fix(bench): record the effective search request body

[jamespsterling]

TL;DR

The search solver's recorded requestBody now includes the Switchyard plugin it sent, so stored runs show which algorithm was requested.

What changed?

• src/benchmarks/search/core/solver.ts: completedState persists { ...request, ...extraBody } as requestBody; the request handed to responses.send is unchanged.
• src/benchmarks/search/core/solver.test.ts: the plugin-order test also asserts state.requestBody equals the sent body plus extraBody.

The effect/Context fix this PR originally carried (moving stripVariantSuffix into an Effect-free module) landed on main in #116 (stripVariantSuffix now lives in harness/constants.ts), so the merge of main (72f82c9) dropped this branch's model-slug.ts copy and its ./model-slug package export.

Why?

Follow-up to #109 found in review of openrouter-web#46340: searchSolver sent the Switchyard plugin through extraBody but recorded requestBody without it, so stored runs did not show which algorithm was requested.

How to test

bun test src/benchmarks/search/core/solver.test.ts

Expected: all pass, including state.requestBody equal to the sent body plus extraBody.

How to test in prod

This repo ships to production through the openrouter-web subtree copy (packages/bench-harness) and the benchmark worker deploy, so verify after the next subtree pull and worker deploy.

• What to verify: a search benchmark run with switchyardAlgorithm set stores plugins: [{ "id": "switchyard-router", "algorithm": "<algorithm>" }] in each sample's requestBody; a run without it stores no plugins key.
• Steps:
  1. Start a search_hle run with model: "nvidia/switchyard", models: ["openai/gpt-6-astra", "google/gemini-3.1-pro"], switchyardAlgorithm: "stage", limit: 2 through the benchmark start endpoint or the bench run CLI.
  2. Open the run's sample results and read requestBody.plugins on any sample.
• Expected: requestBody.plugins contains the switchyard-router entry with algorithm: "stage"; the scored answer and cost fields are unchanged from a run on the previous commit.
• Rollback signal: requestBody.plugins missing on a Switchyard run, or any search run failing at completedState with a serialization error.

Benchmark impact

None to scores. Search runs now store the plugins array in requestBody when a switchyardAlgorithm or web-search plugin is set; runs with neither are byte-identical.

Reviewer focus

• completedState spreads extraBody after request, matching how the SDK applies extraBody on send.
• Net diff against main is the two solver files only.

Checklist

• Tests cover changed behavior
• Public API or configuration changes are backward compatible, or the break is documented
• Benchmark changes document dataset provenance and licensing (n/a)
• No credentials, private results, or restricted dataset contents are included
• Documentation is updated where needed (n/a)

Link to Devin session: https://openrouter.devinenterprise.com/sessions/37ef700865744be284bb0127c4f6a8d0
Open in Devin Desktop: https://openrouter.devinenterprise.com/desktop/session/37ef700865744be284bb0127c4f6a8d0?variant=devin
Requested by: @jamespsterling

---

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access that start with 'DevinAI' or '@devin'.

• Disable automatic comment, CI, and merge conflict monitoring

---

Original prompt from James

Analyze this session: @devin-session:devin-9b303b51c68b4f74a224b28cefa855ae

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.

[perry-the-pr-reviewer]

⚠️ The maintainer app is not installed on OpenRouterTeam, so Perry cannot post an APPROVE. This review is a COMMENT with a full LGTM assessment — a human should approve.

Perry's Review

Verdict: ✅ LGTM

Risk: 🟢 Low

Risk assessment

Dimension	Severity	Risk	Reasoning
Implementation risk	🟩	Low	One-line addition: completedState spreads extraBody into requestBody, matching how mergeExtraBody applies it on the wire. Test verifies the stored body equals the sent body plus extraBody.
Premise risk	🟩	Low	The premise — stored runs should reflect what was actually sent, including the Switchyard plugin — is correct and well-motivated by the linked openrouter-web review.
Estimated impact	🟩	Low	Only affects the stored requestBody field; the wire request is unchanged. Worst case is a stale requestBody that omits plugins, which was the pre-PR state.

Risk Factor	Severity	Risk	Reasoning
Reversibility	🟩	Low	The previous state can be fully restored by reverting the one line.
Detectability	🟩	Low	Tests assert the new behavior; any regression would fail the test suite.
Blast radius	🟩	Low	Affects only the search solver's requestBody storage field.
Data integrity	🟩	Low	No persisted data is corrupted; requestBody is a diagnostic record.
Financial exposure		None	No billing or payment paths touched.
Security and privacy exposure		None	No credentials or sensitive data in requestBody.
Propagation	🟩	Low	Downstream consumers of requestBody (parquet export, run results) read it as opaque JSON.
Availability		None	The change cannot affect serving.
Recovery cost	🟩	Low	Reverting restores the prior behavior with no data migration.
Time to correct	🟩	Low	A one-line revert if needed.

Analysis

The fix is clean and correct. searchSolver builds extraBody containing the Switchyard router plugin via switchyardExtraBody(), passes it to responses.send() as options.extraBody, and mergeExtraBody in the responses client spreads it into the wire request. However, completedState was recording requestBody: { ...request } — without extraBody — so stored runs lost the plugins field. This PR threads extraBody through to completedState and spreads it after request, so requestBody now matches the effective sent body.

When extraBody is undefined (no Switchyard algorithm set), { ...request, ...undefined } is identical to { ...request } — so non-Switchyard runs are byte-identical. When extraBody has a plugins key, it correctly overrides request.plugins because switchyardExtraBody already includes the existing wire plugins (extracted via responsesRequestToJSON) plus the Switchyard plugin, so nothing is lost.

The test captures the sent body and asserts state.requestBody equals { ...sent, ...sentOptions?.extraBody }, which directly verifies the fix.

CI

All checks pass — validate (CI), CodeQL (actions + javascript-typescript), and Devin Review.