feat: add GLM 4.7 and Claude Code CLI provider support

[Aneaire]

Summary

Adds two new provider backends alongside the existing Codex provider:

• GLM 4.7 — Full provider adapter with spawn_agent sub-agent delegation, real-time progress events, and streaming responses. Includes dedicated agent panels in the work log UI with auto-scroll, task descriptions, and status indicators.
• Claude Code CLI — Provider adapter that wraps the Claude Code CLI binary, with session lifecycle management, stream event parsing, and error mapping. Includes health checks and model catalog for Claude 4.5/4.6.

What's included

Server (apps/server)

• GlmAdapter.ts — GLM provider adapter with sub-agent orchestration
• ClaudeAdapter.ts — Claude Code CLI adapter with stream handling
• Provider registry, health checks, and session directory updated for both providers
• Service interfaces for both adapters

Web (apps/web)

• Agent panel UI in MessagesTimeline with grouped entries, streaming responses, and status pills
• Provider/model picker extended with GLM and Claude options + icons
• Settings page for custom GLM/Claude model configuration and Claude binary path
• Composer draft store persists last selected provider/model across threads
• Plan text rendering fixes (ordering before tool calls, per-iteration itemId)

Contracts (packages/contracts)

• ProviderKind extended with "glm" and "claude"
• Model catalogs, aliases, and defaults for both providers
• Provider runtime event types extended for agent events

Shared (packages/shared)

• Model resolution utilities for GLM and Claude

Test plan

• Verify Codex provider still works as before (no regressions)
• Test GLM 4.7 provider with agent delegation workflows
• Test Claude Code CLI provider with a valid Claude binary
• Verify provider/model picker shows all three providers
• Verify custom model settings persist correctly
• Run bun typecheck and bun lint (both pass)

🤖 Generated with Claude Code

Note

Add GLM 4.7 and Claude Code CLI as first-class provider adapters

• Adds full provider adapter implementations for glm (GlmAdapter.ts) and claude (ClaudeAdapter.ts), each exposing session lifecycle, turn execution, and event streaming via the existing ProviderAdapterShape interface.
• The GLM adapter calls the BigModel API with SSE streaming, executes local tools (file ops, shell commands, grep), and supports recursive sub-agent spawning with approval gating; the Claude adapter shells out to the claude CLI, parsing NDJSON output into canonical ProviderRuntimeEvent records.
• Extends ProviderKind to include "glm" and "claude" throughout contracts, routing, and session logic, and registers both adapters in the server layer by default.
• Adds inferProviderFromModel to deduce the provider from a model slug, used to default the composer provider and resolve custom model lists per provider.
• Adds settings UI for the Claude CLI binary path and custom model slug lists for both new providers.
• WorkLogEntry gains agentName, agentSummary, and streamingResponse fields; the MessagesTimeline component renders agent entries in dedicated scrollable panels with independent expand/collapse state.
• Risk: the Claude adapter's readThread and rollbackThread return empty snapshots (no CLI persistence support), and respondToRequest/respondToUserInput are stubs with no behavior.

📊 Macroscope summarized 91e8f10. 27 files reviewed, 9 issues evaluated, 1 issue filtered, 5 comments posted

🗂️ Filtered Issues

apps/web/src/store.test.ts — 0 comments posted, 1 evaluated, 1 filtered

• line 204: Test assertion contradicts test description. The test is named "falls back to the codex default for unsupported provider models without an active session" but the assertion on line 204 checks for DEFAULT_MODEL_BY_PROVIDER.glm instead of DEFAULT_MODEL_BY_PROVIDER.codex. Either the test name is incorrect or the expected value should be DEFAULT_MODEL_BY_PROVIDER.codex. [ Cross-file consolidated ]

[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: 553f061c-9f2d-4fed-b2c8-149996949c04

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

CodeRabbit can use your project's `biome` configuration to improve the quality of JS/TS/CSS/JSON code reviews.

Add a configuration file to your project to customize how CodeRabbit runs biome.

[macroscopeapp]

🟢 Low Layers/GlmAdapter.ts:593

After the while loop exits, any remaining content in buffer that wasn't terminated by a newline is silently discarded. If the SSE stream's final data event doesn't end with a newline character, that event would be lost. Consider processing any remaining non-empty buffer after the loop ends.

