Project Probe

Project Probe v2 — Verification Manifest

Every workflow reads one provider-neutral manifest: .agents/verification.yaml. The probe is read-only; it inventories configuration and never runs application scripts while collecting.

Runtime contract

The collector and Python selector are resources owned by this skill. Resolve the semantic locators below through the active runtime's skill registry to absolute paths before invoking the shell; do not interpret skill: as a literal filesystem path and do not look for these files in the project:

bash "skill:project-probe/scripts/run-python310.sh" "skill:project-probe/scripts/project_probe.py" --root . --check

If either resource cannot be resolved or executed, classify the result as tooling-unavailable and recommend repairing/updating the Skillz installation. This is not project evidence and never justifies a quality waiver.

Python 3.10+ is required. A generated manifest records the selected interpreter and version. If the selector cannot find a compatible runtime it fails before probing with the detected version.

Freshness

1. Run the command above.
2. A zero exit means config_fingerprint is fresh: read the manifest and stop.
3. If it is stale or absent, inspect .agents/project-probe.json when present, then regenerate:

bash "skill:project-probe/scripts/run-python310.sh" "skill:project-probe/scripts/project_probe.py" --root .

fingerprint_schema_version: 2 hashes each repository-relative path and its bytes, sorted deterministically. In Git repositories only indexed files participate, so local user-owned files do not become project truth. The versioned collector covers shell scripts, SKILL.md and command workflows, provider manifests, CI workflows, lockfiles, workspace markers and consumed config. The manifest itself is excluded to avoid a self-invalidating hash.

Command detection

Collect signals in this order:

1. .agents/project-probe.json explicit overrides, if any;
2. package scripts and Python tool configuration;
3. allowlisted shell roles (lint, check, test, verify, build);
4. CI evidence inspected by the agent when it strengthens or contradicts a candidate.

Only lint, typecheck, test and build are verification commands. Reject a candidate whose name or body can deploy, publish, reset a database, destroy state, download remote content or remove files. Never execute clean, seed, migrate, release, deploy or a user-defined near-match as a sanity check. A safe --help, --version or documented dry-run may be used after collection; if it cannot be confirmed, omit the command and record the reason.

Each accepted command remains a string under commands for v1 consumer compatibility. Its exact origin is mandatory under command_sources. Missing capabilities appear both as human-readable absents and keyed absence_reasons; never invent a fallback merely because a stack usually has one.

Manifest contract

fingerprint_schema_version: 2
generated_by: "project-probe/2.0.0"
stack: "mixed"
config_fingerprint: "<sha256>"
fingerprint_sources:
  - ".agents/project-probe.json"
python:
  required: ">=3.10"
  selected_interpreter: "python3.12"
  selected_version: "3.13.7"
commands:
  lint: "bash -n scripts/*.sh"
command_sources:
  lint: ".agents/project-probe.json#commands.lint"
testability:
  harness: "shell+unittest"
  e2e: none
  runtime_verify: none
absents:
  - "no configured typecheck command"
absence_reasons:
  typecheck: "no configured typecheck command"

Optional monorepo metadata records the package manager and workspace globs. The override file may declare commands, stack, Python requirement, testability and absence reasons, but it cannot bypass the destructive-command filter.

Consumer rules

• Run verification only from commands; provenance metadata is evidence, not an executable field.
• Every commands entry must be auto-executable by an agent. Human ceremonies such as a live preproduction rehearsal, manual UAT, Access/OTP/TOTP or a bounded waiver belong in absents with their trigger condition. Rewrite any manifest that incorrectly lists one as a command.
• Report every relevant absence in checkpoints and gate files.
• If --check reports drift or a command fails because its tool disappeared, re-probe before drawing a conclusion.
• Never list a human-gated ceremony in commands; the ship runner executes every command literally.
• Do not create provider-specific mirrors of the manifest.
• Do not write outside the requested manifest path.