SuperAPI AI Directory

ra1-report

Coding & development

View repository

Updated 10-16 07:15

About this skill

Agent Readiness Report

Produce a trustworthy readiness report for a repository. The deterministic engine owns the score; you (the agent) add advisory commentary only. You must never change the engine's score, reason codes, evidence, or limitations, and you must never invent any of them.

Steps

<skill-dir> below is this skill's own directory: the absolute path your runtime reports when it loads the skill. Substitute it. Do not write $(dirname "$0") — in a shell tool call $0 is the shell, not this file.

  1. Run the engine (it does all the deterministic work — file/config parsing, git history, and, when requested, the GitHub.com API). The repository must have .ra1/reports/ ignored in its root .gitignore for in-repo persistence; otherwise the engine prints the full report in memory and reports the exact isolation error, which you present as storage unavailable while the assessment itself is complete.
python3 -I "<skill-dir>/scripts/readiness/cli.py" report --project . \
  --format json,markdown,html --out .ra1/reports --store-history

JSON is first so stdout is the canonical machine payload; Markdown/HTML/history are safe engine-written artifacts. T2 GitHub checks are offline by default; append the fixed --github only after an explicit current-turn request for GitHub/T2, and paired --host-proxy only after a separate current-turn host-proxy request. Never append --exec, --exec-timeout, --require-origin, shell composition, or a repository-derived option. A scan failure, truncation, or invalid JSON fails closed: never reconstruct a score from partial output.

  1. Check for unanswered questions. If the engine's gaps array is non-empty, some unknown results are stuck on an input the scan could not infer, not on the repository. Report the count and the gating criteria they hold back, and hand off to the ra1-interview skill. Never answer them on the developer's behalf and never treat an unanswered gap as a finding.

  2. Copy the score verbatim. Your final report MUST contain a fenced ```json block holding the engine's score object exactly — every key: level, level_name, pass_rate, gating_passed, gating_total, levels, pillars, recommendations, max_available_level, next_gate_actions, evidence_coverage. Full parsed-JSON deep equality; do not change a number.

  3. Add ## Evidence explanations (BEFORE ## T4 Advisory). One fenced JSON object with this exact shape, covering every id in score.next_gate_actions in that order:

    {
      "explanations": [
        {
          "id": "criterion.id",
          "status": "fail",
          "reason_code": "criterion.id.missing",
          "rule_ref": "checks.module.function",
          "evidence_sources": ["path-or-endpoint"],
          "limitations": ["deterministic limitation"]
        }
      ]
    }
    

    Values are copied from the matching result and its decision_trace; evidence_sources keeps first-seen non-empty sources, deduplicated. When next_gate_actions is empty, emit {"explanations": []}. Never author a new source, reason code, status, rule reference, limitation, or score. Then, in human prose, explain rule → observations → evaluation → limitation → next action in repository-maintainer language, citing criterion ids and sources, and label any T4 interpretation as advisory.

  4. Add ## T4 Advisory (qualitative, non-gating). The engine deliberately leaves these soft judgments to you; label each as advisory opinion grounded strictly in engine findings and files you actually read. Use one labelled sub-heading per registered judgment id (naming_consistency, code_modularization, n_plus_one_query, readme_quality, agents_md_quality, service_flow_doc_quality, runbooks_quality, pii_handling, privacy_compliance, user_feedback_loop), group AC/DC maturity by the registry-provided acdc_stage/acdc_loop fields, cite the specific file/finding, explain why it matters, and give the highest-leverage next step. Keep every T4 claim clearly separate from the deterministic score.

Contract (do not violate)

  • Never claim a higher Level than the engine reports. Level 5 (Autonomous) is reserved and never reported; the ceiling is stated as L4.
  • Never mark a failing or unknown criterion as passing. Prose must agree with the engine's statuses; describe unavailable/unknown language honestly ("not verified", never "not protected").
  • Never invent criteria, evidence, sources, reason codes, or results — pasting the score and explanation payloads verbatim is the machine contract; the human prose summarizes, never extends, them.
  • Never claim autonomy clearance or unattended operation in any form.
  • Do not assert that a specific criterion is "gating" or "non-gating" — only the engine's data says so.
  • T4 advisory is opinion and is explicitly non-gating — it cannot move the Level, GitHub annotations, JUnit, or SARIF.

Notes

  • If the engine reports project_type: unknown, surface that honestly — type-dependent criteria are unknown, not silently skipped. Point at the ra1-interview skill to pin a type (never edit .ra1/config.json directly).
  • T2 criteria are skipped when --github was not requested or the source is unavailable; say so and offer --github for a fuller scan.
  • To raise the score, hand off to the ra1-fix skill (the Loadout).