Thanks for the PR!

Template looks good ✓

On direction: This is a straightforward compatibility fix — tools registered without a description were being silently dropped from OpenAI requests, which is clearly wrong. The existing codebase already has test fixtures that create descriptionless tool declarations (e.g., smoke.test.ts uses { name: 'edit' } without a description), confirming this is a real scenario. Aligned.

On approach: The scope is tight — 2 lines changed in production code, test updated to match. This is exactly the minimal fix needed. The func.name guard remains (OpenAI needs a callable name), and ?? '' for the description field is correct per the OpenAI spec where description is optional. Nothing to cut.

Moving on to code review. 🔍

— Qwen Code · qwen3.7-max

2a. Code Review

Independent proposal (before reading the diff): The convertGeminiToolsToOpenAI function should require only func.name (since OpenAI function tools need a callable name) and treat description as optional, defaulting to empty string. The test should verify three cases: normal tool preserved, descriptionless tool preserved, nameless tool skipped.

Comparison with PR: The PR's implementation matches this proposal exactly — two-line change in production code (func.name && func.description → func.name, func.description → func.description ?? ''), and the test now covers all three cases correctly. Nothing missed, nothing extraneous.

Reuse check: No new logic introduced — just a guard relaxation and nullish coalescing. Nothing to extract or consolidate.

Verdict: No blockers. The change is correct, minimal, and well-tested.

2b. Real-Scenario Testing

This is a non-UI converter fix — the behavioral proof is the unit test suite. Ran the focused convertGeminiToolsToOpenAI tests on both main and the PR branch.

Before (main — old test expects descriptionless tools to be dropped)

=== BEFORE (main) ===

 RUN  v3.2.4 /home/github-runner/actions-runner-7/_work/qwen-code/qwen-code/packages/core
      Coverage enabled with v8

 ✓ src/core/openaiContentGenerator/converter.test.ts (145 tests | 144 skipped) 3ms

 Test Files  1 passed (1)
      Tests  1 passed | 144 skipped (145)
   Start at  15:27:30
   Duration  5.68s (transform 2.21s, setup 19ms, collect 3.43s, tests 3ms, environment 0ms, prepare 98ms)

Old test title: "should skip functions without name or description" — expects toHaveLength(1), meaning the descriptionless tool was silently dropped.

After (PR #6243 — updated test expects descriptionless tools to be preserved)

=== AFTER (PR #6243) ===

 RUN  v3.2.4 /home/github-runner/actions-runner-7/_work/qwen-code/qwen-code/packages/core
      Coverage enabled with v8

 ✓ src/core/openaiContentGenerator/converter.test.ts (145 tests | 138 skipped) 7ms

 Test Files  1 passed (1)
      Tests  7 passed | 138 skipped (145)
   Start at  15:29:00
   Duration  10.67s (transform 4.34s, setup 34ms, collect 7.03s, tests 7ms, environment 0ms, prepare 199ms)

Updated test now asserts toHaveLength(2) — the descriptionless tool is preserved with description: ''. All 7 convertGeminiToolsToOpenAI tests pass.

Testing verdict: ✅ Behavior change confirmed — descriptionless tools are now preserved as expected.

— Qwen Code · qwen3.7-max

Stepping back: this is a clean, correct fix for a real bug. Tools without descriptions were silently dropped from OpenAI requests — that's data loss at the API boundary, and the existing codebase already creates descriptionless tool declarations in its own test fixtures (smoke.test.ts), so this isn't a hypothetical edge case.

The implementation is exactly what I would have written: relax the guard to func.name only, default description to ''. Two production lines, one test update. No drive-by refactors, no scope creep, no unnecessary abstractions. The before/after test run confirms the behavioral change.

If I had to maintain this in six months, I'd thank the author for keeping it minimal.

Approving. ✅

— Qwen Code · qwen3.7-max

✅ Maintainer verification — real-binary E2E confirms the fix

I verified this PR by driving the real bundled qwen binary (dist/cli.js, built from PR head 23eb546b78 via npm ci && npm run build && npm run bundle) end-to-end in tmux against a fake OpenAI‑compatible endpoint, with a real stdio MCP server as the descriptionless‑tool source. I then ran a mutation A/B that reverts exactly this PR's two converter lines to prove causality.

