forge

Use this agent when an investigation report or PRD needs adversarial review before implementation. Finds edge cases, missed codebase patterns, gaps in requirements, and ignored conventions. Autonomous — does its own homework instead of asking questions. Triggers: "challenge the plan", "find gaps", "adversarial review", "stress test PRD".

Use this agent when generated code needs evaluation against project criteria. Returns PASS/FAIL with specific findings per file and line. Adversarial — finds real problems, not rubber-stamps. Tool-aware — detects available analysis tools and degrades gracefully. Triggers: "evaluate code", "check implementation", "adversarial review".

Use this agent when code needs writing following the planner's approach and reference patterns. Reads evaluator criteria before implementing, self-checks against BLOCKERs. On subsequent iterations, fixes only what the evaluator flagged. Triggers: "generate code", "implement the plan", "write the implementation".

Use this agent when codebase exploration is needed before implementation. Use this agent when evaluator-criteria.md does not exist and needs auto-generation. Produces a "TODAY IS / EXPECTED IS" report grounded in file paths. Read-only — never writes code or proposes implementations. Exception: writes evaluator-criteria.md when auto-generating on first run. Triggers: "investigate codebase", "explore before implementing", "map patterns", "generate evaluator criteria".

Use this agent when an implementation approach needs designing after investigation and challenge. Produces a file-level plan with references, constraints, and task-specific evaluator criteria. Read-only — analyzes codebase patterns but never writes implementation code. Triggers: "plan implementation", "design approach", "create implementation plan".

Use when finishing a task and cleaning up local branches whose content is already in the default branch. Also use when the user wants to clean up after merging PRs or remove branches that were squash-merged or rebased. Triggers: "finish task", "forge finish", "clean up branches", "prune branches", "clean branches", "delete integrated branches", "remove merged branches", "done with task".

Use when implementing features, fixes, or infrastructure changes that modify code. Use when the user wants adversarial quality checking before committing. Do not use for trivial single-file changes like typos or comment edits. Triggers: "implement", "forge implement", "implement this ticket", "build this feature", "start implementing", "implement feature", "adversarial pipeline", "build with forge".

Use when the user has a clear goal but implementation hasn't started yet. Also use when Claude needs to understand existing codebase patterns, conventions, and constraints before proposing changes. Produces a "TODAY IS / EXPECTED IS" investigation report, then stops. Does NOT write code or propose implementations. Triggers: "investigate first", "explore the codebase", "what exists", "map the patterns", "today is expected is", "before we implement", "what are the conventions", "understand before changing".

Use when starting a new task and wanting a clean working directory on the default branch. Also use when finishing a task, switching context, or returning to main after a PR. Triggers: "start task", "forge start", "reset", "switch to main", "start fresh", "back to main", "done with this task", "clean up workspace", "done with branch".