[Critical] config.getFastModel?.() should be config.getCompactionModel()

The PR adds getCompactionModel() on Config (with the correct fallback chain: compactionModel → fastModel → main model), but the compression service bypasses it entirely and hard-codes the fast model. Setting /model --compaction <id> persists the value and syncs it to runtime Config, but compression always uses the fast model — the feature is dead code end-to-end.

Also update the test in chatCompressionService.test.ts to mock getCompactionModel instead of getFastModel.

— qwen3.7-max via Qwen Code /review

[Critical] Copy-paste bug: (isCompactionModelMode || !m.visionOnly) should be (isCompactionModelMode || !m.voiceOnly)

The compaction filter reuses !m.visionOnly from the vision line above. But getAvailableModelIds (line 219) and the CLI validation (line 613) both use !m.voiceOnly for compaction mode (same as fast mode). The dialog and CLI will show different model sets: the dialog excludes vision-only models (irrelevant to text compression) but includes voice-only models (unsuitable for compression).

— qwen3.7-max via Qwen Code /review

[Critical] handleSelect is missing an isCompactionModelMode handler block

The callback has handlers for isVoiceModelMode, isFastModelMode, and isVisionModelMode but no if (isCompactionModelMode) branch. When a user opens the compaction model picker via /model --compaction (no model-id) and selects a model from the dialog, the code falls through to the main model selection path (config.switchModel(...)), silently changing the user's primary model instead of setting the compaction model.

Add an if (isCompactionModelMode) block mirroring the isFastModelMode pattern:

1. Encode the selector via encodeAuxModelSelector(selected)
2. Persist via settings.setValue(scope, 'compactionModel', fastModel)
3. Call config?.setCompactionModel(fastModel)
4. Add a history item and return

Also add isCompactionModelMode to the handleSelect dependency array (currently missing — the callback won't re-create when compaction mode activates, causing a stale closure).

— qwen3.7-max via Qwen Code /review

[Critical] Test verifies the wrong behavior

This test was renamed from "forwards model, maxAttempts, and thinkingConfig to runSideQuery" to "passes getFastModel to runSideQuery for compression" and asserts model: 'fast-model-v1'. But the compression service should use config.getCompactionModel() (which falls back to the fast model), not config.getFastModel() directly.

After fixing the service to use getCompactionModel():

• Mock getCompactionModel returning 'compaction-model-v1' and assert it's used
• Add a fallback test: when getCompactionModel returns undefined, assert the fast model is used

— qwen3.7-max via Qwen Code /review

[Suggestion] CompressOptions.model is now a dead interface field

The model field remains required (model: string) in the CompressOptions interface and all callers in geminiChat.ts still compute and pass it, but compress() no longer destructures or uses it (the diff removed model from the destructure at line 319). This is a misleading API contract — callers believe they're controlling the compression model, but the value is silently discarded.

Either remove model from CompressOptions and update all callers, or (if keeping the interface for future use) mark it as @deprecated with a comment.

— qwen3.7-max via Qwen Code /review

[Critical] visionOnly is never propagated — the feature is dead on arrival

This PR adds visionOnly?: boolean to AvailableModel and uses it in ModelDialog filters (lines 257-258), but modelRegistry.ts was never updated to propagate the field from model configs. The registry only copies fastOnly and voiceOnly (modelRegistry.ts:233-234). Since no code ever sets visionOnly on a model, m.visionOnly is always undefined and !undefined is true — both vision mode and compaction mode filters that reference visionOnly have zero effect.

Add visionOnly: model.visionOnly, to modelRegistry.ts alongside the existing fastOnly/voiceOnly propagation, and add validation to reject models with both visionOnly and other flags set (mirroring the existing fastOnly + voiceOnly check at modelRegistry.ts:332).

