Skip to content

Claude Code. Any Model. The most powerful AI coding agent now speaks every language.

docs(reports): record the bench single-call run and settings collision

Jack Rudenkoac33c159c821
diff --git a/ai-docs/reports/codex-tool-calls-and-code-mode-2026-10.md b/ai-docs/reports/codex-tool-calls-and-code-mode-2026-10.md
index e4f9f56f..7a83201f 100644
--- a/ai-docs/reports/codex-tool-calls-and-code-mode-2026-10.md
+++ b/ai-docs/reports/codex-tool-calls-and-code-mode-2026-10.md
@@ -75,8 +75,12 @@ request shape from the design, not from the harness.
- Re-run check: `ai-docs/benches/codex-code-mode/` (madbench; `classic`, `auto`, `auto-edit` on
`[email protected]`, subscription only). First run 2026-10-20 on Claude Code 2.1.304: 15/15, 19/19 exec
- scripts resumed, 0 input errors in tool_results. Needs `--concurrency 1` (cause unknown). Claude Code
- 2.1.303, where D1 lived, was not pinned.
+ scripts resumed, 0 input errors in tool_results. Second run with the `single-call` scenario (V3's
+ purpose on the subscription: 32–34 streamed deltas, forwarded once): 20/20 at `--concurrency 4`.
+ Claude Code 2.1.303, where D1 lived, was not pinned.
+- claudish defect found by the bench: `createTempSettingsFile` (`claude-runner.ts:815`) names the file
+ `settings-<Date.now()>.json`, so two sessions started in the same millisecond share one settings file
+ (status line, hooks). The bench spaces launches 100 ms apart until this is fixed.
- Behaviour changes outside code mode, for the PR and release notes: `d6629064` identity-filters
`instructions` on every `cx@`/`openai@` Responses request; a 401 whose forced refresh lands on another
ChatGPT account is answered with the upstream 401, not resent (`1fe1b580`, `a808fab2`).