fix(docs,speed): repoint folded doc link, correct the cap comment

sepion02 · sepion02 · commit d1e97b822081 · 2026-07-28T20:30:18.000+02:00
- llm-providers.md pointed at agent-provider-capabilities.ts, which this PR
  folded into chat-model.ts; the Docs check was failing on the dead link.
- Rename the orphaned test file to match its subject.
- The cap comment claimed a 50x entry 'clamps to 16'. It does not: the parser
  REJECTS it and the field reverts. Note that the bound is editor-input only,
  so agent-authored regions can still exceed it.
- normalizeProjectEditor discarded an out-of-range saved speed and fell back to
  the 1.5x default; clamp it to the cap instead. Lowering MAX_PLAYBACK_SPEED
  widened that silent-reset window from (100,inf) to (16,inf).
diff --git a/electron/ai-edition/deep-agent/chat-model-capabilities.test.ts b/electron/ai-edition/deep-agent/chat-model-capabilities.test.ts
diff --git a/src/components/video-editor/projectPersistence.ts b/src/components/video-editor/projectPersistence.ts
@@ -34,10 +34,8 @@ import {
 	DEFAULT_ZOOM_MOTION_BLUR,
 	MAX_BLUR_BLOCK_SIZE,
 	MAX_BLUR_INTENSITY,
-	MAX_PLAYBACK_SPEED,
 	MIN_BLUR_BLOCK_SIZE,
 	MIN_BLUR_INTENSITY,
-	MIN_PLAYBACK_SPEED,
 	type SpeedRegion,
 	type TrimRegion,
 	type WebcamLayoutPreset,
@@ -320,12 +318,13 @@ export function normalizeProjectEditor(editor: Partial<ProjectEditorState>): Pro
 					const startMs = Math.max(0, Math.min(rawStart, rawEnd));
 					const endMs = Math.max(startMs + 1, rawEnd);
 
-					const speed =
-						isFiniteNumber(region.speed) &&
-						region.speed >= MIN_PLAYBACK_SPEED &&
-						region.speed <= MAX_PLAYBACK_SPEED
-							? clampPlaybackSpeed(region.speed)
-							: DEFAULT_PLAYBACK_SPEED;
+					// Clamp an out-of-range speed rather than discarding it: a saved 25×
+					// should become the cap, not silently reset to the 1.5× default.
+					// Lowering MAX_PLAYBACK_SPEED widened the reset window, so the
+					// range test now only guards against non-numeric values.
+					const speed = isFiniteNumber(region.speed)
+						? clampPlaybackSpeed(region.speed)
+						: DEFAULT_PLAYBACK_SPEED;
 
 					return {
 						id: region.id,
diff --git a/src/components/video-editor/types.ts b/src/components/video-editor/types.ts
@@ -395,8 +395,14 @@ export const MIN_PLAYBACK_SPEED = 0.1;
 // frame-steps by seeking and audio export uses an offline pitch-preserved stretch.
 // ponytail: the editor cap and the native preview cap are the same number —
 // everything above this is the export path's stretch, and the editor's input
-// parser shouldn't accept speeds the preview cannot render. A 50× entry now
-// clamps to 16, not 100.
+// parser shouldn't accept speeds the preview cannot render. A 50× entry is
+// REJECTED (the field reverts to its previous value), not clamped — see
+// `parseCustomPlaybackSpeedInput`, which returns `too-fast` above this bound.
+//
+// NOTE: this bound is enforced on the editor input only. The LLM agent tools
+// (electron/ai-edition/agent-tools.ts) still accept any positive speed, so an
+// agent-authored region can exceed it: preview clamps to MAX_NATIVE_PLAYBACK_RATE
+// while export renders the true speed.
 export const MAX_PLAYBACK_SPEED = 16;
 export const MAX_NATIVE_PLAYBACK_RATE = 16;
 
