Data-only workload profiles

Workload Profiles v1 reduces repeated typing without becoming a dispatcher or policy engine. The canonical command is bundled with the public skill; the repository-root profile.sh is only a compatibility wrapper.

./profile.sh list
./profile.sh show repository-inventory
./profile.sh show diff-review
./profile.sh show bounded-test-backfill

Both commands write one canonical JSON value to stdout. list returns only the fixed maintained names and summaries. show returns one non-executable skeleton that may suggest a maintained plan/accept-edits mode, a maintained persona, and one closed repo-relative path-policy shape.

Every shown profile says that these inputs remain caller-owned:

• explicit approval;
• the exact repository;
• the exact repo-relative path policy;
• the selected tier; and
• executable verification commands.

Profiles never contain a repository or filesystem path, model, tier, effort or thinking value, verifier or shell command, external add-dir, authorization, routing, acceptance, dispatch, or Git action. They do not read a target repository, scan the home directory, inspect environment variables, invoke git or agy, access a provider or network, or accept a caller-selected profile file. A shown profile is therefore not permission to dispatch and cannot satisfy the gate or human diff review.

The v1 manifest binds the exact three canonical data files by SHA-256. The portable runtime rejects missing, extra, reordered, changed, executable, writable, symlinked, oversized, or non-canonical profile artifacts and schema drift. Adding or changing a profile is a versioned source change that must update the semantic allowlist, schema, manifest bindings, paired accept/reject tests, and documentation together.