feat(acp): report standard adapter usage

[atishpatel]

Why

Claude Code and Codex expose standard ACP prompt-response usage, but Buzz only consumed Goose’s private cumulative usage notification. Their token use and Claude’s cumulative cost were therefore absent from NIP-AM metrics.

What

• Read per-turn session/prompt response usage for known Claude and Codex adapters
• Publish Claude’s raw cumulative cost separately from per-turn tokens without changing the NIP-AM schema
• Keep Goose usage exclusive and cover Claude/Codex wire serialization

Risk Assessment

Low-to-medium: changes best-effort observability only and does not affect prompt execution. The adapter-specific mappings preserve source semantics and omit unavailable fields.

References

• Validated with cargo fmt --check, cargo test -p buzz-acp --no-run, and full cargo test -p buzz-acp (678 passed at 652e373a before merge-trailer amendment).

Generated with Codex

[custard-pirate]

I validated this problem independently while building an owner-side NIP-AM usage projection for a mixed Codex/Claude/Goose fleet. This PR covers the important upstream seam, and its separate standard-adapter tracker is a better fit than duplicating the local integration.

Three attribution/coverage details from that deployment may be useful here or in a focused follow-up:

1. Codex exposes the effective model at /_meta/quota/model_usage/0/model in the prompt result. Retaining that value prevents otherwise valid kind 44200 records from landing under an unknown model. When the adapter omits it, the operator-selected model from PromptContext is a useful fallback.
2. Normalize the launched adapter command into a provider-neutral harness label (codex, claude, or goose) before publishing. Absolute paths and wrapper names otherwise fragment the dataset.
3. Consider emitting a coverage-only kind 44200 record with null token fields when a completed turn supplies no usage. That distinguishes unsupported/missing telemetry from agent inactivity without treating unknown values as zero.

Cache-write preservation and the prompt-response token path are already covered here (and overlap #4000), so I do not plan to open a competing ACP usage PR. I can prepare a small dependent follow-up for these three points once the desired behavior and landing order are clear.

[wesbillman]

Reviewing on Wes's behalf.

The new standard-adapter mapping publishes usage.inputTokens directly while also publishing cachedReadTokens/cachedWriteTokens. Both supported adapters define the ACP prompt-response inputTokens as non-cached input: Codex explicitly subtracts cached input in toTokenCount(), and Claude's sessionUsage() keeps fresh input, cache reads, and cache writes as separate accumulators. NIP-AM requires the opposite representation: inputTokens is the inclusive input-side total, and cache fields are informational subsets (docs/nips/NIP-AM.md, Numeric validity and token semantics; also TokenCounts' contract in crates/buzz-core/src/agent_turn_metric.rs).

As written, a Claude turn with inputTokens=100, cachedReadTokens=30, and cachedWriteTokens=25 publishes input=100 rather than 155; the added test currently pins that undercount. Codex has the same issue for cache reads. This corrupts NIP-AM input totals and any aggregation/cost logic built on them.

Please normalize standard usage before constructing TurnUsage: checked-add cache read/write into turn_input_tokens, fail closed on overflow, and update the wire tests to assert the inclusive input value while preserving each cache subset separately.

[atishpatel]

Rebased onto latest main (5e4c05f9) and incorporated Wes's review feedback.

PR #4000 does not make this PR redundant. #4000 adds the cumulative _goose/unstable/session/update path used by Goose/buzz-agent plus the archive/accounting infrastructure around it. It does not ingest the standard ACP session/prompt response usage emitted by Claude/Codex, and it does not ingest Claude's standard session/update cumulative cost notification. Those are the remaining gaps covered here.

Per Wes's review, this revision now normalizes standard-adapter input to NIP-AM's inclusive contract with checked arithmetic:

• Claude: inputTokens + cachedReadTokens + cachedWriteTokens (the test now proves 100 + 30 + 25 = 155 while preserving both cache subsets)
• Codex: inputTokens + cachedReadTokens + cachedWriteTokens (the test now proves 90 + 40 = 130 while preserving Codex's provider total)
• overflow fails closed instead of saturating or publishing an undercount

It also preserves #4000/current-main's fail-closed optional cumulative counters, cache-write support, and pricing identity, while keeping Goose/buzz-agent usage exclusive if both producer paths emit data.

Wes's broader attribution/coverage suggestions—effective model fallback, normalized provider-neutral harness labels, and optional coverage-only records for completed turns with no usage—remain useful follow-up scope; they are not supplied by #4000 and are not folded into this narrowly rebased ingestion change.

Validation: cargo test -p buzz-acp (722 unit + 9 lifecycle tests passed), cargo fmt --all -- --check, and the push hooks' Rust/Desktop Tauri checks.

— Brainy Bumble

[wpfleger96]

🤖 Combined + deduped feedback from the dual review I ran on this (two independent passes; full writeups on our side). The current head already fixes the inclusive-input issue from the earlier review round — checked-add of cache read/write with fail-closed overflow looks right. What's left: two accounting gaps both passes converged on (the cost baseline and the initial_message lifecycle), plus a few nits. Overall the shape is exactly what we hoped for next to #4000: goose stays exclusive, standard adapters fill the gap, and omitting pricingIdentity/model for these records is the correct fail-closed call vs #4629's exact-lookup contract.

feedback addressed

[wpfleger96]

🤖 Verified all five findings from the combined review are addressed at head 001af63a:

• First-turn Claude cost: standard tracker now has per-session cost state seeded to zero via notify_session_spawned() (buzz-spawned only), emitting a checked monotonic turn_cost_usd delta alongside the raw cumulative — with baseline poisoning on counter decrease, and attached/pre-existing sessions correctly left unseeded (attached_claude_session_does_not_invent_first_cost_delta).
• initial_message usage loss: the setup prompt is drained and published under {turn_id}:initial before the user turn on both success and cancel-drain paths; the two-prompt regression proves both monotonic sequences survive.
• Cost-only records now survive missing/malformed prompt usage; overflow can no longer emit an all-null record; and the named fake-adapter test crosses spawn detection, NDJSON dispatch, session/update, and drain.

Re-ran cargo test -p buzz-acp at this exact head: 744 passed, 0 failed. All checks green. Approving.