lib/substantive_fixer.py

@@ -522,55 +522,53 @@ async def substantive_fix_cycle(conn, max_workers=None) -> tuple[int, int]:
     Finds PRs with substantive issue tags that haven't exceeded fix budget.
     Processes up to 3 per cycle (Rhea: 180s interval, don't overwhelm eval).
     """
-    # Build the actionable-tag list from the routing constants so adding a new
-    # tag to FIXABLE_TAGS / CONVERTIBLE_TAGS / UNFIXABLE_TAGS auto-updates the
-    # SELECT filter — no two-place edit footgun.
-    actionable_tags = sorted(FIXABLE_TAGS | CONVERTIBLE_TAGS | UNFIXABLE_TAGS)
-    placeholders = ",".join(["?"] * len(actionable_tags))
-
-    # Push the actionable-tag filter into SQL (was a post-fetch Python loop).
-    # The old shape selected the 3 oldest request_changes PRs and then dropped
-    # ones without actionable tags, so empty-eval_issues rows occupied LIMIT-3
-    # forever (head-of-line). Now LIMIT-3 always returns 3 actionable rows.
-    # Reaper handles the empty-tag PRs after their 24h cooldown.
     rows = conn.execute(
-        f"""SELECT number, eval_issues FROM prs
+        """SELECT number, eval_issues FROM prs
            WHERE status = 'open'
            AND tier0_pass = 1
            AND (domain_verdict = 'request_changes' OR leo_verdict = 'request_changes')
            AND COALESCE(fix_attempts, 0) < ?
            AND (last_attempt IS NULL OR last_attempt < datetime('now', '-3 minutes'))
-           AND json_valid(eval_issues)
-           AND EXISTS (
-               SELECT 1 FROM json_each(eval_issues)
-               WHERE value IN ({placeholders})
-           )
            ORDER BY created_at ASC
            LIMIT 3""",
-        (MAX_SUBSTANTIVE_FIXES + config.MAX_FIX_ATTEMPTS, *actionable_tags),
+        (MAX_SUBSTANTIVE_FIXES + config.MAX_FIX_ATTEMPTS,),  # Total budget: mechanical + substantive
     ).fetchall()
 
     if not rows:
         return 0, 0
 
-    # Defense-in-depth: json_valid(eval_issues) in the SELECT already filters
+    # Filter to only PRs with substantive issues (not just mechanical)
-    # corrupt JSON before json_each runs, so this WARN should be unreachable.
-    # Kept anyway: json_valid and json.loads use technically distinct parsers,
-    # and the journal entry names the failure mode if SQLite ever surfaces a
-    # row that passes json_valid + json_each but fails json.loads.
     substantive_rows = []
+    skipped_no_tags = []
     for row in rows:
         try:
-            json.loads(row["eval_issues"] or "[]")
+            issues = json.loads(row["eval_issues"] or "[]")
         except (json.JSONDecodeError, TypeError):
+            # Corrupt JSON in eval_issues is abnormal (post-merge column drift,
+            # hand-edited row, partial write during crash). WARN so ops can chase
+            # the upstream column-write path. Without this, the row drops out of
+            # both substantive_rows and skipped_no_tags — the third silent path.
             logger.warning(
                 "PR #%d: corrupt eval_issues JSON — skipping in substantive fix cycle",
                 row["number"],
             )
             continue
-        substantive_rows.append(row)
+        if set(issues) & (FIXABLE_TAGS | CONVERTIBLE_TAGS | UNFIXABLE_TAGS):
+            substantive_rows.append(row)
+        else:
+            skipped_no_tags.append((row["number"], issues))
 
     if not substantive_rows:
+        # Visibility for the LIMIT-3 head-of-line block: if the oldest
+        # candidates have no fixer-actionable tags (e.g. eval_issues=[],
+        # broken_wiki_links only), the cycle silently returns 0 — and the
+        # next cycle picks the same head-of-line, forever. Log the eval_issues
+        # of skipped candidates so the journal makes the block visible.
+        if skipped_no_tags:
+            logger.info(
+                "Substantive fix cycle: 0 actionable from %d candidate(s) — head-of-line: %s",
+                len(rows), skipped_no_tags,
+            )
         return 0, 0
 
     fixed = 0