mirror of
https://github.com/block/buzz.git
synced 2026-08-18 06:50:31 +02:00
npub1g8493u0xfsjrvflg4n08ezd7vec99mnwzlv0qgwpr9d7gvjwhuzqx59rhwandWill Pfleger
74d78b5fe9
Every session's first turn was systematically uncounted: the publisher had no prior cumulative snapshot, so it emitted deltaReliable:false and null turn.* fields. For sessions that buzz-acp itself created via session/new, prior usage is zero by definition — the zero baseline IS the correct starting point. Add UsageTracker::seed_zero_baseline(), called from pool.rs immediately after create_session_and_apply_model() succeeds (both channel and heartbeat paths). The zero baseline seeds SessionState with last_input/output=0 and last_cost=Some(0.0), so the first usage_update notification computes turn.*=cumulative.* and emits deltaReliable:true. The re-attach path (session ID already in agent.state.sessions) is unchanged: buzz-acp did not create that session and its prior usage is genuinely unknown, so the existing fail-closed behavior is correct. seed_zero_baseline() is a no-op if a baseline already exists, guarding against accidental double-seeding across session rotation. Regression tests added (a)-(d) per task spec: (a) self-spawned first turn: deltaReliable=true, turn.*==cumulative.* (b) re-attach first turn: unchanged fail-closed behavior (c) second turn unaffected in both modes (d) wire-frame payload field assertions for the spawned path Gates at tip: cargo test -p buzz-acp 645/645, cargo fmt --check clean, cargo clippy -p buzz-acp --all-targets -- -D warnings clean. Co-authored-by: Will Pfleger <pfleger.will@gmail.com> Signed-off-by: Will Pfleger <pfleger.will@gmail.com>