diff --git a/tools/observer-self-assessment-api.mjs b/tools/observer-self-assessment-api.mjs
index ebd0291a..b5960991 100644
--- a/tools/observer-self-assessment-api.mjs
+++ b/tools/observer-self-assessment-api.mjs
@@ -92,8 +92,11 @@ export function readRuntimeFlag(name, { homedir, fsImpl } = {}) {
     if (!fs.existsSync(filePath)) return 'off';
     const raw = fs.readFileSync(filePath, 'utf-8');
     const parsed = JSON.parse(raw);
-    if (typeof parsed.value !== 'string') return 'off';
-    return parsed.value;
+    // Runtime flag files use `mode` (canonical, see all ~/.claude/runtime/*-mode.json);
+    // `value` retained as legacy alias to keep existing test fixtures working.
+    const val = parsed.mode ?? parsed.value;
+    if (typeof val !== 'string') return 'off';
+    return val;
   } catch {
     return 'off';
   }
diff --git a/tools/observer-state-enricher.mjs b/tools/observer-state-enricher.mjs
index 2d0b8267..bf10914f 100644
--- a/tools/observer-state-enricher.mjs
+++ b/tools/observer-state-enricher.mjs
@@ -59,5 +59,17 @@ export function extractClassifierOutput(state) {
     recommended_chain_id: cls.recommended_chain_id ?? null,
     no_skill_found: cls.no_skill_found === true,
     source: cls.source ?? null,
+    // Factor-analysis signal: classifier's stated rationale + confidence.
+    // Field name varies by prompt schema: new (Phase 2) uses `reason_for_choice`,
+    // legacy uses `reasoning`. Null on regex / prefilter paths. Truncated to
+    // keep episode JSONL line size bounded.
+    reasoning: pickReasoning(cls),
+    confidence: typeof cls.confidence === 'number' ? cls.confidence : null,
   };
 }
+
+function pickReasoning(cls) {
+  const v = cls.reasoning ?? cls.reason_for_choice ?? cls.reason ?? null;
+  if (typeof v !== 'string') return null;
+  return v.slice(0, 600);
+}
diff --git a/tools/observer-transcript-parser.mjs b/tools/observer-transcript-parser.mjs
index 028f7fe9..2616d93c 100644
--- a/tools/observer-transcript-parser.mjs
+++ b/tools/observer-transcript-parser.mjs
@@ -406,6 +406,18 @@ export function extractTaskSize(turn) {
  * Defensive: skips entries where `usage` is not a plain object (handles
  * malformed transcript edge cases like `"usage": 42`).
  */
+// Normalize `usage.iterations` to a count.
+// Claude Code transcripts may emit it as: a number (legacy / no extended-thinking),
+// an array of step-objects (extended-thinking turns), or a plain object map.
+// Coerce to a number; non-finite / unknown → 0. Prevents "0[object Object]…"
+// string concatenation that previously poisoned task_cost.iterations.
+function iterationsCount(v) {
+  if (typeof v === 'number' && Number.isFinite(v)) return v;
+  if (Array.isArray(v)) return v.length;
+  if (v && typeof v === 'object') return Object.keys(v).length;
+  return 0;
+}
+
 export function extractTokenUsage(turn) {
   let input = 0, output = 0, cache_read = 0, cache_creation = 0;
   let web_search = 0, web_fetch = 0, iterations = 0;
@@ -416,7 +428,7 @@ export function extractTokenUsage(turn) {
     output         += u.output_tokens || 0;
     cache_read     += u.cache_read_input_tokens || 0;
     cache_creation += u.cache_creation_input_tokens || 0;
-    iterations     += u.iterations || 0;
+    iterations     += iterationsCount(u.iterations);
     if (u.server_tool_use) {
       web_search += u.server_tool_use.web_search_requests || 0;
       web_fetch  += u.server_tool_use.web_fetch_requests  || 0;