chat-silence-diagnosis
Skill0 starsChanged 20 days agoMITView on GitHub
Use when the assistant is silent or late on a chat channel.
In this repository86 other files it gives its agents.
- ai-assistant-editor-setupSkill
- api-and-interface-designSkill
- banner-designSkill
- brainstormingSkill
- caveman-exploreSkill
- caveman-reviewSkill
- ci-cd-and-automationSkill
- code-auditSkill
- codebase-designSkill
- koordinator-team-dispatchSkill
- constraint-driven-developmentSkill
- context-engineeringSkill
- cpanel-shared-hosting-deploySkill
- cross-profile-cron-opsSkill
- deprecation-and-migrationSkill
- documentation-and-adrsSkill
- domain-modelingSkill
- doubt-driven-developmentSkill
- finishing-a-development-branchSkill
- frontend-designSkill
- git-workflow-and-versioningSkill
- graphic-design-laneSkill
- hermes-gateway-web-uiSkill
- hermes-profile-hygieneSkill
- hermes-profile-provisioningSkill
- hermes-service-persistenceSkill
- hermes-stream-stale-diagnosisSkill
- hermes-team-dashboardSkill
- idea-refineSkill
- incremental-implementationSkill
- indonesian-official-lettersSkill
- internal-commsSkill
- interview-meSkill
- investigate-firstSkill
- kanban-rate-limit-resilienceSkill
- karpathy-guidelinesSkill
- large-project-gated-deliverySkill
- lean-buildSkill
- live-agent-office-operationsSkill
- live-react-runtime-qaSkill
- live-web-release-verificationSkill
- markdown-pdf-pipelineSkill
- marketing-strategySkill
- meta-social-publishingSkill
- meta-threads-oauth-lifecycleSkill
- migrationSkill
- multi-profile-project-orchestrationSkill
- network-operationsSkill
- observability-and-instrumentationSkill
- office-skill-distributionSkill
- organization-site-menu-researchSkill
- performance-optimizationSkill
- php-upload-size-limitsSkill
- planning-and-task-breakdownSkill
- ponytail-auditSkill
- ponytail-debtSkill
- ponytail-reviewSkill
- ponytailSkill
- portable-app-handoffSkill
- receiving-code-reviewSkill
- The first 60 shown.
The file
We split this skill into the rules it gives an agent
5 rules. About 0 are common, 2 are distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 15
When the user says "kenapa kamu terdiam?" / "you never replied", prove WHICH layer stalled before saying anything.
- Distinctive: line 33
When the complaint targets a specific profile ("saya chat koordinator, dia terdiam") and NOT the gateway itself, skip straight to proving …
- Very specific: line 121
If no 429, check whether a turn is STILL RUNNING rather than dead: grep -a 'API call #' .../logs/agent.log | tail — recent timestamps mean …
- Very specific: line 179
Don't declare a turn hung without a timestamp: Run date, then compare to the newest API call # line.
- Very specific: line 182
Don't restate the same diagnosis twice: if the user re-asks within the same incident — check whether the earlier backoff/rebuild already …
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Chat silence diagnosis
When the user says "kenapa kamu terdiam?" / "you never replied", prove WHICH layer stalled before saying anything. Three independent layers can each cause apparent silence — inbound transport, model API, outbound send — and they have different fixes. A reply that is late is not a reply that was lost.
When to Use
- The user says the assistant is silent, unresponsive, or "terdiam" on a chat channel (Telegram, TUI, or any messaging platform).
- A reply arrived far later than the message was sent and the user asks why.
- The user re-asks the same complaint within one incident — re-verify state instead of repeating the previous diagnosis.
- Pre-emptive: before claiming "your message never arrived".
Procedure (in order — each step narrows the cause)
0. Profile-level silence (worker/clone silent while gateway lives)
When the complaint targets a specific profile ("saya chat koordinator, dia terdiam") and NOT the gateway itself, skip straight to proving the profile is alive by MESSAGE GROWTH, not by process existence:
# 1) did MY message arrive? (role='user' rows, newest first)
python3 -c "import sqlite3,time; c=sqlite3.connect('~/.hermes/profiles/<profile>/state.db'); now=time.time(); [print(id, int(now-ts),'s', pmid, repr(body)[:120]) for id,ts,pmid,body in c.execute(\"select id,timestamp,platform_message_id,substr(content,1,120) from messages where role='user' order by timestamp desc limit 5\") ]"
# 2) is it producing? sample count twice ~70s apart — growth = alive
python3 -c "import sqlite3,time; c=sqlite3.connect('...state.db'); print(c.execute('select count(*) from messages').fetchone()[0], int(time.time()-c.execute('select max(timestamp) from messages').fetchone()[0]))" # run, wait 70s, run again- Growth (+N messages, last_age < ~60s) = service ALIVE, reply is queued,
not dead. The single most common false alarm: one session = one turn, so
while the profile is finishing an earlier long job, an incoming chat reply
waits its turn. Explain the queue with evidence instead of restarting.
