perf(slack): skip LLM + cache config + drop reaction — thread+Pensando+trigger in ~1 RTT

[JonasJesus42]

Summary

Fast-path the slack-mcp webhook: send thread + "Pensando..." in ~1 Slack RTT, publish the enriched trigger right after, and stop calling the broken LLM path entirely.

Hot path before

webhook → config lookup (Supabase/Redis/KV)
        → addReaction 👀
        → sendThinkingMessage
        → removeReaction 👀
        → buildLLMMessages (with its own getThreadReplies)
        → isLLMAvailable
        → handleLLMCall (broken in prod — always failed)
        → catch → publishMessageReceived (with ANOTHER getThreadReplies)

Hot path after

webhook → memCache hit (sub-ms after first call)
        → sendThinkingMessage           [thread + Pensando]
        → processAttachedFiles
        → publishMessageReceived        [trigger with thinking_message_ts]

Changes

• config-cache.ts — in-process Map<connectionId, { config, expiresAt }> with 24h TTL on top of the existing Supabase → Redis → KV chain. First webhook for a connection still hits storage; subsequent webhooks are sub-ms. Write-through on cacheConnectionConfig, invalidated on removeCachedConnectionConfig.
• eventHandler.ts — handleAppMention / handleDirectMessage / handleThreadReply are now just: sendThinkingMessage → processAttachedFiles → publishMessageReceived/publishAppMention(..., { thinking_message_ts }). Removed: 👀 reactions, buildLLMMessages, isLLMAvailable, handleLLMCall, per-handler user info lookup. handleMessage no longer pre-processes files so the thinking message fires before Whisper/file-download work.
• event-publisher.ts — publishMessageReceived / publishAppMention accept extras.thinking_message_ts. Payload carries thinking_message_ts and a context-aware reply_instruction telling the agent to SLACK_EDIT_MESSAGE the placeholder (no more stacking another reply via SLACK_REPLY_IN_THREAD). thread_messages is still only fetched for thread continuations.
• trigger-store.ts — trigger descriptions document the edit-vs-reply rule + the new thinking_message_ts field.

Latency

• Top-level DM: ~200 ms (single sendThinkingMessage RTT) — thread + "Pensando..." appear, trigger published right after.
• Thread continuation: same + ~500 ms getThreadReplies inside the publisher.

What's NOT in this PR

• llm.ts, llm-handler.ts, and the unused parts of context-builder.ts (buildContextMessages, formatMessagesForLLM, buildCurrentContent) stay as dead code. Leaving them in keeps the path easy to rewire if/when the decopilot endpoint is fixed. A follow-up commit can prune them.

Test plan

• Deploy; send a top-level DM → thread + "Pensando..." appear within ~300 ms; logs show [Triggers] Notified slack.message.received: channel=<id> shortly after; no [LLM] streamAgent / decopilot stream failed in logs.
• Subscriber agent reads thinking_message_ts from the trigger payload and calls SLACK_EDIT_MESSAGE to replace "Pensando..." with the final answer — only one bot message in the thread.
• Send a reply inside the bot's thread → handler routes to handleThreadReply; trigger payload includes thread_messages with full history.
• Send a DM with an audio file (Whisper enabled) → "Pensando..." still appears before transcription; final trigger carries the transcribed text.
• In triggerOnly mode: same flow, no "Pensando..." (preserved by showThinking check).
• Supabase connection_configs query rate drops to ~1 per connection per 24h.

---

Summary by cubic

Make the Slack webhook fast: send the thread plus a “Pensando...” placeholder in ~1 Slack RTT, then publish the enriched trigger. Drop the broken direct LLM flow and add an in-process config cache to cut storage round-trips.

• New Features

  • Trigger payload adds thinking_message_ts and a context-aware reply_instruction: edit the placeholder via SLACK_EDIT_MESSAGE when present; otherwise use SLACK_REPLY_IN_THREAD. Agents should follow this to avoid stacked replies.
  • thread_messages is fetched only for thread continuations to avoid extra Slack RTTs.

• Refactors

  • Simplified handlers in eventHandler.ts: sendThinkingMessage → processAttachedFiles → publishMessageReceived/publishAppMention. Removed 👀 reactions and all LLM-related calls.
  • Added a 24h in-process mem cache in config-cache.ts on top of Supabase → Redis → KV (write-through on save; invalidated on delete).
  • File processing runs after sending the placeholder so feedback is instant; if audio needs Whisper and it’s disabled, a warning is sent and the placeholder is removed.
  • Updated trigger docs in trigger-store.ts to document the edit-vs-reply rule.

^{Written for commit b689b47. Summary will update on new commits.}