| name | adev-read-only-inspection |
| description | Inspect the A-Dev canon and produce a bounded, evidence-aware read-only report. Use when an agent must orient to A-Dev doctrine, Hardness policies, skills, or maturity without modifying files, Git state, external systems, or communications. |
A-Dev Read-Only Inspection
Purpose
Inspect the relevant A-Dev canon and report the requested state, applicable constraints, uncertainty, and evidence boundary without causing effects.
Triggers and non-triggers
Trigger for requests to summarize, inspect, compare, audit, classify, or locate A-Dev framework material when the requested work is read-only.
Do not trigger for edits, commits, branches, pull requests, merges, messages, deployments, external queries, or any task requiring a local or remote mutation. Escalate to a skill whose contract explicitly permits that effect.
Inputs and outputs
- Inputs: a user objective; optional paths, concepts, or acceptance criteria; the repository canon available for inspection.
- Outputs: a concise report with inspected sources, findings, applicable policies, validation performed, uncertainty, and next safe action.
- Out of scope: implementation, authorization decisions for effects, operational verification outside locally available artifacts, and claims not supported by inspected sources.
Authority, permissions, and effects
- Applicable policies: resolve using policy precedence and the human expectations contract.
- Permissions: read tracked repository artifacts needed for the scoped objective.
- Side effects: none. Do not write files, alter Git state, invoke remote mutations, send messages, or expose sensitive material.
- Risk class: R0 — read-only.
Invariants
- Preserve local and remote state.
- Keep findings separate from inference, proposals, and gaps.
- Treat
ADEV.md as doctrine and use the narrower canonical shelf when it exists.
- Do not treat a fixture, example, book passage, or repository-root link as proof of an operational result.
- For any Homedir-derived operational claim, require a direct artifact link under the canonical Homedir repository, or label the claim unverified.
Required sequence
- Capture intent, expected result, constraints, non-goals, acceptance criteria, and uncertainty from the request.
- Check that the requested effect is read-only. If it is not, stop before the effect.
- Read the agent consumption guide and the narrowest canonical sources for the requested concept.
- Identify applicable policy by authority, scope, priority, and the conflict procedure; if a material conflict remains, stop and escalate.
- Inspect only the artifacts needed to answer the objective.
- Validate the report against the stated acceptance criteria and record remaining uncertainty.
Uncertainty, stop, and escalation
- Uncertainty: mark missing source, ambiguous scope, unavailable artifact, or unsupported conclusion as a gap; do not fill it with assumption.
- Stop: stop when the request needs an effect beyond R0, source access is denied, sensitive data would be exposed, or validation cannot support the requested conclusion.
- Escalate: report the blocked action, applicable policy or contract boundary, missing decision or authorization, and the smallest safe next action.
Evidence and validation
- Evidence record: source paths or direct artifact links, policy IDs or named policy sources, commands/checks run, and a concise result. Keep the record privacy-safe.
- Acceptance criteria: all findings trace to inspected sources; effects remain none; uncertainty and evidence gaps are explicit; required local links resolve.
- Validation: run the repository's narrowest relevant read-only checks, such as path, link, or structural review. Do not claim behavioral conformance from this skill alone.
See the synthetic evaluation suite for expected decisions and safe failures.