(No process named after the profile ≠ dead either — the gateway multiplexes;
check PID 1336002-class gateway plus state.db, never
ps | grep <profile>alone.) - Check whether the user's own message exists (id + platform_message_id above). Present-but-unanswered = queued; absent = transport (go to step 1).
- Duplicate platform_message_id rows (same pmid, two ids) = retry duplication, harmless — note it, don't treat as double-processing.
- Sample periods:
messagesolder than ~10 min with no growth while a turn should be running = then continue to steps 2-3 (API layer).
1. Did the message reach the gateway at all?
grep -a 'inbound message' ~/.hermes/profiles/<profile>/logs/gateway.log | tail -25- Per-profile gateway.log, NOT the global
~/.hermes/logs/gateway.log— inbound chat lines live in the profile's log; the global one mostly holds plugin/network lines and can return zero hits (grep the wrong file and you will wrongly conclude every message vanished). - Present with timestamps: this is the evidence that messages arrived. If a message is genuinely absent here, THEN investigate transport (held/redispatch lines, adapter rebuild).
polling heartbeat: N update(s) queued but not consumed (stuck probe)= inbound DID arrive but the adapter loop is not dispatching it — diagnose as a transport/rebuild stall (step 4), not as a missing message.
2. Were replies actually produced?
grep -a 'response ready' ~/.hermes/profiles/<profile>/logs/gateway.log | tailCompare last response ready time vs last inbound message time. The gap IS
the user's perceived silence — quote both timestamps in the report.
3. Check the model API layer (the usual culprit)
For proxy-backed provider errors, distinguish transient upstream failures from credential faults before changing anything. Read the profile's retry settings and auth pool metadata without printing secret values; preserve the configured model and effort. For a bounded one-token probe, read keys from the profile auth store directly into process memory and send one request per distinct credential to its configured base_url, with max_tokens: 1; never put keys in command arguments, output, files, or logs. Space probes briefly and report each HTTP status with a timestamp. A 502/5xx or upstream stream error is transient endpoint/upstream evidence, not proof of invalid credentials; matching 5xx responses from multiple credentials on the same endpoint implicate their shared path, not the individual keys. Do not rotate, reorder, reset, or replace credentials based on a lone 5xx sample. If the bounded probe remains 5xx, report the provider-path escalation as unresolved and re-probe later; do not claim the issue is fixed until a real API call succeeds.
grep -aE '429|RateLimitError|Retrying API call' ~/.hermes/profiles/<profile>/logs/errors.log | grep <date> | tailRetrying API call in 600s (attempt 1/3)means silence is expected, not a hang: default rate-limit backoff is 600s per attempt, up to 3 attempts. Report the remaining wait as an ETA, do not keep polling.Stream stale for 120s ... Killing connection= a hung upstream stream. It retries only a bounded number of times: after ~6 consecutive stale attempts the loop dies withNon-retryable client error: Provider has been unresponsive ... Switch models or start a new sessionand the turn ends without a reply. Count stale lines since the last successfulAPI call #(grep -a 'Stream stale' .../agent.log | grep <date> | wc -l); a rising count means the provider is dead for that model/session, not that progress is slow — recommend a fresh session rather than waiting, and NEVER a model switch: the abort text itself suggests "Switch models or start a new session", but model/effort changes require explicit owner approval, so a new session on the SAME model is the only autonomous remedy. A /new also does not repair the provider path — it only clears the aborted turn.- Behind a proxy endpoint, also read the timeout knobs:
providers.<p>.stale_timeout_seconds/request_timeout_seconds(grep the profile config — current fleet setting 120/120; earlier 300/300 produced minutes of mute per attempt). Every attempt against a stream the proxy holds open burns the full stale window before the retry, so a provider that hangs (not 429s) turns into MINUTES of apparent per-turn silence with no error in chat. When stale counts rise while the proxy answers 200 on a probe, the diagnosis is "provider hangs → each attempt waits out stale_timeout", not "gateway dead" — report the wait as ETA and name the knob. Changing these knobs fleet-wide still needs owner approval first; once approved, set both keys on EVERY operational profile includingdefaultand verify by per-profile readback plus a zero-Stream stalewindow inerrors.log.- A card crash landing at EXACTLY the old stale window (last heartbeat → worker exit =
stale_timeout_seconds) is the watchdog killing a stream-hung run — but compare the run's SPAWN time against your config-change time before crediting your fix: config is read when the worker spawns, so a run spawned before the change keeps the OLD budget and dies once at the old point, while the dispatcher's auto-retry spawns the next run on the new config. Confirm the fix by watching that next run pass the former crash point (heartbeat continuity + zero stale lines), not by re-rolling the knob or re-diagnosing the crash as a new failure.
