fix(fastwire): do not read a non-authoritative tier echo as a Fast downgrade

[lidge-jun]

Summary

Closes #2558.

OpenCodex reported every Codex Fast request as fastOutcome: downgraded /
fastDowngradeReason: response-declined, while the request was correctly sent as
service_tier: "priority" and observed throughput differed from a no-tier request. The
issue's suspicion was right: the ChatGPT-internal Codex backend returns
service_tier: "default" even on turns it scheduled as priority, so its echo was being
read as a verdict it cannot give.

Two consequences, not one. The visible symptom is a false negative in the diagnostics; the
quieter one is cost attribution, because fastOutcome drives priority pricing.

TierObservationContext gains responseTierAuthoritative. Absent means "assume
authoritative", so the public API path — where an echoed default really does mean the
request was declined — is unchanged. core.ts sets it false only for the canonical
ChatGPT-forward destination, using the existing isCanonicalOpenAiForwardProvider check.

The raw echo is still recorded in responseServiceTier. This suppresses a verdict, not the
evidence: an operator can still see exactly what the backend returned.

Verification

Driven red first — removing the guard fails exactly the new case and nothing else:

# guard removed
$ bun test tests/fastwire-observability.test.ts
 31 pass, 1 fail

# with the fix
$ bun test tests/fastwire-observability.test.ts
 32 pass, 0 fail, 95 expect() calls

$ bun x tsc --noEmit
(clean)

The three added cases pin all three states: a non-authoritative destination does not
classify a downgrade, an authoritative one still does, and an omitted flag keeps the
authoritative default.

Checklist

• Targets dev
• Based on the current dev head
• Regression driven red first, then green
• Existing fastwire behaviour unchanged for authoritative destinations
• Typecheck clean
• No secrets or account identifiers in the diff

Summary by CodeRabbit

• Bug Fixes

  • Improved Fast-mode status reporting for routes that cannot reliably confirm the upstream service tier.
  • Prevented echoed default-tier responses from being incorrectly classified as declined Fast requests.
  • Preserved existing downgrade detection for destinations with authoritative tier information.

• Tests

  • Added coverage for authoritative and non-authoritative service-tier responses.

[github-actions]

✅ Deterministic PR hygiene checks passed.

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: 0261532b-4453-46ad-b18f-1cf4151384ab

📥 Commits

Reviewing files that changed from the base of the PR and between 48602ce and aacd24c.

📒 Files selected for processing (4)

• src/providers/fastwire.ts
• src/server/responses/core.ts
• src/types/provider.ts
• tests/fastwire-observability.test.ts

---

📝 Walkthrough

Walkthrough

The change adds destination-aware response-tier authority. Canonical OpenAI forward routes mark echoed tiers as non-authoritative, while other routes retain authoritative behavior. Tests cover non-authoritative echoes, authoritative downgrades, and omitted-flag compatibility.

Changes

Fast tier authority

Layer / File(s)	Summary
Authority contract and classification src/types/provider.ts, src/providers/fastwire.ts	TierObservationContext accepts an optional authority flag. Non-authoritative response tiers cannot confirm or deny Fast.
Route wiring and regression coverage src/server/responses/core.ts, tests/fastwire-observability.test.ts	Canonical OpenAI forward routes pass non-authoritative context. Tests cover non-authoritative echoes, authoritative downgrades, and default behavior when the flag is omitted.

Estimated code review effort: 2 (Simple) | ~10 minutes

Possibly related PRs

• lidge-jun/opencodex#1893: Both changes modify FastWire tier-observation logic and response handling.
• lidge-jun/opencodex#1956: Both changes modify FastWire observation and response-tier metadata behavior.
• lidge-jun/opencodex#1965: This change extends the tier-observation metadata and response handling introduced there.

Suggested labels: review-ready

Suggested reviewers: ingwannu, olddonkey

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch codex/fast-tier-classification-2558

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: aacd24c537

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Exclude non-authoritative echoes from pricing provenance

When a canonical ChatGPT Fast request receives the known service_tier: "default" echo, this guard keeps the outcome as assumed priority, but observeResponseServiceTier still stores "default". serviceTierContextFromOutcome in src/usage/cost.ts:380-384 prioritizes that stored response over the assumed canonical tier, so estimateAttemptCost sees the default tier and skips the OpenAI priority multiplier. Consequently, the cost-attribution regression described by this change remains for every affected request; preserve the echo for diagnostics while ensuring the pricing path ignores it when it is non-authoritative, and add a multiplier regression assertion.

Useful? React with 👍 / 👎.