audit(2026-11-01): Conduit updated to 0.122.2 (memory-timeout patch now upstream, call store re-shaped); document the local-voice-mode fault, the force-drained gateway reload that orphaned a cron run, and both dead Asana paths; 2026-10-31 settled 160/4

This commit is contained in:
Tony0410
2026-11-01 00:30:42 +08:00
parent 74a5b2d117
commit 0c826f2b5f
8 changed files with 488 additions and 22 deletions
@@ -0,0 +1,102 @@
---
title: Asana Is Unreachable From CT460 on Both Paths — MCP Server Never Authenticated, Hosted Connector Rejects the Call Shape
type: incident
status: active
created: 2026-11-01
updated: 2026-11-01
verified_on: 2026-11-01
review_after: 2026-11-08
confidence: high
tags: [hermes, automation, mcp, connector, credential, silent-failure]
sources: [live-mcp-list, live-config-yaml, live-mcp-tokens, live-logs, live-probe]
---
# Asana Is Unreachable From CT460 on Both Paths — MCP Server Never Authenticated, Hosted Connector Rejects the Call Shape
## Headline
The estate has **two independent Asana integrations on CT460, and neither can complete a read** — which
went unnoticed until an interactive session spent 2026-10-31 evening retrying both to answer a single
question ("where does Remembrance Day stand?").
| Path | Surface | Live state 2026-10-31/11-01 |
|---|---|---|
| **MCP server `asana`** | `https://mcp.asana.com/v2/mcp`, `auth: oauth`, `enabled: true` (`hermes mcp list` → ✓ enabled) | **Never authenticated.** Every attempt logs `MCP OAuth setup failed for 'asana': … non-interactive environment and no cached tokens found. Run \`hermes mcp login asana\` interactively first to complete initial authorization.` |
| **Hosted connector** | `connectors__asana__GET_A_TASK`, `…__SEARCH_TASKS_IN_WORKSPACE` | Reaches Asana but returns **`422 INVALID_ARGUMENTS` — "Input should be a valid list on parameter \`opt_fields\`"**. 12+ identical retries in one session. |
## Evidence
**MCP server — no token, ever.** `~/.hermes/mcp-tokens/` holds exactly one file:
```
-rw------- 1 hermes hermes 319 Oct 26 01:07 asana.client.json
```
A registered **client** with **no token file** → the OAuth flow was started (client registered
**2026-10-26 01:07**) and never completed. `config.yaml → mcp_servers.asana` reads
`{auth: oauth, enabled: true, url: https://mcp.asana.com/v2/mcp}`. The earliest Asana trace in retained
logs is an **HTTP 401** from `mcp.asana.com/v2/mcp` at **2026-10-27 16:35:05**, followed immediately by
OAuth discovery calls.
**Failure volume** (log lines across `logs/agent.log*` + `logs/errors.log*`; the failure is written to
both files, so line counts roughly double the number of events):
| Day | `MCP OAuth setup failed for 'asana'` lines |
|---|---|
| 2026-10-21 | 455 |
| 2026-10-22 | 469 |
| 2026-10-23 | 576 |
| 2026-10-24 | 576 |
| 2026-10-25 | 576 |
| 2026-10-26 | 301 |
| 2026-10-27 | 439 |
| 2026-10-28 | 1 150 |
| 2026-10-29 | 1 154 |
| 2026-10-30 | 862 |
| **2026-10-31** | **1 156** |
| 2026-11-01 (to 00:16) | 14 |
| **Total retained** | **7 728** |
The earliest retained line is **2026-10-21 05:09:43** (`errors.log.2`) — the gap **predates the retained
log window**; this is a chronic condition, not a 10-31 regression. The 10-31 peak coincides with the
evening Pulse session hammering the path.
**Hosted connector — the call, not the credential.** From `20261031_210329_b35440` (the Conduit Pulse
"Remembrance Day" session), repeated at 22:17:47, 22:20:42, 22:26:31, 22:31:25, 22:36:45, 23:14:45,
23:26:46, 23:28:55, 23:30:34, 23:39:31, 23:41:48, 23:51:19 — each call burning 11–20 s:
```
tools.managed_gateway_identity: connector connectors.execute 422 INVALID_ARGUMENTS
Tool tool_call returned error: {"results":[{"name":"connectors__asana__GET_A_TASK","error":
{"code":"INVALID_ARGUMENTS","message":"Invalid request data provided\n- Input should be a valid list
on parameter `opt_fields`…"}}]}
```
The connector is **authorised** (it is the MCP path that cannot authenticate), so the block is purely
argument shape: `opt_fields` must be a **JSON list**, and a string/comma-form is rejected.
