TRM Feedback/Report

Standing feedback pass for a TRM ingest batch. Run after trm ingest-dir completes.

Trigger

After any TRM ingest batch, or on request ("feedback on the last ingest batch", "report on <topic>").

Flow

1. Run runTrmCommand(trmRoot, ['validate', topicPath, '--recursive']) and runTrmCommand(trmRoot, ['feedback-stats', topicPath, '--recursive', '--latency-budget-ms', String(latencyBudgetMs)]). Parse both as JSON.
2. Read extracts/extract.json facts for the topic (recursively, matching --recursive) and call findNewTopicCandidates(facts, existingTagsAndPathSegments), where existingTagsAndPathSegments is collected by walking topic.json from the vault root.
3. Select facts with confidence < 0.55 OR flags?.includes('VERIFY'), capped at the top 10 by score.json's promotion_score. For each, call the WebSearch tool, classify each hit's consistency with the fact's claim, then call deriveSignalStrength(hits).
4. Write the narrative report: classifier accuracy verdict, extraction gaps (using fact_density/over_budget/warnings_count_by_type), three required risk labels (extraction risk / classifier drift risk / latency risk, each low/medium/high), new-topic candidates from step 2, cross-check results from step 3. Call buildReportMeta(stats.completeness, hardFailure, webSearchSkipped) for the partial_report/latency_data_stale flags and reportFilename(topicPath) for the filename.
5. Write the file to <trmRoot>/reports/<filename>.
6. Call runTrmCommand(trmRoot, ['crosslink', topicPath, '--tags', 'trm-feedback-report:v1']).

Input Schema

interface Input {
  topicPath: string;
  trmRoot: string;
  recursive?: boolean;
  latencyBudgetMs?: number;
}

Output Schema

interface Output {
  reportPath: string;
  partial_report: boolean;
  latency_data_stale: boolean;
}

---

Full reference: See Skill Operator Guide.

For detailed workflow: See docs/USAGE.md.