Engineer describes feedback loop system for autonomous AI-assisted code review and mission execution

A software development workflow splits into two modes: Iterate (single feature with step-by-step supervision) and Mission (parallel work units in a directed acyclic graph, reviewed at merge gates but not watched step-by-step). Both modes feed the same knowledge base and branch rules.

Mission execution uses isolated worktrees (separate filesystem copies) per task, with workers claiming and executing tasks in parallel (capped at 3–5 concurrent), submitting for review without merging directly. The daemon handles worktree creation, branch management, and merge orchestration via eight RPCs.

Review gates separate deterministic checks (tests pass/fail, builds compile, type-checking, schema validation) from latent judgment (code style, design patterns, presentation). Deterministic failures are mechanical and unambiguous; latent failures require human reading.