fix(openai-shim): recover GLM/Qwen XML tool calls emitted as text

[kevincodex1]

Summary

GLM/Qwen-family models routed through OpenAI-compatible gateways emit tool calls
as XML text (<tool_call><function=…>) instead of structured tool_calls. The
shim had no parser for this, so the XML leaked into visible prose and never
executed — the turn ended with no tool_use block, so the agent appeared to
"forget" and stop mid-task (the reported GLM-5.2 bug).

This adds parseXmlToolCalls() covering the three dialects seen in the wild and
wires it into both the streaming and non-streaming paths, mirroring the existing
Ollama text-tool-call fallback:

• A. <tool_call><function=NAME><parameter=KEY>VALUE</parameter>…</function></tool_call>
• B. <tool_call>NAME<arg_key>KEY</arg_key><arg_value>VALUE</arg_value>…</tool_call> (GLM-native)
• C. <tool_call>{"name":"NAME","arguments":{…}}</tool_call> (Hermes JSON)

The streaming path holds back text from the <tool_call> opener onward (including
an opener split across SSE deltas), converts the buffer to tool_use blocks at
finalize, strips the raw XML, and flips the finish reason to tool_calls.
Structured tool_calls and false-positive prose (a model literally writing
<tool_call>) are both handled losslessly.

Impact

• GLM/Qwen (and DeepSeek-family) tools now actually execute instead of leaking as
  text, fixing the "agent forgets / stops mid-task" symptom.
• No behavior change for Anthropic, Ollama (still uses its JSON fallback), or any
  provider that emits structured tool_calls.
• The XML path is on for all OpenAI-compatible providers — the <tool_call>
  marker is distinctive enough to be safe (JSON text recovery remains Ollama-gated).

Testing

• bun test src/services/api/openaiShim.xmlToolCalls.test.ts — 19/19 pass
  (3 dialects, split-opener, false-positive, structured-call coexistence,
  interleaved-prose flatten case)
• bun test src/services/api/ — 873/873 pass, no regressions
• tsc --noEmit — 0 errors
• bun run deadcode (knip) — clean

Notes

• Known minor trade-off: when a single streamed message has prose between two
  XML tool calls, surviving prose is flattened to prose-then-all-calls. This
  matches the existing Ollama path and doesn't affect agent loops (which act on
  the tool calls regardless of prose order). A test locks in this behavior.
• Longer-term cleaner fix is getting the gateway to request structured tool calls
  from GLM/Qwen, but this client-side fallback is the right defensive layer
  regardless.

Summary by CodeRabbit

• New Features
  • Improved fallback for OpenAI-compatible providers that emit tool calls as XML-formatted text, converting them into structured tool-use blocks while preserving surrounding prose.
  • Added recovery logic for both streamed and non-streamed responses, including richer parsing of XML tool-call parameters.
• Bug Fixes
  • Better handling when the XML tool-call start tag is split across streamed updates.
  • Reduced false positives and improved behavior for incomplete or malformed XML tool-call markup.
• Tests
  • Added unit and streaming coverage for XML tool-call recovery, parsing edge cases, and truncation scenarios.

[coderabbitai]

📝 Walkthrough

Walkthrough

The OpenAI shim now recovers XML-wrapped tool calls from compatible providers in streaming and non-streaming responses. It buffers chunked XML, parses three XML formats, strips recovered text, and adds coverage for the new paths.

Changes

XML tool-call recovery

Layer / File(s)	Summary
XML parser contract src/services/api/openaiShim.ts, src/services/api/openaiShim.xmlToolCalls.test.ts	parseXmlToolCalls handles three XML tool-call shapes, deduplicates calls, returns strip ranges, and the parser tests cover coercion, JSON parsing, truncation, and invalid input cases.
Streaming buffering and opener detection src/services/api/openaiShim.ts	Streaming state buffers XML text, tracks opener holdback across SSE deltas, and suppresses visible output once an XML tool-call region starts.
Streaming finalize and tests src/services/api/openaiShim.ts, src/services/api/openaiShim.xmlToolCalls.test.ts	Buffered XML is converted into tool_use blocks at completion, ranges are stripped from text, finish reasons are remapped, and the streaming tests cover split openers, interleaved prose, and structured tool_calls fallback.
Non-streaming text recovery src/services/api/openaiShim.ts	The shared non-streaming text helper now extracts XML tool calls before pushing remaining text, and both string and joined-array branches call it after think-tag stripping.

