refactor: split god files into domain modules + data-driven model registry

.gitattributes

@@ -0,0 +1,10 @@
+# Normalize every text file to LF in the index and on checkout. This keeps
+# `npm run lint` (prettier + shellcheck, which require LF per .editorconfig)
+# green on Windows, where core.autocrlf=true would otherwise check files out
+# as CRLF. Binary files are untouched (text=auto detects them).
+* text=auto eol=lf
+
+# Shell scripts must stay LF: husky hooks (checked by shellcheck via
+# `npm run lint`) and any POSIX scripts break with CRLF on Windows checkouts.
+.husky/* text eol=lf
+*.sh text eol=lf

CHANGELOG.md

@@ -2,6 +2,28 @@
 
 All notable changes to the **OpenCode Go BYOK Provider** extension are documented here.
 
+## [Unreleased]
+
+### Changed
+
+- **`[Internal]` API keys are configured through the BYOK panel only.** The `OpenCode Go: Set API Key` / `OpenCode Zen: Set API Key` commands and the "Set / Clear API Key" menu items inside `Manage Provider` are removed — keys are entered once via **Chat: Manage Language Models → "+ Add Models"** (the native BYOK flow). `SecretStorage` is no longer a user-facing entry point; it stays as an internal per-vendor mirror (`opencodego.apiKey` / `opencodezen.apiKey`) that the BYOK resolution writes so agent-host variants and cold-start requests inherit the group key. Splitting the secret per vendor also fixes a latent collision where Go and Zen shared a single `opencodego.apiKey` and overwrote each other's key. `Refresh Models` / `Test Connection` now point at the BYOK flow when no key is configured instead of prompting for one.
+
+- **`[Internal]` Per-provider Thinking strategy classes + single config authority.** The thinking/reasoning system is refactored from one monolithic builder into a per-provider strategy (`src/thinking/`): an interface + factory (`provider.ts`), a shared base class, and one class per model family (`deepseek`, `glm`, `kimi`, `minimax`, `openai`, `qwen`, `mimo`, `fallback`). Each provider now owns its reasoning picker schema, its request-payload mapping, and whether its `reasoning_content` is surfaced as chat content. Configuration resolves from a **single authority** — the VS Code per-model configuration (model picker / Manage), with workspace settings and per-family defaults as fallbacks — instead of competing sources (workspace + modelConfiguration + a `globalState` shadow copy + defaults). The shadow copy is removed, so a thinking effort chosen for one model can no longer silently leak onto another model or override an explicit "Off". Model IDs are normalized to `effectiveModelId` (the `::sk-***` fp suffix is gone), which also stops the per-model settings group from being recreated on every pick. Request builders are split out of `extension.ts` into per-endpoint modules (`src/request/{types,schema,shared,openai,anthropic,google}.ts`). Windows tooling fixes: `scripts/lint.ts` runs npm `.cmd` shims through the shell and a new `.gitattributes` enforces LF normalization, so `npm run lint` (prettier + shellcheck) is green on Windows; `scripts/staged-lint.ts` and `isCwdInWorkspace` get the same treatment.
+
+- **`[Internal]` God files split into domain modules (no behavior change).** The three monolithic files — `src/extension.ts` (4653 lines), `src/streaming.ts` (1620) and `src/goUsageTracker.ts` (1510) — are split into domain folders:
+  - `src/usage/` — Go usage domain: `tracker.ts`, `history.ts` (OpenCode CLI SQLite read/aggregation), `pricing.ts`, `formatting.ts`, `dashboard.ts` (status bar + usage webview incl. the HTML template + tooltip SVG), plus the moved `usage.ts` / `usageProfile.ts` / `goUsageSync.ts`.
+  - `src/transports/` — one file per transport (`chatCompletions`, `responses`, `anthropic`, `google`) plus the shared streaming `engine`, pure `sse` parser, `extractors`, `extract` helpers and `thinkTags` filter; the transport contract types live in `src/core/transport.ts` (routing in `src/core/routing.ts`).
+  - `src/provider/` — `OpenCodeProvider` class, `definitions` (PROVIDERS table + model types), `messages`/`tokens` (message conversion + token estimation), `settings` (schema/getSettings/limits/capabilities), `visionProxy`.
+  - `src/models/` — `metadata`, `modelLimits`, `modelCapabilities`, `modelNames`, `pricing`, `metadataFetcher` (models.dev cache).
+  - `src/commands/` — provider, agent-window, diagnostics and thinking-picker command handlers; `src/request/headers.ts` for the OpenCode request headers.
+  - `extension.ts` is now a thin entry (~400 lines) that only wires activation + command registration. The two compat barrels (`streaming.ts`, `goUsageTracker.ts`) are removed and every importer references canonical paths. All behavior-preserving — verified by `npm run compile` + 291 unit tests + mock-server retry E2E (`npm run test-retry`) + `npm run lint`.
+
+- **`[Internal]` Data-driven model registry (`src/core/registry.ts`).** The transport router and the thinking-family detector previously each owned a hardcoded model-prefix table. Both now read ONE data-driven table: `MODEL_REGISTRY` rows map model-family patterns → `{ endpointKind, sdkPackage, thinkingFamily, vendors? }`. `resolveModelRouting()` honors per-vendor restrictions (e.g. MiniMax `m2.x` → Messages on Go, Gemini → Google on Zen); `thinkingFamily()` reads the same table vendor-agnostically. Adding a new model family = adding one row (+ optionally a thinking strategy class). Context limits / capabilities stay metadata-driven (live models.dev) rather than duplicated in a static table. Behavior-preserving — verified by 14 new registry tests (305 total).
+
+### Fixed
+
+- **DeepSeek / Mimo thinking content no longer leaks into the chat transcript.** `treatReasoningAsContent` was mis-detecting native-reasoning families as "no reasoning in body" and echoing their `reasoning_content` as plain chat text. The decision now comes from the provider strategy (always `false` for DeepSeek and Mimo), so chain-of-thought stays in the thinking panel.
+
 ## [0.6.0] — 2026-08-13
 
