evenbetter-validate

evenbetter-validate

Overview

Validate high-severity EvenBetter iOS analyzer findings with a second-pass evidence check. Treat this as the validator half of the orchestrator-workers-with-validators pattern: analyzer workers propose findings, then this skill rechecks source evidence and corpus support before keeping a high-severity violation.

Inputs

• projectPath (required): Absolute path to the analyzed SwiftUI project.
• confidence_threshold (optional): Float from 0.0 to 1.0. Default to 0.7.
• run (optional): Analyzer run number to validate. Default is the latest unvalidated analyzer run in .evenbetter/manifest.json.
• revalidate (optional): Boolean. Default false; when true, allow replacing evenbetter-validate-{N}.json for an already validated run.

If projectPath is missing or not absolute, return a JSON error object with an error key and stop.

Source Safety

Do not edit, delete, format, generate, or execute source/project files inside projectPath. The only permitted writes inside projectPath are creating .evenbetter/ if needed, writing the final validation report to .evenbetter/evenbetter-validate-{N}.json, updating .evenbetter/manifest.json, and setting run.status in .evenbetter/analyze-{N}.json to validated.

Required References

Load only what the current phase needs:

• references/workflow.md: Validation loop, evidence checks, confidence decisions, and compaction-safe invariants.
• references/output-contract.md: Stable JSON envelope and finding result fields.
• references/architecture.md: Orchestrator-worker-with-validators decomposition and citation note.
• scripts/verify_url.py: Deterministic URL verifier for Apple and W3C sources.

Validation Scope

Read projectPath/.evenbetter/manifest.json and select the analyzer run to validate. By default, validate only the newest run whose manifest entry has validated: false or no paired validate file. If run is provided, validate that analyzer run; if it already has a paired validation report, require revalidate: true.

Read projectPath/.evenbetter/analyze-{N}.json for the selected run. Validate only violations where severity is error and state.status is not fixed, rejected, or duplicate_of; this is the current analyzer's high-severity class.

For each high-severity finding:

1. Re-read the cited Swift file and nearby lines.
2. Resolve the corpus clause for rule_id through ../../corpus/index.json, then read the matching H2 section from the indexed corpus markdown file.
3. Verify guideline_reference.url with scripts/verify_url.py.
4. Independently judge whether the cited code violates the rule.
5. Emit confidence, reasoning, and decision.

When the host supports isolated subagents, run the judgment in a fresh validator context that receives only the finding, source excerpt, corpus clause, and URL result. If isolated subagents are unavailable or not permitted, still perform an explicit independent re-evaluation from those artifacts and do not reuse the original auditor's reasoning as evidence.

Output Rules

• Write one JSON object to projectPath/.evenbetter/evenbetter-validate-{N}.json.
• Emit the same JSON object on stdout when running headless.
• Use kept, downgraded, and dropped arrays exactly as defined in references/output-contract.md.
• Keep only findings with confidence >= confidence_threshold, valid source evidence, resolved corpus clause, coherent reasoning, and a verified source URL.
• Include drop_reason for every dropped finding.
• Include validates: "analyze-{N}.json" and analyzer_run: N.
• Update .evenbetter/manifest.json so run N has validate: "evenbetter-validate-{N}.json" and validated: true.
• Do not wrap JSON in Markdown fences or add prose to JSON-only outputs.

Load references/workflow.md next.