UI ops slice: retry & guardrail depth, audit rejections, July event vocabulary in the feed

src/observability/dashboard/state/feed.js

@@ -30,17 +30,26 @@ export function buildActivityFeed(runs) {
 
     for (const ev of run.events ?? []) {
       if (skipInFeed.has(ev.type)) continue;
-      const summary = plainLanguageSummary(ev);
+      // July harness vocabulary (#215) gets first-class summaries here;
+      // everything else keeps the shared plain-language path. Unknown event
+      // types still degrade exactly as before: no summary, no feed line.
+      const summary = opsEventSummary(ev) ?? plainLanguageSummary(ev);
       if (!summary) continue;
-      activityFeed.push({
+      const item = {
         ts: ev.ts,
         summary,
         type: ev.type,
         runId: run.runId,
         projectRoot: run.projectRoot,
         goal: run.manifest?.goal,
         level: eventLevel(ev),
-      });
+      };
+      // Structured detail for the client panels (retry state, guardrail
+      // depth, audit rejections) — only real fields from the event, never
+      // fabricated defaults.
+      const data = opsEventData(ev);
+      if (data) item.data = data;
+      activityFeed.push(item);
     }
   }
 
@@ -49,6 +58,86 @@ export function buildActivityFeed(runs) {
     .slice(0, 200);
 }
 