 ### Added

README.md

@@ -42,7 +42,7 @@
 | 🎯 **Smart routing**             | Each model family auto-routes to its native transport (`/responses`, `/messages`, `streamGenerateContent`, `/chat/completions`)                                                                                                                       |
 | 🖼️ **Vision + PDF + Audio**      | Multimodal models pass through image, PDF, audio, and video inputs. Oversized images auto-resize to 2000×2000 / 5MB to match the gateway contract.                                                                                                    |
 | 📐 **Context-size picker**       | Kimi K3 and other tiered-context models expose `256K` vs full-window selection in the per-model configuration, with the cheaper tier selected by default.                                                                                             |
-| 🔒 **Your key, your control**    | API key stored in VS Code SecretStorage — never leaves your machine                                                                                                                                                                                   |
+| 🔒 **Your key, your control**    | API key entered once in Language Models → **Add Models…** — stored by VS Code, never leaves your machine                                                                                                                                              |
 
 ---
 
@@ -69,7 +69,7 @@
 5. **Click the model picker** (current model name) → **Add Models…**
 6. **Select** **OpenCode Go** or **OpenCode Zen**.
 7. **Press Enter** to accept the default group name.
-8. **Paste your API key** when prompted (stored securely in VS Code SecretStorage).
+8. **Paste your API key** when prompted (stored by VS Code in your language-models configuration).
 9. **Pick the models** you want enabled.
 10. **Select any OpenCode model** from the picker and start chatting. 🚀
 