Estimated code review effort

🎯 4 (Complex) | ⏱️ ~45 minutes

Possibly related PRs

• Gitlawb/openclaude#1076: Also changes src/services/api/openaiShim.ts streaming tool-call recovery, with similar buffering and finish-reason remapping logic.

Suggested reviewers

• jatmn
• gnanam1990

🚥 Pre-merge checks | ✅ 5 | ❌ 2

❌ Failed checks (2 warnings)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 57.14% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.
No Hidden Policy Change	⚠️ Warning	FAIL: Diff adds routing/reasoning policy changes (preferBaseUrlRoute, resolveModelReasoningControl, removeBodyFields) beyond the XML fallback, and they aren't called out in the PR.	Split the policy-default changes into a separate PR or document them explicitly for maintainer sign-off before merging.

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title is concise, scoped, and accurately describes the main change in openai-shim.
Description check	✅ Passed	The description follows the template with clear Summary, Impact, Testing, and Notes sections.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Risk Surface Disclosed	✅ Passed	PASS: The PR explicitly flags the blast radius (all OpenAI-compatible providers) and notes the only trade-off as minor/non-blocking.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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

[coderabbitai]

Actionable comments posted: 1

♻️ Duplicate comments (1)

src/services/api/openaiShim.ts (1)

2141-2191: 🎯 Functional Correctness | 🟠 Major

Streaming finalize inherits the dedup-range leak above.

stripRanges(buffered, toolCallRanges) at Line 2150 strips only recorded ranges, so a duplicated <tool_call> block survives into strippedVisible and is emitted as text. No separate fix needed here — resolving the root cause in parseXmlToolCalls covers this path.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/services/api/openaiShim.ts` around lines 2141 - 2191, The streaming
finalize path in openaiShim.ts inherits the duplicate-range issue from
parseXmlToolCalls, so a repeated <tool_call> block can still survive
stripRanges(buffered, toolCallRanges) and be emitted as text. Fix the root cause
in parseXmlToolCalls so it returns correct, non-overlapping toolCallRanges for
duplicated XML tool calls; this emit path in the finalize block should then work
unchanged.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@src/services/api/openaiShim.ts`:
- Around line 1338-1343: The dedup logic in openaiShim’s tool-call parsing is
skipping range tracking for repeated named blocks, which leaves duplicate
<tool_call> XML visible in stripped output. In the parsing path around the
results/ranges handling, record every valid named block’s range before any dedup
check, and apply dedup only when pushing into results so stripRanges can remove
all XML occurrences. Also extend the openaiShim.xmlToolCalls test to assert the
duplicate block is removed from the final stripped text, not just that calls are
deduped.

---

Duplicate comments:
In `@src/services/api/openaiShim.ts`:
- Around line 2141-2191: The streaming finalize path in openaiShim.ts inherits
the duplicate-range issue from parseXmlToolCalls, so a repeated <tool_call>
block can still survive stripRanges(buffered, toolCallRanges) and be emitted as
text. Fix the root cause in parseXmlToolCalls so it returns correct,
non-overlapping toolCallRanges for duplicated XML tool calls; this emit path in
the finalize block should then work unchanged.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro

Run ID: 333af17f-0833-49ae-a135-06b76bf51d4f

📥 Commits

Reviewing files that changed from the base of the PR and between a5a8ccc and 4fc76f5.

📒 Files selected for processing (2)

• src/services/api/openaiShim.ts
• src/services/api/openaiShim.xmlToolCalls.test.ts

📜 Review details

🧰 Additional context used

📓 Path-based instructions (13)

src/**/*.{ts,tsx}

📄 CodeRabbit inference engine (AGENTS.md)

Use TypeScript with strict mode and ESM imports

Files:

• src/services/api/openaiShim.xmlToolCalls.test.ts
• src/services/api/openaiShim.ts

{src/commands/**/*.ts,src/services/**/*.ts,src/entrypoints/**/*.ts}

📄 CodeRabbit inference engine (AGENTS.md)

Use chalk for terminal color in CLI code

Files:

• src/services/api/openaiShim.xmlToolCalls.test.ts
• src/services/api/openaiShim.ts