- A card crash landing at EXACTLY the old stale window (last heartbeat → worker exit =
- If no 429, check whether a turn is STILL RUNNING rather than dead:
grep -a 'API call #' .../logs/agent.log | tail— recent timestamps mean healthy, just slow.Turn endedafter the last inbound = turn finished; absence after inbound = still in flight.
4. Network/transport layer (secondary — verify before blaming it)
grep -acE 'ConnectTimeout' ~/.hermes/logs/gateway.log # count
# adapter rebuilds = fatal, they delay polling:
grep -aE 'Fatal telegram adapter error|queued for background reconnection' ~/.hermes/logs/gateway.log | tailTelegram ConnectTimeout flapping is common and self-healing — messages still
arrive through it (proved by step 1). Only cite it as the cause when step 1
shows messages MISSING or held. An adapter rebuild (updater.stop() did not
finish) delays polling but queues rather than drops updates.
5. Check for send-side blocks (only when producing but not delivering)
Forbidden: bot was blocked by the user on startup notifications = user
blocked the bot at some point; harmless for inbound, fatal for outbound. Note
it, ask the user to unblock — do not treat it as the cause of missed messages
that DID get inbound-logged.
6. Report format the user expects
Evidence-first: a compact timeline (inbound vs response-ready timestamps), root cause named with its log line quoted, ETA or current state, then remedy if any. Separate confirmed facts from hypothesis.
7. Laporan terjadwal (cron) tidak sampai ke chat
Laporan periodik yang "jalan" (cron aktif, exit succeeded) belum tentu sampai ke user — lapisan ini berbeda dari inbound.
- Cek target delivery, bukan status job. Job cron dengan
deliver: origintapi tanpa chat target yang tersimpan mencatat deliverydelivereddengancontentKOSONG: pesan terkirim tanpa isi. Perbaiki dengan target eksplisit —hermes cron edit <job-id> --deliver telegram:<chat_id>— laluhermes cron run <job-id>untuk menguji sekali. - Tabel delivery bukan bukti isi. Kolom
contentbisa kosong pada SEMUA baris walau pengiriman nyata terjadi, jadi panjang kolom itu tidak boleh dipakai untuk menyimpulkan "tidak terkirim" maupun "terkirim". Receipt yang sah ada di log gateway: carimessage_idyang dikirim ke chat target. Kalau log tidak punyamessage_iduntuk target itu, pengiriman ke user belum terbukti sama sekali. - Konsekuensi laporan: kalau isi laporan tidak bisa diverifikasi diterima, kirim laporan tahap langsung di chat dan jadikan cron hanya cadangan.
Pitfalls
- Never diagnose from one log. Inbound lives in profile gateway.log, API errors in errors.log, turn progress in agent.log, network in global gateway.log — grep a single file and you get a plausible wrong story.
- Network noise ≠ model failure. Telegram ConnectTimeout is the transport to Telegram, not the LLM (see kanban-rate-limit-resilience); but dismissing it as "just noise" without step 1 is equally wrong — prove messages arrived.
- A rate-limited turn is not a dead turn. While
Retrying API call in 600sis pending, the turn holds the session and will resume; restarting the gateway or re-sending the message mid-backoff wastes quota and duplicates work. - Don't declare a turn hung without a timestamp. Run
date, then compare to the newestAPI call #line; API latencies of 20–70s per call are normal for large-context turns, and multi-call turns legitimately run minutes. - Don't restate the same diagnosis twice if the user re-asks within the
same incident — check whether the earlier backoff/rebuild already resolved
(look for
recovered/confirmed healthylines) before repeating the old cause. - Don't judge liveness from
ps. A profile can be fully busy with zero processes matching its name (gateway hosts the turn); message-count growth over a ~70s window is the only reliable liveness proof.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