+// Plain-English lines for the loop-engineering signals shipped 2026-07
+// (#156 remainder + #215): checkpoints, context pressure, approval audit
+// rejections, memory write decisions, metrics drift, retry decisions and
+// goal-loop evaluations. Returns null for anything it does not own.
+function opsEventSummary(ev) {
+  switch (ev.type) {
+    case 'stage_checkpoint_before_saved':
+      return checkpointSummary('before', ev);
+    case 'stage_checkpoint_after_saved':
+      return checkpointSummary('after', ev);
+    case 'context_pressure_warning': {
+      const size = fmtCount(ev.size);
+      const threshold = fmtCount(ev.threshold);
+      const metric = ev.metric ?? 'chars';
+      return `Context pressure — ${ev.source ?? 'context'} at ${size} ${metric} (threshold ${threshold}); warning only, nothing was pruned`;
+    }
+    case 'approval_audit_failed':
+      return `Approval record rejected by audit — ${ev.artifact ?? 'unknown artifact'} treated as absent; the gate stayed closed`;
+    case 'episode_memory_written':
+      return `Episode memory written for ${ev.task_id ?? 'run'}${ev.trusted === true ? ' (trusted)' : ev.trusted === false ? ' (stored untrusted — never injected into prompts)' : ''}`;
+    case 'episode_memory_skipped_untrusted':
+      return `Episode memory skipped for ${ev.task_id ?? 'task'} — ${ev.reason ?? 'write policy refused the episode'}${ev.write_policy ? ` (policy: ${ev.write_policy})` : ''}`;
+    case 'metrics_write_failed':
+      return `Metrics write failed${ev.operation ? ` (${ev.operation})` : ''} — persisted totals are behind the events; run totals fall back to event recompute`;
+    case 'retry_decision':
+      // Pinned #123 emitter contract: { task_id, stage_id, attempt,
+      // max_attempts, retry_recommendation, action, next_status, reason,
+      // issues[] } with action ∈ complete|retry|exhausted|human_context|block.
+      return `Retry decision for ${ev.task_id ?? 'task'}: ${ev.action ?? '?'}${ev.next_status ? ` → ${ev.next_status}` : ''}${ev.reason ? ` — ${ev.reason}` : ''}`;
+    // goal_evaluated keeps its shared plain-language summary (goal_id /
+    // status / score / reason — the real pipeline-loop.js emitter fields);
+    // the ops layer only attaches the structured data below.
+    default:
+      return null;
+  }
+}
+
+function checkpointSummary(phase, ev) {
+  const verified = ev.verified === true
+    ? 'verified restorable'
+    : ev.verified === false ? 'NOT verified — restore may fail' : 'verification unknown';
+  return `Checkpoint saved ${phase} ${ev.stage_id ?? 'stage'} (${verified})`;
+}
+
+function fmtCount(value) {
+  const num = Number(value);
+  return Number.isFinite(num) ? num.toLocaleString('en-US') : '?';
+}
+
+// Structured per-type detail carried on feed items so the Alerts &
+// Guardrails / Approvals panels can render depth (limit vs value, attempt
+// counts, audit issues) without a second data path. Only fields that exist
+// on the event are copied — absent fields stay absent.
+const OPS_EVENT_DATA_FIELDS = {
+  guardrail_triggered: ['task_id', 'limit_name', 'current_value', 'limit_value', 'reason'],
+  guardrail_overridden: ['task_id', 'artifact'],
+  task_retry_scheduled: ['task_id', 'stage_id', 'attempt', 'max_attempts', 'retry_recommendation', 'reason'],
+  task_retry_exhausted: ['task_id', 'stage_id', 'attempt', 'max_attempts', 'retry_recommendation', 'reason'],
+  task_human_context_required: ['task_id', 'stage_id', 'attempt', 'max_attempts', 'retry_recommendation', 'reason'],
+  retry_decision: ['task_id', 'stage_id', 'attempt', 'max_attempts', 'retry_recommendation', 'action', 'next_status', 'reason'],
+  context_pressure_warning: ['task_id', 'source', 'metric', 'size', 'threshold', 'blocking', 'stage_id'],
+  approval_audit_failed: ['record_id', 'artifact', 'status', 'issues', 'reason'],
+  stage_checkpoint_before_saved: ['stage_id', 'task_id', 'verified'],
+  stage_checkpoint_after_saved: ['stage_id', 'task_id', 'verified'],
+  episode_memory_written: ['task_id', 'trusted'],
+  episode_memory_skipped_untrusted: ['task_id', 'reason', 'write_policy'],
+  metrics_write_failed: ['task_id', 'operation', 'error'],
+  goal_evaluated: ['iteration', 'max_iterations', 'goal_id', 'status', 'score', 'critical_count', 'failing_stages', 'recommended_rerun_stages'],
+};
+
+function opsEventData(ev) {
+  const fields = OPS_EVENT_DATA_FIELDS[ev.type];
+  if (!fields) return null;
+  const data = {};
+  for (const field of fields) {
+    if (ev[field] !== undefined) data[field] = ev[field];
+  }
+  return Object.keys(data).length ? data : null;
+}
+
 function eventLevel(ev) {
   if (ev.type === 'task_validated' && ev.status === 'FAIL') return 'fail';
   if (ev.type === 'validation_failed') return 'fail';
@@ -59,7 +148,23 @@ function eventLevel(ev) {
   if (ev.type === 'task_retry_exhausted') return 'fail';
   if (ev.type === 'task_human_context_required') return 'blocked';
   if (ev.type === 'task_blocked_by_validator') return 'blocked';
+  if (ev.type === 'retry_decision') {
+    // next_status is the task transition the harness actually made (#123):
+    // PASS | FAIL (re-claimable) | BLOCKED | NEEDS_CONTEXT.
+    const nextStatus = String(ev.next_status ?? '').toUpperCase();
+    if (nextStatus === 'BLOCKED' || nextStatus === 'NEEDS_CONTEXT') return 'blocked';
+    if (nextStatus === 'PASS') return 'pass';
+    return 'warn';
+  }
   if (/^retry_/.test(String(ev.type ?? ''))) return 'warn';
+  if (ev.type === 'stage_checkpoint_before_saved' || ev.type === 'stage_checkpoint_after_saved') {
+    return ev.verified === false ? 'warn' : 'info';
+  }
+  if (ev.type === 'context_pressure_warning') return 'warn';
+  if (ev.type === 'approval_audit_failed') return 'fail';
+  if (ev.type === 'episode_memory_written') return 'info';
+  if (ev.type === 'episode_memory_skipped_untrusted') return 'warn';
+  if (ev.type === 'metrics_write_failed') return 'warn';
   if (ev.type === 'goal_evaluated') return ev.status === 'PASS' ? 'pass' : 'warn';
   if (ev.type === 'loop_iteration_retrying_stages') return 'warn';
   if (ev.type === 'loop_completed') return 'pass';