         }
       }
     }
   } finally {
     reader.releaseLock();
   }
+  // Process any remaining data in buffer that wasn't terminated by newline
+  const trimmed = buffer.trim();
+  if (trimmed && trimmed.startsWith("data:")) {
+    const data = trimmed.slice(5).trim();
+    if (data !== "[DONE]") {
+      try {
+        yield JSON.parse(data) as SSEChunk;
+      } catch {
+        // Skip malformed chunk
+      }
+    }
+  }
 }

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

In file apps/server/src/provider/Layers/GlmAdapter.ts around lines 593-626:

After the `while` loop exits, any remaining content in `buffer` that wasn't terminated by a newline is silently discarded. If the SSE stream's final data event doesn't end with a newline character, that event would be lost. Consider processing any remaining non-empty buffer after the loop ends.

Evidence trail:
apps/server/src/provider/Layers/GlmAdapter.ts lines 593-625 at REVIEWED_COMMIT: The parseSseStream function (lines 594-626) shows:
- Line 603: `while (!signal.aborted)` loop
- Lines 608-609: `const lines = buffer.split("\n"); buffer = lines.pop() ?? "";`
- Lines 611-620: processes only complete lines from the split
- Line 605: `if (done) break;` exits the loop
- Lines 623-625: `finally { reader.releaseLock(); }` - no processing of remaining buffer content after loop exits

[macroscopeapp]

🔴 Critical src/session-logic.test.ts:679

The test expects PROVIDER_OPTIONS to contain a glm entry as its first element, but the implementation does not include glm at all. This causes the array length and contents assertion to fail when the test runs against the actual PROVIDER_OPTIONS export.

     expect(PROVIDER_OPTIONS).toEqual([
-      { value: "glm", label: "GLM (z.ai)", available: true },
       { value: "codex", label: "Codex", available: true },
       { value: "claude", label: "Claude Code", available: true },
       { value: "cursor", label: "Cursor", available: false },
     ]);

Also found in 1 other location(s)

apps/web/src/store.test.ts:204

Test assertion contradicts test description. The test is named "falls back to the codex default for unsupported provider models without an active session" but the assertion on line 204 checks for DEFAULT_MODEL_BY_PROVIDER.glm instead of DEFAULT_MODEL_BY_PROVIDER.codex. Either the test name is incorrect or the expected value should be DEFAULT_MODEL_BY_PROVIDER.codex.

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

In file apps/web/src/session-logic.test.ts around lines 679-684:

The test expects `PROVIDER_OPTIONS` to contain a `glm` entry as its first element, but the implementation does not include `glm` at all. This causes the array length and contents assertion to fail when the test runs against the actual `PROVIDER_OPTIONS` export.

Evidence trail:
- apps/web/src/session-logic.test.ts lines 680-686 (REVIEWED_COMMIT): Test expects `PROVIDER_OPTIONS` to include `{ value: "glm", label: "GLM (z.ai)", available: true }` as the first element.
- apps/web/src/session-logic.ts lines 23-32 (REVIEWED_COMMIT): `PROVIDER_OPTIONS` is defined with only three entries (`codex`, `claude`, `cursor`), with no `glm` entry.

Also found in 1 other location(s):
- apps/web/src/store.test.ts:204 -- Test assertion contradicts test description. The test is named "falls back to the codex default for unsupported provider models without an active session" but the assertion on line 204 checks for `DEFAULT_MODEL_BY_PROVIDER.glm` instead of `DEFAULT_MODEL_BY_PROVIDER.codex`. Either the test name is incorrect or the expected value should be `DEFAULT_MODEL_BY_PROVIDER.codex`.

[macroscopeapp]

🟡 Medium src/session-logic.ts:458

Lines 458–460 assign payload.itemType to entry.itemType without validation, bypassing extractWorkLogItemType which uses isToolLifecycleItemType to ensure the value is a valid ToolLifecycleItemType. When payload.itemType is an invalid string, this code overwrites the correctly-rejected undefined with the invalid string, breaking the type constraint.

-      if (payload && typeof payload.itemType === "string") {
-        entry.itemType = payload.itemType;
-      }

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

In file apps/web/src/session-logic.ts around lines 458-460:

Lines 458–460 assign `payload.itemType` to `entry.itemType` without validation, bypassing `extractWorkLogItemType` which uses `isToolLifecycleItemType` to ensure the value is a valid `ToolLifecycleItemType`. When `payload.itemType` is an invalid string, this code overwrites the correctly-rejected `undefined` with the invalid string, breaking the type constraint.

