[askalf]

What does this PR do?

The thinking share from #1783 now covers GPT traffic. Every Codex row was recorded with thinkingTokens: 0 (hardcoded in reportCodexOutcome), so dario usage, --by-key and /analytics showed no reasoning on any ChatGPT-subscription lane, however much the model reasoned.

• reasoningTokensFromResponses() in codex-backend.ts reads output_tokens_details.reasoning_tokens from the Responses usage block. OpenAI counts it inside output_tokens, the same way Anthropic counts thinking inside output, so it fits the existing field as is. Rounded, capped at the output it is part of, 0 when absent or malformed.
• splitResponsesUsage() carries it as reasoning; CodexForwardOutcome gains reasoningTokens, set on both report sites (the translated Messages route and the Responses passthrough); reportCodexOutcome records it as thinkingTokens.
• Nothing on the wire changes; this is accounting only. Ledger rows written before read as before.

Why: on one deployment's last 7 days, 22% of all output tokens were GPT through the Codex engine, and the thinking line could say nothing about any of it. With the per-key default effort from #1786 able to set reasoning.effort on those lanes, the share is the number that says which of them need it.

How to test

• npm run build
• node test/ledger-thinking-codex.mjs (new, 12 checks: the reader's tolerance and cap, splitResponsesUsage, then a real startProxy with a loopback Codex stub on both GPT routes, checked on the window record, the ledger cell, the consumer summary with its price, and the Thinking: line; 7 of them fail against the proxy without the record change). It is its own file because the proxy reads the Codex base URL and the accounts folder once per process.
• Unchanged and green: codex-backend, ledger, ledger-wiring, ledger-thinking, ledger-per-key-tokens, analytics-recording, analytics-stream, responses-inbound-wiring, anthropic-responses-translate.test, anthropic-responses-stream.test, key-effort, keys-budget-proxy, metrics, tui-*.

Checklist

• npm run build passes
• npm test passes (CI; the suites above were run locally)
• Touches src/? Then a bullet under ## [Unreleased] in CHANGELOG.md is in this PR
• Version left alone; no ## [x.y.z] heading added
• No issue auto-close wording
• Touches proxy.ts: one field on the Codex analytics record; test/compat.mjs in CI
• No new runtime dependencies added
• No tokens/secrets in code or logs
• README and site test count 287 -> 288 (one new test file)

[github-actions]

📦 This PR will not cut a release

version stays at 6.20.0 but 5 shipping file(s) changed: README.md, docs/pool-keys-and-analytics.md, site/index.html (+2 more)

cc-drift-auto-release tags and publishes only when package.json's
version moves on master. As it stands this PR merges green and ships
nothing — the change sits on master until some later PR's bump carries it
out.

If that is intended (revert, docs-only refactor, or a stacked PR whose
partner carries the bump), ignore this. Otherwise bump all three slots
that scripts/preflight.mjs checks, in one commit:

• package.json → 6.20.1
• package-lock.json → .version and .packages[""].version
• CHANGELOG.md → a ## [6.20.1] - 2026-11-02 section under ## [Unreleased]

This comment is advisory and updates itself on every push; it never
blocks the merge.

[github-actions]

Live test: ✅ PASS

Head b04ef19cad0a318967ade3c856395c8f646749be on the self-hosted runner: All stages PASS: install/build/cc-detect/cli/doctor/live-probe(pong)

Stages

install: PASS
build: PASS
cc-detect: PASS
cli: PASS
doctor: PASS (exit 1, 19 rows; no credentials, reported cleanly)
live: PASS (pong)

Full workflow run

[sprayberry-redline]

Verdict: approve. Codex Responses usage now carries bounded reasoning-token counts through both forwarding routes into analytics and the ledger, with coverage for the reader, routes, summaries, and displayed share.

Minor:

• test/ledger-thinking-codex.mjs:2: This introductory block narrates the motivation and structure of the patch rather than documenting a test-specific constraint. It adds a large maintenance burden without helping the assertions establish their behavior.

[sprayberry-redline]

Verdict: approve. This carries Responses reasoning token usage through both Codex forwarding paths into analytics and ledger thinking totals. The reader validates and caps the value against output, and the new end-to-end coverage exercises translated and passthrough routes, absent details, and the ledger summaries.