Agent Autopsy

Agent Autopsy is a versioned archive of model-facing prompts, tool schemas, and steering surfaces from coding-agent CLIs. It shows what each product sends to a model, which release supports each artifact, and what is still unknown.

Start with the cross-provider catalog. It lists every known surface for Codex, Claude Code, Qwen Code, Grok, Antigravity, and Amp, including explicit gaps and dynamic inputs.

Providers

How to read the archive

• prompts/ contains system prompts or source-derived prompt templates.
• tools/ contains exact observed tool schemas and model or mode variants.
• misc/ contains steering and other model-facing artifacts.
• VERSION records provider-level capture facts.
• SURFACES.json records per-surface coverage, capture release, status, and hashes.

Runtime values use <harnessVariable>{{name=example}}</harnessVariable>. A current surface was captured at the observed release. verified-unchanged was checked at that release. stale and frozen identify older artifacts. gap means that a known surface is absent. dynamic identifies a typed runtime input that must not be mistaken for fixed product text.

Trust and privacy

Raw requests are private evidence and can be used by the analysis model. They are not repository artifacts. The author agent semantically removes actual PII and secrets. Model-facing machine, repository, tenant, staging, request, and user context can remain when it contains no actual PII. Synthetic placeholder paths such as /Users/example are safe. An independent reviewer agent must attest that PII was removed before publication.

Run the local integrity checks with:

node .github/scripts/validate-surfaces.cjs
node .github/scripts/generate-catalog.cjs --check

See artifact format and status rules, automation, capture credentials, and contributing.