agent-harness-kit

Build a lightweight skill index and load full skill instructions only on demand to reduce startup context pressure.

Select and run a multi-agent workflow pattern for work that exceeds one agent's reliable scope. Use for parallel research, independent reviews, cross-domain changes, or high-risk implementation planning.

Select and run a multi-agent workflow pattern for work that exceeds one agent's reliable scope. Use for parallel research, independent reviews, cross-domain changes, or high-risk implementation planning.

Use composable middleware behaviors for provider/tool execution: retry, caching, timeout, telemetry, and budget guards.

Use this skill whenever the agent makes a mistake, the user observes an avoidable failure, a pattern recurs, or someone says "the agent keeps doing X". Files an "Engineer the Harness" entry — Mitchell Hashimoto's discipline: every failure becomes a permanent prevention mechanism. Always invoke this instead of just fixing the immediate symptom.

Use composable middleware behaviors for provider/tool execution: retry, caching, timeout, telemetry, and budget guards.

Use this skill whenever the agent makes a mistake, the user observes an avoidable failure, a pattern recurs, or someone says "the agent keeps doing X". Files an "Engineer the Harness" entry — Mitchell Hashimoto's discipline: every failure becomes a permanent prevention mechanism. Always invoke this instead of just fixing the immediate symptom.

Build a compact read-only context packet for a natural-language codebase question. Use before editing unfamiliar code, when tracing task evidence, contracts, validation, or proof paths, or when the relevant files are not obvious.

Build a compact read-only context packet for a natural-language codebase question. Use before editing unfamiliar code, when tracing task evidence, contracts, validation, or proof paths, or when the relevant files are not obvious.

Create a Story Packet for normal/high-risk features. Use after /feature-intake classifies work as normal or high-risk, or when the user asks to break a feature into acceptance criteria, test expectations, and agent-sized work units.

Use this skill whenever the user asks to "add a feature", "implement X", "build Y", or before invoking /add-feature. Classifies the request into tiny/normal/high-risk based on estimated complexity and blast radius, then routes to the appropriate workflow — straight to code (tiny), story packet (normal), or ADR + story packet + mandatory review (high-risk). Prevents "wrong direction" sessions by forcing a pause before implementation. Pattern from harness-experimental and OpenAI's "implementation prompts do not go straight to code" discipline.

Use this skill after a trace-backed agent failure or repeated harness friction. Turns the failure into a ranked harness change, records a prediction, applies the smallest prevention, and reruns the relevant eval/regression gate. This is the agent-harness-kit AHE-lite loop.

Use this skill whenever the user asks to "test the UI", "check what the app looks like", "inspect the page", "verify the dev server is up", or before claiming a UI feature is done. Boots the dev server via .harness/scripts/dev-up.sh and drives the failing flow through Playwright MCP if installed (else falls back to curl + lightweight HTML capture). Mirrors the OpenAI Chrome-DevTools-Protocol-into-runtime pattern at solo scale — verify the running app, don't trust the type checker alone.

Run Browser Validation with Playwright golden-path checks, screenshots, console/network capture, and HTML reporting.

Use this skill whenever a decision is made about architecture, dependencies, frameworks, naming conventions, or layer order. Creates a numbered ADR (Architecture Decision Record) in `.harness/docs/adr/` in the canonical Nygard format. Always invoke this before changing layer order, adding a layer, swapping a major dependency, or introducing a new external service.

Run Mini SWE-bench style harness regression tasks and A/B comparisons to measure harness improvement objectively.

Inspect context usage, token budget, compaction history, and overflow risk. Use when sessions get long, before large changes, after compaction, or when cost/context drift is suspected.

Use this skill weekly, before releases, or when the user mentions "stale docs", "doc drift", "docs are wrong", or "the README is out of date". Cross-checks every code path, file path, and command referenced in `.harness/docs/` and `CLAUDE.md` against the current repo state and produces a list of stale references — the doc-gardening agent pattern.

Use this skill when adding or changing harness eval tasks. Writes deterministic checks first, then optional rubric dimensions with JSON output, so evals grade outcome, process, style, and efficiency without becoming vague prompt feedback.

Use this skill whenever a skill, subagent, or hook is changed, before merging to main, or when the user mentions "eval", "regression test for the harness", or "is the harness still working". This skill is a thin wrapper — it runs a single shell command (`npm run harness:eval` or `python -m harness.eval_runner`) and summarizes the JSONL output. Do not implement the eval logic yourself; the runner is already deterministic.

Use this skill to scaffold a new human-language locale for the kit's skills/agents/CLAUDE.md. Mirrors every existing SKILL.md.hbs into a SKILL.md.<locale>.hbs stub so a translator (or LLM) can edit copy without touching machine-readable frontmatter. Default locale codes — vi, ja, fr, es, de — but accepts any 2-5 char code.

Use this skill to render the harness's domain/layer config as a mermaid diagram + check for drift between .harness/config.json#domains and the actual filesystem. Surfaces "the config says layers A→B→C but the repo has folders A, B, X" — drift that silently invalidates the structural-test contract.

