feat: capture the token usage Providers already report (0.15.0)

[chivopic]

Stage 12. Stage 11 made "can this agent do the job" answerable, but the report had no cost in it — pass rate, turns, wall time. A prompt change that lifts the pass rate from 70% to 75% while tripling the tokens is usually a bad trade, and it looked like a pure win. Providers return usage on every call and chivgent was throwing it away.

What it does

Usage flows from the Provider response through to the run: per turn on message_end, per run on agent_end and the run result, per attempt in the eval report, and cumulatively behind /session.

task                  pass  turns  tokens  tools                       p50
find-auth-logic       4/5   2.0    3.4k    list_files,search_text,...  6.1s
rename-symbol         3/5   5.4    12.1k   read_file,edit_file         9.8s

overall  7/10 (70%)  58.3k tokens

Compaction's own summarising call counts too: it trades a call now against smaller inputs later, and that trade cannot be judged with half of it hidden.

Tokens only, never money

Converting to a currency needs a model-to-price table, and that table goes stale silently when a Provider changes its rates — producing a figure that looks precise and is wrong. That is worse than no figure, because nobody questions a number with a decimal point in it. Token counts are what the Provider itself reported and do not expire.

A missing figure stays missing

usage is undefined rather than zero when a Provider reports nothing, and a total containing such a call is flagged incomplete. Zeroes would be indistinguishable from a genuinely free call and would quietly understate every total containing them.

Compatibility note

Streamed Chat Completions only report usage when asked, so stream_options: { include_usage: true } is now sent. A self-hosted OpenAI-compatible endpoint predating that field may reject the request; --no-stream is the documented fallback. This is the only change here that can affect an existing user.

A bug only running it could find

Every turn added the context manager's usage, which is undefined when no compaction was needed — so every real CLI run was flagged incomplete, since the CLI always has a context manager. The unit test asserting complete: true used an agent with no context manager, so it never reached that line. Caught by watching agent_end.usage.complete in a live end-to-end run. Not spending anything and not reporting anything are different, and only the second is a gap; the test that would have caught it is now there.

Verification

• npm run release:check passes: type-check, 379 tests, build, npm pack --dry-run.
• The compaction-cost test was verified in reverse: removing that accumulation makes it fail.
• End to end against a stub that reports usage: non-streaming and streaming both carry it through, including cachedInputTokens, into the eval table and the JSON report.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FFMvrN7qqBmVMNX26XtRDy

---

Generated by Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.