diff --git a/technical-documentation/architecture/llm-providers.md b/technical-documentation/architecture/llm-providers.md
@@ -8,7 +8,7 @@ The provider layer defines model metadata, protects credentials, discovers model
 | [`electron/ai-edition/llm-config-store.ts`](../../electron/ai-edition/llm-config-store.ts) | `LlmConfigStore` — plain JSON for selection, `safeStorage` blob for credentials. |
 | [`electron/ai-edition/llm-provider-auth.ts`](../../electron/ai-edition/llm-provider-auth.ts) | Model-list discovery per provider. Despite the filename it performs no authentication any more — see [Known gaps](#known-gaps). |
 | [`electron/ai-edition/deep-agent/chat-model.ts`](../../electron/ai-edition/deep-agent/chat-model.ts) | `createOpenScreenChatModel` — the single transport. Picks a `@langchain/*` chat model class per provider. |
-| [`electron/ai-edition/deep-agent/agent-provider-capabilities.ts`](../../electron/ai-edition/deep-agent/agent-provider-capabilities.ts) | Per-provider reasoning-effort capability and its LangChain wire options. |
+| [`electron/ai-edition/deep-agent/chat-model.ts`](../../electron/ai-edition/deep-agent/chat-model.ts) | Per-provider reasoning-effort capability and its LangChain wire options. |
 | [`electron/native-bridge/services/aiEditionService.ts`](../../electron/native-bridge/services/aiEditionService.ts) | IPC surface: connect / disconnect, snapshot, `llmListProviderModels`. |
 | [`src/components/ai-edition/ProviderSettings.tsx`](../../src/components/ai-edition/ProviderSettings.tsx) | Renders cards and forms directly from `PROVIDER_DEFINITIONS`. |
 
@@ -111,14 +111,14 @@ Everything returns `{models, error?}` rather than throwing, so the settings UI c
 
 1. Add a complete `ProviderDefinition` in `provider-registry.ts` (auth kind, env keys, default model, base URL, `wireProtocol`, reasoning support). Widen `authKind` if the provider is not API-key-based.
 2. Add a branch in `createOpenScreenChatModel` if none of the three existing adapters fits.
-3. Add a capability branch in `agent-provider-capabilities.ts` if the provider exposes reasoning, and constrain `getReasoningEffortOptions` if its scale is not the full six tiers.
+3. Add a capability branch in `chat-model.ts` if the provider exposes reasoning, and constrain `getReasoningEffortOptions` if its scale is not the full six tiers.
 4. Add a discovery branch in `aiEditionService.llmListProviderModels`, plus its fetch helper in `llm-provider-auth.ts`.
 5. Extend the native-bridge contracts if the provider needs operations the existing IPC surface doesn't cover.
 6. Confirm `ProviderSettings.tsx` renders the right fields from the registry metadata alone, then add registry, transport, and UI tests.
 
 ## Known gaps
 
-- **`normalizeReasoningEffort` in `provider-registry.ts` is dead.** It is exported but has no caller anywhere in the repo; `normalizeReasoningEffortForCapability` in `agent-provider-capabilities.ts` is the live one. The two also disagree — the dead copy's strategy union knows `custom-openai-account` but not `minimax-thinking`. Delete it rather than fixing it.
+- **`normalizeReasoningEffort` in `provider-registry.ts` is dead.** It is exported but has no caller anywhere in the repo; `normalizeReasoningEffortForCapability` in `chat-model.ts` is the live one. The two also disagree — the dead copy's strategy union knows `custom-openai-account` but not `minimax-thinking`. Delete it rather than fixing it.
 - **`custom-openai-account` is a phantom strategy.** It appears in `ReasoningCapability["strategy"]` but no branch of `getReasoningCapability` returns it, and no branch of `buildLangChainReasoningOptions` handles it. Left over from the removed ChatGPT provider.
 - **`llm-provider-auth.ts` is misnamed.** It performs no authentication since the device flows were removed — it is purely model-list discovery. `model-discovery.ts` would say what it does.
 - **MiniMax discovery spends the user's key.** Nine probe requests per discovery click, uncached, at `max_tokens: 1`. Cache per key if it ever moves to a hot path.