## Impact
- **Non-interactive sessions (cron, background jobs, Conduit background investigations) cannot read
Asana at all** — the MCP path needs an interactive browser login that cron cannot provide.
- **Interactive sessions** get the hosted connector, but the agent re-issued the same mis-shaped call
a dozen times instead of adapting — ~3 minutes of model time spent failing identically.
- Asana-backed user requests (task status behind the Conduit Pulse flow) degrade to guesses.
## Remediation (not performed — this audit records, it does not re-authorise)
1. Run **`hermes mcp login asana`** from an interactive session to complete the OAuth flow; confirm
`~/.hermes/mcp-tokens/asana.json` (or equivalent token file) appears afterwards, then re-probe the
MCP server.
2. **Or** disable the `asana` MCP server if the hosted connector is the intended path — two Asana paths
on one host is itself the source of the confusion, and one of them has never worked.
3. For the connector: pass **`opt_fields` as a JSON list**, not a comma-separated string.
4. Watch for the silent-cost pattern: a path that fails **every** time still burns a tool call each
time. 1 100+ failures/day means the failure is being retried, not fixed.
## Related
- [[systems/hermes-agent]] — MCP server inventory and auth state
- [[systems/conduit]] — the Pulse flow that needed the Asana task status
- [[systems/scheduled-tasks]] — cron fleet (the surface that cannot authenticate interactively)
- [[incidents/2026-08-08-gitea-config-backup-push-dead]] — the same "green row, dead destination" class
@@ -0,0 +1,103 @@
---
title: Conduit "voice mode: local" never applies — Mac mini speech server's LLM proxy is disabled
type: incident
status: active
created: 2026-11-01
updated: 2026-11-01
verified_on: 2026-11-01
confidence: high
tags: [conduit, voice, local-live, mac-mini, speech-server, investigation]
sources:
- session 20261031_193709_40b9c2 (Conduit voice session, started 2026-10-31 19:37:09 AWST)
- ~/.hermes/conduit-settings.json, ~/.hermes/conduit-calls.json, ~/.hermes/config.yaml
- live probes of the Mac mini speech server (LLM endpoint, /v1/usage, realtime socket)
---
# Conduit "voice mode: local" never applies
**Status: OPEN — investigation in progress. Nothing was changed on the Mac mini speech server.**
On 2026-10-31 Anthony reported (from the phone): switching the Conduit app's voice
setting to **"local"** shows a **successful** connection test, but the change never
takes effect. A background job launched from that voice conversation
(`20261031_193709_40b9c2`, started **19:37:09 AWST**, 30 messages / 17 tool calls) was
still open at the 2026-11-01 00:2x audit.
## Why the test succeeds and the change still does nothing
The "test connection" button checks **socket reachability only** — it never exercises
the model path. So it passes while every local turn is still impossible:
| # | Finding | Evidence |
|---|---------|----------|
| 1 | The Mac mini speech server's **LLM proxy is disabled** | its LLM endpoint answers **HTTP 501** — *"The LLM proxy is disabled. Start the server with `--enable_llm_proxy` to enable it."* |
| 2 | The Mac has never served a local turn | `/v1/usage` → **`llm_proxy.requests: 0`** across **25** observed turns |
| 3 | The realtime handshake is not understood | the Mac's realtime socket **accepts the connection** but rejects **`session.update`** as an **unknown event** |
| 4 | The app's test button proves nothing about the model path | it returns success on socket reachability alone |
Findings 1–3 are the strongest single explanation: **even with `voice.mode = local`
set on the phone, the Mac cannot call a model back.** Finding 4 is why the UI reports
success.
## `voice.mode` is phone-side, not host-side
The host never reads a `voice.mode` key — there is none to read:
- `~/.hermes/config.yaml → voice` holds **`auto_tts`**, **`local_live`**,
**`speech_server`** — **no `mode`**.
- `~/.hermes/conduit-settings.json` contains **no `voice.*` keys at all**; its only
device state is `devices/<UUID>/mode: small` for three devices.
- The device itself reports **`voice.mode: "classic"`**.
So "set mode to local and the change does not apply" cannot be diagnosed on the host
config — the selector lives in the app, and what the host does with it is a
**`local_live`** path decision.
## A misread that was caught mid-session (worth keeping)
`local_live.py:882` is the **only** writer of `_local_live_last_request`, and it sits
inside the **chat-completions route**. A "recent `last_request_at`" reading during the
investigation therefore indicated **this investigation's own probe**, not phone
traffic — the agent initially read it as phone activity and corrected the call in the
same session. Live check: `conduit-calls.json → last_request_at` is **`null`**.
Lesson: on this integration, `last_request_at` is proof of *some* HTTP request hitting
the local-live route, **not** proof of a phone call.
