The agent asked for confirmation because the system prompt told it to, and read
the answer itself. Both halves are now enforced.
Each MCP toolset is wrapped in ApprovalRequiredToolset whose predicate reads the
readOnlyHint the server publishes — the gateway decides what counts as a write,
and the agent keeps no list of tool names in step with it.
output_type=[str, DeferredToolRequests] lets a turn end holding a proposed call
instead of making it; the executor parks the A2A task in input-required and
renders the arguments the model actually sent, not what it said it would do.
The reply is matched against a closed set of words, so "oui mais change le prix"
denies the call and the correction comes back for approval on its own merits —
verified end to end, along with a draft the user approved that was still refused
because msrp sat below buyprice.
Also wired: a ProductDraft validated after approval and before the call leaves;
UsageLimits on every run; FilteredToolset denying sales' write tools inside the
agent as well as at the gateway; PrefixedToolset namespacing the two servers;
dynamic instructions naming the caller; Agent.instrument_all() for a span per
model call and per tool call; and pydantic_ai.retries on inference only, so a
429 from the gateway is waited out rather than failing mid-reasoning.
Inference gets its own httpx client: validate_response raises on any 4xx, and
applying that to the shared client would turn a token-exchange refusal into an
exception carrying no body — which is where Keycloak says what was wrong. Its
transport is built explicitly, because a transport passed to the client makes
verify= inert and the lab CA would otherwise go undeclared.
Tests grew from 7 to 31, covering what approval decides and what a draft must
satisfy. evals/catalogue_evals.py adds a behavioural suite, kept out of pytest
because every case spends real tokens; its first case checks that a count comes
from the tool rather than from what the model remembers of a public dataset.
CLAUDE.md corrected: instrument=True is not an Agent parameter in 2.35.3,
pydantic_ai.retries needs tenacity, TemplateStr needs pydantic-handlebars, and
approval does not remove the in-memory history — resuming still needs the
messages, so replicas: 1 stands until they live in the A2A task store.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
An agent talks to URLs that honour a contract: an MCP endpoint, an OIDC token
endpoint, an OpenAI-shaped LLM endpoint. Whatever serves those URLs is
replaceable, and naming it in an agent's code or docs implies a coupling that
does not exist.
Every product name is gone from prose, comments and identifiers, replaced by
what the agent actually depends on. Where a behaviour of the current middleware
leaks into the contract — the second audience scope, the tool ACL — the text
now says what the caller must do and why, not who imposes it.
One string survives and cannot be fixed here: the broker client an agent must
name in its audience scopes is itself named after the middleware. Closing that
means renaming it in Keycloak and following the rename through the gateway
configuration and LibreChat. Recorded as a blocker.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The gateway already authorised writes for a product_manager — ai-mcp-proxy
filters tool discovery by consumer group, so create-product is simply absent
from a read-only user's turn. Only the system prompt was holding the agent
back, and it now describes how to write instead: restate the record, ask for
confirmation, never invent a missing field, report the identifier, and stop on
a refusal rather than working around it.
That flow cannot terminate without history. Each turn built a fresh Agent with
no message_history, so the user's "yes" arrived at an agent that had no idea
what it was agreeing to. History is now kept per A2A contextId in a bounded
in-memory map — enough for confirm-then-act, deliberately not durable, and
documented as needing sticky routing beyond one replica.
Verified end to end: pm-user creates S99_TEST1 across two turns and the row
appears in MySQL; ro-user asking for the same thing is told no write tool
exists, because for that token none does. Both test rows removed afterwards.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two things, both discovered by instrumenting the failure rather than trusting
the message it printed.
The MCP session never failed on protocol. macOS resolves any `.local` name
through mDNS first and takes a flat ~5.0s doing it, which sits exactly on
httpx's 5s default connect timeout; fastmcp reported the resulting
ConnectTimeout only as "Failed to initialize server session". Earlier tests
used `curl --resolve`, which bypasses DNS entirely and made it look like DNS
had been eliminated. Settings.http_timeout (60s) now covers the shared client,
the client fastmcp builds, and MCPToolset's init_timeout. list-products,
get-product and create-product now enumerate over the real gateway.
Inference used a static "unused" api_key against a gateway whose
openid-connect plugin requires audience ["llm-proxy"]. The model call now
carries a token exchanged for llm-proxy-audience, so it is attributable to the
same (agent, end user) pair as every tool call.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
First of the six agents, and the one that gives the A2A facade for the
others to interoperate with. A turn exchanges the user's token once per
audience, builds one MCP toolset per server with those tokens, runs the
model through the Kong AI gateway, and drops the toolsets.
Per-turn MCP clients are the point, not an implementation detail: a token
that outlives its turn can be replayed after the session ended. Several
narrow tokens rather than one broad one, so a prompt injection that talks
the agent into the wrong domain is refused at Kong rather than succeeding
quietly.
a2a-sdk rather than FastA2A, revising the May decision. That decision
rested on Agent.to_a2a() being removed in PydanticAI v2 — which has now
happened, so the argument no longer separates them. With four other
runtimes due to interoperate, the protocol author's implementation is the
safer base. httpx2 rather than httpx, which PydanticAI deprecates for
OpenAI-compatible providers and drops in v3.
Four things about a2a-sdk 1.1 that no documentation gave and that were
established against the running library, all recorded where they bite:
a2a.types are protobuf, not pydantic; DefaultRequestHandler speaks 0.3 and
V2 is needed for 1.0; a client must send A2A-Version: 1.0 or the server
assumes 0.3 and answers VERSION_NOT_SUPPORTED on a perfectly valid
request; and an executor must enqueue a Task before any status update.
Verified: the app builds with no deprecation warnings, serves its Agent
Card, and fails a turn cleanly with "no bearer token" rather than
answering from the model's own knowledge — which would hide a broken
identity chain behind a plausible reply. Seven tests cover the exchange
shape (scope not audience, per-audience caching, refusal) and the card.
The Keycloak client does not exist yet, so no end-to-end run has happened.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Stéphane Tailland