feat(slack-app phase 3b): Anthropic agent loop + read-only memory tools

[TS00]

Summary

Phase 3b — replaces the Phase 3a canned reply with the actual agent. The bot now answers Slack messages by reading and searching the requesting user's Reflect memories.

Full design: `docs/eng-plan-slack-app-v1.md`. Parent memory `d959bc61`.

New modules

• `src/slack-conversation-service.ts` — per-thread state in `slack_conversation_state`. 24h TTL, MAX 20 user/assistant turn pairs (older truncated). Stores text-only — tool-use turns aren't persisted across messages (they bloat fast and don't help on the next turn). Lazy `pruneExpiredConversations` on every save (cheap, indexed).
• `src/slack-agent-tools.ts` — 8 read-only memory tools: `get_memory_briefing`, `search_memories`, `read_memories`, `get_memories_by_tag`, `get_memory_by_id`, `read_team_memories`, `read_thread`, `get_graph_around`. Each is a thin shim over the existing memory-service / memory-graph functions, scoped to the resolved Reflect user, vendor='slack', limits clamped to 25.
• `src/slack-agent.ts` — `runSlackAgentTurn`. Builds the system prompt (mode + speaker), runs the Anthropic SDK loop, executes tool calls in parallel per step, caps at 6 steps. Default model `RM_SLACK_AGENT_MODEL` or `claude-sonnet-4-7`. Anthropic client is injectable for tests.

Wiring

`slack-events-handler.ts`: handler now loads the LLM key for the workspace's scope, loads conversation history for the thread, runs the agent turn, persists the new history, posts the reply. Falls back to a polite "set an LLM key in dashboard" message when no key is configured. Audit metadata gains `tool_calls` / `agent_steps` / `stop_reason`.

Tests (+20, 367/367 total)

slack-conversation: roundtrip, update-not-duplicate, MAX_TURNS truncation, prune of expired rows, fail-closed on corrupted JSON, defensive filtering of malformed message shapes.

slack-agent: every tool's dispatch path against a real DB + end_turn + tool_use loop with a stubbed Anthropic client + empty-text fallback + system prompt encoding (DM vs channel, speaker).

Manual prereq for end-to-end validation

Before the agent can reply on dev, you need to set the Anthropic API key:

1. Open https://dev.reflectmemory.com/dashboard/connections/slack
2. In the LLM provider key section, paste your Anthropic key (sk-ant-...) and click Save.
3. Without this, the bot will reply with a polite "set an LLM key in dashboard" message instead.

Test plan

• CI green
• Dev deploy succeeds
• Set Anthropic key in dashboard
• DM the bot something open-ended ("what have I been working on?") → expect a real response that references your memories by title
• @mention in a channel with a specific question ("what did we decide about X?") → expect a tool call to search_memories or get_memory_briefing then a focused reply
• Continue the same thread with a follow-up → expect the agent to remember context (24h TTL on conversation state)
• DM without an LLM key set → expect the polite fallback message

Made with Cursor