fix: 3 critical bugs + routing prefix strip

[dpbmaverick98]

Summary

• Bug 1 — Channel response drops: Send-then-ack flow in channel clients was non-atomic, causing random response loss after the SQLite migration. Added delivering status to responses table, claim/unclaim API endpoints, and claim-before-send pattern in Telegram client with retry tracking (max 3 attempts) and periodic recovery for stuck deliveries.

• Bug 2 — Inter-agent mention failures: Teammate mentions ([@agent: msg]) were silently dropped due to case sensitivity, missing validation logging, and a weak regex. Added case-insensitive agent ID lookup, detailed failure logging in isTeammate() and extractTeammateMentions(), improved regex for bracket handling, and a validateAgentResponse() helper.

• Bug 3 — Multi-agent reply loss (race condition): When multiple agents completed simultaneously, concurrent conv.pending-- operations caused the conversation to never complete. Added withConversationLock() promise-chain mutex, safe incrementPending/decrementPending operations, and automatic conversation state recovery.

• Bug 4 — @agent routing broken by message prefix: The messages API route prepends [channel/sender]: to incoming messages, which caused parseAgentRouting() to fail since messages no longer start with @agent_id. All messages fell back to the first configured agent. Fixed by stripping the prefix before parsing.

Files Changed

File	Changes
src/lib/db.ts	delivering status, claimResponseForDelivery(), unclaimResponse(), recoverStuckDeliveringResponses()
src/server/routes/queue.ts	/api/responses/:id/claim and /unclaim endpoints
src/lib/conversation.ts	withConversationLock(), incrementPending(), decrementPending(), state validation + recovery
src/queue-processor.ts	Atomic pending counter ops, periodic stuck delivery recovery
src/lib/routing.ts	Case-insensitive matching, detailed logging, improved regex, prefix stripping in parseAgentRouting()
src/channels/telegram-client.ts	Claim-before-send, retry tracking, stale delivery cleanup
docs/bug-fixes/	Full bug analysis and solution documentation
CLAUDE.md	Claude Code guidance for working in this repo

Test plan

• npm run build compiles cleanly
• ./tinyclaw.sh start launches queue processor + API server successfully
• /api/queue/status returns correct stats
• /api/responses/:id/claim and /unclaim endpoints respond correctly
• @agent_id routing works correctly from Telegram (no longer falls back to first agent)
• @team_id routing activates the team leader
• Tested locally with Telegram channel connected

🤖 Generated with Claude Code

[dpbmaverick98]

[jlia0]

thanks this is my bad, the last commit was too aggressive, what are your thoughts on sqlite migration?

[dpbmaverick98]

SQLite fixes durability, but the conversations Map is still in-memory—agent handoffs break on restart.

Proposal: Keep SQLite for the queue + add NATS for agent-to-agent coordination (request-reply pattern). When @coder calls @writer, it just waits for a response instead of juggling the Map. SQLite isn't wasted; it becomes our agent memory layer (RAG, history). Bonus: NATS makes adding a Web UI trivial (built-in WebSocket).

This fits TinyClaw's multi-agent vision without the fragility. Worth me implementing the NATS layer as a follow-up?

Also: Is this PR good to merge as-is? It's my first public contribution—happy to address any feedback!