support retry with attempt tracking (default max 3 attempts).

Decisions

pm decide <id> "what" [--reasoning "why"]
pm why "search term"

Management

pm show <featureId>        Feature detail with all IDs
pm update <id>             Update properties [--title "..."] [--description "..."] [--priority ...]
pm cleanup                 Reset stuck tasks [--errors] [--drafts] [--all] [--quiet]
pm init                    Initialize pm in current directory
pm hook <type>             Run a specific hook (used internally by Claude Code)

Data storage

All data lives in .pm/data.json — a single JSON file in your project root. No database, no server, no network calls.

The file contains features (with phases and tasks), standalone issues, and an append-only log of all actions. Add .pm/ to your .gitignore — task state is local to each developer.

Other files in .pm/:

• session.json — edit tracking (files touched, scope warnings), resets between tasks
• identity.json — agent/model identity persisted at session start, used by prompt-context to remind the agent which flags to pass

Requirements

• Node.js 18+
• Claude Code or OpenCode — pm supports both AI agents via hooks

License

MIT