[LinuxDevil]

What

• All 28 createGitHubTools() tools and all 20 createJiraTools() tools are built with defineTool instead of ai's tool(). Names, descriptions, zod schemas, sandbox flags and factory signatures are unchanged. Each tool now carries canonical inputSchema / execute, and the legacy .tool stays.
• The change is mechanical (a regex codemod over the this.register('name', { displayName, tool: tool({ ... }) }) pattern). Inner lines keep their indentation, so only each tool's header and footer lines changed. parameters: became input:.
• GitHubTools.register() disables the out-of-scope tools by registering a copy of the descriptor whose canonical execute and legacy tool.execute both throw the "out of scope" error. A canonical execute takes priority, so overriding only tool.execute would have left the real implementation reachable. This is the same hazard D24 fixed in examples/ops-pipeline/fixer.ts.
• JiraTools.register() now routes a defineTool result through routeFetchThroughSandbox as well. Before, the register(DefinedTool) overload skipped it, so a descriptor registered that way would not have been flagged requiresSandbox.
• src/tools/noAiToolImport.test.ts lists github.ts and jira.ts. CHANGELOG has an entry under Changed (not breaking).

Why

Step 4 of the ai v4 -> v6/v7 upgrade chain (after D22 #79, D23 #103 and D24 #108). With the last tool() call under src/tools gone, the legacy .tool shape can be removed in D26.

ai imports remaining under src/:

• Production code: src/types/tool.ts (type import), src/providers/* (OpenRouterProvider, AnthropicProvider, OpenAIProvider, OllamaProvider, aiSdkProvider), src/execution/structuredOutput.ts and src/execution/errors.ts.
• Tests only: src/tools/toolContract.test.ts and src/tools/defineTool.test-d.ts (they import ai to assert that defineTool no longer calls ai.tool()), plus provider and execution tests.

No production file under src/tools imports ai any more. A literal grep -rn "from 'ai'" src/tools still matches those two D22 test files. I left them alone because they are out of this ticket's scope and they test the absence of tool().

Acceptance criteria

• Both files build every descriptor with defineTool (28 + 20), import nothing from ai, and keep names, descriptions, schemas and factory signatures unchanged. Neither file sets needsApproval or requiresSandbox itself, so no such flags changed. requiresSandbox and sandboxExecute are still set in register() via routeFetchThroughSandbox.
• Scope and disable logic works through the canonical execute: a new test calls every out-of-scope GitHub tool through getToolExecute() and descriptor.execute and expects the "out of scope" error with no fetch call.
• Existing github.test.ts and jira.test.ts pass with no edits to existing assertions. I only added new tests, and the existing tests still call .tool.execute, which now equals the canonical execute.
• Both files added to noAiToolImport.test.ts.
• The diff is mostly header and footer lines, not reformatting.
• CHANGELOG entry under Changed (not breaking). No docs changes.
• Only the in-scope files were touched.

How verified

• npx tsc --noEmit: pass, no output.
• npm run lint: 0 errors, 402 warnings. For the files I touched, github.ts + jira.ts still have 23 warnings, the same as on origin/main, and the two test files have 18, with none added.
• npm run test:types: 30 passed, no type errors.
• npx vitest run after npm run build (the create-loushy-agent e2e test needs dist): 160 files passed, 2337 tests passed, 3 skipped.
• npm run build && npm run build --workspace=packages/create-loushy-agent: success.
• npm run test:coverage: exit 0.
• npm run fallow: exit 0, "No issues found" in dead code, and no refactor targets above threshold. It lists one existing duplicate block, AnthropicProvider.ts and OpenAIProvider.ts, which this PR does not touch.
• docs:verify-snippets and docs:llms:check were not run, since README and docs are untouched.

🤖 Generated with Claude Code

https://claude.ai/code/session_016bY29ibQgv3MDdjAbUT4YR

---

Generated by Claude Code