## Configuration at the time of writing
Changed during the 2026-10-31 session (`config.yaml` mtime **21:02:30**):
- `voice.local_live.front_model` = `{base_url: http://100.78.63.20:20128/v1, model: mistral/voxtral-small-latest}`
— was `mistral/ministral-14b-latest`.
Unchanged: `voice.speech_server.url = wss://…:8443/v1/realtime`, `voice.auto_tts: false`,
`tts.provider: ollama_local`, `stt.provider: local`.
The Conduit dashboard gateway was restarted at **19:35:02** (pid 176771), minutes
before the session opened.
## Prompting context (candidate, not proven)
The plugin **self-updated to `0.122.2` on 2026-10-31 18:05**, which drove a gateway
reload at **18:06:41** → [[incidents/2026-10-31-gateway-reload-force-drain-orphaned-run]].
A newly shipped local-live path is a reasonable regression candidate, but the
disabled LLM proxy would break local mode on **any** plugin version, so the version
change is not yet implicated.
## Open items
- **Enable the Mac mini's LLM proxy** (`--enable_llm_proxy`) or confirm the intended
Local Live architecture — with the proxy off, finding 1 alone explains the symptom.
- Confirm the **realtime event vocabulary** the plugin expects vs. what the Mac's
socket implements (`session.update` is currently rejected as unknown).
- Decide whether `front_model` should be a **text** or **audio** model — a
chat-completions `base_url` pointed at an audio model may itself be the mismatch.
- Re-test from the phone after any change; the app's own test button is not evidence.
## Related
- [[systems/conduit]]
- [[incidents/2026-10-31-gateway-reload-force-drain-orphaned-run]]
- [[incidents/2026-10-29-conduit-voice-silence]]
- [[incidents/2026-10-22-conduit-memory-start-timeout]]
@@ -0,0 +1,101 @@
---
title: Gateway Reload on 2026-10-31 Force-Drained After Its Full 1 800 s Cap and Orphaned a Live Cron Run — Mnemosyne Health Monitor (6 h 38 m in)
type: incident
status: active
created: 2026-11-01
updated: 2026-11-01
verified_on: 2026-11-01
review_after: 2026-11-08
confidence: high
tags: [gateway, cron, scheduler, lifecycle, restart, memory, silent-failure]
sources: [live-logs, live-gateway-ledger, live-sqlite-cron-executions, live-sqlite-cron-incidents, live-sqlite-deliveries, live-probe]
---
# Gateway Reload on 2026-10-31 Force-Drained After Its Full 1 800 s Cap and Orphaned a Live Cron Run — Mnemosyne Health Monitor (6 h 38 m in)
## Headline
On **2026-10-31 at 18:06:41 AWST** the gateway received a **`SIGUSR1` reload** (`systemctl reload` /
`hermes gateway restart`). It deferred its own stop for in-flight work — and this time it **waited the
whole deferral cap**, hit it at **18:37:38** with *"drain timed out with interrupted agents"*, and
force-restarted. The casualty was `ea79a2107e36` **Mnemosyne Health Monitor** (12-hourly `0 */12 * * *`),
whose 12:01:06 slot had been running for **6 h 38 m 09 s**. It closed **`failed`** at 18:39:15 on
**`RuntimeError: Cron job 'Mnemosyne Health Monitor' lost its durable fire claim ownership`**.
This is the **third variant of the orphan class** — after the 2026-10-20 *unclean crash* exit and the
2026-10-26 *planned `SIGTERM` whose drain timed out after 1.35 s* — and the first where the drain
mechanism behaved as designed and still ended in a lost run.
