fix(server): Codex usage keeps requests from interrupted turns#29415

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 108.5 KiB
• Claude decoded thread snapshot: 108.8 KiB

Codex parsing now buffers request-level usage events and reconciles them with later cumulative snapshots. The scan-cache format advances to version 11, and older Codex cache entries restart scanning from the beginning.

Changes: Codex usage reconciliation

Layer / File(s)	Summary
Buffer and reconcile Codex requests apps/server/src/provider/Drivers/codexUsage.ts, apps/server/src/provider/Drivers/codexUsage.test.ts, apps/server/src/usage/usageTranscriptReader.test.ts	The parser buffers token_usage_record events and emits their usage when a later cumulative snapshot exactly reconciles the counter gap. Emitted records use the timestamp, model, service tier, and other attribution captured when each request was read. Tests cover interrupted turns, fork copies, and buffered and streaming parsing.
Migrate Codex scan-cache state apps/server/src/usage/usageScanCache.ts, apps/server/src/usage/usageScanCache.test.ts, apps/server/src/usage/UsageService.test.ts	The cache format advances from version 10 to version 11. Older Codex entries restart scanning from offset zero. Tests cover cache compatibility and an upgrade after an interrupted request.

Estimated code review effort: 4 (Complex) | ~45 minutes

Change: Bug fix · Severity of issue fixed: Low

Sequence Diagram(s)

mermaid

1	sequenceDiagram
2	participant Transcript as Codex transcript
3	participant Parser as parseCodexLine
4	participant State as CodexScanState
5	participant Output as UsageRecord output
6	Transcript->>Parser: token_usage_record event
7	Parser->>State: buffer attributed request
8	Transcript->>Parser: cumulative token_count snapshot
9	Parser->>State: reconcile buffered usage with counter gap
10	Parser->>Output: emit reconciled request records

Possibly related issues

• pingdotgg/t3code#29363 — The changes update Codex cumulative-snapshot reconciliation, but do not establish that they fix the reported baseline rewind.
• pingdotgg/t3code#29365 — The parser buffers token_usage_record entries and uses their exact totals to recover requests across interrupted turns.

Possibly related PRs

• pingdotgg/t3code#29392 — Both PRs reconcile buffered request usage with later cumulative snapshots; this PR also changes attribution, fork handling, and cache behavior.

Suggested reviewers: pixpmusic, fixfon, maria-rcks

A fast request after a fork may be omitted from usage totals. This timing-dependent issue should be addressed or accepted before merging.

Check name	Status	Explanation
Title check	Passed	The title "fix(server): Codex usage keeps requests from interrupted turns" matches the main change. The diff adds parsing of token_usage_record lines so requests from interrupted turns are billed wh…
Description check	Passed	The description covers the template's main parts. Problem: it explains that interrupted-turn requests appear only as token_usage_record lines, that the reader drops them, and that the next `token_co…
Linked Issues check	Passed	Issue #29365 requires that Codex usage bill a request recorded only as top-level token_usage_record when its turn is interrupted. It also requires that the next token_count reconcile, and that the…
Out of Scope Changes check	Passed	Whole-PR changes map to #29365. codexUsage.ts implements the bridging logic. codexUsage.ts format and projection changes expose token_usage_record payloads. Fork-copy handling applies the same b…

Full details: Title check

Explanation

The title "fix(server): Codex usage keeps requests from interrupted turns" matches the main change. The diff adds parsing of token_usage_record lines so requests from interrupted turns are billed when a later cumulative snapshot reconciles, and it updates the scan cache to v11 to re-parse affected rollouts. The title follows the conventional-commit form the template recommends and names the primary fix without listing files.

Full details: Description check

Explanation

The description covers the template's main parts. Problem: it explains that interrupted-turn requests appear only as token_usage_record lines, that the reader drops them, and that the next token_count then fails reconciliation. It gives reproduction steps, a synthetic rollout, and before/after results (120 vs 510 tokens). Change: it lists the parser, reducer, format-gate, fork-copy, and scan-cache updates and explains why counters are not replaced. Verification: it names the specific failing-then-passing tests, reports a real-data regression replay, and includes the runtime before/after check with screenshots. It also states what was not checked. Scope and approval: it links the issue (#29365) and says it was triaged through the automated issue flow. The template asks for a triaged issue or discussion with explicit maintainer approval, including the approval comment. The description links the issue but does not show a maintainer approval comment or explain why the fix qualifies without one. The model/harness attribution is present. Minor gaps: the template's HTML comments remain in the template, and the description does not explicitly confirm that the reference link is an approved triage. Overall the description is mostly complete.

Full details: Linked Issues check

Explanation

Issue #29365 requires that Codex usage bill a request recorded only as top-level token_usage_record when its turn is interrupted. It also requires that the next token_count reconcile, and that the warning clear. The whole-PR change in codexUsage.ts buffers pending records since the last cumulative snapshot. It uses them only when cumulative - previous equals the pending sum on every field, then bills last after the unreported entries. Records never replace counters. Per-entry model, tier, speed and timestamp are kept. Tests cover the reporter's rollout through the buffered and streaming reader paths (510 tokens, 0 malformed), multiple interrupted turns with attribution changes, fork-copied history, and a control case with several records between reconciling snapshots. The usageScanCache.ts v11 change forces re-parse of affected pre-v11 Codex entries, so existing caches pick up the fix. The incremental diff since the previous reviewed head (f3bc74f2) leaves codexUsage.ts unchanged, so the earlier conclusion still holds for the parser. The verification results were not rerun in this review. Assessment is based on source and summary evidence.

Full details: Out of Scope Changes check

Explanation

Whole-PR changes map to #29365. codexUsage.ts implements the bridging logic. codexUsage.ts format and projection changes expose token_usage_record payloads. Fork-copy handling applies the same burst boundary to the new record type, which prevents double billing. usageScanCache.ts moves to v11 and re-parses affected Codex entries; the diff matches this purpose. The UsageService.test.ts, usageScanCache.test.ts and usageTranscriptReader.test.ts changes cover these behaviors. The incremental diff since the previous reviewed head adds Cursor credential and shared-account tests in UsageService.test.ts. These appear to come from the merge of main (merge base f6503f0e). The whole-PR summary for that file reports +103/-3 lines and describes only usage-cache tests, which does not include those Cursor additions. Those Cursor changes are therefore not attributed to this PR. This conclusion relies on the summary line counts and merge history, not on a full base-to-head diff of that file.

Generate docstrings

• Commit to this branch
• Create a new PR

Generate unit tests (beta)

• Commit to this branch
• Create a new PR