fix(tool): record locally executed tools in OpenCode V2

[pedropombeiro]

Description

With OpenCode V2, the plugin drops telemetry for every tool that OpenCode runs locally, such as shell, read, edit, MCP tools, Code Mode, and subagents:

• opencode.tool.duration is never recorded.
• No opencode.tool.* spans are created.
• opencode.commit.count and the commit log event never fire.
• Every tool_result log event reports duration_ms: 0.

Root cause: In V2 tool events, executed means the model provider ran the tool, not that OpenCode ran it. OpenCode publishes session.tool.called with executed: tool.providerExecuted (publish-llm-event.ts#L458-L476). It then publishes the terminal event for a locally executed tool with the same false value (#L573-L591). Only provider-hosted tools, such as OpenAI web_search or Anthropic code_execution, report executed: true (#L477-L513). handleToolCalled and finishTool gated spans, durations, and commit detection on executed, so the plugin measured only hosted tools.

Fix: A tool counts as executed once the plugin observes session.tool.called. OpenCode starts a local execution only after that event is published (step.ts#L100-L128). Calls that never ran don't publish it:

• Malformed-JSON rejections (#L315-L338).
• Calls cancelled before publication (#L360-L375).

Neither case is recorded as an execution.

• Failures after the call: A tool that fails after session.tool.called, including a permission rejection or user decline (step.ts#L199-L207), is recorded as a failed execution (success=false). Its duration includes any permission wait.
• Provider execution: The flag is now the tool.provider_executed span attribute. It accepts both the current top-level executed and the nested provider.executed shape, which OpenCode dev uses (session-event.ts#L312-L372). That covers only the payload shape: dev also renames the events to session.next.tool.*, which this PR doesn't address.
• Subagents: Subagent dispatch spans now start at session.tool.called instead of at the first progress event that carries a child session ID. Child-session linking is unchanged.

Tests now cover local success and failure, local git commit detection, permission rejection, malformed input without a call, provider-hosted success and failure, the nested provider.executed shape, and a subagent that fails before creating a child. Two existing tests encoded the old assumption that executed: false meant the tool didn't run, so I updated them.

I also checked it end to end with OpenCode 2.0.20 and a local OTLP/HTTP JSON sink. A read call produced a 33 ms span, histogram sample, and duration_ms. A shell call running sleep 2 recorded 2,834 ms. A git commit through shell incremented opencode.commit.count and emitted the commit log event.

Type of change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update
• Refactoring (no functional changes)
• Chore (dependency updates, etc.)

Checklist

• I have read the CONTRIBUTING.md document
• My code follows the style guidelines of this project
• bun run lint passes with no errors
• bun run check:jsdoc-coverage passes with no errors
• bun run typecheck passes with no errors
• bun test passes with no errors
• I have added tests that prove my fix is effective or that my feature works
• I have updated the documentation accordingly
• My commits follow the Conventional Commits specification

Related issues

Independent of #112, which covers V1 message.part.updated timestamp restamping and orphan spans. This PR fixes V2 tool events that were never measured.

Additional context

The executed gating was introduced in #132, which treated executed: false calls as invalid. That holds for malformed calls but also matches every successful local call.

[dialupdisaster]

Non-blocking test coverage suggestion: strengthen the disabled tool-tracing test in tests/handlers/disabled.test.ts. It currently only sends session.tool.input.started, which does not create a span even when tracing is enabled.

Add a local tool sequence (input.started → called with executed: false → success) with tool tracing disabled, and assert that:

• No tool span is created.
• The tool duration histogram and tool_result.duration_ms still record the elapsed time.
• For a successful shell git commit call, the commit counter and commit log event still fire.

This would protect the behavior changed here: execution metadata and command capture must happen before the tracing-disabled early return, so disabling tool spans does not also disable metrics or commit detection.

[pedropombeiro]

Thanks @dialupdisaster 🏓 Good catch: the old test only sent session.tool.input.started, so it passed whether or not tracing was enabled.

I replaced it in 23a6c22 and 7ae54b2 with a full local shell git commit call (input.started → called with executed: false → success) with tool tracing disabled. It asserts that:

• No tool span is created after called or success.
• The duration histogram and tool_result.duration_ms record 250 ms, measured from called.
• The commit counter and commit log event don't fire after called, then fire exactly once after success.

I confirmed the test fails if the tracing-disabled early return moves above the metadata capture in handleToolCalled.

[dialupdisaster]

Thanks for the contribution!