Sweep small cross-implementation parity gaps

[shibayan]

Summary

This issue tracks ten small, independent parity fixes. It is an umbrella checklist, not permission to combine unrelated work into one PR. Each coherent concern may land independently; follow the repository's one-concern-per-PR rule.

Work items

• core / coalesce — code interpreter — Coalesce streamed code-interpreter fragments by call id #132 (packages/core/src/types/coalesce.ts)

  • Add the Python-authoritative merge rule for code-interpreter call/result fragments keyed by type plus callId or additionalProperties.item_id.
  • Merge non-adjacent fragments with the same key at the first occurrence; do not merge different types/keys or keyless items.
  • Cover adjacent, interleaved, different-key, item_id fallback, and keyless cases.

• core / explode — terminal metadata — Leave the finish reason off the trailing metadata-only update #133 (packages/core/src/types/response.ts)

  • Keep response-level finishReason on message updates, but do not copy it to the trailing metadata-only update used for usage/additional properties/continuation data.
  • Test the direct update sequence for both chat and agent responses, not only the re-folded result.

• core / missing user-input literal — Emit the UserInputRequiredException marker on user-input results #134, follow-up An argument failure is reported to the model as an execution failure #131 (packages/core/src/client/function-execution.ts)

  • Emit the exact wire-visible exception literal UserInputRequiredException.
  • Preserve detailed-error on/off behavior and add an exact-match test.

• anthropic / toolChoice without tools — Send the Anthropic tool choice independently of tool declarations #135 (packages/anthropic/src/chat-client.ts)

  • Send configured tool_choice independently of local tools/MCP server presence; omit it only when no choice was configured.
  • Cover no-tools, MCP-only, local-tools, and undefined-choice requests.

• anthropic / empty-message comment — Say why the empty-message guard is there #136

  • Keep the local guard but replace the inaccurate Python-parity comment with the actual rationale: TypeScript turns a later service 400 into an earlier clear error. No runtime change.

• mcp / isError text — Put each text block of an MCP error result on its own line #138, follow-up FoundryToolbox assembles MCP error text differently from the MCP client #137 (packages/mcp/src/client.ts)

  • Join MCP error text blocks with "\n" in MCP-side assembly only. Do not change shared textOfContents behavior.
  • Cover multiple text blocks, structured-plus-text content, and the no-text fallback.

• mcp / blank skill resource names — Refuse a blank skill resource name before it becomes a read #139 (packages/mcp/src/skills.ts)

  • Reject empty and whitespace-only names before I/O on the direct Skill API path.
  • Prove resources/read is not called and retain existing traversal/decoding guards.

• core / run_skill_script schema — Declare the run_skill_script args default and name its script runner #140, follow-up SkillScriptArguments excludes null but a script can receive one #143 (packages/core/src/skills/provider.ts)

  • Add default: null to the args schema and the Python-equivalent “or configured runner.” description.
  • Add an exact public tool-schema assertion; runtime execution semantics do not change.

• agentserver / request-ID precedence — Resolve the agentserver request id from the trace before the incoming header #141 (packages/agentserver/src/context.ts and trace-context plumbing)

  • Resolve a valid active/extracted trace ID first, then a non-empty incoming x-request-id, then a generated UUID.
  • Cover valid/invalid/zero traceparent, competing request ID, and identical resolved values in the response header and error body.

• openai / unanswered function-call wire tests — Pin the send-time filtering of unanswered OpenAI function calls #142

  • Pin the intentional send-time filtering: a lone unanswered local function_call is omitted without throwing; later user input is retained while only the call is removed.
  • Cover JSON session round-trip without relying on raw representation. This item adds tests only and does not preserve orphan calls.

Completion criteria

• Every behavior change has a reproduction-first failing test; comment-only and test-only items are exempt.
• Wire-visible changes are listed in CHANGELOG.
• Each PR identifies the checkbox it completes. A split successor issue is linked from this body or a comment.
• pnpm check passes for every implementation PR.
• This umbrella remains open until all ten items are complete or moved to linked successor issues that are themselves closed.