{src/services/**/*.ts,src/utils/**/*.ts}

📄 CodeRabbit inference engine (AGENTS.md)

Use execa for child processes

Files:

• src/services/api/openaiShim.xmlToolCalls.test.ts
• src/services/api/openaiShim.ts

{src/integrations/**/*.ts,src/services/**/*.ts}

📄 CodeRabbit inference engine (AGENTS.md)

Test the exact provider/model path you changed when possible for provider modifications

Files:

• src/services/api/openaiShim.xmlToolCalls.test.ts
• src/services/api/openaiShim.ts

**/*.{ts,tsx,js,jsx,py,json,md}

📄 CodeRabbit inference engine (CONTRIBUTING.md)

Follow the existing code style in the touched files

Files:

• src/services/api/openaiShim.xmlToolCalls.test.ts
• src/services/api/openaiShim.ts

**/*.test.{ts,tsx,js,jsx}

📄 CodeRabbit inference engine (CONTRIBUTING.md)

Add or update tests when the change affects behavior

Files:

• src/services/api/openaiShim.xmlToolCalls.test.ts

**/*.test.{ts,tsx,js}

📄 CodeRabbit inference engine (CONTRIBUTING.md)

Test the exact provider/model path you changed when possible

Files:

• src/services/api/openaiShim.xmlToolCalls.test.ts

**/*.{ts,tsx,js,jsx,py}

📄 CodeRabbit inference engine (CONTRIBUTING.md)

Keep comments useful and concise

Files:

• src/services/api/openaiShim.xmlToolCalls.test.ts
• src/services/api/openaiShim.ts

**/*.{ts,tsx}

📄 CodeRabbit inference engine (CONTRIBUTING.md)

Follow TypeScript strict mode and type safety practices by running typecheck before submitting

Files:

• src/services/api/openaiShim.xmlToolCalls.test.ts
• src/services/api/openaiShim.ts

**

⚙️ CodeRabbit configuration file

**: # AGENTS.md - AI Agent Coding Guide

This guide is for AI coding agents working in the OpenClaude repository. Read it before changing code, and also follow CONTRIBUTING.md for contributor policy, PR expectations, review follow-up, and project scope.

Project Snapshot

OpenClaude is a coding-agent CLI for cloud and local model providers. It supports OpenAI-compatible APIs, Anthropic, Gemini, DeepSeek, Ollama, MCP, local backends, slash commands, tools, agents, and a React/Ink terminal UI.

The installed CLI runs on Node.js >=22.0.0. Bun is used for source builds, scripts, dependency management, and tests.

Work Style

• Keep changes focused on one problem.
• Prefer existing patterns in the file or nearby module.
• Avoid unrelated formatting, renames, dependency changes, or broad rewrites.
• Add or update tests when behavior changes.
• Update docs when setup, commands, provider behavior, or user-facing behavior changes.
• For new features, larger refactors, dependencies, or runtime changes, follow the issue-first guidance in CONTRIBUTING.md.

Stack And Conventions

• TypeScript with strict mode and ESM imports.
• React + Ink for terminal UI.
• Bun lockfile and Bun scripts for development workflows.
• Node runtime for the built CLI.
• Python exists for legacy/local-provider helper code. Do not add new Python code or expand Python-based features unless a maintainer explicitly approves that direction.

Common libraries and patterns:

• chalk for terminal color.
• commander for CLI argument parsing.
• execa for child processes.
• Existing service, provider, settings, permission, and UI patterns over new abstractions.

Repository Map

• src/commands/ - slash and CLI command implementations.
• src/components/ - React/Ink UI components.
• src/services/ - API, MCP, OAuth, wiki, voice, and other service integrations.
• src/tools/ - tool implementations.
• src/utils/ - shared utilities.
• `src/integration...

Files:

• src/services/api/openaiShim.xmlToolCalls.test.ts
• src/services/api/openaiShim.ts

**/*

⚙️ CodeRabbit configuration file

**/*: Apply the OpenClaude maintainer review rubric from AGENTS.md. Review the current diff, not stale discussion context. Separate real blockers from suggestions. Do not request changes for vague style churn. Treat approval as merge-ready from CodeRabbit's side, pending required human review and GitHub Checks. If checks are failing or unavailable, say so clearly instead of implying the PR is fully ready.