Use this skill when choosing, upgrading, or comparing models for a harness lane. Runs the same task set across candidate model profiles and reports pass rate, latency, token cost, intervention count, and failure class movement before changing defaults.

Use this skill when the user asks for project status, roadmap, MVP phase, scope, checklist, risks, handoff, harness control-plane health, or a project-management report. Renders project state, memory, orchestration, and session isolation into a self-contained HTML dashboard.

Use this skill when restructuring a feature in .harness/feature_list.json — splitting steps, merging steps, renaming, or marking a previously-failing step done. The side-car diffs .harness/feature_list.json#steps before/after and rejects the edit when a step.done transition is not accompanied by a test reference. Forces "no done without proof".

Run Tier 2 real-world-style regression benchmarks with isolated workspaces, hidden checks, A/B comparison, cost/time metrics, and HTML dashboard reporting.

Use this skill when a decision, risk, scope change, handoff note, or durable project fact should be shared across future humans, AI agents, and teammates. Writes a semantic event to `.harness/memory/ledger.jsonl` and refreshes `.harness/memory/current-summary.md`; do not use it for raw transcripts, secrets, or ordinary git diff details.

Use this skill to run a deterministic review of the current branch against its base — git diff base...HEAD, structural-test, baseline-monotonic check, and a markdown summary that lists each violating file with its layer rule. Replaces the "ask the agent to review the diff" pattern, which routinely misses cross-file drift.

Use this skill when the user wants to schedule the harness eval to run every night, asks "how do I set up nightly evals", "schedule the eval", "run evals on a cron", or "nightly regression for the harness". The kit already ships a GitHub Actions workflow at .github/workflows/eval-nightly.yml — this skill walks the user through enabling it (secret setup, smoke run via workflow_dispatch, verifying the first scheduled run). Do NOT use this skill to "remind me to run eval every night in this Claude session" — that is the /loop skill or CronCreate (which is session-only), a different concern.

Use this skill when an eval, regression run, hook, or long agent session fails and the next step depends on trace evidence. Classifies the failure from telemetry, transcripts, eval JSONL, git diff, and hook output before proposing a fix. Prevents prompt-only guesses by requiring a trace-backed failure class.

Use this skill whenever the user asks to "create a skill", "add a slash command", "package a workflow", or "make X reusable across sessions". Generates a SKILL.md with valid YAML frontmatter (name regex, description ≤ 1024 chars, body ≤ 500 lines) and supporting .harness/scripts/references/assets. Tests the skill by simulating an auto-discovery prompt.

Run Mini SWE-bench style harness regression tasks and A/B comparisons to measure harness improvement objectively.

Inspect context usage, token budget, compaction history, and overflow risk. Use when sessions get long, before large changes, after compaction, or when cost/context drift is suspected.

Create a Story Packet for normal/high-risk features. Use after /feature-intake classifies work as normal or high-risk, or when the user asks to break a feature into acceptance criteria, test expectations, and agent-sized work units.

Use this skill weekly, before releases, or when the user mentions "stale docs", "doc drift", "docs are wrong", or "the README is out of date". Cross-checks every code path, file path, and command referenced in `.harness/docs/` and `CLAUDE.md` against the current repo state and produces a list of stale references — the doc-gardening agent pattern.

Use this skill when adding or changing harness eval tasks. Writes deterministic checks first, then optional rubric dimensions with JSON output, so evals grade outcome, process, style, and efficiency without becoming vague prompt feedback.

Use this skill whenever a skill, subagent, or hook is changed, before merging to main, or when the user mentions "eval", "regression test for the harness", or "is the harness still working". This skill is a thin wrapper — it runs a single shell command (`npm run harness:eval` or `python -m harness.eval_runner`) and summarizes the JSONL output. Do not implement the eval logic yourself; the runner is already deterministic.

Use this skill whenever the user asks to "add a feature", "implement X", "build Y", or before invoking /add-feature. Classifies the request into tiny/normal/high-risk based on estimated complexity and blast radius, then routes to the appropriate workflow — straight to code (tiny), story packet (normal), or ADR + story packet + mandatory review (high-risk). Prevents "wrong direction" sessions by forcing a pause before implementation. Pattern from harness-experimental and OpenAI's "implementation prompts do not go straight to code" discipline.

Use this skill after a trace-backed agent failure or repeated harness friction. Turns the failure into a ranked harness change, records a prediction, applies the smallest prevention, and reruns the relevant eval/regression gate. This is the agent-harness-kit AHE-lite loop.

Use this skill to scaffold a new human-language locale for the kit's skills/agents/CLAUDE.md. Mirrors every existing SKILL.md.hbs into a SKILL.md.<locale>.hbs stub so a translator (or LLM) can edit copy without touching machine-readable frontmatter. Default locale codes — vi, ja, fr, es, de — but accepts any 2-5 char code.