feat: add unanswered Slack mention detection to heartbeat

[vltbaudbot]

Summary

Adds periodic checking for Slack app_mention events that didn't receive a reply, helping catch messages lost during agent restarts.

Problem

During control-agent restarts, messages forwarded to dead sockets are lost forever because the Slack bridge acks them before confirming agent receipt. This creates a window where user messages can be silently dropped.

Example incident: On 2026-02-27, 3 messages were lost during a 45-minute restart window (05:03-05:48 UTC).

Solution

This PR adds an unanswered mention check to the heartbeat extension that:

1. Scans bridge logs for recent app_mention events (last hour)
2. Checks for replies by searching:
   • Bridge logs for POST /send or /reply with the thread_ts
   • Control-agent session logs for thread_ts references
3. Reports unanswered mentions older than 5 minutes (grace period for processing)
4. Runs automatically every 10 minutes via existing heartbeat mechanism
5. Zero-token when healthy - only prompts LLM when issues are found

Changes

• Add checkUnansweredMentions() function
• Add hasRepliedToThread() helper for reply detection
• Integrate into main heartbeat flow
• Add HEARTBEAT_CHECK_UNANSWERED_MENTIONS env var (enabled by default)
• Add configuration output for new settings

Testing

Tested locally during incident recovery - successfully detected the 3 lost messages from the restart.

Configuration

# Enabled by default, disable with:
HEARTBEAT_CHECK_UNANSWERED_MENTIONS=0

# Adjust grace period (default 5 min) by editing UNANSWERED_MENTION_THRESHOLD_MS constant

Follow-up Work

This is a detection/monitoring solution. Long-term reliability improvements would require changes to the bridge architecture (two-phase ack, persistent queue, etc) and potentially the broker service itself.

[greptile-apps]

Greptile Summary

Added periodic unanswered Slack mention detection to the heartbeat extension to catch messages lost during agent restarts. The implementation scans bridge logs for app_mention events and checks for replies by searching both bridge and session logs.

Key changes:

• Scans last 500 lines of bridge log for mentions within the past hour
• Checks for replies via POST /send or /reply in bridge logs, or thread_ts references in control-agent session logs
• Reports mentions older than 5 minutes without replies
• Configurable via HEARTBEAT_CHECK_UNANSWERED_MENTIONS env var (enabled by default)
• Zero-token when healthy, only prompts LLM when unanswered mentions are found

Issues found:

• Documentation: New env var HEARTBEAT_CHECK_UNANSWERED_MENTIONS is not documented in CONFIGURATION.md (violates guideline: "When behavior changes, update docs in the same PR")
• Tests: No tests added for new functions (isUnansweredMentionsCheckEnabled(), checkUnansweredMentions(), hasRepliedToThread()) despite existing comprehensive test suite
• Comment on line 18 incorrectly states the env var needs to be set to "1" to enable when it's actually enabled by default

Confidence Score: 4/5

• Safe to merge with minor documentation and testing improvements recommended
• Implementation is solid with proper error handling and follows existing patterns. Deducted one point for missing documentation in CONFIGURATION.md and lack of tests for new functionality. The code is safe and will work correctly, but doesn't meet the project's documentation and testing standards.
• Check that CONFIGURATION.md is updated with the new env var before merging

Important Files Changed

Filename	Overview
pi/extensions/heartbeat.ts	Added unanswered Slack mention detection with log parsing and reply checking. Minor documentation and test coverage issues.

_{Last reviewed commit: ed2c48d}

[greptile-apps]

_{1 file reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

Comment is misleading - should say "enabled by default, set to 0/false/no to disable"

Suggested change

	* HEARTBEAT_CHECK_UNANSWERED_MENTIONS — set to "1" or "true" to enable (default: enabled)
	* HEARTBEAT_CHECK_UNANSWERED_MENTIONS — enabled by default, set to "0" or "false" to disable

Prompt To Fix With AI

This is a comment left during a code review.
Path: pi/extensions/heartbeat.ts
Line: 18

Comment:
Comment is misleading - should say "enabled by default, set to 0/false/no to disable"

```suggestion
 *   HEARTBEAT_CHECK_UNANSWERED_MENTIONS — enabled by default, set to "0" or "false" to disable
```

How can I resolve this? If you propose a fix, please make it concise.

[benvinegar]

Updated this comment for accuracy: HEARTBEAT_CHECK_UNANSWERED_MENTIONS is enabled by default, and can be disabled with 0, false, or no.

_{Responded by pi-coding-agent using openai/gpt-5.}

[greptile-apps]

Simple string search could have false positives if threadTs appears in error messages or logs without being an actual reply. Consider checking for more specific patterns like the thread being mentioned in a tool call or message send context.

_{Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!}

Prompt To Fix With AI

This is a comment left during a code review.
Path: pi/extensions/heartbeat.ts
Line: 411

Comment:
Simple string search could have false positives if `threadTs` appears in error messages or logs without being an actual reply. Consider checking for more specific patterns like the thread being mentioned in a tool call or message send context.

<sub>Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!</sub>

How can I resolve this? If you propose a fix, please make it concise.

[benvinegar]

Agreed — I tightened hasRepliedToThread() to avoid broad file-level string matches. It now:

• parses slack-reply-log.jsonl line-by-line and matches exact thread_ts
• scans recent control-agent JSONL sessions for assistant bash tool calls that are explicit outbound curl ... /send commands with matching thread_ts

I also added tests covering outbound detection and false-positive avoidance.

_{Responded by pi-coding-agent using openai/gpt-5.}