Support Copilot as a text generation provider

[mdelura]

Problem

GitHub Copilot is my only agent provider and I can't choose a text generation model. Settings shows "No text generation providers available", so thread titles aren't generated.

Current behavior

Copilot runs through the generic ACP Registry driver, which doesn't support text generation:

• The driver reports supportsTextGeneration: false, and its generation methods all fail (packages/provider-acp-registry/src/server/driver.ts).
• The Settings pickers hide providers that report this (SettingsPanels.tsx, SourceControlWritingSettings.tsx).
• The server also rejects ACP Registry instances for this setting (selectionSupportsTextGeneration in apps/server/src/serverSettings.ts).

Proposal

Add text generation for Copilot only, opt-in through the existing text generation and source control pickers. It follows the Antigravity precedent (AntigravityTextGeneration.ts): a one-off ACP session in an empty temp directory that decodes a JSON reply.

I have a working implementation, which you can review here: compare with main

I haven't opened a PR, because I'd like your agreement on the direction and scope first.

What I found testing Copilot CLI 1.0.105

• Blocking tools needs a specific flag value. --available-tools= (empty) restricts nothing: the model read a planted file without a permission request. An allowlist naming only an unknown tool (--available-tools=t3_no_tools) leaves it with no tools. Copilot then adds "Info: Disabled tools..." notices to the reply, which the JSON extraction tolerates.
• Cost drops a lot with tools off. Input went from about 30,000 tokens to about 2,100 per request.
• Helper sessions pollute history. They appear in session/list alongside real sessions, and Copilot has no session/delete. A throwaway COPILOT_HOME per request avoids this, and also keeps user-level MCP servers and hooks out of the helper. Sign-in still worked on my machine, where the token is in the credential store.

How the implementation works

• Each request uses a temp cwd, a temp COPILOT_HOME and tools disabled, plus --no-custom-instructions, --no-ask-user, --disable-builtin-mcps, --no-auto-update and --sync-session=off.
• Any tool call, permission request or user question fails the request.
• Only reasoning effort is applied from saved model options, so mode and allow_all can't reach the helper.
• The server accepts a Copilot selection but never falls back to one automatically.
• Settings shows "Choose a model" until a selection is in effect. Otherwise the picker previews Copilot while the server is still using its default.
• Review guides stay unsupported, because they need to read the checkout.

Testing so far

• Unit tests cover the isolation arguments, cleanup, tool and permission rejection, option filtering, and parsing the real reply format.
• A real title was generated through the real driver and Copilot CLI 1.0.105 on Windows, with no helper session left in the real Copilot home.
• Not yet tested: the picker in a running app, and Linux or macOS.

Questions for maintainers

1. Scope: Is Copilot-only the right starting point? Isolation depends on each agent's own flags, so I don't think this generalizes safely to all ACP Registry agents.
2. Sign-in without a credential store: Copilot's help says it falls back to a plain-text token in ~/.copilot/ when there's no credential store, which seems likely on headless servers. I copy config.json into the temp home to cover this, but I haven't tested it with a real plain-text token. Are you comfortable with that copy, or would you prefer a different approach?
3. Flag stability: No-tools relies on how Copilot treats an unknown allowlist entry. If that changes, tool activity fails the request, but a tool could run first inside the empty temp directory. Is that acceptable?
4. Cost: Each generated title is a real Copilot prompt that counts against the user's plan. Model multipliers vary widely (0.33x to 14x in my model list). Is opt-in with a note in the docs enough?

Alternatives

• Keep the current behavior and require another provider for text generation.
• Add a generation method that doesn't use an agent provider. That seems like a much bigger change than this needs.

[ben-w-smith]

I hit the same three blocks with Z.ai GLM (glm-acp-agent), another ACP Registry agent, and opened #31715 for it.

My implementation follows the same approach you describe: opt-in per agent, a temp working directory, only the model and reasoning level applied from saved options, any tool call, permission request, or question failing the request, no automatic fallback, and no review guides. GLM needs different isolation settings than Copilot (a temp ACP_GLM_SESSION_DIR and ACP_GLM_MAX_TURNS=1 instead of a temp COPILOT_HOME and tool flags).

On your first question, I agree that this doesn't generalize safely to every registry agent. If maintainers want one pattern for both agents, I'm happy to match yours so the two changes don't conflict, especially the selectionSupportsTextGeneration change in serverSettings.ts, which we both need.

[JosephLteif]

I opened an issue for this, I would love for this to be implemented