dropped-messages

Dropped Messages — Diagnosis & Recovery

Operator contract (harness clarity)

• Not config drift: Changes to this script’s heuristics live in scripts/dropped-thread-followup.sh (git-tracked). They do not rewrite config.yaml, ~/.cursor/mcp.json, or Slack tokens unless an operator edits those files.
• Overrides: Prefix the script with env vars (DROP_LOOKBACK_HOURS, DROP_THREAD_REPLY_LIMIT, DROP_EXCLUDE_CHANNELS, DROP_JEFFREY_ONLY_CHANNELS, …). LaunchAgent / plist can set them for stable “my settings.”
• Audit trail: git log -p -- scripts/dropped-thread-followup.sh is the source of truth for why default behavior changed.

When to invoke this skill

• Jeffrey asks "why didn't you respond to X?" or "did you miss my message?"
• The dropped-thread-followup script found drops (actioned > 0)
• You want to proactively audit for coverage gaps
• A thread shows a very long gap before the agent replied

Coverage model

The script ~/.hermes/scripts/dropped-thread-followup.sh monitors two classes:

Class A — Threaded (reply_count > 0)

Messages that started a thread. These ARE visible to the script. A drop occurs when:

• Jeffrey's message is the last in the thread with no bot follow-up (≥30 min)
• Thread went cold: >2h old, no result indicators (PR link, "done", etc.)
• Bot admitted it didn't execute ("did not execute", "only sent an ack", etc.)
• Bot replied with a timeout/overload error (request timed out, timed out before, deadline exceeded, high load / 2064-style messages) — counted as timeout-failure (dropped run), not a resolved answer

Class B — Standalone (reply_count == 0)

Messages Jeffrey posted without any thread forming. These were invisible before the fix. A drop occurs when:

• Jeffrey posted a root message
• Bot added only a reaction (?) but no text reply in the channel within 30 min
• Message is >30 min old

Class C — DM channel

D0AFTLEJGJU is not in config.yaml channel maps; the script appends $JLEECHAN_DM_CHANNEL (default D0AFTLEJGJU) to the scan list after resolving C-channels.

Class D — #all-$USER-ai (C09GRLXF9GR)

Not in the default C-channel fallback (C0AKALZ4CKW + C0AJQ5M0A0Y); add IDs via config.yaml or DROP_CHANNELS. When C09GRLXF9GR is in the resolved channel list, it is skipped unless you override DROP_EXCLUDE_CHANNELS (default is that ID, to avoid nudging a high-churn channel). Clearing exclusions: DROP_EXCLUDE_CHANNELS="" excludes nothing (the script handles empty explicitly, unlike bash :-).

Quick diagnosis steps

1. Run dry-mode audit (non-destructive)

DRY_RUN=1 DROP_LOOKBACK_HOURS=48 bash ~/.hermes/scripts/dropped-thread-followup.sh

Look for DRY_RUN: would nudge lines. Each is a candidate drop (review for false positives, e.g. boilerplate intros).

1b. Include #all-$USER-ai in the audit

DRY_RUN=1 DROP_LOOKBACK_HOURS=48 \
  DROP_CHANNELS="C0AKALZ4CKW C0AJQ5M0A0Y C09GRLXF9GR" DROP_EXCLUDE_CHANNELS="" \
  bash ~/.hermes/scripts/dropped-thread-followup.sh

2. Check DM channel specifically

DRY_RUN=1 DROP_CHANNELS="D0AFTLEJGJU" DROP_LOOKBACK_HOURS=48 \
  bash ~/.hermes/scripts/dropped-thread-followup.sh

3. Manual Slack scan (when script can't reach a channel)

Use MCP: mcp__slack__conversations_history(channel_id="C09GRLXF9GR", limit="2d") Filter for rows where UserID=U09GH5BR3QU (Jeffrey) and ThreadTs==MsgID (root messages).

4. Check session lock drops (silent HTTP 200 drops)

grep "session file locked" ~/.hermes/logs/gateway.err.log | tail -20

If found: the gateway processed the HTTP request but silently dropped the message before the agent saw it. Fix: clear stale .lock files and restart gateway.

Root causes & fixes

Symptom	Root cause	Fix
DM never scanned	D-prefix not in resolve_channels()	Set $JLEECHAN_DM_CHANNEL (hardcoded D0AFTLEJGJU fallback)
Standalone msg missed	reply_count==0 invisible to thread scan	fetch_standalone_user_messages() — now active
Thread replied with timeout	Gateway timeout (timeoutSeconds too low)	Keep timeoutSeconds≤600, maxConcurrent≤3
Session lock	Dead PID holding .lock file	Clear locks + restart (see CLAUDE.md gateway section)
Work done, no ack	Agent completed task but didn't post thread reply	Always post in-thread ack after completing work
Script LOOKBACK too short	Default 8h window misses older drops	Set DROP_LOOKBACK_HOURS=48 for full audits
#all-$USER never audited	Not in default fallback; also excluded when listed	DROP_CHANNELS including C09GRLXF9GR and DROP_EXCLUDE_CHANNELS=""
Long threads misclassified	Default was 20 replies fetched	DROP_THREAD_REPLY_LIMIT (default 200)
Bare word “done” suppressed nudges	Substring done matched tiny replies	Strong phrases + weak words only if agent text is ≥80 chars
Bug hunt / automation noise	Root posts look like unanswered work	Roots matching *Daily Bug Hunt*, *Repos scanned:*, etc. skip cold/stale
#all-$USER teammate threads	Nudges on threads Jeffrey never joined	DROP_JEFFREY_ONLY_CHANNELS (default C09GRLXF9GR); set to "" to disable
Nudge cites wrong line	Used first human message (often boilerplate)	Nudge Original request prefers Jeffrey’s latest message, else latest human

Recovery workflow

When you find a dropped thread, reply in-thread immediately:

1. Acknowledge: "I missed this message at [time] — sorry for the delay."
2. Do the work OR explain the blocker
3. Post proof (PR link / commit / result)
4. Record nudge to prevent double-nudging: the script handles this via state file

Preventing future drops

The cron for dropped-thread-followup.sh runs every 4h by default. After this fix:

• DM channel is always scanned
• Standalone messages are caught if bot doesn't reply within 30 min
• The state file ~/.hermes/logs/dropped-thread-state.json tracks nudged threads

To verify the cron is wired:

grep -r "dropped-thread" ~/.hermes/cron/