Repo Agent Context Audit

Repo Agent Context Audit

Overview

Assess whether a repository has a small, usable agent context stack: a short top-level instruction file, task-specific skills, and behavior/implementation specs for substantial work. Default to a read-only audit and minimal recommendations; create or edit high-context files only when the user explicitly asks.

Default Standard

Prefer this three-layer shape:

1. AGENTS.md or repo-equivalent: 80-150 lines, top-level routing only.
2. .agents/skills/<task>/SKILL.md: reusable workflows for common fragile tasks.
3. specs/<id>/PRODUCT.md and specs/<id>/TECH.md: checked-in contracts for substantial features.

Do not force this exact layout when a repo already has a coherent equivalent, such as WARP.md, CLAUDE.md, CONTRIBUTING.md, or framework-specific instruction files. Map existing files to the layers first, then fill only the real gaps.

Workflow

1. Discover

Run the read-only scanner when possible:

# From this skill directory:
python3 scripts/scan_repo_context.py <repo-root>

Then inspect the important files directly. Always search before creating:

• AGENTS.md, CLAUDE.md, WARP.md, .claude/instructions.md
• CONTRIBUTING.md, README.md, .github/copilot-instructions.md
• .agents/skills/*/SKILL.md
• specs/**/{PRODUCT,product,TECH,tech}.md

If multiple instruction files overlap, record their scopes and precedence instead of merging them by default.

2. Classify

Classify the repo into one of these states:

• Healthy: short top-level guidance, task workflows, and specs are discoverable.
• Missing top-level router: useful docs exist but agents lack an entrypoint.
• Overloaded top-level file: one file mixes rules, workflows, architecture, and reference data.
• Specless complex repo: substantial work happens without PRODUCT/TECH contracts.
• Stale or divergent: instructions contradict code, scripts, or observed repo conventions.
• Unsafe to modify: high-context files are generated, externally owned, or conflict across scopes.

3. Score

Use the rubric in references/standards.md for:

• top-level routing quality
• progressive disclosure
• procedural workflows
• decision gates
• production examples
• spec quality
• validation mapping
• stale or conflicting guidance risk

4. Recommend

Lead with the smallest useful change. Good recommendations usually look like:

• Add a short AGENTS.md that points to existing docs instead of duplicating them.
• Split a long top-level file into a router plus references or skills.
• Add .agents/skills/write-product-spec and .agents/skills/write-tech-spec only if spec writing is repeated.
• Add specs/<id>/PRODUCT.md and TECH.md templates only if the repo ships substantial features.
• Remove or rewrite stale instructions only after citing the conflict.

5. Scaffold Only On Request

When the user explicitly asks to create files, read references/templates.md and adapt the templates to the repo. Before editing:

• identify every AGENTS.md or equivalent whose scope covers the target path
• preserve existing high-context files unless the user asked for a rewrite
• keep generated docs small
• include actual repo commands, paths, and test gates
• leave placeholders only when the repo truly lacks the fact

Decision Gates

Operating Contract

Direct actions:

• Run read-only discovery, scoring, and scanner commands.
• Produce an audit report with cited files and smallest useful changes.
• Draft exact scaffold content when the user asks for proposed text.

Escalate before:

• Creating or editing AGENTS.md, CLAUDE.md, WARP.md, hooks, settings, or generated docs.
• Rewriting existing repo instructions instead of adding a short router or pointer.
• Batch-normalizing multiple repositories.

Evidence-backed pushback:

• Challenge new skill/spec scaffolding when the repo is small, has no repeated workflow, or already has a coherent equivalent.
• Challenge edits when the only evidence is style preference rather than a real agent failure, review bottleneck, or missing workflow.

Feedback loop:

• Promote repeated audit findings into the target repo's router, a task skill, or a spec template.
• Keep this skill's rubric and templates updated when multiple repos expose the same false positive or missing decision gate.

Gotchas

• Do not treat every README as an agent instruction file. Top-level README files are usually entrypoints; nested README files are supporting evidence unless a repo explicitly routes agents there.
• Do not add AGENTS.md just because it is missing. If CLAUDE.md, WARP.md, or CONTRIBUTING.md already works as a coherent router, recommend a pointer or no change.
• Do not create PRODUCT/TECH specs for small bugfix repos. Specs are for ambiguity, cross-module risk, and repeated review churn.
• Do not silently normalize naming across repos. First report drift such as PRODUCT.md vs product.md, then ask before changing conventions.
• Do not edit high-context files during an audit. Scaffold only after the user asks to create or apply files.

Report Format

Return concise findings:

## Agent Context Audit
- state: <classification>
- top-level router: <present/missing/overloaded>
- reusable skills: <present/missing/not needed>
- specs: <present/missing/inconsistent/not needed>
- main risk: <one sentence>

## Smallest Useful Change
1. <change> - <why> - <estimated effort>

## Evidence
- <file>:<line> - <what it proves>

## Optional Scaffold
- <files to create or update, only if requested>

Resources

• scripts/scan_repo_context.py: read-only repo scanner for high-context files, skills, and specs.
• references/standards.md: scoring rubric and design rules. Read for audits.
• references/templates.md: minimal AGENTS.md, PRODUCT.md, and TECH.md templates. Read only when scaffolding or proposing exact file contents.