fix(session): preserve tool metadata across pending→running transition

[rmk40]

Issue for this PR

Closes #20184

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

The AI SDK fires tool execute() as a detached promise before the processor handles the tool-call stream event. When execute() calls ctx.metadata({sessionId}), the processor's tool-call handler overwrites the DB with a fresh {status: "running", input, time} state that drops the metadata. This makes subagent sessions unclickable in the TUI because the TUI reads state.metadata.sessionId for navigation.

The fix adds a synchronous toolMetadata map on the processor context. setToolMetadata() on the Handle writes to this map immediately when the metadata callback fires. The tool-call handler then reads from the map and merges title/metadata into the running state. prompt.ts's metadata callback calls setToolMetadata() synchronously before its existing async DB update, so metadata is captured regardless of event ordering.

This PR also introduces two reusable test fixtures:

• test/fixture/anthropic.ts — Fake Anthropic HTTP server (Bun.serve({port:0})) with SSE response helpers (toolResponse, textResponse, waitRequest, deferred). Reusable for any test that needs to exercise the real AI SDK streamText pipeline against a fake provider.
• test/fixture/prompt-layers.ts — Full SessionPrompt Effect layer stack with real LLM.defaultLayer and no-op stubs for MCP/LSP/FileTime. Reusable for any test that needs the complete prompt pipeline (tool registration, resolveTools, processor event handling) without mocking the LLM.

Related: #20183

How did you verify your code works?

• E2E regression test (test/session/metadata-race.test.ts) that goes through the full production path: ToolRegistry.register → SessionPrompt.loop → resolveTools → ctx.metadata() → setToolMetadata → processor tool-call handler. Uses the fake Anthropic server fixture with real AI SDK streamText. Asserts metadata.sessionId is present on the running tool part in the DB.
• Test passes 3/3 consistently (~1s each), no flakiness.
• Test does not compile without the fix (setToolMetadata missing from Handle), confirming it catches the regression.
• Existing prompt-effect.test.ts (23 tests) and processor-effect.test.ts (10 tests) still pass.
• Manual testing confirmed subagent sessions are clickable again.

Screenshots / recordings

N/A — data layer fix, not UI change.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Based on my search results, I found one related PR:

Related PR Found:

• PR fix(session): preserve metadata in tool-call handler for subagent click navigation #20183: "fix(session): preserve metadata in tool-call handler for subagent click navigation"
  • fix(session): preserve metadata in tool-call handler for subagent click navigation #20183

Why it's related: PR #20183 addresses the same issue (preserving metadata during session state transitions) and is explicitly mentioned in the current PR description as a related fix. Both are addressing the problem of metadata being lost when a session transitions from pending to running state, which prevents subagent session navigation from working in the TUI.

The current PR (#20198) appears to be a more comprehensive fix using a side-channel approach, while #20183 may be an earlier attempt or partial fix for the same root cause.

[github-actions]

Thanks for updating your PR! It now meets our contributing guidelines. 👍

[rmk40]

Split into two PRs: #20208 (test fixtures) and a follow-up PR (fix + regression test).