feat(editors): add github-copilot-vscode adapter

docs/copilot-vscode/README.md

@@ -0,0 +1,132 @@
+# GitHub Copilot in VS Code → Opper
+
+The CLI now ships an adapter that does the Stable-channel setup for you:
+
+```bash
+opper editors github-copilot-vscode
+```
+
+It detects the OAI Compatible community extension, prompts before
+installing it (with a marketplace link and a clean cancel path), then
+writes the Opper provider block into your VS Code user `settings.json`.
+Remove with `opper editors github-copilot-vscode --remove`.
+
+The artifacts in this directory are the manual recipe — useful as a
+reference, for users who don't have the CLI installed, or for the
+Insiders channel where the native BYOK path isn't currently usable.
+
+## What this gets you
+
+- **In scope:** Copilot **Chat** and **Agent mode** in VS Code, answered by
+  Opper-routed models from the curated picker (`src/config/models.ts`).
+- **Out of scope:** Inline ghost-text completions still go to GitHub's own
+  Copilot service. BYOK does not redirect them. Embeddings, repository
+  indexing, intent detection and a few other side queries also keep
+  hitting GitHub's service.
+- **Auth:** API key entered once via VS Code's "Manage Models" UI; stored
+  in the OS keychain (Insiders) or by the community extension (Stable),
+  not in `settings.json`.
+- **Subscription gates:** Copilot Free / Pro have BYOK on by default.
+  Copilot Business / Enterprise users need their org admin to enable the
+  "Bring Your Own Language Model Key in VS Code" policy.
+
+## Files
+
+- `generate.ts` — derives both JSON snippets from `PICKER_MODELS`. Run
+  `npx tsx docs/copilot-vscode/generate.ts` after any model-list change.
+- `insiders-settings.json` — paste-ready block for **VS Code Insiders**
+  using the native `github.copilot.chat.customOAIModels` setting.
+- `stable-settings.json` — paste-ready block for **VS Code Stable**, used
+  with the community extension `johnny-zhao.oai-compatible-copilot`.
+
+## VS Code Insiders setup (UI path)
+
+> **Status (May 2026):** The native settings-driven path
+> (`github.copilot.chat.customOAIModels`) is **deprecated** in the Copilot
+> Chat extension shipped with Insiders 1.120 — it lives under
+> `ConfigKey.Deprecated.CustomOAIModels` in `vscode-copilot-chat` source
+> with a "remove after 6 months" TODO. It runs a one-shot migration into
+> a new BYOK storage system, then ignores subsequent edits.
+>
+> So Insiders has no working declarative configuration today. Use the
+> UI flow below; we'll revisit declarative setup once Microsoft ships
+> the array-based replacement for `customOAIModels`.
+
+The `insiders-settings.json` snippet in this directory is preserved as a
+reference but should **not** be merged into Insiders user settings —
+it'll be ignored.
+
+1. Open Copilot Chat → click the model picker → **Manage Models**.
+2. Click **+ Add Models…** → **OpenAI Compatible**.
+3. When prompted:
+   - Base URL: `https://api.opper.ai/v3/compat`
+   - API key: your `OPPER_API_KEY`
+   - Model id: `claude-opus-4-7` (start here; add more once one works)
+4. Repeat step 2 for any additional model id from `PICKER_MODELS` you
+   want surfaced.
+5. Pick the model in the chat picker → send a message → confirm the
+   trace lands on `https://platform.opper.ai/traces`.
+
+## VS Code Stable setup (community extension)
+
+Until the native custom-OAI provider lands in Stable, the community
+extension is the most painless path.
+
+1. Install the extension:
+   `code --install-extension johnny-zhao.oai-compatible-copilot`
+   (or search for "OAI Compatible Provider for Copilot" in the
+   Extensions sidebar).
+2. Open user `settings.json` and merge the contents of
+   `stable-settings.json` into it (top-level keys: `oaicopilot.baseUrl`
+   and `oaicopilot.models`).
+3. Reload the window.
+4. Open Copilot Chat → model picker → **Manage Models** → choose **OAI
+   Compatible** → enter your `OPPER_API_KEY` when prompted → select the
+   Opper models to surface in the picker.
+5. Send a chat message using one of the new models and verify the trace
+   appears on Opper.
+
+## Testing checklist
+
+For each model worth validating (start with `claude-opus-4-7` and one
+non-Anthropic model — `gpt-5.5` or `gemini-3.1-pro-preview`):
+
+- [ ] Plain chat reply renders end-to-end.
+- [ ] Streaming chunks arrive incrementally (not just at the end).
+- [ ] **Agent mode** runs a multi-step task that uses tools (e.g.
+      `read_file`, `apply_patch`).
+- [ ] Long-context request (paste a 20k-token prompt) doesn't truncate.
+- [ ] Vision: drag in an image — for models flagged `vision: true` it
+      should be accepted; for `vision: false` it should be rejected
+      cleanly rather than silently ignored.
+- [ ] Thinking / reasoning content surfaces (or is suppressed cleanly)
+      for models flagged `thinking: true`.
+- [ ] Trace lands on `https://platform.opper.ai/traces` with the right
+      model id.
+
+## What to flag back
+
+If anything in the table below is wrong, ping me and I'll regenerate:
+
+- A capability flag (`toolCalling`, `vision`, `thinking`) that doesn't
+  match what Opper actually reports for a model.
+- A `maxInputTokens` / `maxOutputTokens` that VS Code rejects or that the
+  upstream model can't honour.
+- A model id that `/v3/compat` returns "unknown model" for.
+- Any model that needs a different `apiMode` than `"openai"` (Stable
+  extension only — Insiders pins to OpenAI shape).
+
+## Promotion to phase 2
+
+Once the JSON is stable, the adapter lives in
+`src/agents/github-copilot-vscode.ts` as a configure-only adapter (mirror
+of `claude-desktop.ts`):
+
+- `configure()` writes a sentinel-bracketed `customOAIModels` entry into
+  the user `settings.json` for whichever channel is detected.
+- `unconfigure()` strips it.
+- No `spawn()` — VS Code is launched by the user, not by us.
+
+The `generate.ts` logic here becomes the body of `configure()`; the
+`VISION` set either graduates onto `PickerModel` in `src/config/models.ts`
+or sits beside the new adapter.