@@ -421,11 +421,10 @@ The easiest way to manage your key is **Settings → Language Models** (gear ⚙
 
 | Command                                                   | Description                                                     |
 | --------------------------------------------------------- | --------------------------------------------------------------- |
-| `OpenCode Go: Manage Provider`                            | Manage legacy API key, refresh models, test connection          |
-| `OpenCode Go: Set API Key`                                | Store/update legacy OpenCode Go API key                         |
+| `OpenCode Go: Manage Provider`                            | Test connection, refresh models, configure utility models       |
 | `OpenCode Go: Refresh Models`                             | Force a fresh model-list fetch (bypasses the Manage menu)       |
 | `OpenCode Go: Diagnostics`                                | Report of Go models + request history                           |
-| `OpenCode Zen: Manage Provider`                           | Manage Zen API key, refresh models, test connection             |
+| `OpenCode Zen: Manage Provider`                           | Test connection, refresh models, configure utility models       |
 | `OpenCode Zen: Refresh Models`                            | Force a fresh Zen model-list fetch (bypasses the Manage menu)   |
 | `OpenCode Zen: Diagnostics`                               | Report of Zen models + request history                          |
 | `OpenCode: Model Picker Diagnostics`                      | All registered models (Go + Zen + Copilot) side-by-side         |
@@ -474,7 +473,7 @@ Inline suggestions, next-edit suggestions, semantic search, and embedding-backed
 <details>
 <summary><b>Where is my API key stored?</b></summary>
 
-In VS Code's **SecretStorage** — the same encrypted store used by GitHub auth. It never leaves your machine and is never sent anywhere except directly to `opencode.ai`.
+In your VS Code **language-models configuration** — add it via **Chat: Manage Language Models → Add Models… → OpenCode Go / OpenCode Zen**. VS Code stores the key in its encrypted language-models storage, it never leaves your machine, and it is only sent to `opencode.ai`.
 
 </details>
 

docs/architecture/01-20260514-open-code-provider-architecture.md

@@ -97,16 +97,15 @@ The native provider configuration schema is declared in `package.json` under `co
 
 | Command                                                   | Purpose                                                                    |
 | --------------------------------------------------------- | -------------------------------------------------------------------------- |
-| `OpenCode Go: Manage Provider`                            | Legacy fallback key management, refresh, and connection test               |
-| `OpenCode Go: Set API Key`                                | Legacy fallback key storage                                                |
+| `OpenCode Go: Manage Provider`                            | Refresh models, test connection, configure utility models                  |
 | `OpenCode Go: Remove/Re-add Provider in Language Models`  | Toggle `opencodego.enabled` (remove/re-add the provider, requires reload)  |
 | `OpenCode Zen: Remove/Re-add Provider in Language Models` | Toggle `opencodezen.enabled` (remove/re-add the provider, requires reload) |
 | `OpenCode Go: Diagnostics`                                | Go model and transport diagnostics                                         |
 | `OpenCode Zen: Diagnostics`                               | Zen model and transport diagnostics                                        |
 | `OpenCode: Model Picker Diagnostics`                      | Cross-provider model metadata comparison                                   |
 | `OpenCode: Set Thinking Effort...`                        | Global thinking-mode helper for supported families                         |
 
-The recommended setup path is still VS Code's native **Language Models** UI. The legacy commands remain for diagnostics and fallback compatibility.
+The recommended — and only — setup path is VS Code's native **Language Models** UI ("+ Add Models"). The `Set API Key` command and the legacy key-management menu items were removed (the old single `opencodego.apiKey` secret could not represent both Go and Zen keys); the remaining manage commands cover refresh, connection testing, and diagnostics.
 
 ---
 
@@ -119,14 +118,14 @@ For model discovery (`provideLanguageModelChatInformation`), the extension resol
 1. Read `options.configuration.apiKey` (the native BYOK value) if VS Code supplied one.
 2. If step 1 produced nothing, fall back to `SecretStorage` unconditionally.
 
