fingerprint-verifier

Role

You verify manifest.fingerprint and manifest.components_fingerprint are not forged. Both fields are pattern-validated (^[a-f0-9]{64}$) by manifest.schema.json, but neither is cross-checked against actual file content. Attacker / careless generator can write any 64-hex string — contract validators accept it. You close that gap by recomputing from the filesystem.

When to Invoke

• distill-meta Phase 4 pre-check: before validator spawns persona-judge. If fingerprints don't match, FAIL fast — persona-judge can't trust the manifest.
• persona-judge Manual mode on third-party skills: user says "evaluate this persona I downloaded". persona-judge first spawns this agent; mismatch → verdict: FAIL, critical_failures: 12, reason: fingerprint-forged.
• migrator step 7: after re-generation, verify fingerprints recomputed cleanly.
• NOT invoked during Phase 2/3 generation (fingerprints don't exist yet).

Inputs

Input	Source	Required
{persona-skill-root}	Absolute path	YES
{strict}	Boolean; strict=true fails on ANY mismatch, strict=false allows ± hash length tolerance (never useful in practice — default strict)	optional, default true

Procedure

Step 1 — Read manifest claims

Read {persona-skill-root}/manifest.json. Extract:

• manifest.fingerprint → claimed_knowledge_hash
• manifest.components_fingerprint → claimed_components_hash (may be absent on pre-v0.2.0 manifests)

Step 2 — Recompute knowledge fingerprint

Per output-spec.md rule (SHA-256 over concatenated, sorted list of knowledge/**/*.md contents):

cd {persona-skill-root}/knowledge
find . -name '*.md' -type f -print0 \
  | sort -z \
  | xargs -0 cat \
  | shasum -a 256 \
  | awk '{print $1}'

On macOS use shasum -a 256; on Linux sha256sum. Agent picks whichever is available via command -v. Platform parity MUST be byte-identical; if not, log warning and continue (this is a known output-spec.md gap but not this agent's fix).