agent-harness
FeaturedLinlin's curated AI agent harness configuration: workflow rules, skills, hooks, plugin recommendations, tooling preferences, and project templates. Loadable into any new project so a fresh /init can pick the relevant subset.
Install
Plugins install via a marketplace, in two steps.
This plugin isn't listed in a marketplace we've indexed. Install it directly from its GitHub repository — the README has the setup steps.
View on GitHubBundles
Everything this plugin ships — skills, agents, commands, hooks, and MCP servers it bundles.
Skills (34)
new-hook
Scaffold a new hook under hooks/<kebab-name>/. Creates README.md (what/when/install/variants) and settings.snippet.json (drop-in JSON), plus updates INVENTORY.md and hooks/README.md in the same edit batch. Hooks are sensitive — pipe-test the raw command before wrapping.
new-rule
Scaffold a new workflow rule under rules/<kebab-name>/. Creates RULE.md (frontmatter + body template) and snippet.md (drop-in for downstream CLAUDE.md), plus updates INVENTORY.md and INVENTORY.zh.md in the same edit batch.
publish
Tag a new version of agent-harness, push to GitHub, and create a GitHub release. Use when shipping a meaningful update (new rules, new templates, etc.). Asks for SemVer bump and writes release notes from git log.
init-agent-config
Scaffold a new project with the relevant subset of agent-harness rules, hooks, skills, recommendations, and templates. Asks about project type, language preferences, and context tags, then composes a project-specific CLAUDE.md, .claude/settings.json, and starter files. Use in a fresh or existing project to apply agent-harness conventions.
feature-audit-template
1. Inventory ALL features: - Skills (\`ls skills/*/SKILL.md\`) - Rules (\`ls rules/*/RULE.md\`) - Hooks — Claude (\`ls hooks/*/README.md\`) AND Codex (\`cat hooks.json\`) - Scripts (\`ls scripts/*.{sh,js}\`) - Templates (\`ls templates/*/TEMPLATE_README.md\`) 2. For each feature, check: - Does it have a Claude Code implementation? (source) - Does it have a Codex implementation? (wrapper, or mapping) - Is the Codex wrapper correct? (tool names mapped, hooks present, auto-load configured) - Does i
workflow-fix-template
1. List ALL workflow files: `ls .github/workflows/` 2. Check CI run logs (gh CLI or web UI): `gh run list --limit 5; gh run view <id> --log` 3. Identify EVERY failing job — do not stop at the first one 4. For each failure: read the log output, identify root cause 5. Common failure patterns: - Gitleaks false positives → needs .gitleaks.toml - npm ERR_OSSL_EVP_UNSUPPORTED → needs NODE_OPTIONS=--openssl-legacy-provider - Missing binary → check bin field in package.json - Permission denied → check c
agent-config-adapter
Adapt an existing agent configuration or plugin to a new agent or model route. Use when moving agent-harness or another agent setup between Claude Code, Codex, Gemini, Cursor, local models, or non-native model backends such as DeepSeek routed through another agent.
agent-update-watcher
Use to keep up with the fast-moving agent ecosystem (new/updated CLIs, plugins, skills) WITHOUT burning tokens on constant checks. Declares the sources to watch in a config, checks them only when a minimum interval has elapsed, and reports ONLY what actually changed vs the recorded version — you then decide whether to adopt.
autopilot
Set up or manage the autopilot daily autonomous project-driver. Configures a daily timer that runs a fresh autorun session to advance a project (features, bug fixes, refactors, design, docs). Use when the user wants to "set up autopilot", schedule autonomous daily work, or configure/inspect the autopilot tool.
autoresearch-toolfinder
Find the right autonomous-research or research-agent tool from Codex using the source agent-harness autoresearch-toolfinder skill. Use when choosing, comparing, or setting up autoresearch tools, AI-scientist systems, research-agent loops, platform ports, or evaluation harnesses.
claude-desktop-sessions
Use when Claude Code sessions started in the terminal CLI don't appear in the Claude Desktop app's Code tab, when you want to hand a running CLI session to Desktop (and reach it from the phone via Remote Control), or when asked to "import/migrate my CLI sessions into the desktop app". Registers existing sessions in Desktop's sidebar by writing its per-session registry files, which point at the transcript store both apps already share — no transcript is copied or modified. Lists candidates numbered so you can import a chosen subset.
doc-writing
Use when writing or updating any document a human will read — a design or architecture spec, a feature manual and its in-app guide, an audit / bugfix / test report, a README, release notes, a plan, or a session handoff. Encodes the document requirements mined from two real project sessions (define every term in place, prove every claim and say how a tester re-proves it, full clickable links, mermaid diagrams, nothing stale, remaining tasks highlighted, no secrets) and ships `scripts/doccheck.py` to lint the finished file against the ones a machine can check.
