Repository navigation
Skills and Agents
build-loop ships ~18 commands, ~40 skills, ~28 agents from one source. Commands are what you type; skills carry procedural detail; agents are bounded roles the orchestrator dispatches (you do not run agents directly).
Normal coding work, auto-routed to build, fix, refactor, optimize, research, or test:
/build-loop:run add billing settings with tests
Deep iterative debugging (also auto-invoked by the loop on review failures):
/build-loop:debug tests pass locally but fail in CI
Direct advanced modes:
/build-loop:optimize-run reduce API latency
/build-loop:research-run compare queue providers
/build-loop:test --strict
/build-loop:self-improve
/build-loop:self-improve runs Phase 6 Learn alone against recent runs without a new build. The repo also ships native debugger commands: /build-loop:debugger, /build-loop:debugger-detail, /build-loop:debugger-scan, /build-loop:debugger-status, and /build-loop:assess.
Full command set: run, debug, assess, optimize-run, research-run, test, self-review, verify-plan, review-knowledge, compose-handoff, setup-memory, start-prd, promote-experiment, rally-point, debugger, debugger-detail, debugger-scan, debugger-status.
None of these are commands you run directly. An agent is core when a pipeline step is contingent on its verdict. Model tier follows role; it does not define authority. Deterministic judge surfaces also exist outside agents/ (scripts/plan_verify.py, scripts/judgment_gate.py, release verifiers).
| Agent | Role | Tier |
|---|---|---|
build-orchestrator |
Lead owner of Assess→Plan→Execute→Review→Iterate→Learn; dispatch, transitions, commits, report. | Thinking |
assessment-orchestrator |
Multi-domain debugging coordinator for unclear symptoms (db / frontend / API / perf). | Thinking |
optimize-runner |
Optimization-loop coordinator for metric-driven experiments. | Code |
| Agent | Role | Tier |
|---|---|---|
advisor |
Frontier planning author / re-planner for deeper Phase 2 synthesis. | Frontier |
plan-critic |
Plan critique — dependencies, scope drift, ownership, alternatives, MECE. | Frontier |
scope-auditor |
Plan→Execute boundary + public-signature caller coverage. | Frontier |
independent-auditor |
Independent adversarial review of chunk/build completion claims. | Frontier |
fix-critique |
Root-cause and regression pressure-test after a proposed fix. | Frontier |
fact-checker |
Claim, metric, and rendered-data provenance before completion. | Frontier |
security-reviewer |
Auth, secrets, trust boundaries, injection, adjacent risks. | Frontier |
overfitting-reviewer |
Optimization review for test gaming / Goodhart / overfitting. | Frontier |
promotion-reviewer |
Review of proposed skill/agent/enforcement promotions. | Frontier |
synthesis-critic |
Advisory coherence review for synthesis-heavy outputs. | Code |
alignment-checker |
Advisory queue-item alignment vs intent / goal / non-goals. | Code |
implementer (bounded coding worker), architecture-scout, design-contract-specialist, ui-validator, mock-scanner, root-cause-investigator, the domain assessors (api-assessor, database-assessor, frontend-assessor, performance-assessor), and the Learn-phase agents (recurring-pattern-detector, self-improvement-architect, retrospective-synthesizer, transcript-pattern-miner).
The ~40 skills carry the procedural detail the orchestrator fires into. Notable ones:
-
Core loop:
build-loop,spec-writing,plan-verify,model-tiering,capabilities. -
Debugging:
debug-loop,debugging,debugging-memory,root-cause-analysis,logging-tracer. -
Knowledge / learning:
knowledge,knowledge-review,auto-decision-capture,self-improve,recursive-retrospective. -
UI / design:
ui-design,runtime-parity-verification,ibr-bridge. -
Coordination:
agent-rally-point,agent-rally-watcher,handoff,verify-dispatch. -
Build infra:
plugin-builder,plugin-tests,mcp-builder,loop-builder,security-methodology,authentication.
Bundled skills own procedural detail; bridges are optional escalation hops. logging-tracer-bridge and defenseclaw-bridge delegate to the standalone Coding Debugger supporting plugin for extended capability (cross-project incident memory, additional assessor coverage). NavGator integration is via a cherry-pick bridge. build-loop bundles its own debug-loop / debugging-memory / logging-tracer skills, so no build-loop MCP server is required for core debugging.