[LinuxDevil]

What

• New exported ToolExecutionContext in src/types/tool.ts: { toolCallId: string; messages: readonly Message[]; abortSignal?: AbortSignal; sessionId?: string; onDelegatedUsage?: (usage: RunUsage) => void }. It merges the old ToolRunContext extras into one public type.
• defineTool (execute), DefinedTool.execute, ToolDescriptor.execute, sandboxExecute (third argument) and buildToolRunContext() (return type, no more as unknown as cast) use it. ToolRunContext stays as a deprecated alias (Partial<Omit<ToolExecutionContext, 'messages'>>, so existing imports and ai tool() execute options in DelegationTool still fit).
• src/types/agent.ts: AgentExecutionOptions.messages is Message[] instead of CoreMessage[] (no CoreMessage is exported, so no alias; the field is unused inside the SDK).
• src/tools/defineTool.ts no longer imports anything from ai (the legacy .tool is typed via ToolDescriptor['tool']['execute'], with a // legacy (.tool), removed in D26 cast). src/types/tool.ts keeps its one import type { Tool as AITool } from 'ai' for the legacy .tool field, marked // legacy (.tool), removed in D26.
• Type tests, docs/tools.md "Execute context", regenerated llms-full.txt, CHANGELOG (Changed, non-breaking).

Remaining from 'ai' imports in src/ (non-test), after this change:

• src/types/tool.ts: Tool as AITool (legacy .tool, D26)
• src/execution/DelegationTool.ts: tool (value, builds the legacy delegate tool)
• src/execution/structuredOutput.ts: zodSchema (value)
• src/execution/errors.ts: APICallError, LoadAPIKeyError, RetryError (values, used for error classification)
• src/providers/{OpenRouter,Anthropic,OpenAI,Ollama}Provider.ts: LanguageModel (type); src/providers/aiSdkProvider.ts (excluded by ticket)
• src/tools/built-in/* (slack, http, currentDate, jira, github, email, dayName): tool (D24/D25); src/tools/mcp/McpToolLoader.ts: tool
• tests: src/providers/toolCallTurns.contract.test.ts (coreMessageSchema, CoreMessage), src/tools/defineTool.test-d.ts (tool, for the legacy-descriptor type test)

Why

Step 2 of the ai v4 -> v6/v7 upgrade chain (D22-D29): the public tool types must not leak ai's ToolExecutionOptions, whose messages shape (CoreMessage[]) changes between ai majors. We already pass our own Messages at runtime (U15).

Acceptance criteria

• ToolExecutionContext exported from src/types/tool.ts (via src/types and the root), merged with the ToolRunContext extras; ToolRunContext kept as a deprecated alias
• defineTool, ToolDescriptor.execute, sandboxExecute, DefinedTool, buildToolRunContext() use it (workspace tools infer ctx from defineTool; no workspace/built-in source needed changes)
• No src/types/*, src/tools/defineTool.ts or src/execution/* type import from ai except the legacy AITool in src/types/tool.ts (the ticket's one allowed legacy import; defineTool.ts needs none)
• src/types/agent.ts: CoreMessage replaced by Message
• Type tests: defineTool ctx is ToolExecutionContext and args still infer; a legacy ai tool() descriptor still registers in ToolRegistry.register
• Docs (docs/tools.md), llms files regenerated, CHANGELOG entry

Notes: sessionId is declared and passed through buildToolRunContext() but the executor does not set it yet (wiring it needs src/execution/toolCallExecution.ts, out of scope here; docs say so).

How verified

• npx tsc --noEmit: clean
• npm run lint: 0 errors, 404 warnings (no any added; the change removes a cast)
• npm run test:types: 6 files, 28 tests passed, no type errors
• npx vitest run (after npm run build): 2223 passed, 3 skipped, 1 failed: src/deploy/adapters/cloudflare.test.ts "wrangler dev serves the worker on the local workerd runtime" (needs the real wrangler/workerd; fails in isolation too, unrelated to this change). src/execution/guardrails.test.ts failed once under coverage and passed in the other runs (flaky).
• npm run build && npm run build --workspace=packages/create-loushy-agent: success
• npm run docs:verify-snippets -- --skip-build: all 115 snippets type-check, 6 also run
• npm run docs:llms:check: passes
• vitest run --coverage --coverage.reportOnFailure --config vitest.coverage.config.ts then npm run fallow: exit 0, 0 above threshold (the plain test:coverage does not write the coverage file because of the cloudflare failure above)

🤖 Generated with Claude Code

https://claude.ai/code/session_016bY29ibQgv3MDdjAbUT4YR

---

Generated by Claude Code