— qwen3.7-max via Qwen Code /review

[Suggestion] Zero test coverage for /model --compaction command handler

85+ lines of new command handling code with no tests. The existing --fast and --vision flags have extensive test coverage that should be mirrored here. Missing tests for:

• Setting a valid compaction model: /model --compaction <model-id>
• Rejecting unavailable compaction models (both bare and auth-qualified)
• Non-interactive mode showing current compaction model status
• Dialog action when /model --compaction is run with no args in interactive mode
• Config.getCompactionModel() / setCompactionModel() / resolveCompactionModelSelector() methods
• isCompactionModelMode state management in useModelCommand

— qwen3.7-max via Qwen Code /review

[Critical] The newly added filter condition (isCompactionModelMode || !m.visionOnly) silently prevents vision-only models from appearing in the /model --vision dialog.

availableModelEntries filters registry models with two back-to-back conditions:

(isVisionModelMode || !m.visionOnly) &&   // ← existing line: shows all models in vision mode
(isCompactionModelMode || !m.visionOnly), // ← new line added by this PR

Both conditions are ANDed. When isVisionModelMode = true and isCompactionModelMode = false (i.e., the user runs /model --vision):

• Condition 1 evaluates to true || !m.visionOnly = true (correct — include vision-only models)
• Condition 2 evaluates to false || !m.visionOnly = !m.visionOnly (wrong — excludes vision-only models)

The combined AND result is !m.visionOnly, so vision-only models can never appear in the selection list, defeating the purpose of condition 1. The fix is either to remove the new line (since condition 1 already handles the vision filter) or to add isVisionModelMode to the new guard:

(isCompactionModelMode || isVisionModelMode || !m.visionOnly),

— claude-sonnet-4-6 via Qwen Code /review

Generated by Claude Code

[Critical] handleSelect has no isCompactionModelMode branch. Selecting a model while in compaction mode falls through to config.switchModel(), which switches the primary chat model instead of persisting a compaction model.

The fast/voice/vision handlers each early-return with mode-specific persistence (settings.setValue(scope, 'fastModel', ...), 'visionModel', etc.). The compaction mode is missing an equivalent:

if (isCompactionModelMode) {
  const compactionModel = encodeAuxModelSelector(selected);
  const scope = getPersistScopeForModelSelection(settings);
  settings.setValue(scope, 'compactionModel', compactionModel);
  config?.setCompactionModel?.(compactionModel);
  uiState?.historyManager.addItem({ type: 'success', text: `Compaction Model: ${compactionModel}` }, Date.now());
  onClose();
  return;
}

Without this, /model --compaction + Enter overwrites the user's primary model setting, which is a destructive side effect of what should be a narrowly-scoped compaction model selection.

— claude-sonnet-4-6 via Qwen Code /review

Generated by Claude Code

[Critical] i18n test failure — missing translations for new description

The new description string with --compaction for chat compression model has no entries in zh.js or zh-TW.js. Both are strict-parity locales, so mustTranslateKeys.test.ts fails, blocking CI.

Add the translation key to both Chinese locale files (packages/cli/src/i18n/locales/zh.js and zh-TW.js).

— qwen3.7-max via Qwen Code /review

[Suggestion] getCompactionModel() returns undefined instead of falling back when configured model is unavailable

When compactionModel is set but the model is no longer available (e.g., provider changed, credentials expired), this returns undefined instead of falling through to getFastModel() ?? getModel(). The documented priority chain (compactionModel → fastModel → main model) breaks at the first step.

— qwen3.7-max via Qwen Code /review

[Suggestion] Three isCompactionModelMode integration gaps in this component

isCompactionModelMode is destructured here but not used in several places where the other mode flags are checked:

1. Dialog title (~line 790): No isCompactionModelMode branch in the title ternary — falls through to generic "Select Model" instead of "Select Compaction Model".

2. Left-arrow escape (~line 491): useKeypress handler checks isFastModelMode || isVoiceModelMode || isVisionModelMode but omits isCompactionModelMode — left arrow won't dismiss the compaction dialog.

3. Preferred entry (~line 430-483): No preferredCompactionModelEntry computed — dialog opens with cursor at the wrong position instead of highlighting the currently configured compaction model.

Each should mirror the pattern used by isFastModelMode/isVisionModelMode.

— qwen3.7-max via Qwen Code /review

[Suggestion] Misleading "clear" instruction — no clear mechanism exists

This message tells users to use "/model --compaction " to clear the override, but when an empty/whitespace model name is passed, the code only shows the info message or opens the dialog — it never calls setCompactionModel(undefined) or persistSetting(settings, 'compactionModel', '').

Either implement the clear path (mirroring how --fast could work), or remove the false instruction from this message.

— qwen3.7-max via Qwen Code /review

[Critical] Missing auth-type validation gate

The --fast handler (line 463) explicitly checks config.getContentGeneratorConfig()?.authType and returns an error when unavailable. This --compaction handler skips that gate entirely, going straight to resolveModelId(modelName).

