feat: add GLM (Zhipu AI) provider support

[Aneaire]

Summary

• Adds GLM (Zhipu AI) as a new provider alongside Codex, enabling users to use GLM models (glm-4.7, glm-4.7-flash, glm-5) via the OpenAI-compatible chat completions API
• Full HTTP-based adapter with SSE streaming, tool use (read/write/edit files, run commands, list directories, search files), and an agent loop
• GLM health check validates GLM_API_KEY / ZAI_API_KEY at startup with a status banner in the UI
• Web UI updates: provider picker, model selector, settings page for custom GLM models

Files changed

Contracts (packages/contracts/):

• ProviderKind extended from "codex" to ["codex", "glm"]
• Added GLM models, aliases, and model options

Shared (packages/shared/):

• Added inferProviderFromModel() utility and GLM model slug set

Server (apps/server/):

• New GlmAdapter service + layer (HTTP adapter with SSE streaming and 6 tools)
• Registered in ProviderAdapterRegistry and serverLayers
• GLM health check in ProviderHealth
• Updated ProviderSessionDirectory and ProviderCommandReactor to recognize "glm"

Web (apps/web/):

• Settings page with GLM custom model management
• Provider picker with GLM icon
• Store and composer draft store updated for GLM provider

Setup

Set GLM_API_KEY (or ZAI_API_KEY) as an environment variable before starting the server:

export GLM_API_KEY=your-api-key

The API key can be obtained from z.ai.

Test plan

• Verify Codex provider still works unchanged (no regressions)
• Set GLM_API_KEY and confirm the GLM health check shows "ready" in the UI
• Without GLM_API_KEY, confirm the health banner shows the missing key message
• Select GLM provider and a GLM model, send a message, verify streaming response
• Verify tool use works (file read/write, command execution)
• Verify custom GLM models can be added/removed in Settings
• Verify model picker shows GLM models when GLM provider is selected

🤖 Generated with Claude Code

Note

Add GLM (Zhipu AI) as a provider alongside Codex

• Adds a full GLM provider adapter in GlmAdapter.ts that streams OpenAI-compatible chat completions from the GLM API, executes tool calls locally (filesystem + shell), and emits lifecycle/runtime events matching the Codex provider contract.
• Registers GLM in the provider registry, health checks (keyed on GLM_API_KEY/ZAI_API_KEY), session directory, and orchestration layers so GLM sessions are treated as first-class alongside Codex.
• Extends contracts and shared model utilities in model.ts with GLM model options, default models, slug aliases, and reasoning effort maps.
• Adds GLM to the web UI: provider picker, model picker with GlmIcon, settings page for custom models, and health banner.
• Risk: tool calls (write, edit, run_command) execute with local filesystem and shell access; approval-required mode gates these but must be explicitly enabled per session.

^{Macroscope summarized e6899af.}

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro

Run ID: bdb48ca5-85f0-48c5-8f9b-6c48ef3ffd97

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Post copyable unit tests in a comment

📝 Coding Plan

• Generate coding plan for human review comments

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

Tip

You can disable the changed files summary in the walkthrough.

Disable the reviews.changed_files_summary setting to disable the changed files summary in the walkthrough.

[macroscopeapp]

🟡 Medium Layers/GlmAdapter.ts:361

getSession throws ProviderAdapterSessionNotFoundError synchronously inside Effect.sync() blocks, so the exception becomes a Cause.Die defect instead of a typed Cause.Fail. Methods like sendTurn, interruptTurn, respondToRequest, readThread, and rollbackThread therefore die unrecoverably rather than returning a typed ProviderAdapterError that callers can handle. Consider using Effect.fail(new ProviderAdapterSessionNotFoundError(...)) or Effect.try with catch to convert the throw into a typed failure.

-      const getSession = (threadId: ThreadId): GlmSession => {
-        const session = sessions.get(threadId);
-        if (!session) {
-          throw new ProviderAdapterSessionNotFoundError({
-            provider: PROVIDER,
-            threadId,
-          });
-        }
-        return session;
-      };

🚀 Reply "fix it for me" or copy this AI Prompt for your agent:

In file apps/server/src/provider/Layers/GlmAdapter.ts around lines 361-370:

`getSession` throws `ProviderAdapterSessionNotFoundError` synchronously inside `Effect.sync()` blocks, so the exception becomes a `Cause.Die` defect instead of a typed `Cause.Fail`. Methods like `sendTurn`, `interruptTurn`, `respondToRequest`, `readThread`, and `rollbackThread` therefore die unrecoverably rather than returning a typed `ProviderAdapterError` that callers can handle. Consider using `Effect.fail(new ProviderAdapterSessionNotFoundError(...))` or `Effect.try` with `catch` to convert the throw into a typed failure.

Evidence trail:
1. GlmAdapter.ts lines 361-370: `getSession` function throws `ProviderAdapterSessionNotFoundError` synchronously
2. GlmAdapter.ts line 740-741: `sendTurn` calls `getSession` inside `Effect.sync()`
3. GlmAdapter.ts line 788-789: `interruptTurn` calls `getSession` inside `Effect.sync()`
4. GlmAdapter.ts line 806-807: `respondToRequest` calls `getSession` inside `Effect.sync()`
5. GlmAdapter.ts line 858-859: `readThread` calls `getSession` inside `Effect.sync()`
6. GlmAdapter.ts line 867-868: `rollbackThread` calls `getSession` inside `Effect.sync()`
7. Effect-TS docs (https://effect-ts.github.io/effect/effect/Effect.ts.html): "The provided function (thunk) must not throw errors; if it does, the error will be treated as a 'defect'."

[macroscopeapp]

🟢 Low Layers/GlmAdapter.ts:307

In parseSseStream, when reader.read() returns done: true, the function breaks immediately without processing any remaining data in buffer. If the server sends a final data: line without a trailing newline before closing the connection, that chunk is lost and never yielded.

🚀 Reply "fix it for me" or copy this AI Prompt for your agent:

In file apps/server/src/provider/Layers/GlmAdapter.ts around line 307:

In `parseSseStream`, when `reader.read()` returns `done: true`, the function breaks immediately without processing any remaining data in `buffer`. If the server sends a final `data:` line without a trailing newline before closing the connection, that chunk is lost and never yielded.

Evidence trail:
apps/server/src/provider/Layers/GlmAdapter.ts lines 306-337 at REVIEWED_COMMIT - specifically line 319 (`if (done) break;`), lines 323-324 (`const lines = buffer.split("\n"); buffer = lines.pop() ?? "";`), and the absence of any buffer processing after the while loop or in the finally block (line 336-337)

[zortos293]

Julius will add claude soon, and I don’t think we should integrate his change until he releases his own Claude adapter. so you can build ontop of it.
See also: #179