feat: add auto approval and auto preset selection

[HugoRCD]

🔗 Linked issue

None.

📚 Description

Adds 'auto' modes for approval and presets to the SDK and the eve extension. An evaluation model decides per message through AI SDK experimental_evaluate. The default is TypeSafe's Jev ('typesafe-ai/jev' through AI Gateway).

createGithubAgent({ model: 'anthropic/claude-opus-5.5', preset: 'auto', requireApproval: 'auto' })
githubExtension({ preset: 'auto', requireApproval: 'auto' })

requireApproval: 'auto' (createGithubTools, createGithubAgent, eve extension)

• Before a call to a low-risk write tool (AUTO_APPROVAL_TOOLS: labels, assignees, reactions, comments, review-thread replies, reviewer requests, notification reads, workflow re-runs), the model rates the risk (0–2) and whether the latest user message asked for this exact call. The call runs without approval when risk <= 1 and intent >= 0.6; otherwise the user is asked.
• Other write tools keep requiring approval. Per-tool values accept 'auto': { updateIssue: 'auto', mergePullRequest: true }.

preset: 'auto' (createGithubAgent, eve extension)

• One evaluation call per user message, one yes/no question per preset. Presets at or above 0.7 are used, at most two. Read-only repo-explorer is only used when no other preset qualifies. When none qualifies, the most likely preset is used. The full catalog is never exposed.
• In the eve extension, routing runs once per user message, and tools already called in the session stay registered so a parked approval resumes after routing changes. eve's [Pending approvals] resume message is not treated as a user request.

Failure handling: when the evaluation call fails (model not enabled on AI Gateway, no credits, outage), 'auto' approval asks and preset: 'auto' uses repo-explorer. Both log a github_tools.EVALUATION_FAILED warning with the gateway error as cause.

Configuration: evaluation: { model, maxRisk, minIntent, minPresetProbability, maxPresets } tunes both modes (defaults 1, 0.6, 0.7, 2).

Fix: eve tool output formatters no longer crash on structured GitHub API errors (Cannot read properties of undefined in the next model step for listPullRequestFiles, getFileContent, getRepositoryTree, getPullRequestContext, getCommit, compareCommits).

Compatibility

• 'auto' needs ai 7.0.105 or later; without experimental_evaluate it throws EVALUATION_UNAVAILABLE. The root entry still loads on ai 6. ai is now a static namespace import, because eve could not bundle the extension with the dynamic import('ai') ("Expected one bundled authored module").
• createDurableGithubAgent does not accept 'auto'.
• ToolOptions.needsApproval is widened to the AI SDK type, which also accepts a function. Non-breaking.

Benchmark (examples/eve-bench): eve evals on 12 GitHub tasks with anthropic/claude-opus-5.5 and Jev, writes stubbed. Against the full catalog, preset: 'auto' + requireApproval: 'auto' passed 12/12 tasks with 42% fewer tokens per task (107k to 62k), 34% lower cost ($0.164 to $0.108), and half the approval prompts (8 to 4). On 20 direct approval cases, Jev sent 10/10 unsafe writes to a human and auto-ran 8/10 safe ones. The defaults above were tuned on these runs. One run per configuration; costs vary 2–4× between runs.

Docs: eve extension, presets and approval-control guides, API reference, READMEs, published skill, AGENTS.md, examples/eve now uses both 'auto' modes. Changesets: minor for @github-tools/sdk and @github-tools/eve-extension, patch for the formatter fix.

Verification: pnpm lint, pnpm typecheck, and pnpm test (130 tests) pass. Locally pnpm build fails only in examples/eve, where eve falls back to microsandbox without Docker running; main fails the same way on this machine.

📝 Checklist

• I have linked an issue or discussion.
• I have updated the documentation accordingly.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
github-tools-docs	Ready	Preview, v0	Sep 24, 2026 8:15am UTC
github-tools-test-agent	Ready	Preview, v0	Sep 24, 2026 8:15am UTC

[github-actions]

Thank you for following the naming conventions! 🙏