[sanchitrk]

Summary

Tools can now be added or removed mid-conversation with a transcript SystemMessage (toolsAdded, toolsRemoved, metadata). Context.tools is the initial declaration. Transcript helpers resolve the current tool set and decide, per provider, whether a change is declared in place or folded into the top-level tool list.

Codex and Copilot Responses

• Both run on one shared ResponsesLanguageModel (providers/responses/, moved from openai-codex/). A ResponsesRequestProfile captures each backend's differences:
  • where the system prompt goes;
  • whether max_output_tokens is sent;
  • when encrypted reasoning is requested;
  • zstd compression;
  • tool-field defaults.
• Models that support it declare every tool in additional_tools items, starting at the head of the input. The incremental-update notice precedes later declarations, and removals are announced as text. The request input only grows, so the cached prefix survives tool changes.
• Models that only support tool search get later additions as a tool-search call/output pair. Everything else gets the current tools top-level.
• Stream items are tracked by output_index, because Copilot rotates item ids on every event (which split one tool call into an empty call plus a call with arguments).

Copilot

• model.list shows only the models the login's availableModelIds allows, matching each model's API id.
• A Claude model that can't disable thinking no longer gets thinking: disabled, which Copilot rejected with a 400.
• GPT models that accept reasoning.effort: "none" get it when thinking is off (checked live).
• tool_choice and parallel_tool_calls are sent only when the call names a tool choice; Grok rejected them on requests without tools.
• The model generator's long-context and thinking override lists for Copilot are updated.

Other changes

• Errors: when an adapter can't parse a provider's error, the message and code are read from the raw response body. This fixes Copilot errors that showed only "Bad Request".
• Grammar tools: calls are replayed using the declaration of every tool the conversation declared, so calls to a tool that was later removed still replay correctly.
• activeTools: a request with activeTools sends its tools top-level instead of declaring them in place.
• Removed: ToolCall.addedToolNames; the token estimate now counts toolsAdded.
• Harness: stored context messages are narrowed to user and assistant, and decoding a stored system message fails with an explicit error. Storing system messages comes later.

Test plan

• vp check and vp run typecheck
• aikit unit tests, including the new transcript.test.ts (prefix stability across removals and loads, the tool-search boundary, redefinition, re-declaration) and the Copilot Responses tests (rotating ids, tool fields, off effort)
• Live E2E tools.added.e2e.test.ts on Codex, Copilot Responses, Copilot Messages, Copilot Chat, Anthropic, OpenAI, Google and OpenRouter, with an artifact for each
• Codework model.list Copilot filter test
• Five unrelated live tests failed in the last full run, on OpenAI gpt-5.4-nano, OpenRouter glm images, and the Anthropic and OpenRouter thinking-level artifacts. One nano test passed when re-run.