Files:

• src/services/api/openaiShim.xmlToolCalls.test.ts
• src/services/api/openaiShim.ts

{src/services/api/**,src/integrations/**,src/utils/model/**,src/utils/provider*.ts,src/commands/provider/**}

⚙️ CodeRabbit configuration file

{src/services/api/**,src/integrations/**,src/utils/model/**,src/utils/provider*.ts,src/commands/provider/**}: Review provider routing, model selection, env precedence, auth/token handling, OpenAI-compatible shims, retries, proxy behavior, and outbound HTTP behavior with high scrutiny. Block on silent default changes, hidden fallback expansion, credential reuse mistakes, hardcoded provider assumptions, or new network reach that is not intentional and documented.

Files:

• src/services/api/openaiShim.xmlToolCalls.test.ts
• src/services/api/openaiShim.ts

{src/**/*.test.ts,src/**/*.test.tsx,tests/**,scripts/**/*.test.ts,vscode-extension/**/*.test.js}

⚙️ CodeRabbit configuration file

{src/**/*.test.ts,src/**/*.test.tsx,tests/**,scripts/**/*.test.ts,vscode-extension/**/*.test.js}: Review tests for meaningful coverage of the changed behavior, isolation of global/env/config state, async cleanup, fake timers, provider profile leaks, and Windows-compatible assumptions. Block when risky runtime changes lack focused regression coverage or tests assert implementation details while missing the user-visible behavior.

Files:

• src/services/api/openaiShim.xmlToolCalls.test.ts

🔇 Additional comments (2)

src/services/api/openaiShim.ts (1)

1630-1635: LGTM!

Also applies to: 1877-1879, 1895-1928, 2204-2206, 3746-3782, 3795-3795, 3810-3810

src/services/api/openaiShim.xmlToolCalls.test.ts (1)

1-316: LGTM!

[coderabbitai]

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

src/services/api/openaiShim.ts (1)

3766-3777: 🎯 Functional Correctness | 🟠 Major | 🏗️ Heavy lift

Preserve recovered XML content ordering.

stripRanges(...).trim() merges all non-XML prose and pushes it before every recovered tool_use. For mixed content like before <tool_call>A</tool_call> after <tool_call>B</tool_call> tail, the output becomes one text block (before after tail) followed by both tools, moving text that originally followed a tool ahead of execution. Emit text/tool blocks by walking the parsed XML ranges in source order, or have the parser return ordered parsed blocks so duplicates can still be stripped without reordering content.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/services/api/openaiShim.ts` around lines 3766 - 3777, The XML recovery
logic in openaiShim’s tool-call parsing is reordering mixed content by
collecting all non-XML text with stripRanges(...).trim() and appending it before
every tool_use, which moves text that originally appeared after a tool call
ahead of it. Update the parsing flow around parseXmlToolCalls and the content
assembly so text and recovered tool_use blocks are emitted in source order,
either by iterating over parsed ranges sequentially or by having the parser
return ordered blocks while still deduplicating stripped XML segments.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Outside diff comments:
In `@src/services/api/openaiShim.ts`:
- Around line 3766-3777: The XML recovery logic in openaiShim’s tool-call
parsing is reordering mixed content by collecting all non-XML text with
stripRanges(...).trim() and appending it before every tool_use, which moves text
that originally appeared after a tool call ahead of it. Update the parsing flow
around parseXmlToolCalls and the content assembly so text and recovered tool_use
blocks are emitted in source order, either by iterating over parsed ranges
sequentially or by having the parser return ordered blocks while still
deduplicating stripped XML segments.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro

Run ID: 5742d81f-3351-49a2-b6b8-5122fc60c08f

📥 Commits

Reviewing files that changed from the base of the PR and between 4fc76f5 and 58be780.

