harness-kit

Plugin harness-kit from marketplace harness-kit in Dfintz/Fintz-harness-kit. Commands read from the repo's manifest — 19 days ago.

Skills (27)

• architect

  Runs the harness Architect stage and produces a settled Architecture Brief with gate decisions, boundaries, constraints, and validation. Use when a non-trivial task needs its design settled before implementation.

• feedback

  Runs the harness Feedback stage, adjudicating reviewer, stakeholder, or author challenges point by point and updating the Architecture Brief when decisions change. Use after Review Breadth and Review Depth, or when a settled decision is challenged.

• implement

  Runs the harness Implement stage, delivering the change an Architecture Brief describes with proof and self-review artifacts. Use when a Brief is settled, or for hotfixes and narrow bug fixes via the surgical variant.

• remember

  Persists reusable lessons and Architecture Briefs to the harness memory surfaces. Use when a non-obvious, evidence-backed lesson or a settled Brief should carry forward to future sessions.

• review-breadth

  Runs the harness Review Breadth stage, producing a severity-ordered findings ledger for correctness, standards, safety, completeness, and proof quality. Use after Implement on any non-trivial change.

• review-depth

  Runs the harness Review Depth stage, re-running architectural gates and checking ownership, boundaries, reuse, and Architecture Brief conformance. Use after Review Breadth on an implemented change.

• run-loop

  Executes a harness loop natively, following the loop JSON contract, bounds, rubric, and guardrails. Use when asked to run a convergence, workflow, or experiment loop, or to iterate until checks pass.

• ai-techniques-radar

  Track and evaluate external AI techniques and engineering trends. Use when deciding whether to adopt a new approach, library, or method.

• architect-challenge

  Review and challenge an Architecture Brief before implementation. Returns APPROVED, REVISE, or BLOCKED verdict.

• budget-aware-execution

  Cost-aware tool and model selection, token budget tracking, and bounded execution. Use for long-running or loop-heavy tasks.

• context-engineering

  Session memory hygiene, task-switch checkpointing, and compact handoff authoring. Use at session start and when switching tasks to preserve and recover context.

• deterministic-validation

  Define and enforce objective exit criteria, proof selection, and completion checks. Use when proving a change is done and during loop convergence checks.

• doubt-driven-development

  Security and correctness skepticism for high-stakes changes, and disciplined diagnosis for hard bugs. Use for security changes, irreversible operations, cross-model review of production behavior, and debugging complex failures.

• eval-first-tuning

  Establish baselines and run comparative evaluations before adopting a technique or library. Use when tuning retrieval, prompts, or agent quality.

• harness-feature

  Run the full harness stage machine for a feature task (Understand → Architect → Architect

• harness-prompt-pack

  Generate a harness prompt pack with orchestrator, stage prompts, cycle memory files, and optional sidecars.

• observability-and-instrumentation

  Telemetry, RED metrics, structured logging, and trace analysis. Use when adding or debugging observability in a service.

• pr

  PR creation, verification, and review-before-ship workflow. Use when opening a pull request or running a verification sub-agent before human review.

• prototype

  Build a throwaway Logic prototype to validate a state model or data shape before committing to an Architecture Brief. Use when the design question is about business logic, state transitions, or data shape — the kind that looks reasonable on paper but only feels wrong once you push it through real ca

• retrieval-quality-ops

  A/B evaluation of retrieval stacks comparing vector-only against contextual+BM25+rerank. Use when measuring recall@K, token cost, and task success to inform retrieval stack adoption decisions.

• setup-harness-bootstrap

  Adopt the harness in a new repository. Use when initializing stage workflows, skill trees, registry, and validation for a fresh project.

• teach-agent

  Convert repository domain knowledge into machine-operational guidance for agents. Use when curating lessons, authoring skill docs, or gating knowledge promotion.

• to-questionnaire

  Optional user-invoked conversion of freeform requests into a compact questionnaire for async stakeholder clarification.

• understand-process

  Graph-first impact analysis and dependency discovery for any non-trivial change. Use at stage 0 (Understand) to run the graph freshness gate, map changed/affected components, and identify blast radius.

• wait-what

  Optional user-invoked re-pitch guidance for unclear responses. Use only when the user asks to re-explain or simplify a previous answer.

• grill

  Alignment-first interrogation that closes the gap between what a human wants and what the agent builds, BEFORE planning or code. The agent relentlessly interviews the user one question at a time until every branch of the decision tree is resolved, challenges the plan against the project's shared lan

• harness

  Project-agnostic AI agent harness for driving non-trivial work to done. Provides a workflow stage machine (Understand → Architect → Implement → Review-Breadth → Review-Depth → Feedback) with five architectural gates, plus three loop kinds: convergence (iterate until lint/type/build/test are green),

Similar plugins