Fix GLM-5 and other reasoning models appearing to hang via OpenAI shim

[otaviocarvalho]

Summary

• Added reasoning_content support to the OpenAI shim's streaming and non-streaming response handlers
• Reasoning models (GLM-5, GLM-5.1, DeepSeek) stream chain-of-thought in reasoning_content while content stays empty "", causing the shim to emit nothing and the UI to appear frozen indefinitely

Impact

• user-facing impact: GLM-5/5.1 and other reasoning models now work correctly via CLAUDE_CODE_USE_OPENAI=1. Reasoning is displayed as thinking blocks, actual reply as text. Previously these models appeared to hang forever.
• developer/maintainer impact: OpenAIStreamChunk and non-streaming response types now include reasoning_content. The streaming parser has a new thinking/content phase transition. No API changes to the shim's public interface.

Testing

• bun run build
• bun run smoke
• focused tests: bun test src/services/api/openaiShim.test.ts — 6/6 pass

Notes

• provider/model paths tested:
  • OpenCode Zen: OPENAI_BASE_URL=https://opencode.ai/zen/v1 + OPENAI_MODEL=glm-5 — streaming (-p flag) and non-streaming verified
  • z.ai Coding Lite plan: OPENAI_BASE_URL=https://api.z.ai/api/coding/paas/v4 + OPENAI_MODEL=GLM-5.1 — streaming verified
  • z.ai also tested with GLM-4.7 — confirmed working with reasoning_content
  • Note: z.ai issue GLM unable to connect #214 reporter used pass instead of paas in the URL, which caused the original 500/404 — unrelated to this fix
• screenshots attached: N/A (no UI changes, terminal output only)
• follow-up work or known limitations:
  • Models that only emit reasoning_content with no content phase will show thinking as the visible reply (fallback behavior, matches user expectation)
  • Tool calling with GLM-5/5.1 not yet verified — may need separate testing
  • Benefits any OpenAI-compatible provider using the reasoning_content field (DeepSeek, QwQ, etc.)

Fixes #214

[kevincodex1]

this looks great, just one more thing can we add unit test for this change?

[gnanam1990]

Good fix for a real hang — the streaming state machine is sound and the root cause is correctly identified. Two things to address before merge:

1. Non-streaming path: || fallback treats content: "" as missing

const rawContent = choice?.message?.content || choice?.message?.reasoning_content

If a model returns content: "" (empty string) alongside reasoning_content, the || will use reasoning_content as the visible text — so the full chain-of-thought becomes the user-visible reply. The streaming path correctly uses !== '' guards; the non-streaming path should match:

const rawContent = (choice?.message?.content || '') !== ''
  ? choice?.message?.content
  : choice?.message?.reasoning_content

or simply use ?? with an explicit empty-string check.

2. No test for the non-streaming path

The 6 focused tests cover streaming only. Please add at least one test for the non-streaming case where content is null/"" and reasoning_content carries the reply — this is the exact scenario that caused the original hang.

---

Minor note (not a blocker): if tool calls arrive while a thinking block is open (hasEmittedThinkingStart && !hasClosedThinking), the thinking block won't be closed before the tool call block starts, producing malformed stream output. Tool calling with GLM-5 is acknowledged as untested — worth a follow-up issue to track.

[otaviocarvalho]

@kevincodex1 @gnanam1990 Hi, team! Addressed your PR feedbacks, please check again: b5b0cda

I also have some ipv6 changes needed (I am running openclaude in a VM w/o ipv4), but I will open a subsequent PR for that.

p.s. Please note the second commit is already done via interacting with GLM-5.1 already (using Z.ai Lite plan for now): Co-Authored-By: GLM-5.1 <noreply@openclaude.dev>

[Vasanthdev2004]

Rechecked the latest head 1295d5f215539a22a85ae2f9f39244cb7d25366d.

The two requested changes from the earlier review are fixed on this revision, and I did not find a remaining blocker.

What I verified on this head:

• non-streaming fallback now uses the explicit empty-string/null check instead of the old || behavior
• dedicated non-stream tests were added for:
  • content: null + reasoning_content
  • content: "" + reasoning_content
  • real content taking precedence over reasoning_content
• direct repro of the original hang path now works:
  • a GLM-style streaming response with only reasoning_content emits thinking deltas and completes cleanly instead of yielding an empty/frozen stream
• direct repro of the mixed reasoning/tool path is also sane:
  • streaming reasoning_content followed by a tool call closes the thinking block before the tool_use block starts
  • non-streaming responses with reasoning + tool call now produce structured content instead of hanging
• bun test src/services/api/openaiShim.test.ts -> 12 pass
• bun run build -> success
• bun run smoke -> success

Residual risk is mostly scope rather than a demonstrated bug: the PR still doesn't add broad provider-matrix coverage for every reasoning-model/tool-calling variant, but the current head does fix the reported GLM/DeepSeek-style reasoning-content hang and the earlier review blockers.

[kevincodex1]

this looks good already @otaviocarvalho . kindly please rebase to main and fix failing tests

[otaviocarvalho]

this looks good already @otaviocarvalho . kindly please rebase to main and fix failing tests

Thanks, @kevincodex1! Looks like some flaky tests were causing the build failure. Adjusted them here, please check again: 8410977

[Vasanthdev2004]

Rechecked the latest head 84109778504826c51f52d1d347eb7a6523e38de5 against current origin/main aff2bd89e22161f57ef330737b7f149208b96097.

I did not find a remaining blocker on this revision.

What I verified on this head:

• direct repro of the original streaming hang path now works:
  • GLM-style stream with only reasoning_content emits a thinking block and completes cleanly
  • current origin/main emits no content block at all for the same repro
• direct streaming reasoning -> text transition works:
  • reasoning chunk emits thinking_delta
  • later content chunk closes the thinking block and opens a text block
• non-streaming reasoning fallback works for the requested edge cases:
  • content: null + reasoning_content -> thinking block plus visible text
  • content: "" + reasoning_content -> thinking block plus visible text
  • real content still takes precedence over reasoning_content
• bun test src/services/api/openaiShim.test.ts src/commands/provider/provider.test.tsx -> 22 pass
• bun run build -> success
• bun run smoke -> success

Residual risk / note:

• the extra reconciler.ts / staticRender.tsx changes do alter failure behavior for render-helper paths: on uncaught Ink render errors, the helpers now time out and return captured error output after ~3s instead of hanging indefinitely. I didn't find a concrete branch-specific correctness regression from that, but it is a broader failure-mode change beyond the original OpenAI shim fix.