📒 Files selected for processing (1)

• src/services/api/openaiShim.ts

📜 Review details

🧰 Additional context used

📓 Path-based instructions (10)

src/**/*.{ts,tsx}

📄 CodeRabbit inference engine (AGENTS.md)

Use TypeScript with strict mode and ESM imports

Files:

• src/services/api/openaiShim.ts

{src/commands/**/*.ts,src/services/**/*.ts,src/entrypoints/**/*.ts}

📄 CodeRabbit inference engine (AGENTS.md)

Use chalk for terminal color in CLI code

Files:

• src/services/api/openaiShim.ts

{src/services/**/*.ts,src/utils/**/*.ts}

📄 CodeRabbit inference engine (AGENTS.md)

Use execa for child processes

Files:

• src/services/api/openaiShim.ts

{src/integrations/**/*.ts,src/services/**/*.ts}

📄 CodeRabbit inference engine (AGENTS.md)

Test the exact provider/model path you changed when possible for provider modifications

Files:

• src/services/api/openaiShim.ts

**/*.{ts,tsx,js,jsx,py,json,md}

📄 CodeRabbit inference engine (CONTRIBUTING.md)

Follow the existing code style in the touched files

Files:

• src/services/api/openaiShim.ts

**/*.{ts,tsx,js,jsx,py}

📄 CodeRabbit inference engine (CONTRIBUTING.md)

Keep comments useful and concise

Files:

• src/services/api/openaiShim.ts

**/*.{ts,tsx}

📄 CodeRabbit inference engine (CONTRIBUTING.md)

Follow TypeScript strict mode and type safety practices by running typecheck before submitting

Files:

• src/services/api/openaiShim.ts

**

⚙️ CodeRabbit configuration file

**: # AGENTS.md - AI Agent Coding Guide

This guide is for AI coding agents working in the OpenClaude repository. Read it before changing code, and also follow CONTRIBUTING.md for contributor policy, PR expectations, review follow-up, and project scope.

Project Snapshot

OpenClaude is a coding-agent CLI for cloud and local model providers. It supports OpenAI-compatible APIs, Anthropic, Gemini, DeepSeek, Ollama, MCP, local backends, slash commands, tools, agents, and a React/Ink terminal UI.

The installed CLI runs on Node.js >=22.0.0. Bun is used for source builds, scripts, dependency management, and tests.

Work Style

• Keep changes focused on one problem.
• Prefer existing patterns in the file or nearby module.
• Avoid unrelated formatting, renames, dependency changes, or broad rewrites.
• Add or update tests when behavior changes.
• Update docs when setup, commands, provider behavior, or user-facing behavior changes.
• For new features, larger refactors, dependencies, or runtime changes, follow the issue-first guidance in CONTRIBUTING.md.

Stack And Conventions

• TypeScript with strict mode and ESM imports.
• React + Ink for terminal UI.
• Bun lockfile and Bun scripts for development workflows.
• Node runtime for the built CLI.
• Python exists for legacy/local-provider helper code. Do not add new Python code or expand Python-based features unless a maintainer explicitly approves that direction.

Common libraries and patterns:

• chalk for terminal color.
• commander for CLI argument parsing.
• execa for child processes.
• Existing service, provider, settings, permission, and UI patterns over new abstractions.

Repository Map

• src/commands/ - slash and CLI command implementations.
• src/components/ - React/Ink UI components.
• src/services/ - API, MCP, OAuth, wiki, voice, and other service integrations.
• src/tools/ - tool implementations.
• src/utils/ - shared utilities.
• `src/integration...

Files:

• src/services/api/openaiShim.ts

**/*

⚙️ CodeRabbit configuration file

**/*: Apply the OpenClaude maintainer review rubric from AGENTS.md. Review the current diff, not stale discussion context. Separate real blockers from suggestions. Do not request changes for vague style churn. Treat approval as merge-ready from CodeRabbit's side, pending required human review and GitHub Checks. If checks are failing or unavailable, say so clearly instead of implying the PR is fully ready.

Files:

• src/services/api/openaiShim.ts

{src/services/api/**,src/integrations/**,src/utils/model/**,src/utils/provider*.ts,src/commands/provider/**}

⚙️ CodeRabbit configuration file

{src/services/api/**,src/integrations/**,src/utils/model/**,src/utils/provider*.ts,src/commands/provider/**}: Review provider routing, model selection, env precedence, auth/token handling, OpenAI-compatible shims, retries, proxy behavior, and outbound HTTP behavior with high scrutiny. Block on silent default changes, hidden fallback expansion, credential reuse mistakes, hardcoded provider assumptions, or new network reach that is not intentional and documented.

Files:

• src/services/api/openaiShim.ts

🔇 Additional comments (1)

src/services/api/openaiShim.ts (1)

1339-1339: LGTM!