-The unconditional fallback (since 0.5.0, [#86](https://github.com/ltmoerdani/opencode-copilot-chat/issues/86)) covers users who stored the key via the extension command `OpenCode Go: Set API Key` instead of the native BYOK flow. It mirrors Copilot's own `AbstractLanguageModelChatProvider`, which always falls back to its own storage when `configuration.apiKey` is absent. A per-vendor flag (`hasConfiguredByokGroup`) suppresses the groupless call once a native BYOK group exists, so models are not listed twice ([#106](https://github.com/ltmoerdani/opencode-copilot-chat/issues/106)). A group call whose `configuration` is present but carries no API key is treated as a **per-model configuration group** (only `settings`, created when the user picks e.g. `reasoningEffort` in the model picker) and returns no models, so the groupless call remains the single source; per-model settings still apply at request time via `modelConfiguration` ([#131](https://github.com/ltmoerdani/opencode-copilot-chat/issues/131)).
+The fallback is an internal mirror of Copilot's own `AbstractLanguageModelChatProvider`, which always falls back to its own storage when `configuration.apiKey` is absent. Since the `Set API Key` command was removed, the only writer is the BYOK group resolution itself: when a non-agent provider resolves a key it persists it into its **per-vendor** secret (`opencodego.apiKey` / `opencodezen.apiKey`, resolved via `secretKeyFor()` in `src/config.ts`), so agent-host variants and cold-start requests inherit it. A per-vendor flag (`hasConfiguredByokGroup`) suppresses the groupless call once a native BYOK group exists, so models are not listed twice ([#106](https://github.com/ltmoerdani/opencode-copilot-chat/issues/106)). A group call whose `configuration` is present but carries no API key is treated as a **per-model configuration group** (only `settings`, created when the user picks e.g. `reasoningEffort` in the model picker) and returns no models, so the groupless call remains the single source; per-model settings still apply at request time via `modelConfiguration` ([#131](https://github.com/ltmoerdani/opencode-copilot-chat/issues/131)).
 
 Security rules:
 
 - Real API keys are never written to repository files.
 - Documentation must use placeholders only.
-- API keys should be entered through VS Code's native secret-backed provider configuration.
-- Legacy `SecretStorage` support remains only as a fallback path.
+- API keys are entered through VS Code's native secret-backed provider configuration ("+ Add Models"); there is no command-palette key entry.
+- `SecretStorage` remains only as an internal fallback (per-vendor, mirroring the BYOK group key for agent variants and cold-start requests).
 
 Safe placeholder example:
 

docs/architecture/02-20260809-provider-adapter-architecture.md

@@ -226,9 +226,11 @@ Each phase gate: `npm run compile` must pass + a targeted test with at least 1 m
 
 ## Timeline
 
-| Date       | Status    | Change                                                                        |
-| ---------- | --------- | ----------------------------------------------------------------------------- |
-| 2026-08-09 | 🟢 Active | Initial research + analysis document created. Proposal only; no code changed. |
+| Date       | Status    | Change                                                                                                                                                                                                                                                                                                                                                                                           |
+| ---------- | --------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| 2026-08-09 | 🟢 Active | Initial research + analysis document created. Proposal only; no code changed.                                                                                                                                                                                                                                                                                                                    |
+| 2026-08-14 | ✅ Done   | God-file split executed on `refactor/split-god-files` (usage/ · transports/ · provider/ · models/ · core/ · commands/); `extension.ts` 4653 → ~414 lines. See CHANGELOG `[Unreleased]`.                                                                                                                                                                                                          |
+| 2026-08-14 | ✅ Done   | **Data-driven registry implemented** (`src/core/registry.ts`). `resolveModelRouting()` (transport) and `thinkingFamily()` both read the `MODEL_REGISTRY` table; `ModelEndpointKind` type moved to the registry. Scope note: context limits/capabilities stay metadata-driven (live models.dev) — not duplicated as a static table. The full `ModelTransport` port interface remains future work. |
 
 ---