Files
agent-estate-wiki/incidents/2026-10-24-conduit-call-request-refused.md
T

4.8 KiB

title, type, status, created, updated, verified_on, review_after, confidence, tags, sources
title type status created updated verified_on review_after confidence tags sources
Conduit Call Requests Refused by the Agent incident open 2026-10-25 2026-10-25 2026-10-25 2026-11-08 high
conduit
voice
tool-calling
agent-conduct
incident
unresolved
live-sqlite-sessions
live-sqlite-conduit
plugin-tree
live-probe

Conduit Call Requests Refused by the Agent — 2026-10-24

Window: 2026-10-24 12:53 → 12:56 AWST (three attempts) Status: Open — behavioural failure in the agent, not a broken tool or config Impact: Anthony asked to be phoned three times to test Conduit calls. No call was placed, and he was told the capability does not exist.

What happened

Three desktop sessions inside three minutes, all asking for a call:

Session Started Model Tool calls Reply
20261024_125329_08c71f 12:53:34 kilo-auto/free, switched mid-session to smart-route (custom:omni-101) 0 “I don't have the ability to make phone calls …”, then “There's no calling capability in my toolset”, then “I don't have a phone calling capability”
20261024_125538_d054bb 12:55:54 smart-route 0 “Hey, I'm here and listening. What can I do for you?” — the greeting a call would open with, but no call was placed

Anthony's turns in …08c71f, verbatim: “Call me now to test calls.” → “Yes. You do” → “CAL ME”. No tool was invoked in either session (tool_call_count: 0 — no tool_search, no skill_view, no conduit_call_user). Whether the app itself rang (tap-to-call) is not visible in the plugin's records; what is proven is that the agent never attempted a call.

The capability does exist

  • conduit_call_user (toolset conduit, 📞) is registered by the plugin for every session (register_tool in call_tool.py) and is reachable through Hermes's lazy tool discovery (tool_search) — which is exactly how the nightly wiki-audit session can see it.
  • The plugin ships the skill calling-the-user (skill_view("conduit_push:calling-the-user")), written for this question.
  • Anthony's call settings were permissive at the time (conduit-calls.json → settings): enabled, when_asked, decides, alerts all true; min_gap_s 120 / 6 per hour / 20 per day; avoid_unknown_place false. A user-asked call would have gone out.
  • The device was paired and the gateway (started 2026-10-23 18:54:43, carrying the memory-timeout patch) was up throughout.

So nothing outside the agent blocked the call. The tool was simply never called.

Root cause

Capability asserted instead of discovered. Both models answered from priors (“I'm a text-based AI agent”) rather than searching the tool registry or loading the bundled skill. The first reply came from a free model, but the mid-session switch to smart-route did not change the answer — so this is not one weak model's fault. The agent treated “I don't know of such a tool” as “there is no such tool”, in a session where the tool was one tool_search away.

Evidence correction (affects the 10-22 memory-timeout incident)

~/.hermes/conduit-calls.json is not a call log:

  • history — real call records — is empty (0 entries) at 2026-10-25 00:2x.
  • ends (ring buffer, MAX_RECENT_ENDS = 64 in calls_store.py) records every turn-end that fires a watch (fire() appends {session_id, outcome, at}). At this audit its two entries are cron sessions — local-afac8b… (Mnemosyne Embedding Coverage Monitor, 10-25 00:01) and local-8cfe41… (Mnemosyne Health Monitor, 10-25 00:04).

Therefore the 10-24 note that “two call sessions closed cleanly at 00:03:23 / 00:06:57” was a false positive: those timestamps match the two Mnemosyne cron sessions that ran that night, not calls. Do not read ends as “a call happened” — read history. The correction is also stamped into incidents/2026-10-22-conduit-memory-start-timeout.

Open items

  • Verify a real call end-to-end. The memory-timeout fix has still never been exercised by a call (history empty since it landed). Cheapest path: tap-to-call in the Conduit app; a chat-route call needs the agent to actually invoke conduit_call_user.
  • The failure is behavioural: worth checking whether the plugin's tool appears in a desktop session's headline tool set at all (sessions.tool_names stores only a hash, so it could not be read from state.db — the deferred-discovery path is the likely culprit for the miss).
  • Consider surfacing calling-the-user more loudly (e.g. a system-prompt hint) so a model does not have to discover it while a user is waiting.