Support Z.ai GLM (glm-acp-agent) as a text generation provider

[ben-w-smith]

Problem

I use Z.ai GLM through the ACP Registry driver (glm-acp-agent) and can't pick it as the text generation model, so thread titles, commit messages, PR text, and branch names need a second provider.

Current behavior

It's blocked in the same three places described in #31680:

• The registry 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.
• The server replaces a registry selection with the fallback model (selectionSupportsTextGeneration in apps/server/src/serverSettings.ts), so even a saved selection never reaches GLM.

Proposal

Add text generation for glm-acp-agent only, opt-in through the existing pickers. It follows the Antigravity precedent (AntigravityTextGeneration.ts), like #31680 does for Copilot. Other registry agents keep today's behavior, GLM is never picked as an automatic fallback, and review guides stay unsupported because they need to read the checkout.

I have a working implementation in #31716. I opened it so the code is visible while you decide on direction and scope, and I'm happy to change it or wait.

How the implementation isolates GLM (glm-acp-agent 1.14.0)

• Each request runs in an empty temp working directory, and fails if any file appears there.
• The session runs in GLM's default mode. Only the model and thinking level are applied from the saved selection, so a saved bypass_permissions mode can't reach the helper.
• Permission requests, questions, file reads and writes, terminals, and extension requests are refused and fail the request. Any tool_call or tool_call_update also fails it.
• ACP_GLM_SESSION_DIR points at a temp folder for that request. GLM supports closing a session but not deleting one, so the helper closes the session and then deletes the folder. Nothing is added to ~/.local/state/glm-acp-agent/sessions.
• ACP_GLM_MAX_TURNS=1. In default mode GLM runs reads and web lookups without asking, so the tool-call check can only stop the request after a tool has started. With one turn, a tool's output is never sent back to the model.
• It runs through the agent with the instance's own command, environment, and sign-in. Nothing calls the Z.ai API directly.

Testing so far

• Unit tests cover the guards, the mode allowlist, temp folder cleanup, review guide rejection, the registry flag, and the server keeping a GLM selection while still rejecting other registry agents.
• I ran it live on macOS with glm-5.3-flash at Low thinking and generated 5 titles, about 6 seconds each. Each request's session file appeared in its temp folder and was deleted afterwards. Nothing was written to the workspace. The mode stayed default even after I picked "Bypass all permissions" in the text generation picker.
• Not yet tested: Linux and Windows. I only generated titles live; commit messages, PR text, and branch names use the same request path.

Questions for maintainers

1. Scope: Is GLM-only right, next to Copilot in Support Copilot as a text generation provider #31680? Each agent needs its own isolation settings, so I don't think one rule fits every registry agent.
2. Duplication: The helper copies Antigravity's guard code instead of sharing it. Is that acceptable for a first version, or would you rather share it first?
3. Reads in default mode: GLM can read a file before the tool-call check stops the request. The one-turn limit keeps that content from reaching the model. Is that acceptable?
4. Cost: Each generated title is a real GLM request against the user's Z.ai plan. Is opt-in with a note in the docs enough?

Alternatives

• Keep the current behavior and require another provider for text generation.