Verdict: the fix does exactly what it claims. With the base guard func.name && func.description, a descriptionless MCP tool is silently dropped from the OpenAI request payload before the model ever sees it. With this PR's guard func.name, the tool is preserved (description: "") and becomes callable end-to-end. The control tool (with a description) is present in both runs, so this is not a general MCP breakage — it is specifically the descriptionless case.

Why this is a faithful real-world reproduction

The descriptionless shape is genuinely reachable end-to-end, not a synthetic unit-test input:

MCP tools/list  (tool has name + inputSchema, NO description key)
  → @google/genai mcpToTool()      → funcDecl.description === undefined   [verified empirically]
  → mcp-client.ts:1039             → DiscoveredMCPTool(…, description ?? '')  → description = ''
  → tools.ts:231 schema getter     → { name, description: '', parametersJsonSchema }
  → pipeline.ts:539                → convertGeminiToolsToOpenAI(request.config.tools)
  → OLD guard `name && description` → '' is falsy → TOOL DROPPED
  → NEW guard `name`               → TOOL KEPT with description: ''

convertGeminiToolsToOpenAI() is the single chokepoint for every OpenAI-compatible request, so any descriptionless tool — MCP or custom adapter — hits this guard.

A/B result (only variable = the 2-line converter guard)

The two binaries differ only in the converter guard — confirmed by grepping the flipped text in both packages/core/dist/.../converter.js and the bundled dist/chunks/*.js. Same MCP server, same fake endpoint, same settings.json, same prompt.

Hard evidence (captured, not narrated)

Descriptionless tool as serialized into the FIXED request payload:

{"type":"function","function":{"name":"mcp__descserver__read_nodesc","description":"","parameters":{"type":"object","properties":{"path":{"type":"string"}},"required":["path"]}}}

Genuine execution (not fabricated): a nonce known only to the MCP server (injected via its env, never in the prompt or the fake) round-trips back through the model request on the FIXED binary:

• MCP server log (FIXED): {"event":"tools/call","name":"read_nodesc","args":{"path":"probe.txt"}}
• request[1] role:tool message: MCP_TOOL_EXECUTED name=read_nodesc … NONCE=nonce-LIVE-2799Q7
• On the MUTANT binary the MCP server logs no tools/call — the tool never reached the model, so it could not be called.

Unit tests (independently re-run at head): converter.test.ts → 135 passed, including the renamed regression should preserve functions without description and skip functions without name.

Scope / notes for merge

• alwaysLoadTools: true in my test config is observability-only. It puts the MCP tools directly in request[0] instead of behind tool_search; the converter is the same single chokepoint either way, so the fix behaves identically for deferred MCP tools too.
• Nameless declarations are still skipped — the PR preserves that. (MCP can't produce a nameless tool; qwen already skips those at discovery, so this branch is covered by the unit test, not the E2E.)
• description: "" is spec-valid for OpenAI-compatible providers (function.description is optional), so no downstream regression. The documented tradeoff (a descriptionless tool gives the model less guidance) is strictly better than silently dropping a tool the user registered.
• No merge conflicts (MERGEABLE); the change is minimal and low-risk.

MCP tools/list（工具有 name + inputSchema，没有 description 字段）
  → @google/genai mcpToTool()      → funcDecl.description === undefined   【已实测确认】
  → mcp-client.ts:1039             → DiscoveredMCPTool(…, description ?? '')  → description = ''
  → tools.ts:231 schema getter     → { name, description: '', parametersJsonSchema }
  → pipeline.ts:539                → convertGeminiToolsToOpenAI(request.config.tools)
  → 旧 guard `name && description`  → '' 为 falsy → 工具被丢弃
  → 新 guard `name`                → 工具保留，description 为 ''

{"type":"function","function":{"name":"mcp__descserver__read_nodesc","description":"","parameters":{"type":"object","properties":{"path":{"type":"string"}},"required":["path"]}}}

_{Verified locally with the real bundled binary + a real stdio MCP server + fake OpenAI endpoint; mutation A/B isolates the two-line converter change. Reproduction harness available on request.}