{"id":51591654,"url":"https://github.com/trac3r00/agent-skills","last_synced_at":"2026-07-11T15:02:04.351Z","repository":{"id":370028546,"uuid":"1292867760","full_name":"trac3r00/agent-skills","owner":"trac3r00","description":"Agent Skills for long-lived AI agents: context-budget (audit per-turn token weight, fail CI on bloat) + claim-audit (flag unverified factual claims). Standalone tools + CI-ready, Claude Code / Codex / OpenCode compatible.","archived":false,"fork":false,"pushed_at":"2026-07-08T00:43:13.000Z","size":69,"stargazers_count":0,"open_issues_count":0,"forks_count":0,"subscribers_count":0,"default_branch":"main","last_synced_at":"2026-07-08T02:08:05.219Z","etag":null,"topics":["agent-skills","ai-agents","ci","claude-code","codex","context-engineering","hallucination","llm-tools","token-optimization"],"latest_commit_sha":null,"homepage":"","language":"Python","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"mit","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/trac3r00.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":"LICENSE","code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null,"notice":null,"maintainers":null,"copyright":null,"agents":null,"dco":null,"cla":null}},"created_at":"2026-07-07T23:22:55.000Z","updated_at":"2026-07-08T00:43:16.000Z","dependencies_parsed_at":null,"dependency_job_id":null,"html_url":"https://github.com/trac3r00/agent-skills","commit_stats":null,"previous_names":["trac3r00/agent-skills"],"tags_count":null,"template":false,"template_full_name":null,"purl":"pkg:github/trac3r00/agent-skills","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/trac3r00%2Fagent-skills","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/trac3r00%2Fagent-skills/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/trac3r00%2Fagent-skills/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/trac3r00%2Fagent-skills/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/trac3r00","download_url":"https://codeload.github.com/trac3r00/agent-skills/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/trac3r00%2Fagent-skills/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":35365809,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-05-26T15:22:16.424Z","status":"online","status_checked_at":"2026-07-11T02:00:05.354Z","response_time":104,"last_error":null,"robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":true,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["agent-skills","ai-agents","ci","claude-code","codex","context-engineering","hallucination","llm-tools","token-optimization"],"created_at":"2026-07-11T15:02:00.338Z","updated_at":"2026-07-11T15:02:04.346Z","avatar_url":"https://github.com/trac3r00.png","language":"Python","funding_links":[],"categories":[],"sub_categories":[],"readme":"# agent-skills\n\nRunnable [Agent Skills](https://agentskills.io) for people running **real, long-lived AI agents** — and for people who just want their own recurring costs audited. The agent skills tackle the messes agents make: they accrete context, assert things without evidence, and talk across surfaces that don't share a brain. The everyday skills tackle the messes *humans* let slide — like a subscription stack nobody re-reads. All ship as small, dependency-light Python tools you can run standalone, drop into CI, or load as a skill in Claude Code, Codex, OpenCode, and any agent that follows the open [Agent Skills standard](https://agentskills.io/specification).\n\nMost skill directories are full of prompt wrappers. These are **runnable tools that return an exit code** — they hold a line, not just a vibe.\n\n| Skill | What it does | Why it's rare |\n|-------|--------------|---------------|\n| [`context-budget`](skills/context-budget) | Audits an agent's per-turn context window like a cost center — ranks what actually eats tokens (system prompt, skills, memory, tool schemas, gate code) and **fails CI when the agent gets fatter**. | Everyone talks about \"context engineering\"; almost nobody gives you a ruler with a budget you can enforce. |\n| [`claim-audit`](skills/claim-audit) | A linter for AI answers: separates **grounded / hedged / bare** factual claims and surfaces the unverified hard assertions most likely to be hallucinations. Gates a reply before it ships. | Fact-checkers are heavy and online; this is an offline, instant triage of *which* claims to verify. |\n| [`open-loops`](skills/open-loops) | Extracts the **unresolved commitments, deferrals, decisions, and open questions** from a thread into a JSON ledger — so a cron job or fresh session picks them up instead of dropping them. | The conversation surface and the scheduled surface never share memory; this is the missing handoff between them. |\n| [`subscription-audit`](skills/subscription-audit) | Finds the **recurring charges hiding in a bank/card CSV** — clusters repeat payments by merchant and cadence, reports the true monthly and yearly cost, and flags the **forgotten** ones (stale, or a free trial that started charging) worth cancelling. | Every \"find my subscriptions\" product wants a bank login; this reads a CSV you already have, fully offline, and returns an exit code you can put on a monthly cron. |\n| [`gate-graph`](skills/gate-graph) | Builds an AST overlap matrix for validator/middleware gate modules, flags duplicate checks and dead gates, and fails CI when overlap or count thresholds are exceeded. | Gate layers grow quietly; this keeps validator stacks from becoming bloated without a measurable signal. |\n| [`skill-decay`](skills/skill-decay) | Cross-references a declared skill/tool inventory (SKILL.md files or a name list) against **real usage logs** and classifies each capability as **never / stale / live** — so you prune the dead weight taxing every prompt with receipts, not vibes. Fails CI when decay candidates exceed a budget. | `context-budget` tells you how heavy a capability is; this tells you whether anyone actually uses it. A file that is both is the first to cut. |\n\n## Quick start\n\n```bash\ngit clone https://github.com/trac3r00/agent-skills\ncd agent-skills\n\n# 1) Where do my agent's tokens actually go?\npython skills/context-budget/scripts/context_budget.py \\\n    ~/.agent/system_prompt.md ~/.agent/skills ~/.agent/memory \\\n    --budget 40000 --top 15\n\n# 2) Which claims in this answer are unsupported?\necho \"The capital of Australia is Sydney. See https://example.com for details.\" \\\n  | python skills/claim-audit/scripts/claim_audit.py -\n\n# 3) What did the thread leave open for a later job to pick up?\nprintf '[me] charge the car tonight\\n[bot] set the charge later\\n' \\\n  | python skills/open-loops/scripts/open_loops.py -\n\n# 4) Which subscriptions am I paying for on repeat?\npython skills/subscription-audit/scripts/subscription_audit.py statement.csv --budget 80\n\n# 5) Which validator gates are duplicate, orphaned, or over threshold?\npython skills/gate-graph/scripts/gate_graph.py ./skills --max-gates 49 --max-overlap 0.5 --json\n```\n\n`context-budget` uses `tiktoken` when available and falls back to a calibrated\n`chars/4` estimate otherwise, so it never hard-fails on a fresh box. `claim-audit`,\n`open-loops`, and `gate-graph` are pure standard library — zero dependencies.\n\n## Use them in CI\n\nBoth tools return a non-zero exit code when a threshold is crossed, so they wire\nstraight into a pre-commit hook or GitHub Actions:\n\n```yaml\n# .github/workflows/agent-guards.yml\n- name: Context budget\n  run: python skills/context-budget/scripts/context_budget.py PATHS --budget 40000\n- name: Claim audit on golden answers\n  run: python skills/claim-audit/scripts/claim_audit.py answers/*.txt --fail-over 0.4\n- name: Open-loop handoff hygiene\n  run: python skills/open-loops/scripts/open_loops.py thread.jsonl --max-open 8\n- name: Gate overlap hygiene\n  run: python skills/gate-graph/scripts/gate_graph.py ./skills --max-gates 49 --max-overlap 0.5\n```\n\n\"The agent got fatter\" and \"the answer is mostly unsupported assertions\" now fail\nthe build instead of silently costing money or shipping hallucinations.\n\n## Install straight into Claude Code or Codex\n\nThis repo ships as a **plugin marketplace** for both agents — one line, no manual copying:\n\n```bash\n# Claude Code\nclaude plugin marketplace add trac3r00/agent-skills\nclaude plugin install agent-guards@agent-guards\n\n# Codex\ncodex plugin marketplace add trac3r00/agent-skills\ncodex plugin add agent-guards@agent-guards\n```\n\nAll five skills (`context-budget`, `claim-audit`, `open-loops`, `subscription-audit`, `gate-graph`)\nload together as the `agent-guards` plugin. In Claude Code the plugin reports its own\nalways-on token cost via `claude plugin details agent-guards` — the same\ncontext-budget discipline the skills enforce, applied to themselves.\n\n## Or load as a plain Agent Skill\n\nEach skill folder is also a standalone Agent Skill (`SKILL.md` + `scripts/`). Point\nyour agent's skills directory at `skills/`, or copy a folder into `~/.claude/skills/`,\n`~/.codex/skills/`, or your Hermes `skills/` dir.\n\n## Design notes\n\n`context-budget`, `claim-audit`, and `gate-graph` came out of running a production autonomous agent whose gate layer\ngrew to **214,588 tokens across 83 files** — a single quality gate eating 10.7% of\nthe window on every request. `context-budget` is the instrument that made that\nvisible; `claim-audit` is the reflex that keeps its answers honest; `gate-graph` is\nthe maintenance lens for duplicate and orphaned guards. The heuristics are deliberately\nsimple and transparent — you can read the whole thing in one sitting and trust what it flags.\n\n## License\n\nMIT — see [LICENSE](LICENSE).\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Ftrac3r00%2Fagent-skills","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Ftrac3r00%2Fagent-skills","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Ftrac3r00%2Fagent-skills/lists"}