feat: Implement Safe Commit Pipeline for Autonomous Agent Execution

Problem (A2A Context — Claude Opus 4.6 via Antigravity)

When the Autonomous Orchestrator (#9842) executes Golden Path directives, the agent can generate code (via QA sub-agent), create tickets (via GitHub Workflow MCP), and run tests. However, there is currently no safety mechanism preventing a malformed autonomous commit from landing directly on the dev branch.

The existing human-in-the-loop workflow (Antigravity/Gemini sessions) provides implicit safety through human approval of each commit. The autonomous loop removes this gate.

Solution

Implement a CommitGate utility in ai/agent/CommitGate.mjs that enforces:

1. Branch Isolation

All autonomous agent commits MUST target a dedicated agent/<ticket-id> branch, never dev directly. The CommitGate creates and checks out this branch before any file modifications.

2. Pre-Commit Validation

Before executing git commit, the CommitGate runs:

const gates = [
    () => this.runUnitTests(modifiedFiles),    // npm run test-unit -- <affected specs>
    () => this.runE2ETests(modifiedFiles),      // ButtonBaseNL-style regression suite
    () => this.validateJSDoc(modifiedFiles),    // Contextual Completeness Gate (AGENTS.md §3)
    // () => this.runLintCheck(),               // Future: pending linter R&D (see companion ticket)
];

All gates must pass. If any gate fails, the commit is aborted and the failure is recorded in the dead letter queue (Loop.failedEvents).

Note: Neo.mjs does not currently have a linter. The JSDoc validation gate enforces structural quality (missing @summary, undocumented configs) using AST-level checks rather than style linting. A future linter (tracked separately) would slot into this gate array.

3. PR Creation with Auto-Review Request

After a successful commit + push, the CommitGate automatically creates a PR via the GitHub Workflow MCP, targeting dev, with:

• Title: feat: <ticket title> (#<ticket-id>) (Conventional Commits)
• Body: Fat Ticket context from the original issue
• Label: agent-task:review
• Assignee: tobiu (human review required for v1)

4. Integration with Loop.reflect()

The CommitGate result (success/failure/PR URL) feeds into the Reward Signal pipeline (#9843) for Golden Path recalibration.

Architectural Context

• ai/agent/Loop.mjs (L476-508): executeTools() — the tool execution layer where CommitGate would intercept git-related tool calls
• AGENTS.md Section 3 (Gate 1): The existing "Ticket Gate" protocol that must be enforced programmatically
• AGENTS.md Section 7: Git Protocol — commit message format requirements (feat: ... (#TICKET_ID))

Avoided Pitfalls

• Do NOT allow direct commits to dev — always use feature branches
• Do NOT skip tests "because they passed last time" — every commit must re-validate
• Do NOT auto-merge PRs in v1 — human review is the safety net until the Reward Signal proves the agent's reliability over time
• The CommitGate should be a composable utility, not baked into the Loop, so it can be used by any agent profile independently
• Do NOT assume a linter exists — the JSDoc gate uses AST parsing, not style rules

Verification

• Unit test: test/playwright/unit/ai/agent/CommitGate.spec.mjs
  • Assert: Commits are rejected if unit tests fail
  • Assert: Commits target agent/<id> branch, never dev
  • Assert: PR is created with correct labels and assignee
  • Assert: Failure result is recorded in dead letter queue