Agent PM

A Slack-native bot that turns product ideas into PRDs and tickets. Post an idea (app mention or DM) → a pi agent drafts a structured PRD in the thread → a human approves with a ✅ reaction → the bot files one GitHub issue per user story and posts the links.

This is a full rewrite of the previous Python/FastAPI service, which is preserved on the archive/python-legacy branch.

How it works

1. Draft — mention the bot (@Agent PM an idea for…) or DM it. The bot posts a "drafting…" placeholder, runs the agent, and updates the message with the PRD draft plus a "React with ✅ to file N issues" line. The bot seeds its own ✅ reaction for one-click approval.
2. Revise (optional) — reply in the same thread with follow-ups ("make the non-goals stricter"). The bot fetches the thread history and produces a revised draft, which replaces the pending one. Only the latest draft in a thread is approve-able; approving a superseded draft gets a "stale draft" note.
3. Approve — a human adds ✅ to the latest draft. The bot files one GitHub issue per user story (deterministic code, not the LLM) and replies with the results: "filed X of N", issue links, and any per-story failures.

Nothing is ever filed without an explicit human approval. DRY_RUN (default true) is a global safety override on top of that: even approved drafts only produce would-be issue payloads in the thread. Set DRY_RUN=false to enable real issue creation.

Caveat: pending drafts live in memory. A process restart loses them — old draft messages will simply no longer be approve-able.

Quick start

1. Create the Slack app (at api.slack.com/apps):

   • Enable Socket Mode and create an app-level token with the connections:write scope → SLACK_APP_TOKEN (xapp-...).
   • Under OAuth & Permissions, add bot scopes: app_mentions:read, chat:write, reactions:read, reactions:write, im:history, im:read, im:write. Install the app to your workspace → SLACK_BOT_TOKEN (xoxb-...).
   • Under Event Subscriptions, subscribe to the bot events app_mention, message.im, and reaction_added.

2. GitHub token: a token (fine-grained PAT or GitHub App token) with Issues: write on the target repository.

3. Configure and run:

   bun install
   cp .env.example .env   # fill in the values
   bun run dev            # or: bun start

Configuration

Variable	Required	Default	Description
SLACK_BOT_TOKEN	yes	—	Slack bot token (xoxb-...)
SLACK_APP_TOKEN	yes	—	Slack app-level token for Socket Mode (xapp-...)
PROVIDER	no	openai	LLM provider: openai, anthropic, or google
provider API key	yes	—	OPENAI_API_KEY, ANTHROPIC_API_KEY, or GEMINI_API_KEY depending on PROVIDER
MODEL	no	per provider	gpt-4o-mini (openai), claude-sonnet-4-6 (anthropic), gemini-2.5-flash (google); startup fails on unknown models
AGENT_TIMEOUT_MS	no	180000	Abort an agent run after this many milliseconds
GITHUB_TOKEN	yes	—	Token with Issues:write on GITHUB_REPO
GITHUB_REPO	yes	—	Target repository as owner/name
DRY_RUN	no	true	Global safety override: approved drafts produce would-be payloads only. Only false/0/no disables.

Architecture

• src/slack.ts — Bolt app in Socket Mode. Draft → approve → file flow with an in-memory pending-draft store, per-conversation run serialization, event-id dedup, and structured JSON run logs. Formatting and state helpers are pure, Bolt-free, and unit-tested.
• src/agent.ts — the pi agent only drafts (sole tool: terminal submit_prd); filing lives in deterministic fileIssues(prd, github, dryRun). LLM proposes, code disposes.
• src/github.ts — a minimal GitHub issues client built on plain fetch (no octokit).

Development

bun test             # unit tests (no network or API keys needed)
bun run typecheck    # tsc --noEmit
bun run lint         # biome check .

License

MIT — see LICENSE.