When no auth type is configured, /model --compaction <id> produces a confusing error path instead of the clear "Authentication type not available." message that --fast shows.

— qwen3.7-max via Qwen Code /review

[Critical] Cross-provider model ambiguity — raw modelName persisted without authType qualifier

The --vision handler (line 614-617) builds qualifiedModelName = \${authType}:${modelId}`before persisting. This handler persists the baremodelName. When the same model ID (e.g., gpt-4o`) is configured under multiple providers, compression traffic could silently route to a different provider than the user intended.

— qwen3.7-max via Qwen Code /review

[Suggestion] Missing test coverage for compaction mode in useModelCommand.test.ts

The hook adds isCompactionModelMode with mutual-exclusion logic (compaction suppresses fast/voice/vision) and closeModelDialog resets it, but useModelCommand.test.ts has no test verifying:

• openModelDialog({ compactionModelMode: true }) sets isCompactionModelMode and suppresses other modes
• closeModelDialog() resets isCompactionModelMode to false

Existing tests cover voice/vision modes; compaction should mirror that coverage.

— qwen3.7-max via Qwen Code /review

[Suggestion] Duplicated resolution logic — resolveCompactionModelSelector() (line 3085) is a verbatim copy of resolveFastModelSelector() (line 3021)

Both methods share identical control flow: check field → resolveModelId → bare-name fallback via currentAuthType → getAllConfiguredModels. The only differences are this.fastModel vs this.compactionModel and the debug log string.

With three auxiliary model types (fast/vision/compaction), extracting a shared helper would prevent drift:

private resolveAuxModelSelector(fieldName: 'fastModel' | 'visionModel' | 'compactionModel', label: string) {
  const value = this[fieldName];
  if (!value) return undefined;
  // ... shared logic ...
}

— qwen3.7-max via Qwen Code /review

[Suggestion] Left-arrow dismiss and preferred-entry gaps for isCompactionModelMode

Two integration points missed beyond the existing gaps already noted on this component:

1. Left-arrow handler (~line 506): (isFastModelMode || isVoiceModelMode || isVisionModelMode) omits isCompactionModelMode. In all other specialized pickers, left-arrow dismisses the dialog.

2. Preferred entry (~line 440): The preferredKey resolution chain has no preferredCompactionModelEntry, so the current compaction model is never highlighted in the picker.

— qwen3.7-max via Qwen Code /review

[Suggestion] Model filter inconsistency across three compaction code paths

The model filter for compaction mode differs between the tab-completion path, the interactive dialog, and the execute handler:

A user can tab-complete a fastOnly or visionOnly model with /model --compaction <TAB>, but the dialog would never show those models, and the execute handler would accept them.

Suggested fix: Align all three paths. If compaction is a text-summarization task that doesn't need specialized models, all three should exclude fastOnly, voiceOnly, and visionOnly:

— qwen3.7-max via Qwen Code /review

[Suggestion] Missing test for compaction-model dialog dispatch

Parallel tests exist for 'voice-model' and 'vision-model' dialog cases in slashCommandProcessor.test.ts, verifying that openModelDialog is called with the correct mode flag. This new 'compaction-model' case is structurally identical but has no corresponding test.

Suggested fix: Add a test case that mocks a command returning { type: 'dialog', dialog: 'compaction-model' } and asserts openModelDialog is called with { compactionModelMode: true }.

— qwen3.7-max via Qwen Code /review

[Critical] --project/--global scope flags silently dropped in all three --compaction code paths

Unlike the --fast, --voice, and --vision handlers, the --compaction handler never propagates scopeOverride to persistence or the dialog:

1. Line 758 — This dialog return is missing ...persistScopeSpread(scopeOverride) (compare --fast at line 552, --voice at line 483, --vision at line 660).
2. Line 808 — persistSetting(settings, 'compactionModel', modelName) is missing the 4th argument scopeOverride (compare --fast at line 609: persistSetting(settings, 'fastModel', modelName, scopeOverride)).
3. slashCommandProcessor.ts:850 — actions.openModelDialog({ compactionModelMode: true }) is missing persistScope: result.persistScope (compare vision-model case which passes it).

Impact: Running /model --compaction gpt-4o --project silently saves to the default scope instead of workspace. The success message also omits scopeSuffix, so the user receives no feedback that --project was ignored.

— qwen3.7-max via Qwen Code /review