Skip to content

Put the AG-UI Bots back on a model that answers - #175

Merged
davidmckayv merged 2 commits into
mainfrom
fix/the-router-model-call-works-on-a-current-model
Aug 22, 2026
Merged

Put the AG-UI Bots back on a model that answers#175
davidmckayv merged 2 commits into
mainfrom
fix/the-router-model-call-works-on-a-current-model

Conversation

@davidmckayv

Copy link
Copy Markdown
Contributor

Everything the from-scratch gate turned up. All five driven, not reasoned about.

Risk Analyst answered nothing at all

Every question, tools or not. The run reached the server and returned 200; no text ever came back, nothing was logged, nothing appeared. Silence is the hardest failure to read, and it took most of the gate to find.

It is the framework Bot on :4201, not the hand-written one on :4200. I defaulted it to gpt-5.6-terra and inferred the Responses API for it in #168, on the strength of a comment in this repo saying that integration could take 5.6 "one line". I never drove it. Called directly:

gpt-5.6-terra → RUN_STARTED … RUN_FINISHED        (no text)
gpt-5.5       → RUN_STARTED … "Hello" … RUN_FINISHED

Both AG-UI Bots now default to gpt-5.5, each for its own measured reason, written where the default is. The built-in Bots keep gpt-5.6-terra — the one place 5.6 is proven, and where the bug that started all this lived. Setting BOT_MODEL to a 5.6 model still works and still switches the Responses API on by itself.

The router was dead

It sent temperature: 0, which reasoning models refuse outright — "Only the default (1) value is supported" — and this call treats a throw as "not sure". Every routing decision quietly became the default coworker; the roster was never consulted. A question naming Google Drive went to a Bot holding no Drive tools.

A fallback that lands on somebody who cannot answer

Every path into the fallback is a guess. When the message names a system exactly one coworker can reach, it isn't one — the others cannot answer at all. Only when unambiguous, and a confident match still wins.

agent-bot was handed a model it cannot use

Same blanket edit. It now refuses to start on a model whose function tools its endpoint rejects, rather than failing one silent tool call at a time, and it has its own AGENT_BOT_MODEL so the other Bot's model cannot take its tools away.

Two silent-failure classes closed

Its history builder no longer trusts the order it is given: read back from the durable store, a tool result arrives before the call it answers and the call's name is missing — a payload no provider accepts, answered with silence.

joinWithin waited on a detached connect forever. A connect that ignores the detach left the gate shut and every later message waiting on it. Bounded now, with a test that hangs without it.

The gate, end to end

  • Drive: an untagged question routes to the coworker holding Drive, searches, reads, and answers with all five proxy metrics and exact targets — 95%, 99%, non-zero, 90%, under 3s — checked against the document read out of band. Source and modified time cited.
  • Browser: real Hacker News rendered in the container; both titles match an independent curl exactly.
  • Policy: page.host == "example.org" refused, computer.action_refused recorded with the rule, and the Bot says which rule stopped it.
  • Built-in Bot: answers on gpt-5.6-terra.
  • 1254 tests, lint, format, typecheck clean across every package.

The router's model call sent temperature 0. Reasoning models refuse the
setting outright, and this call treats a throw as "not sure", so every
routing decision quietly became the default coworker and the roster was
never consulted. A question naming Google Drive went to a Bot holding no
Drive tools. Omitted rather than set per model, because a list of which
models accept it is a list that goes stale.

The fallback now asks whether the message named a system exactly one
coworker can reach, before it reaches for the default. Every path into
that fallback is "we are not sure", and the default is a guess; when one
coworker holds the named system and the others cannot answer at all, it
is not. Only when unambiguous, and a confident match still wins, because
a specialist with no connectors is the right answer to a question about
its specialism.

agent-bot was running gpt-5.6-terra. I wrote that it stays on gpt-5.5 and
then changed both compose lines with one edit, so the Bot in the box was
handed a model whose function tools that endpoint refuses. It has its own
variable now, and it refuses to start on a model it cannot use rather
than failing one tool call at a time with no reply and no reason.

Its history builder no longer trusts the order it is given. Read back
from the durable thread store a tool result arrives BEFORE the assistant
message that made the call, and the call's name is missing: a payload no
provider accepts, answered with silence rather than an error. Calls and
results are paired here instead.
Risk Analyst went silent. Every question, tools or not: the run reached
the server, returned 200, and no text ever came back. No error, nothing
logged, nothing on screen.

It is the framework Bot, on 4201, not the hand-written one on 4200. I
defaulted it to gpt-5.6-terra and inferred the Responses API for it, on
the strength of a comment in this repo saying that integration could take
5.6 in one line. I never drove it. Called directly it answers
RUN_STARTED, then RUN_FINISHED, and nothing in between. On gpt-5.5 the
same call answers.

So both AG-UI Bots default to gpt-5.5, each for its own measured reason,
and both reasons are now written where the default is. The built-in Bots
keep gpt-5.6-terra, which is the one place 5.6 is proven and the place the
bug that started this was. Setting BOT_MODEL to a 5.6 model still works
and still switches the Responses API on by itself.

Also here, from the same gate run:

The router sent temperature 0, which reasoning models refuse outright,
and this call treats a throw as "not sure". Every routing decision quietly
became the default coworker and the roster was never consulted: a question
naming Google Drive went to a Bot holding no Drive tools.

Its fallback now asks whether the message named a system exactly one
coworker can reach, before reaching for the default.

agent-bot was handed gpt-5.6-terra too. It refuses to start on a model
whose function tools its endpoint rejects, rather than failing one silent
tool call at a time, and it has its own variable so the other Bot's model
cannot take its tools away.

Its history builder no longer trusts the order it is given. Read back from
the durable store, a tool result arrives before the call it answers and
the call's name is missing: a payload no provider accepts, answered with
silence.

joinWithin waited on a detached connect forever. A connect that ignores
the detach left the gate shut and every later message waiting on it, which
is silence of a different kind. Bounded now.

Driven end to end: an untagged question about a Drive document routes to
the coworker holding Drive, searches, reads, and answers with all five
proxy metrics and their exact targets, checked against the document read
out of band.
@davidmckayv
davidmckayv merged commit 3f5f409 into main Aug 22, 2026
8 checks passed
@davidmckayv
davidmckayv deleted the fix/the-router-model-call-works-on-a-current-model branch August 22, 2026 18:05
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant