Praktor

Boule deliberates; Praktor acts.

Praktor (Greek πράκτωρ, "the doer / agent that accomplishes") is an autonomous CLI that implements the Tasks produced by Boule. It claims a ready Task from GitHub, drives a Claude Agent SDK coder to write the change and its tests, runs the repo's own checks, opens a PR, and coordinates with peer runners through GitHub Discussions — all GitHub-native, no server or database.

How it fits with Boule

| | Boule | Praktor | |---|---|---| | Role | Product/program management | Software engineering | | Output | Designs → Requirements → Tasks (typed Issues) | Code + tests + Pull Requests | | Writes | GitHub Issues/Projects/Discussions only | The working tree (Read/Write/Edit/Bash), then a PR |

Praktor respects Boule's process: it only picks up Tasks that are status:accepted, whose native blocked_by dependencies are all closed, and that aren't already taken (praktor:in-progress). Each Task's Verifies: #<REQ> link gives the coder the requirement(s) — with their Gherkin acceptance criteria — that the change must satisfy. It honors Boule's boule:halt kill-switch.

Commands

praktor doctor              # validate config, creds, repo + coordination category
praktor next                # list Tasks ready to implement (accepted + unblocked)
praktor status              # snapshot: accepted / ready / in-progress / claims
praktor implement [task]    # claim one ready task, implement it, open a PR

[task] is a #number or a boule-id; omit it to take the first ready task.

Configuration

Praktor has its own identity and config — it never borrows Boule's credentials:

• PRAKTOR_REPO / GITHUB_REPOSITORY — target repo owner/name
• Auth: Praktor's own GitHub App via the PRAKTOR_APP_* trio (PRAKTOR_APP_ID, PRAKTOR_APP_INSTALLATION_ID, PRAKTOR_APP_PRIVATE_KEY), or a fine-grained PAT (PRAKTOR_GITHUB_TOKEN / GITHUB_TOKEN)
• CLAUDE_CODE_OAUTH_TOKEN (or ANTHROPIC_API_KEY) for the agent

Global flags: --repo, --project, --budget <usd>, --max-turns <n>, --dry-run, --json, -v.

Coordination (Discussions)

Before working a task, Praktor posts a claim in the Agent Handoffs Discussion category (configurable) with a machine-readable marker, and re-checks after claiming to yield on a race. Claims expire after a TTL so a crashed run never deadlocks the backlog. This is a cooperative lock, not a transaction — sized for a handful of parallel runners.

License

MIT