An agent run keeps its tokens in and out, and its cost-ledger entry carries them #12274

A finished agent run now keeps its tokens in and out, and its cost-ledger entry carries them. bun x tsc --noEmit and bun scripts/gate.ts are clean, and test/cost-ledger.test.ts and test/schema-migrations.test.ts pass. I didn't run bun scripts/ledger-verify.ts; the new test calls verifyDir, which that script wraps. I also didn't run scripts/design-lint.ts, since nothing in src/web/ changed.

• Migration: 20261114080000-run-tokens adds nullable tokens_in and tokens_out to run, with its pin. Old rows, checks, deploys and agent runs that reported no usage stay NULL, never 0.
• finishRun: it takes tokensIn and tokensOut and stores them only when they are positive whole numbers. Anything else is stored as NULL, and the code added to the finish path can't throw.
• entryOf: it writes tokens_in and tokens_out only when the run row has them. The hashed body doesn't change otherwise, so existing chains still verify.
• Runner: the finish call in src/crew/runner.ts passes the meter's running totals, not result.tokensIn. That way a wrap-up or resume invocation adds to the total instead of replacing it. A run that reported nothing has a total of 0, which is stored as unknown.
• Test: test/cost-ledger.test.ts has a new test. A check's entry has neither field, an agent run with usage writes both, and an agent run with 0/0 writes neither. The chain still verifies. This is the file that tests the cost ledger; test/ledger.test.ts is the decision ledger.
• Not done: src/web/ and src/flow/ are untouched, and I didn't backfill old runs.

To try it, finish an agent run and read the newest line in the ledger's <yyyy-mm>.jsonl. It should show tokens_in and tokens_out.

Closes #12270

A finished agent run now keeps its tokens in and out, and its cost-ledger entry carries them. bun x tsc --noEmit and bun scripts/gate.ts are clean, and test/cost-ledger.test.ts and test/schema-migrations.test.ts pass. I didn't run bun scripts/ledger-verify.ts; the new test calls verifyDir, which that script wraps. I also didn't run scripts/design-lint.ts, since nothing in src/web/ changed.

• Migration: 20261114080000-run-tokens adds nullable tokens_in and tokens_out to run, with its pin. Old rows, checks, deploys and agent runs that reported no usage stay NULL, never 0.
• finishRun: it takes tokensIn and tokensOut and stores them only when they are positive whole numbers. Anything else is stored as NULL, and the code added to the finish path can't throw.
• entryOf: it writes tokens_in and tokens_out only when the run row has them. The hashed body doesn't change otherwise, so existing chains still verify.
• Runner: the finish call in src/crew/runner.ts passes the meter's running totals, not result.tokensIn. That way a wrap-up or resume invocation adds to the total instead of replacing it. A run that reported nothing has a total of 0, which is stored as unknown.
• Test: test/cost-ledger.test.ts has a new test. A check's entry has neither field, an agent run with usage writes both, and an agent run with 0/0 writes neither. The chain still verifies. This is the file that tests the cost ledger; test/ledger.test.ts is the decision ledger.
• Not done: src/web/ and src/flow/ are untouched, and I didn't backfill old runs.

To try it, finish an agent run and read the newest line in the ledger's <yyyy-mm>.jsonl. It should show tokens_in and tokens_out.

The gate failure is fixed and pushed; I rewrapped the comment on the new token field in src/core/runs.ts to under 100 columns.

bun scripts/gate.ts now exits 0 and bun x tsc --noEmit is clean. I didn't re-run the tests after this change, since it only touches a comment. My first gate run in this turn still printed the old error, because it ran in the same command as the edit. I pushed before I saw that. The second run was clean, and the pushed commit contains the fix.

The earlier summary of the change still applies:

• Migration: 20261114080000-run-tokens adds nullable tokens_in and tokens_out to run, so unknown stays NULL and never becomes 0.
• Finish path: finishRun stores the totals only when they are positive whole numbers.
• Ledger: entryOf writes tokens_in and tokens_out only when the run has them.
• Runner: it passes the meter's totals across all of a run's invocations.
• Test: the new test is in test/cost-ledger.test.ts.