Evidence trail:
apps/web/src/session-logic.ts lines 458-460 (bypass validation), lines 435 and 453-455 (correct validation usage), lines 542-548 (extractWorkLogItemType definition with isToolLifecycleItemType check). packages/contracts/src/providerRuntime.ts lines 111-113 (isToolLifecycleItemType validates against TOOL_LIFECYCLE_ITEM_TYPES).

[macroscopeapp]

🟡 Medium Layers/ClaudeAdapter.ts:701

interruptTurn sets session.activeProcess = null and session.status = "ready" immediately after proc.kill("SIGINT"), before the process has actually exited. This creates a race where sendTurn can spawn a new process while the old one is still running, causing interleaved events from both processes. Consider waiting for the close event before clearing the session state.

-          session.activeProcess = null;
-          session.status = "ready";
-          session.activeTurnId = null;
-          session.updatedAt = nowIso();
-
-          emit(
-            makeEvent(threadId, "session.state.changed", {
-              state: "ready",
-            }),
-          );
+          // State updates deferred to close handler to avoid race conditions
+          // where a new turn could be started while the old process is still running
+          session.updatedAt = nowIso();
+          // session.activeProcess, status, and activeTurnId will be cleared in close handler

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

In file apps/server/src/provider/Layers/ClaudeAdapter.ts around lines 701-710:

`interruptTurn` sets `session.activeProcess = null` and `session.status = "ready"` immediately after `proc.kill("SIGINT")`, before the process has actually exited. This creates a race where `sendTurn` can spawn a new process while the old one is still running, causing interleaved events from both processes. Consider waiting for the `close` event before clearing the session state.

Evidence trail:
apps/server/src/provider/Layers/ClaudeAdapter.ts lines 672-708 (interruptTurn function): line 678 calls `proc.kill("SIGINT")`, lines 687-689 register an async `close` event handler, but lines 699-701 immediately set `session.activeProcess = null`, `session.status = "ready"`, and `session.activeTurnId = null` without waiting for the close event.

apps/server/src/provider/Layers/ClaudeAdapter.ts lines 640-668 (sendTurn function): calls `runClaudeProcess` without checking if a previous process is still running.

apps/server/src/provider/Layers/ClaudeAdapter.ts lines 490-510 (runClaudeProcess function): spawns a new process and sets `session.activeProcess = proc` without guarding against an existing process.

[macroscopeapp]

🟡 Medium Layers/GlmAdapter.ts:647

getSession throws ProviderAdapterSessionNotFoundError directly, but it's called inside Effect.sync() at lines 1083, 1131, 1149, 1201, and 1210. This converts the typed error into an untyped defect that bypasses Effect.catchTag and similar handlers, causing sendTurn, interruptTurn, respondToRequest, readThread, and rollbackThread to die with a defect instead of returning a typed error in the Effect error channel.

  const getSession = (threadId: ThreadId): Effect.Effect<GlmSession, ProviderAdapterSessionNotFoundError> =>
    sessions.get(threadId)
      ? Effect.succeed(sessions.get(threadId)!)
      : new ProviderAdapterSessionNotFoundError({
          provider: PROVIDER,
          threadId,
        });

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

In file apps/server/src/provider/Layers/GlmAdapter.ts around lines 647-656:

`getSession` throws `ProviderAdapterSessionNotFoundError` directly, but it's called inside `Effect.sync()` at lines 1083, 1131, 1149, 1201, and 1210. This converts the typed error into an untyped defect that bypasses `Effect.catchTag` and similar handlers, causing `sendTurn`, `interruptTurn`, `respondToRequest`, `readThread`, and `rollbackThread` to die with a defect instead of returning a typed error in the Effect error channel.

Evidence trail:
GlmAdapter.ts lines 647-655 show `getSession` throwing `ProviderAdapterSessionNotFoundError` directly. Lines 1083, 1131, 1149, 1201, 1210 show calls to `getSession` inside `Effect.sync()`. Effect-TS documentation confirms that exceptions thrown inside `Effect.sync()` become defects, not typed errors: https://github.com/ethanniser/effect-workshop/blob/main/CHEATSHEET.md, https://github.com/PaulJPhilp/EffectPatterns

[tejas-hosamani]

Ref: #179

[Noojuno]

@Aneaire You should probably split this up into 2 PR's, since it's two distinct changes (adding Claude Code support and adding GLM support)

It's also worth noting @juliusmarminge already has a PR to add Claude Code (#179), so I imagine when it comes down to it they will be using that :)