chore(eval): remove Gemini and Yutori agents

[Felarof (felarof99)]

Summary

• Remove the Gemini Computer Use and Yutori Navigator eval agent implementations and sample configs.
• Prune active registry, schema, dashboard, docs/report, validation, package, and lockfile references to those eval agent types.
• This branch also contains the existing dashboard test-run/save-to-reference commit that was already on the local branch before this removal was pushed.

Test plan

• PATH="$HOME/.bun/bin:$PATH" bun install --frozen-lockfile
• PATH="$HOME/.bun/bin:$PATH" bun run --filter @browseros/eval typecheck
• PATH="$HOME/.bun/bin:$PATH" bun run lint
• rg found no remaining Gemini/Yutori eval-agent references outside ignored/generated files

🤖 Generated with Claude Code

[greptile-apps]

Greptile Summary

This PR completes the removal of the Gemini Computer Use and Yutori Navigator eval agent implementations — deleting ~2,900 lines of agent code, configs, and validation — and bundles two dashboard enhancements: a Test Run mode (1 worker, 1 task) and a Save to Reference button backed by a new PUT /api/config/:name endpoint with solid path-traversal protection.

Confidence Score: 4/5

Safe to merge; only P2 findings present

All issues found are P2 — a minor UI text-state bug and a backwards-compatibility edge case for legacy result files. The removal is thorough, typecheck and lint pass per the test plan, and the new server endpoint is well-guarded.

packages/browseros-agent/apps/eval/src/types/result.ts — narrowed AgentConfigMetaSchema may silently fail on old Gemini/Yutori result directories if re-grading is attempted

Important Files Changed

Filename	Overview
packages/browseros-agent/apps/eval/src/agents/gemini-computer-use/agent.ts	Deleted — Gemini Computer Use agent implementation fully removed
packages/browseros-agent/apps/eval/src/agents/yutori-navigator/agent.ts	Deleted — Yutori Navigator agent implementation fully removed
packages/browseros-agent/apps/eval/src/types/config.ts	Removed GeminiComputerUseConfigSchema and YutoriNavigatorConfigSchema from AgentConfigSchema discriminated union
packages/browseros-agent/apps/eval/src/types/result.ts	AgentConfigMetaSchema type enum narrowed to only 'single'/'orchestrator-executor'; may silently break re-grading of old Gemini/Yutori result directories
packages/browseros-agent/apps/eval/src/dashboard/server.ts	Adds PUT /api/config/:name (Save to Reference) with robust path-traversal protection and schema validation; adds testRun mode limiting to 1 worker + 1 task
packages/browseros-agent/apps/eval/src/dashboard/index.html	Adds Test Run button and Save to Reference button; minor UI bug — test button text stays "Starting…" while eval runs
packages/browseros-agent/apps/eval/src/utils/config-validator.ts	Removed gemini-computer-use API key env-var check; remaining logic is clean
packages/browseros-agent/apps/eval/scripts/weekly-report.ts	Removed gemini-computer-use label mappings; yutori-navigator never had a dedicated label, so no change needed there

Sequence Diagram

sequenceDiagram
    participant UI as Dashboard (index.html)
    participant Server as server.ts
    participant FS as File System

    Note over UI,FS: Save to Reference flow
    UI->>Server: PUT /api/config/:name (JSON body)
    Server->>FS: stat(filepath) — must already exist
    Server->>Server: EvalConfigSchema.safeParse(body)
    Server->>FS: Bun.write(filepath, JSON)
    Server-->>UI: { status: 'saved', name }

    Note over UI,FS: Test Run flow
    UI->>Server: POST /api/run { testRun: true }
    Server->>FS: loadTasks(datasetPath)
    Server->>Server: tasks = tasks.slice(0, 1)
    Server->>Server: ParallelExecutor({ numWorkers: 1 })
    Server-->>UI: { status: 'started', taskCount: 1, testRun: true }
    UI->>UI: setEvalRunningUI(true)

Loading

Prompt To Fix All With AI

This is a comment left during a code review.
Path: packages/browseros-agent/apps/eval/src/dashboard/index.html
Line: 944

Comment:
**Test button text stuck on "Starting..." during run**

After a successful submission, `setEvalRunningUI(true)` is called but only sets `testBtn.disabled = true` — it never resets `testBtn.textContent`. This means the test button shows "Starting..." for the entire duration of the eval until `setEvalRunningUI(false)` is called on completion. Consider adding `testBtn.textContent = 'Test Run'` (or "Running…") inside the `if (running)` branch.

How can I resolve this? If you propose a fix, please make it concise.

---

This is a comment left during a code review.
Path: packages/browseros-agent/apps/eval/src/types/result.ts
Line: 14-19

Comment:
**Narrowed enum breaks parsing of legacy result files**

`AgentConfigMetaSchema` now only accepts `'single'` and `'orchestrator-executor'`. Any previously-stored `metadata.json` file with `agent_config.type: 'gemini-computer-use'` or `'yutori-navigator'` will fail the `TaskMetadataSchema.parse()` call in `trajectory-saver.ts#loadMetadata`. The `hasExistingGraderResults` helper swallows the Zod error (returns `exists: false`), but `loadMetadata()` in `updateGraderResults` has no try/catch, so re-grading an old Gemini/Yutori result directory would throw. Consider using `z.string()` for the `type` field (the `.passthrough()` already allows unknown fields) if backwards compatibility with existing result files matters.

How can I resolve this? If you propose a fix, please make it concise.

_{Reviews (1): Last reviewed commit: "chore(eval): remove Gemini and Yutori ag..." | Re-trigger Greptile}

[greptile-apps]

Narrowed enum breaks parsing of legacy result files

AgentConfigMetaSchema now only accepts 'single' and 'orchestrator-executor'. Any previously-stored metadata.json file with agent_config.type: 'gemini-computer-use' or 'yutori-navigator' will fail the TaskMetadataSchema.parse() call in trajectory-saver.ts#loadMetadata. The hasExistingGraderResults helper swallows the Zod error (returns exists: false), but loadMetadata() in updateGraderResults has no try/catch, so re-grading an old Gemini/Yutori result directory would throw. Consider using z.string() for the type field (the .passthrough() already allows unknown fields) if backwards compatibility with existing result files matters.

Prompt To Fix With AI

This is a comment left during a code review.
Path: packages/browseros-agent/apps/eval/src/types/result.ts
Line: 14-19

Comment:
**Narrowed enum breaks parsing of legacy result files**

`AgentConfigMetaSchema` now only accepts `'single'` and `'orchestrator-executor'`. Any previously-stored `metadata.json` file with `agent_config.type: 'gemini-computer-use'` or `'yutori-navigator'` will fail the `TaskMetadataSchema.parse()` call in `trajectory-saver.ts#loadMetadata`. The `hasExistingGraderResults` helper swallows the Zod error (returns `exists: false`), but `loadMetadata()` in `updateGraderResults` has no try/catch, so re-grading an old Gemini/Yutori result directory would throw. Consider using `z.string()` for the `type` field (the `.passthrough()` already allows unknown fields) if backwards compatibility with existing result files matters.

How can I resolve this? If you propose a fix, please make it concise.