agent-harness

Systematically analyze project structure, dependencies, and patterns for team onboarding and codebase understanding. 3-tier mode — quick (inline) or deep/thorough (parameterized lens analysts + a completeness-critic pass + synthesis via a plugin-shipped native Workflow segment, opt-in gated), with incremental analysis support. Use when joining a new project or generating reproducible codebase documentation.

Opt-in gated 3-Phase orchestrator (Plan -> Gate -> Generate -> Verify -> Evaluate). Runs plugin-shipped native Workflow segment scripts on the workflow path (ultracode or --mode opt-in) with schema-validated returns; inline single path otherwise. Use for development tasks (feature work, bug fixes, maintenance) AND non-development tasks that benefit from structured planning, implementation, and 3-layer review. (formerly /workflow)

Automated test generator with mutation-based quality verification — single (inline) or multi (parallel coverage analysts + propose-only mutation skeptics via plugin-shipped native Workflow segments, opt-in gated; test generation and mutation execution stay orchestrator-inline). Generates unit and integration tests, runs them, and validates meaningfulness via simplified mutation testing. Supports coverage-gap detection and regression test generation from debug reports.

Systematic, bias-free code review for PRs, branches, or file changes. Quick mode (1 agent, 5-perspective checklist, inline) or deep/thorough modes (2-3 specialist reviewers + adversarial cross-verification + synthesis via a plugin-shipped native Workflow segment, opt-in gated). Optional --comment (inline PR comments) and --fix (gated apply). The deeper, bias-reduced complement to the built-in /code-review. (formerly /code-review)

Execute framework/library version upgrades, language transitions, and dependency replacements with staged, verified execution — single (inline) or multi (parallel external-research + codebase-impact analysts + synthesis and an isolated evaluator via plugin-shipped native Workflow segments, opt-in gated). WebSearch-based migration guide research, per-step test verification, and session recovery.

Safe, behavior-preserving code structure improvement with 3-tier mode — single (inline) or multi/comprehensive (2-3 analysts + cross-critique + synthesis via plugin-shipped native Workflow segments, opt-in gated). Atomic changes, test after each step, stop on failure. Use when improving code structure without changing behavior.

Hypothesis-driven debugger with mandatory executable verification. Classifies build/compile vs runtime/logic errors, attempts reproduction, generates falsifiable hypotheses verified by code search, git history, and test execution. Quick mode (orchestrator only, inline) or deep mode (2 specialist analysts + adversarial cross-verification via a plugin-shipped native Workflow segment, opt-in gated). Use when facing a bug, unexpected behavior, or error that needs systematic root cause analysis.

Requirements specification writer with multi-round Q&A discovery. Transforms vague ideas into structured specs compatible with /harness input. Modes — quick (orchestrator only, inline) / deep (4 analysts + Critic via plugin-shipped native Workflow segments, opt-in gated). Use when you need a well-defined spec before starting implementation.

Analyze project source code and generate/enhance CLAUDE.md for effective Claude Code development. Detects conventions, dependencies, build/test commands, architecture patterns, and produces actionable documentation. Use on new projects or when existing CLAUDE.md is thin. Richer superset of built-in /init — adds Smart Routing, a confirmation gate, an isolated bias-free evaluator pass, monorepo sub-CLAUDE.md generation, and enhancement (not just first-time creation).

[DEPRECATED -> use /team-memory] Team knowledge base manager. Renamed to /team-memory to end the trigger ambiguity with Claude Code's built-in personal auto-memory / the # shortcut / CLAUDE.md. This stub forwards to /team-memory.

Q&A-based universal release pipeline orchestrator. Guides version bump (2-pass), changelog generation (Conventional Commits), build/test verification, code review summary, git operations (commit/tag/push), and GitHub release — with HARD-GATEs before every irreversible action. Auto-detects environment, skips unavailable stages. Session recovery with substep tracking. Distinct from built-in /code-review — /ship orchestrates the whole release; its Stage 5 is a release-readiness diff summary, not a line-by-line review.

Team knowledge base manager. Save, show, search, and clean git-committed team knowledge records (decisions, bugs, patterns, todos, conventions) stored in docs/harness/memory/. Distinct from Claude Code's built-in personal auto-memory and the `#`/CLAUDE.md surface — this is shared, version-controlled, and team-wide. Human-gated CRUD only; never escalates to background agents or the Workflow engine.

[DEPRECATED -> use /deep-review] Systematic, bias-free code review. Renamed to /deep-review to end the literal collision with Claude Code's built-in /code-review (2.1.147+, --comment/--fix). This stub forwards to /deep-review.

[DEPRECATED -> use /harness] 3-Phase Planner/Generator/Evaluator workflow. Renamed to /harness to end the concept collision with Claude Code's native Workflow engine. This stub forwards to /harness.

Optimize CLAUDE.md and project .md files for token efficiency. Applies Dual-Zone model (Inline/Index), deduplication, and structural compression. Use on any project to reduce context token cost.