ralphi

ralphi — Inspection

Step 1: Mode Determination

Utterance contains "고쳐", "수정해", "fix", "처리해" → Autonomous (proceeds to fix) Otherwise → HITL (report only and wait)

Step 2: Scope Determination

• File path given → File scope → Step 3A
• Directory given → Directory scope → Step 3B
• Nothing given → Ask the user: "Provide a file path or directory path. (everything / all are not valid scopes — pick a concrete target.)" No guessing. Do not run git diff on your own.

Step 3A: File Inspection

1. Detect type → determine criteria file:
   • docs/prd-*.md (check §section structure) → criteria-prd.md
   • */SKILL.md (check frontmatter name:) → criteria-skill.md
   • *.sh (check shebang #!/) → criteria-script.md
   • .py .ts .js .go etc. source code → criteria-code.md
   • Unclear → Ask the user: "Detected file {path} — what type? (prd / skill / script / code / other?)" No guessing.
2. Load only 1 criteria from references/. Do not read other criteria.
3. Static analysis (structure, format, contract violations)
4. Dynamic execution (if script/code)
5. Necessity check (Socratic) — see method below.
6. → Proceed to Step 4

Step 3B: Project/Directory Inspection

1. Run tests. Tests before reading code. If test runner is unknown, ask the user: "What is the test command? (e.g. pytest, npm test, go test ./..., cargo test, custom?)" No guessing. Tool not installed (command not found) → SKIP that check + warn. Do not attempt installation.
2. List changed files (git diff)
3. Analyze only the diff of each file. Do not read entire files.
4. Scenario walkthrough (intent vs implementation). Read only the relevant function. No call graph tracing.
5. Coverage gap exploration
6. Necessity check (Socratic) — see method below. Apply per diff hunk instead of per component.
7. → Proceed to Step 4

Necessity Check (Socratic) — shared method

Used by Step 3A.5 (per component) and Step 3B.6 (per diff hunk).

Ask the four questions:

1. Why is this here? — (purpose)
2. What is it composed of? — (composition, scaffold for Q1/Q4)
3. What is it really, stripped of framing? — (truth, scaffold for Q1/Q4)
4. What happens if it's removed? — (necessity)

Verdict: Q1 and Q4 are the only gates. Q2 and Q3 are scaffolding to reach honest answers. Items that fail Q1 or Q4 → [warning] unjustified existence (kept only by convention / inheritance / authority).

Step 4: Mode Branch

HITL Case

Report in Step 5 format → wait for user judgment.

User Response Interpretation Rules

Clear instructions (execute immediately):

• "Fix #1" → fix that issue only → test → report.
• "Fix 1, 3" → fix those issues sequentially, test after each.
• "Fix all" → fix all sequentially, test after each.
• "Ignore" / "Skip" → record the issue and end.

Ambiguous responses (must re-ask):

• "Yeah", "Got it", "Sure", "OK" → Do not interpret as permission to fix. → Re-ask: "Which issues should I fix? Specify numbers or say 'fix all'."
• Response unrelated to issues → re-ask.

Autonomous Case

Fix immediately in critical→warning→coverage order → test after each fix → report results in Step 5 format when all done.

• Test FAIL → rollback that fix, unfixed "test failure"
• Intent unclear (code deletion, logic change) → unfixed "intent unclear"
• Structural change needed (file move, interface change) → unfixed "structural change needed"

Fix prefers subtraction. Default action for [unjustified existence] is removal. For [critical/warning] violations, minimal additive patches are allowed (null check, missing import, type fix, error guard) when needed to satisfy the criteria. Do not introduce new abstractions, modules, or features. If a fix would require new abstraction/module/feature, mark as [unfixed] structural change required.

If there are unfixed items, hand off to user judgment.

Step 5: Report

Severity: critical (behavioral error) / warning (potential issue) / coverage (test gap)

HITL:

## ralphi Inspection Results
Target: {path | "user-specified scope"}
### Findings ({N} items)
1. [{severity}] {file:line} — {one-line cause}
Which issues should I fix first?

Autonomous:

## ralphi Inspection + Fix Results
Target: {path | "user-specified scope"}
### Fixed ({M}/{N} items)
1. [fixed] {file:line} — {one-line fix description}
### Unfixed ({K} items) ← only when present
1. [{severity}] {file:line} — {cause} → {reason unfixed}
Tests: {PASS | FAIL (details)}

No issues found → ralphi inspection complete, no issues found. Single line. No enumeration.

Context Budget

When	What is read
Step 3A (file inspection)	Exactly 1 criteria file from references/, matched to detected type. No others.
Step 3B (project inspection)	git diff only. No full file reads. Read only the function relevant to a scenario walkthrough.
Necessity check	Operates on already-loaded content. No new file reads.
Step 5 (report)	No file reads. Output only.

Prohibited

• Reading unchanged files
• Loading 2 or more criteria simultaneously
• Fixing without instructions in HITL mode
• Reading entire files in project scope (diff only)
• "FYI..." style supplementary information
• Verbose reports