Tracemantle

TraceMantle is a local static analyzer for SKILL.md files. It validates frontmatter structure, scores advisory description features, checks file references, and extracts a capability graph from heading and backtick-reference patterns. Run it before committing a skill, plugging it into CI, or publishing to a skill registry.

Prerequisites

• path - the SKILL.md file or directory to validate
• prompt.txt - file path for a saved agent prompt (critique or graph extraction)
• response.json - file path for an agent-provided JSON response

Validate a skill

Pass path to tracemantle for a full structural check against the agentskills.io specification. Validates name constraints, description quality, body sizing, file references, and cross-agent compatibility. The validation report is written to stdout. Exit codes: 0 no errors, 1 errors found, 2 input error (missing file), 3 symbolic passed but semantic critique flagged errors.

tracemantle <path>
tracemantle <path> --format json

Review a skill with an agent

tracemantle generates a structured critique prompt for path and ingests response.json as the agent's reply. Save the emitted prompt to prompt.txt. The merged validation report combines symbolic diagnostics and semantic findings from the critique.

tracemantle <path> --emit-critique-prompt > prompt.txt

Hand prompt.txt to your agent. The agent returns JSON. Save the response as response.json, then ingest it.

tracemantle <path> --ingest-critique response.json

Pass --critique-agent codex or --critique-agent cursor to select a prompt variant tuned for that platform. The critique JSON schema is identical across all agent targets.

Extract a capability graph

tracemantle extracts the capability graph from heading structure and backtick references in the body of path. Use prompt.txt for emitted graph prompts and ingest response.json from your agent for richer extraction.

Run graph analysis using the heuristic extractor (no agent needed):

tracemantle <path> --analyze-graph

For richer extraction, emit the graph prompt, run it through your agent, and ingest the response:

tracemantle <path> --emit-graph-prompt > prompt.txt
tracemantle <path> --ingest-graph response.json

Print the graph to stdout:

tracemantle <path> --emit-graph
tracemantle <path> --emit-graph --format json

When both heuristic and agent graphs are available, tracemantle compares them and emits graph.contradiction.heuristic_disagreement at error severity for agent-claimed edges that the heuristic does not confirm.

Check validation history

Append a record to the validation ledger for each --history run on path. The validation history uses immutable per-run records outside the skill bundle, under its parent evidence directory. Legacy history is read without being moved or deleted.

tracemantle <path> --history
tracemantle <path> --show-history
tracemantle <path> --show-history --format json

Each record stores a timestamp, TraceMantle version, bundle/configuration digests, validation modes and per-file results. Comparable prior passing records can produce a history regression warning. Legacy text hashes alone do not establish revision equivalence.

Compare release evidence

Use path with the manifest command to identify skill content and all packaged resources. Import evaluator output into a selected evidence store outside the bundle. Compare baseline and candidate bundles with policy and approved evidence from a trusted base commit. Missing evidence blocks release as unknown; imported model judgments are not proof of observed skill invocation. TraceMantle does not run agents or imported checkers.

tracemantle manifest <bundle> --format json
tracemantle compare <baseline> <candidate> --trusted-root <trusted-repo> --base-revision <full-commit-sha> --format json

Outputs

• validation report - diagnostic output written to stdout
• capability graph - graph of capabilities, inputs, and outputs extracted from the skill body
• validation ledger - immutable validation records outside the evaluated bundle