docs/copilot-vscode/generate.ts

@@ -0,0 +1,86 @@
+/**
+ * Phase-1 scratch generator for the GitHub Copilot in VS Code BYOK
+ * integration. Reads PICKER_MODELS and emits the two settings.json snippets
+ * sitting next to this file:
+ *
+ *   - insiders-settings.json: native `github.copilot.chat.customOAIModels`
+ *     block (works in VS Code Insiders 1.104+).
+ *   - stable-settings.json: `oaicopilot.*` block consumed by the
+ *     "OAI Compatible Provider for Copilot" community extension on stable.
+ *
+ * Run from the repo root:
+ *   npx tsx docs/copilot-vscode/generate.ts
+ *
+ * When phase 2 promotes this to a real adapter, the logic here moves into
+ * `src/agents/github-copilot-vscode.ts` and the capability table either
+ * graduates onto `PickerModel` or sits beside the adapter.
+ */
+import { writeFile } from "node:fs/promises";
+import { dirname, join } from "node:path";
+import { fileURLToPath } from "node:url";
+import { PICKER_MODELS } from "../../src/config/models.js";
+
+const PROVIDER_URL = "https://api.opper.ai/v3/compat";
+const PROVIDER_NAME = "Opper";
+const DEFAULT_MAX_OUTPUT = 32_768;
+
+// Conservative vision allowlist — flip during testing as we confirm each
+// model's compat behaviour. Keeping non-vision off by default avoids the
+// picker advertising image support that the upstream model can't honour.
+const VISION = new Set<string>([
+  "claude-opus-4-7",
+  "claude-sonnet-4-6",
+  "claude-haiku-4-5",
+  "gpt-5.5",
+  "gemini-3.1-pro-preview",
+]);
+
+const here = dirname(fileURLToPath(import.meta.url));
+
+// Insiders' native customOAIModels uses an OBJECT keyed by model id, with
+// each entry repeating the provider URL. The grouped-array shape (one
+// provider, nested `models`) is a still-open feature request
+// (microsoft/vscode#277102), not the implemented schema — using it
+// registers the provider name but silently drops the model list.
+const insidersBlock = {
+  "github.copilot.chat.customOAIModels": Object.fromEntries(
+    PICKER_MODELS.map((m) => [
+      m.id,
+      {
+        name: `${PROVIDER_NAME} · ${m.id}`,
+        url: PROVIDER_URL,
+        requiresAPIKey: true,
+        toolCalling: true,
+        vision: VISION.has(m.id),
+        thinking: m.reasoning,
+        maxInputTokens: m.contextWindow,
+        maxOutputTokens: DEFAULT_MAX_OUTPUT,
+      },
+    ]),
+  ),
+};
+
+const stableBlock = {
+  "oaicopilot.baseUrl": PROVIDER_URL,
+  "oaicopilot.models": PICKER_MODELS.map((m) => ({
+    id: m.id,
+    owned_by: "opper",
+    displayName: m.id,
+    apiMode: "openai",
+    context_length: m.contextWindow,
+    max_tokens: DEFAULT_MAX_OUTPUT,
+    vision: VISION.has(m.id),
+    enable_thinking: m.reasoning,
+  })),
+};
+
+await writeFile(
+  join(here, "insiders-settings.json"),
+  `${JSON.stringify(insidersBlock, null, 2)}\n`,
+);
+await writeFile(
+  join(here, "stable-settings.json"),
+  `${JSON.stringify(stableBlock, null, 2)}\n`,
+);
+
+console.log("wrote insiders-settings.json and stable-settings.json");

