feat: draft → approve → file interaction, thread context, robustness

[haasonsaas]

Summary

Reworks the bot's interaction model and hardens the runtime, per the approved improvement list. Nothing is filed by the LLM anymore; filing requires an explicit human ✅ and (for real issues) DRY_RUN=false.

A. Interaction model: draft → approve → file

• Deterministic filing — the create_github_issue tool is gone; the agent's only tool is the terminal submit_prd. New fileIssues(prd, github, dryRun) loops prd.userStories and files one issue per story (title = story title, body = description + acceptance-criteria checklist). LLM proposes, code disposes.
• Approval gate — the bot posts the PRD draft with a "React with ✅ to file N issues" line, seeds its own ✅ for one-click approval, and listens for reaction_added. A human ✅ on the latest draft triggers filing; the bot's own reaction and bot users are ignored.
• DRY_RUN stays a global override — with DRY_RUN=true (default) the approval flow works end-to-end but only posts would-be payloads.
• Partial failure visibility — per-story errors are caught, filing continues, and the results reply reports "filed X of N" plus each failure's error message.

B. Thread context

• Follow-ups in a thread (or DM) fetch history (conversations.replies; conversations.history for flat DMs), bot messages marked as assistant context, and pass it into runPrdAgent(idea, deps, history?) — so "make the non-goals stricter" revises the pending draft.
• A revised draft supersedes the pending one (posted as a new reply, no editing). A ✅ on a superseded draft gets a "stale draft" note. Pending drafts are in-memory only — a restart loses them (documented in README).

C. Robustness

• Run timeout — agent.abort() after AGENT_TIMEOUT_MS (default 180000), surfaced as a "timed out" reply. Deviation: pi-agent-core 0.82.0's prompt() has no maxTurns option (checked dist/agent.d.ts — prompt(input, images?) only), so no turn cap; timeout only.
• Truncation — replies capped at 3900 chars + marker (Slack's ~4000 limit).
• Event dedup — in-memory FIFO set of event_ids (cap 1000); duplicates dropped silently.
• Serialization — runs chained per conversation (channel+thread); different conversations run in parallel.
• Placeholder ts is now checked before chat.update (the old ts ?? "" wart is gone).

D. Config + logging

• PROVIDER = openai (default) | anthropic | google → pi-ai's openaiProvider/anthropicProvider/googleProvider (export names verified in dist). API key env per provider (OPENAI_API_KEY/ANTHROPIC_API_KEY/GEMINI_API_KEY); model defaults gpt-4o-mini / claude-sonnet-4-6 / gemini-2.5-flash — all three verified to resolve non-undefined via models.getModel() at runtime.
• Structured run logging — one JSON console.log per completed run: ts, conversation, durationMs, storyCount, dryRun, outcome (drafted|filed|failed), error.

E. Docs

• New root AGENTS.md: pi conventions (Type from pi-ai, errors-don't-throw, fresh Agent per run, lockstep version pinning), dry-run discipline (approval AND DRY_RUN=false), and the no-subsystem-without-producer-and-consumer rule.
• README rewritten for the new flow, new env vars, in-memory caveat, and the two new Slack scopes (reactions:read, reactions:write) + reaction_added event subscription the approval flow needs.

Test plan

bun test — 47 pass / 0 fail, no network or keys: filing loop incl. partial failure and dry-run, history prepend into the outgoing prompt, run timeout via abort, truncation, dedup incl. FIFO eviction, queue serialization + post-failure recovery, draft store supersede/stale semantics, provider config + model defaults + timeout parsing. bunx tsc --noEmit, bunx biome ci ., and bun install --frozen-lockfile all clean. src/ is 698 lines.

Not verified

Live Slack/GitHub E2E still hasn't been run (no workspace/tokens here). First real run should keep DRY_RUN=true and exercise draft → ✅ approve → would-be payloads before enabling real filing.