Show all 34 bundled skills Showing all 34 bundled skills
end-of-turn-marker
Every turn MUST end with a visible divider header followed by numbered summary items. Use [END:FINAL], [END:WAIT], or [END:NEEDS_USER] embedded in the divider.
figma-authoring-constraints
Use when a designer asks how to structure a Figma file so it produces clean code, when a design keeps yielding pixel-snapshot output, or when get_variable_defs comes back empty — the 20 Figma-side authoring constraints that make a design cleanly code-able via the Figma MCP.
figma-design-fetch
Use when the user shares a figma.com URL, wants to implement/mock a UI from Figma, or do design-to-code — connect the Figma MCP, fetch the design (code/assets/screenshot) to disk, then rebuild with existing design-system components and visually self-verify against the Figma screenshot.
agent-harness-general
Audit the source agent-harness general skill catalog. Use only when maintaining or adapting this repository's original reusable skill set.
code-verifier
Use BEFORE claiming any code/test/script runs successfully. Detects FAKE-RUN patterns (hardcoded results, assert True, mocks-only tests, swallowed exceptions, fabricated numbers, dead-code short-circuits). Apply automatically whenever about to claim "test passes", "code works", "results show X", "training converges", or commit/push. Complements superpowers:verification-before-completion (which enforces real-run discipline) by auditing whether the run itself is genuine.
linux-freeze-triage
Use when a Linux desktop/laptop goes black-screen, frozen, or unresponsive (often after idle, or when running a GPU/video app like Zoom/Chrome), and you must find the REAL cause instead of guessing. Covers ruling out suspend/sleep, NVIDIA driver kernel-vs-userspace version mismatch from auto-upgrades, OOM meltdowns, PCIe link errors, and DPMS/display-wake hangs. Bundles a near-zero-cost watchdog to capture the next freeze and a read-only diagnostic battery. Apply before recommending any fix for a freeze/black-screen.
research-critic
Use AT EVERY RESEARCH STEP (hypothesis design, experiment setup, result interpretation, conclusion writing) to challenge soundness BEFORE committing the step. Catches confirmation bias, p-hacking, leakage, weak baselines, ungrounded claims, ad-hoc thresholds, survivorship bias. Apply automatically whenever about to write a paper claim, design an ablation, interpret a number, or update a results doc. Pairs with code-verifier (which audits artifact authenticity).
init-codex-config
Scaffold or migrate a project to use agent-harness from Codex. Use when applying agent-harness in Codex, converting Claude Code setup to AGENTS.md/.codex/.agents, or making a project ready for Codex skills and hooks without changing the original Claude setup.
long-running-tasks
Manage long-running commands and multi-step work in Codex. Use when a task may exceed normal turn time, needs a dev server, watcher, training run, crawl, benchmark, or repeated polling.
memory-flywheel
Use to persist and recall per-project cross-session working memory so long or multi-project sessions don't lose detail to context compaction. Record each round of progress (verbatim I/O + reasoning + metadata) into a project memory dir, read a coarse INDEX first, then open only what keyword recall points at. Deterministic, grep-native, LLM-as-component.
preview-template
Start or identify the local preview environment from Codex. Use for UI, docs, static site, Vite, Next.js, Storybook, MkDocs, or visual verification workflows.
privacy-redact
Scan and redact private details before publishing or committing text. Use for usernames, absolute paths, secrets, project codenames, private URLs, and local machine details.
prompt-library
Use to save and reuse good prompts across projects and agents, and to MINE past prompts out of local Claude Code / Codex / Copilot / opencode history. Many prompts recur (a detailed feature spec, a "design the whole thing" brief, a manual/onboarding prompt) and are worth reusing verbatim or as a reference. This curates them as browsable, greppable Markdown — each with the original, an optimized rewrite, and when to use / when NOT to use it — behind a PRIVACY GATE that refuses to store anything still containing paths/emails/tokens/usernames/codenames, so the library stays publishable.
system-cleanup
Diagnose and plan Linux disk cleanup from Codex using the source agent-harness system-cleanup skill. Use when disk space is low, the user asks to free space, or a data disk or mount is read-only. Advises first and requires confirmation before destructive or sudo cleanup.
task-orchestrator
Meta-skill that enforces a structured Research-Design-Plan-Execute-Verify pipeline for every atomic task. Compensates for model/agent weaknesses (especially DeepSeek, non-native models) by adapting plan template strictness based on detected capabilities. Self-improving: records better approaches back into templates.
task-relationship-analysis
Use BEFORE executing any multi-step or multi-feature request (3+ tasks). Forces a pre-execution pass that maps how the tasks relate — synergies, conflicts, shared substrate, ordering — so you don't mechanically execute them in listed order and miss that several should be built on one shared piece, or that two conflict. Scaffolds a pairwise matrix + synthesis checklist.
tui-installer
Use to install (or just plan) the recommended terminal/TUI stack for driving multiple AI coding agents on Ubuntu — zellij + claude-squad + lazygit + delta. Dry-run by default (prints the plan, installs nothing); --apply asks y/N per tool. When onboarding agent-harness, offer to run --check and ask before installing anything.
verify-template
Run project verification from Codex before claiming work is done. Use before marking a phase complete, opening a PR, or saying tests/lint/build pass.
i18n-sync
Check key parity across locales/*.json. If divergence is detected, list missing keys per locale and offer to add stubs. Use after editing any locale file.
preview
Start a local dev server for this static site. Required before /verify-visual since file:// breaks fetch() calls for locales/*.json.
verify-visual
Use chrome-devtools MCP to screenshot the local site (or a specific element) and visually verify it matches the design intent. Required after every UI-affecting change before marking work done.
verify
Run the project's CI gate locally — ruff lint, ruff format check, mypy, and fast pytest (excluding slow/download markers). Use before opening a PR or marking work done.
Commands (2)
Hooks (9)
Quality Score: 73/100
Details
- Author
- jajupmochi
- Repository
- jajupmochi/agent-harness
- Created
- 2 months ago
- Last Updated
- 2 days ago
- Language
- Python
- License
- MIT