docs/copilot-vscode/insiders-settings.json

@@ -0,0 +1,104 @@
+{
+  "github.copilot.chat.customOAIModels": {
+    "claude-opus-4-7": {
+      "name": "Opper · claude-opus-4-7",
+      "url": "https://api.opper.ai/v3/compat",
+      "requiresAPIKey": true,
+      "toolCalling": true,
+      "vision": true,
+      "thinking": true,
+      "maxInputTokens": 1000000,
+      "maxOutputTokens": 32768
+    },
+    "claude-sonnet-4-6": {
+      "name": "Opper · claude-sonnet-4-6",
+      "url": "https://api.opper.ai/v3/compat",
+      "requiresAPIKey": true,
+      "toolCalling": true,
+      "vision": true,
+      "thinking": true,
+      "maxInputTokens": 1000000,
+      "maxOutputTokens": 32768
+    },
+    "claude-haiku-4-5": {
+      "name": "Opper · claude-haiku-4-5",
+      "url": "https://api.opper.ai/v3/compat",
+      "requiresAPIKey": true,
+      "toolCalling": true,
+      "vision": true,
+      "thinking": false,
+      "maxInputTokens": 200000,
+      "maxOutputTokens": 32768
+    },
+    "gpt-5.5": {
+      "name": "Opper · gpt-5.5",
+      "url": "https://api.opper.ai/v3/compat",
+      "requiresAPIKey": true,
+      "toolCalling": true,
+      "vision": true,
+      "thinking": true,
+      "maxInputTokens": 1050000,
+      "maxOutputTokens": 32768
+    },
+    "gemini-3.1-pro-preview": {
+      "name": "Opper · gemini-3.1-pro-preview",
+      "url": "https://api.opper.ai/v3/compat",
+      "requiresAPIKey": true,
+      "toolCalling": true,
+      "vision": true,
+      "thinking": true,
+      "maxInputTokens": 1048576,
+      "maxOutputTokens": 32768
+    },
+    "deepinfra/kimi-k2.6": {
+      "name": "Opper · deepinfra/kimi-k2.6",
+      "url": "https://api.opper.ai/v3/compat",
+      "requiresAPIKey": true,
+      "toolCalling": true,
+      "vision": false,
+      "thinking": true,
+      "maxInputTokens": 262144,
+      "maxOutputTokens": 32768
+    },
+    "deepinfra/glm-5.1": {
+      "name": "Opper · deepinfra/glm-5.1",
+      "url": "https://api.opper.ai/v3/compat",
+      "requiresAPIKey": true,
+      "toolCalling": true,
+      "vision": false,
+      "thinking": true,
+      "maxInputTokens": 202752,
+      "maxOutputTokens": 32768
+    },
+    "fireworks/minimax-m2p7": {
+      "name": "Opper · fireworks/minimax-m2p7",
+      "url": "https://api.opper.ai/v3/compat",
+      "requiresAPIKey": true,
+      "toolCalling": true,
+      "vision": false,
+      "thinking": false,
+      "maxInputTokens": 196608,
+      "maxOutputTokens": 32768
+    },
+    "deepinfra/deepseek-v4-pro": {
+      "name": "Opper · deepinfra/deepseek-v4-pro",
+      "url": "https://api.opper.ai/v3/compat",
+      "requiresAPIKey": true,
+      "toolCalling": true,
+      "vision": false,
+      "thinking": true,
+      "maxInputTokens": 1048576,
+      "maxOutputTokens": 32768
+    },
+    "deepinfra/deepseek-v4-flash": {
+      "name": "Opper · deepinfra/deepseek-v4-flash",
+      "url": "https://api.opper.ai/v3/compat",
+      "requiresAPIKey": true,
+      "toolCalling": true,
+      "vision": false,
+      "thinking": true,
+      "maxInputTokens": 1048576,
+      "maxOutputTokens": 32768
+    }
+  }
+}