fix(claude): parent an in-stream subagent's spans under its Agent call

[drewstone]

Review finding on row 7 (contract scoping): the stream-json reader never read parent_tool_use_id. Claude Code writes this field on every event belonging to a subagent running inline in the same stream (not a separate subagents/*.jsonl file), naming the Agent/Task tool_use id that started it — but every such event was parented to the trace root regardless.

A contract scoped to {kind: TOOL, tool: Agent} therefore saw only the Agent span itself: tools.required rules failed on tools the subagent really called, and tools.forbidden/maxCalls rules passed vacuously on calls that were actually there — both verdicts backwards.

Fix

Read parent_tool_use_id off assistant and user stream-json events. When it resolves to a TOOL span already seen in this stream, parent the new llm.turn/user.prompt span there (its own tool calls nest under that span, so the whole sidechain moves as a unit), and stamp agent.parent.confidence: explicit, reusing the confidence vocabulary #133 introduced for the cross-file case. An id that resolves to nothing falls back to the trace root, same as before this field was read at all.

Proof (real trace, no unit tests)

$ node proof-parent.mjs subagent.jsonl   # model calls Agent; subagent calls Read
before: Read span ancestor chain: [ 'tool.Read's own span id' ]        (flat, not under Agent)
        Read is under Agent: false
after:  Read span ancestor chain: [tool.Read's llm.turn, tool.Agent, root]
        Read is under Agent: true
        llm.turn housing Read: parentConfidence = explicit

$ node proof-parent-contract.mjs subagent.jsonl   # via agent-eval's checkTraceContracts
scope Agent, tools.required[Read]:
  before: FAIL "no span matches eventually(kind=TOOL,tool=Read) over 1 span(s)"
  after:  PASS 1/1
scope Agent, tools.forbidden[Read]:
  before: PASS 1/1   (vacuous — Read was invisible in scope)
  after:  FAIL "span ... matches never(kind=TOOL,tool=Read)"

The contract-level proof ran against a scratch build using agent-eval's trace-contracts API (agent-eval#846), since this branch's published agent-eval pin (0.186.2) predates that API; the span-hierarchy proof (proof-parent.mjs) ran directly in this worktree with no substitution.

Gates: pnpm install --frozen-lockfile, pnpm typecheck, pnpm build, pnpm check:source all clean. Full existing suite: 866/866 tests across 62 files pass, including tests/claude-subagents.test.ts, tests/claude-task-scope.test.ts and tests/claude-workflow.test.ts. No unit tests added or changed.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.