@qwen-code /triage

Thanks for the PR!

Template is loosely followed — "What" / "Why" / "Reviewer Test Plan" / "Risk" sections are present and clear. Missing the "Tested on" matrix, "Linked Issues", and "中文说明" sections from the template. Not blocking for a focused data fix like this.

On direction: solid. This is the same class of fix as #5268 (DeepSeek presets) — modalityDefaults.ts only marks qwen3.6-plus and kimi-k2.5 as multimodal, so the Token Plan and Idealab presets shouldn't override that with image: true, video: true on text-only models. Overstating modalities can mislead the vision bridge into sending unsupported media types. Aligned.

On approach: minimal and focused. Three modalities lines removed from two preset files, plus regression assertions ensuring the Plus/2.5 variants stay multimodal. I verified completeness — grep across all presets confirms these are the only entries with the mismatch; alibaba-coding-plan.ts and minimax.ts are consistent with modalityDefaults.ts. No scope creep, exactly the right size.

Moving on to code review. 🔍

— Qwen Code · qwen3.7-max

2a. Code Review

Independent assessment before reading the diff: for text-only models like qwen3.6-flash and kimi-k2.6, the preset shouldn't declare modalities: { image: true, video: true } when modalityDefaults.ts correctly resolves them to text-only via the catch-all patterns (^qwen → {}, ^kimi- → {}). The fix should remove those overrides and add regression guards.

The PR's approach matches exactly: three modalities lines removed from two preset files, regression assertions added for both the text-only models (qwen3.6-flash, kimi-k2.6, bailian/kimi-k2.6) and the genuinely multimodal ones (qwen3.6-plus, kimi-k2.5). No correctness issues, no security concerns, no AGENTS.md violations. Clean.

2b. Testing

Non-UI change (preset config data) — tmux not applicable. Unit test verification instead:

 ✓ src/providers/__tests__/presets/idealab.test.ts (6 tests) 8ms
 ✓ src/providers/__tests__/presets/alibaba-token-plan.test.ts (2 tests) 6ms

 Test Files  2 passed (2)
      Tests  8 passed (8)
   Duration  7.89s

All 8 tests pass — the updated assertions correctly verify text-only configs for qwen3.6-flash / kimi-k2.6, and the new regression assertions guard the multimodal Plus/2.5 variants. The idealab test gained a dedicated it() block for bailian/kimi-k2.6.

— Qwen Code · qwen3.7-max

Stepping back: this is a clean, surgical data fix. Three lines of overstated modality flags removed from presets that modalityDefaults.ts already resolves correctly as text-only. The PR author clearly understood the problem — same class of mismatch as #5268, just in presets that were missed.

My independent proposal before seeing the diff was exactly what the PR does: remove the modalities overrides from qwen3.6-flash, kimi-k2.6 in Token Plan, and bailian/kimi-k2.6 in Idealab, then add regression tests. The PR matches the minimal correct fix. I verified completeness by grepping all preset files — no other entries have this mismatch.

Tests confirm the fix works: 8/8 pass, including new regression guards for the genuinely multimodal Plus/2.5 variants. If I had to maintain this in six months, I'd thank the author for keeping it small and well-tested.

Approving. ✅

— Qwen Code · qwen3.7-max

✅ Local build & verification — LGTM, ready to merge

Built the real artifacts from this PR's HEAD and verified the change end-to-end, including the shipped CLI chunk and the runtime modality fallback it relies on. Correct, complete, well-guarded by tests. Recommending merge.

What this PR does

Removes the overstated modalities: { image: true, video: true } from three text-only preset entries — qwen3.6-flash & kimi-k2.6 (Token Plan) and bailian/kimi-k2.6 (Idealab) — so they fall through to the text-only default in modalityDefaults.ts. The genuinely multimodal qwen3.6-plus / kimi-k2.5 entries are left untouched. Same class of mismatch as #5268 (DeepSeek presets).

Environment

• macOS (Darwin 25.5.0), Node v22.22.2, npm 10.9.7
• Worktree on PR HEAD b987ad9465; base origin/main (12 commits behind but MERGEABLE; the 4 changed files are clean against origin/main).

Why it's correct (data flow)

buildGenerationConfig (provider-config.ts) only emits generationConfig.modalities when the ModelSpec declares it with a truthy value:

if (spec.modalities && Object.values(spec.modalities).some(Boolean)) {
  parts.modalities = spec.modalities;
}

So dropping the field makes the key absent, and at runtime the model resolves through defaultModalities(model). I exercised that real function for the affected ids:

The fix simply makes the presets agree with defaultModalities. This is already the idiomatic pattern in the repo — e.g. alibaba-standard.ts lists qwen3.6-plus with no explicit modalities and relies on the default.

Verification matrix

Fix confirmed in the shipped artifacts

Inspected the compiled @qwen-code/qwen-code-core dist and the CLI's code-split chunk (dist/chunks/chunk-*.js, where the provider presets actually ship — they're tree-shaken out of the main dist/cli.js):

• qwen3.6-flash, kimi-k2.6, bailian/kimi-k2.6 → contextWindowSize + enableThinking, no modalities ✅
• qwen3.6-plus, kimi-k2.5 → still carry modalities: { image: true, video: true } ✅

These presets are consumed by the real CLI provider-setup flow (useAuth.ts, useProviderSetupFlow.ts, acpAgent.ts, runQwenServe.ts), so the fix has genuine user-facing effect.

Notes (non-blocking)

• Test layer: packages/core tests resolve @qwen-code/qwen-code-core to the compiled dist/ (no vitest src alias), so a source-only edit needs a core rebuild to be reflected — the mutation A/B above accounts for this (mutate src → rebuild → test fails).

Verdict: LGTM ✅ — a clean, complete data-consistency fix; presets now match modalityDefaults, multimodal variants preserved, and the new assertions guard against regression.

if (spec.modalities && Object.values(spec.modalities).some(Boolean)) {
  parts.modalities = spec.modalities;
}

@qwen-code /triage