## Timeline (live gateway log, `~/.hermes/logs/agent.log`)
| Time (AWST) | Event |
|---|---|
| 18:06:41.772 | `SIGUSR1 received (systemctl reload / hermes gateway restart): performing a graceful gateway restart — drain active turns, exit, supervisor relaunches.` |
| 18:06:41.788 | `Restart requested with 2 active work unit(s); deferring stop() until they finish (cap=1800s) so in-flight turns are not amputated (#77184)` |
| 18:06:41.899 | Deferred on **1** unit: the cron unit `ea79a2107e36` (`elapsed_s: 21950.7`, `external: True`, `wedged: False`, **`restart_safe: True`** — *excluded* from the wait) plus `{'kind': 'managed', 'pid': 397}` |
| 18:07:11 → 18:11:12 | Countdown logged every 30 s: 1770 s → 1529 s remaining, same single `managed pid 397` unit |
| 18:11:23 | **Second `SIGUSR1`** arrives while the first restart is still deferred |
| 18:20:13 | `Restart deferred: … (988s remaining before force drain): [{'kind': 'managed', 'pid': 397}]` |
| **≈18:36:41** | **Cap reached** (18:06:41 + 1 800 s) |
| 18:37:38.341 | `Skipping .clean_shutdown marker — drain timed out with interrupted agents; next startup will recover their interrupted turns.` |
| 18:38:43.284 | New gateway starts; 18:39:14 reconnection watcher for 1 failed platform (`relay`) |
| 18:39:15.208 | Failure **delivered** (delivery row `9e7da0fa…`, `for_failure=1`, status `delivered`) |
| 18:39:15.563 | Execution `9e7da0fafd2847b69643c6f235cb8651` closes **`failed`** — 6 h 38 m 09 s |
| 18:39:23 / 18:45:23 / 19:00:22 | Three derivative `96eb95ce12f2` (log-error watchdog) reds quote the cron error; the watchdog clears on its 19:15:21 slot |
| 19:23:39 → 19:24:40 | **Second restart, clean**: `drain took 0.00s, timed_out=False`; `Previous gateway exited cleanly — skipping session suspension` |
Current gateway: `hermes-gateway.service` active since **19:23:56**, PID `175269` (4 h 43 m at the
2026-11-01 00:07 probe). Nothing failed for the remaining ~5 h of 10-31 (every `96eb95ce12f2` and
`0c771463c3cd` slot from 19:15:21 onward `completed`).
## Incident-ledger effect
| Row | Job | State | First seen | Last seen |
|---|---|---|---|---|
| `ea79a2_9197df32320b` | `ea79a2107e36` | **`alerted` (open)** | 2026-10-31T18:06:44 | 2026-10-31T18:06:44 |
| `ea79a2_a4661554a3c2` | `ea79a2107e36` | **`alerted` (open, reused row)** | **2026-10-26**T14:35:59 | 2026-10-31T18:39:15 |
| `0e753e_e7f8194a7d06` | `0e753e9720fd` | `detected` (open, unchanged) | 2026-10-01T09:00:55 | 2026-10-01T09:00:55 |
The `error_sig` match re-opened the **2026-10-26 row** rather than filing a fresh one, so two rows now
describe one event: the earlier *"Cron execution unverified; reconcile `admissions/292056e5742c1ba4…`"*
(`failure_type: unknown`) and the definitive *fire-claim* row. **Unclosed incident rows 1 → 3.**
## What is genuinely new here
1. **The deferral cap is reachable in normal operation, not just under a stuck process.** The 10-26
incident drained in **1.35 s**; this one waited the full **1 800 s** because a `managed` unit
(`pid 397`) stayed registered. Force drain is therefore a *routine* outcome of any reload that
overlaps a long-running cron job.
2. **`restart_safe: True` did not protect the execution.** The scheduler explicitly classified the cron
unit as restart-safe and *excluded* it from the wait — and the run still died with a lost claim. The
flag protects the *scheduler's* bookkeeping, not the running agent's lease.
3. **A `SIGUSR1` reload is not a lighter-weight restart.** The pre-restart log shows plugin/update
machinery active minutes earlier (`.update-<id>/plugin` capability checks at 18:05:41 and 18:19:53,
`plugin scan passed` at 18:20:22) — the reload behaved exactly like a full stop/start for the
in-flight job.
## Open questions
- Why is a `managed` (non-cron) work unit allowed to hold the gateway for the entire 30-minute cap?
What is `pid 397`, and should it be exempt from the "wait" set (or bounded separately)?
- The 12:01 slot had already run **6 h 38 m** against **10 h 01 m** for its 10-30 catch-up. Should a
reload deferral weight long cron runs differently — or should the cron lease survive a supervisor
relaunch, as the `restart_safe` flag implies?
- Two `SIGUSR1`s 4 m 42 s apart produced a **single** interruption and then a second clean restart at
19:23 — is the double signal expected?
## Impact
- One job failure + three derivative watchdog reds = **all four** of 2026-10-31's failed executions
(day closed **160 `completed` / 4 `failed`**).
- `ea79a2107e36` now reads `last_status: error`, **`failure_streak: 2`** — the only non-zero streak in
the fleet; non-clean `last_status` **3 → 4**, failing (run failed outright) **0 → 1**.
- Its next slot (2026-11-01 **00:00:41**) was already `running` at audit time.
## Related
- [[systems/scheduled-tasks]] — fleet ledger and the 10-31 window numbers
- [[incidents/2026-10-26-planned-restart-orphaned-cron-run]] — variant 2, the 1.35 s drain timeout
- [[incidents/2026-10-20-gateway-unclean-exit-orphaned-run]] — variant 1, the unclean crash exit
- [[systems/hermes-agent]] — gateway lifecycle record
- [[incidents/2026-10-30-mnemosyne-health-monitor-paused]] — the same job, the day before