Tool mapping bugs causing task failures and missing functionality

[BNasraoui]

Issue: Tool mapping bugs causing task failures and missing functionality

Environment

• OpenCode 1.3.17
• opencode-claude-bridge 1.9.0
• Claude Sonnet 4.6

Bug 1: Agent subagent_type not required — causes task validation errors

Observed: When the model calls the Agent tool without subagent_type, OpenCode's task tool rejects it:

The task tool was called with invalid arguments: [
  {
    "expected": "string",
    "code": "invalid_type",
    "path": ["subagent_type"],
    "message": "Invalid input: expected string, received undefined"
  }
]

Root cause: src/claude-tools.ts:217 has required: ["description", "prompt"] — missing subagent_type. OpenCode's task tool requires it.

Fix: Add "subagent_type" to the required array.

Bug 2: general-purpose maps to build instead of general

Observed: When the model picks general-purpose, the bridge translates it to build. But OpenCode has a dedicated general subagent (per official docs). This sends general-purpose tasks to the wrong agent with different permissions.

Also, general was missing from the outbound map, so OpenCode's general type couldn't translate back to general-purpose for the API.

Root cause: src/index.ts:749 maps "general-purpose": "build". Outbound map at src/index.ts:592 doesn't include general.

Fix: Map "general-purpose": "general" inbound, add general: "general-purpose" outbound.

Bug 3: AskUserQuestion has no mapping to OpenCode's question tool

Observed: The model tries to call AskUserQuestion to ask clarifying questions, but the bridge treats it as a stub tool with no mapping to OpenCode's question tool. All clarification attempts fail.

Root cause: AskUserQuestion is only in STUB_TOOLS, not in INBOUND_TOOL_NAME_MAP or OUTBOUND_TOOL_NAME_MAP.

Fix: Add AskUserQuestion: "question" to inbound map, question: "AskUserQuestion" to outbound map.

Additional finding: SSE chunk boundary fragility

The inbound stream parameter translation (e.g. "file_path" → "filePath", "activeForm" → "priority") uses regex on raw SSE text chunks. If a chunk boundary splits a key name, the translation silently fails. This causes intermittent failures on Edit, Read, Write, and TodoWrite tools. This is harder to fix (requires buffering) but worth noting.

PR

PR submitted with fixes for bugs 1-3.

[dotCipher]

Pending #8 for now

[dotCipher]

Fixed and released in v1.10.3.

Root cause

The bridge was translating inbound Claude tool names and tool arguments by applying regex replacements directly to streamed SSE chunks. That breaks when Anthropic splits a tool name or argument key across chunk boundaries, for example:

• AskUserQuestion split across chunks, so OpenCode saw the unmapped Claude tool name instead of question
• file_path split across chunks, so OpenCode never received filePath
• activeForm split across chunks, so TodoWrite items never got priority

That is why tool calls could fail with schema errors or unknown-tool errors even though the mappings were correct in source.

What changed in 1.10.3

• inbound tool translation now buffers input_json_delta fragments per tool-use block and translates the fully assembled JSON once at content_block_stop
• translation is now structural (JSON.parse / object-walk / JSON.stringify) instead of regex on raw stream bytes
• this makes chunk-boundary splits safe for:
  • tool name mapping like AskUserQuestion -> question
  • key renames like file_path -> filePath
  • nested field mappings like multiSelect -> multiple and activeForm -> priority
• the previously merged shared-tool mapping fixes are preserved on top of this refactor

Verified

I reproduced the broken AskUserQuestion behavior on the old implementation, then verified end-to-end on the PR branch build that the question UI renders correctly once the stream translation fix is in place.

Upgrade

npm install -g opencode-claude-bridge@1.10.3