mirror of
https://github.com/block/goose.git
synced 2026-06-02 06:19:33 +02:00
Matt Toohey
4889539ebb
- Always populate Message.metadata.inference in reply_internal so the
client has a per-message ground truth even when fetch_model_info()
cannot resolve a canonical model (resolved_model: None is now a
legitimate carried value).
- Add gen_ai.provider and gen_ai.request.model fields to the existing
tracing::instrument on stream_response_from_provider, which wraps
every Provider::stream call. (Instrumenting the trait method
directly is not possible: it has no body for tracing-attributes to
wrap and the in-tree comment at base.rs explicitly forbids it.)
- Add structured tracing::info! at the four reply hand-off points,
tagged with session.id for end-to-end correlation:
* /reply entry, after the session is loaded
(goose-server/src/routes/reply.rs)
* restore_provider_from_session, with provider_source and
model_source = "session" | "config_default"
(goose/src/agents/agent.rs)
* update_provider, with source = "explicit_update"
(goose/src/agents/agent.rs)
* just before provider.stream(...)
(goose/src/agents/reply_parts.rs)
* MessageEvent::Message emit, with the message's inference metadata
(goose-server/src/routes/reply.rs)
- ACP path: add symmetric entry and emit logs in
GooseAcpAgent::on_prompt (goose/src/acp/server.rs). The actual LLM
call site is shared via Agent::reply, so the agent-level changes
cover both /reply and ACP.
Together these allow correlating one log chain per session and
locating where the requested model and the actually-used model
diverge.
Verified: cargo build -p goose -p goose-server, cargo clippy
--all-targets -p goose -p goose-server -- -D warnings, and
cargo test -p goose --lib conversation:: all pass. Pre-existing
sqlx-runtime-feature test failures under cargo test -p goose --lib
are unrelated to this change (confirmed via git stash).
Signed-off-by: Matt Toohey <contact@matttoohey.com>