Method writeRunReceipt

src/agent/loop.ts:876–901

Write an audit receipt to QODEX_RECEIPT_FILE (set by the scheduler) from ground-truth * signals: the git diff + the verify ledger. Status/PR come from the model's headline (a * checkable claim); the FACTS (files, checks) are what QodeX measured. Best-effort.

(opts: { reachedFinal: boolean; merged: boolean; finalContent: string; changedFiles: string[]; signal?: AbortSignal })

              );
              await writeCandidate(candidate);
              capturedThisRun = true;
              await recordLearningEvent({ event: 'capture', name: candidate.name, confidence });
              // Code-graph grounding: how much of this skill references symbols that really exist here.
              let fitSuffix = '';
              try {
                const { getCodeGraphDB } = await import('../codegraph/tools.js');
                const db = getCodeGraphDB();
                if (db) {
                  const { extractSymbolHints, codebaseFitScore } = await import('../skills/learning/codebase-fit.js');
                  const fit = codebaseFitScore(extractSymbolHints(candidate.skillMd), n => db.findSymbolsByName(n, undefined, 1).length > 0);
                  if (!fit.noSignal) fitSuffix = ` · codebase-fit ${Math.round(fit.score * 100)}%`;
                }
              } catch { /* code graph optional */ }
              const draftSuffix = draft ? ` · ${draft.steps.length}-step draft` : '';
              yield { type: 'notice', data: { message: `🎓 Captured candidate skill "${candidate.name}" (confidence ${confidence}/100${fitSuffix}${draftSuffix}) — review with \`qodex skill candidates\`, promote with \`qodex skill promote ${candidate.name}\`.` } };

              // Auto-Evaluation (opt-in): immediately replay the captured skill in a clean
              // worktree and record whether it produces verified code. Costs a model call +
              // worktree, so it's behind learning.autoEval. Best-effort — the task already
              // succeeded; an eval failure here never affects it.
              if (learningCfg.autoEval) {
                try {
                  const { evalSkillMd } = await import('../skills/learning/eval.js');
                  const { formatEvalSection, upsertEvalSection, skillContentHash } = await import('../skills/learning/eval-record.js');
                  const { candidatesDir } = await import('../skills/learning/candidate-store.js');
                  const { result } = await evalSkillMd(this.cwd, candidate.skillMd, { noCache: true });
                  if (result) {
                    const fsmod = await import('fs');
                    const pathmod = await import('path');