feat(agent): add guarded PR preparation subagent

[ncolesummers]

Summary

• add the isolated Terra/xhigh pr-preparer sibling with deny-all framework capabilities and bounded evidence readers
• persist a strict evidence-bound loopworks.pr_preparation_result.v1 before the guarded PR writer runs
• bind external-write approval to both repository changes and the exact prepared-intent digest
• add deterministic issue PR preparation subagent for PR intent content #50 eval coverage and update the governing architecture records
• remove the invalid validation-reviewer write_production_files sentinel that prevented Eve 0.22.5 runtime graph resolution

Why

The PR stage previously composed its own intent and was still human-owned, so model-authored PR narrative could not enter the existing approval-gated writer through a typed, durable handoff. The new sibling can author only bounded title and summary text; the root reconstructs links and evidence from persisted state, validates every digest, and remains the only persistence boundary. GitHub mutation stays inside the existing guarded writer.

The adjacent Eve fix is included with maintainer approval because disableTool() only accepts framework tool slugs. A custom write_production_files sentinel was unnecessary—the absent tool is already unavailable—and caused runtime graph compilation to fail.

Safety and behavior

• PR-preparer has no shell, file-write, network, GitHub, or durable-transition capability.
• UI-affecting runs require exact mobile, laptop, and desktop validation captures; non-UI runs require an explicit empty list.
• Unsafe URLs, stale or conflicting evidence, duplicate screenshots, changed digests, raw output, prompts, bytes, and secret-like content fail closed.
• Prepared-intent persistence is idempotent and conflicting replay is rejected before GitHub.
• Existing draft-PR reconciliation and GitHub App ownership remain unchanged.

Validation

• bun run validate
  • formatting, linting, agent-doc sync, Markdown, and TypeScript passed
  • 407 unit/integration tests passed
  • Storybook production build passed
  • 22 Playwright tests passed, including mobile and light/dark accessibility coverage
• bun run build passed
• bunx eve info resolved the runtime graph with no diagnostics
• bunx eve eval --list discovered the issue PR preparation subagent for PR intent content #50 case

The provider-backed eval was not invoked from the development environment because external model data egress was denied. Its deterministic fixture, parser, discovery path, and local Eve runtime graph were verified.

Closes #50

Co-authored-by: OpenAI Codex GPT-5 noreply@openai.com

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
loopworks	Ready	Preview, Comment	Jul 21, 2026 6:07am

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.