# Agent Compatibility

> Audit a public-candidate repository for fresh-agent onboarding, startup, focused verification, execution continuity, recovery, and documentation reliability. Use before opening a public repository or after major tooling changes.

- Skill: `seasoncake/agent-compatibility` (Agent Skill, multi-file: 4 files)
- Install (CLI): `npx skillmds@latest add seasoncake/agent-compatibility`
- Raw SKILL.md: https://api.skillmd.com/api/skills/seasoncake/agent-compatibility/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: SeasonCake (https://skillmd.com/u/seasoncake)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/seasoncake/agent-compatibility

---


# Agent Compatibility

Determine whether a fresh coding agent can understand the repository and safely verify a
small change without hidden private context or an unnecessarily heavy loop.

## Boundaries

- Target a public-candidate tree or explicitly approved read-only simulation.
- Do not publish, push, open issues, install dependencies, mutate Git, or run unrelated
  heavy product gates.
- Missing private inputs are expected; judge whether docs provide a safe public substitute.
- This is a compatibility report, not a security certification or release verdict.

## Surfaces

Inspect the following surfaces sequentially unless the user explicitly authorizes parallel
work. Start from tracked files and bounded reads, not a recursive workspace search:

1. Inventory: tracked content, ignored residue, and intended export are distinct.
2. Orientation: README, license, contribution path, guidance, architecture, vocabulary,
   and public/private boundary.
3. Startup: one documented, non-interactive fixture path without credentials, GUI clicks,
   or machine-specific paths.
4. Validation: one focused positive/negative check that finishes quickly.
5. Docs reliability: commands, paths, versions, expected output, and failure guidance agree.
6. Continuity: a selected reversible endpoint proceeds without repeated authorization;
   runtime input is not misclassified as authority, while external/destructive stages stop.
7. Recovery: stale jobs recover from durable receipts, process absence uses exact identity,
   staged new files receive checks, and links are validated from the Git index/fresh clone.

Archived chat, screenshots, local caches, and ignored files must not be required for a fresh
agent. When execution is prohibited, statically trace dependencies -> entrypoint -> input
provenance -> expected output without claiming runtime proof.

Check both a material intent fork that should elicit a brief question and a clear selected
endpoint that should proceed without repeated questions. If a suitable native question tool
is unavailable, a plain-text fallback should work without adding a new service.

## Output

Return `READY`, `READY WITH FRICTION`, or `NOT READY`, followed by at most six top fixes.
Each fix includes evidence, affected workflow, smallest improvement, and verification.
List all heavy or external stages not run; do not hide evidence behind a single score.

For a fresh-clone and stale-job recovery example, read
[the synthetic example](references/synthetic-example.md).