From e7da32843e22c032e39ef8826e736a9973234fbf Mon Sep 17 00:00:00 2001
From: Story Crater Bot <19826264+Riotpiaole@users.noreply.github.com>
Date: Wed, 19 Aug 2026 17:13:49 -0700
Subject: [PATCH] fix(agent-pod): force judge to actually call tools instead of
 narrating

Observed live: phase-judge (on homelab-reasoning) wrote a full page of
'I should check X, then Y' reasoning, declared VERDICT: PASS, and showed
the touch command as a fenced code block in its own text -- never ran
git diff, never wrote the result file, never touched the sentinel.
Coordinator timed out waiting on a file that was never going to appear.
---
 k8s/apps/agent-pod/coordinator-configmap.yaml | 24 +++++++++++++++----
 1 file changed, 20 insertions(+), 4 deletions(-)

diff --git a/k8s/apps/agent-pod/coordinator-configmap.yaml b/k8s/apps/agent-pod/coordinator-configmap.yaml
index 04324da..840dc41 100644
--- a/k8s/apps/agent-pod/coordinator-configmap.yaml
+++ b/k8s/apps/agent-pod/coordinator-configmap.yaml
@@ -107,6 +107,22 @@ data:
       "Read and follow ~/.pi/agent/skills/karpathy-guidelines/SKILL.md and " +
       "~/.pi/agent/skills/caveman/SKILL.md as hard rules for this entire task, before anything else. ";
     
+    // judge (routed to homelab-reasoning) has been observed narrating an
+    // entire review in prose -- "I should run git diff, then check X..." --
+    // and then writing a verdict based on that narration without ever calling
+    // a real tool. Live example: a phase-judge call produced a page of
+    // "I would check..." reasoning, declared VERDICT: PASS, and showed the
+    // touch command as a fenced code block IN ITS OWN TEXT rather than
+    // executing it. Coordinator just timed out waiting on a sentinel that was
+    // never going to appear, since nothing was ever actually run. Spelled out
+    // explicitly since "use the judge skill" alone apparently isn't enough to
+    // rule this out.
+    const REQUIRE_REAL_TOOL_CALLS =
+      "Do not narrate what you would check -- actually run the commands via a real tool call and read their real " +
+      "output before writing anything. A verdict based on describing checks instead of executing them is invalid. " +
+      "Writing the verdict file and touching the sentinel are themselves tool calls you must execute, not text to " +
+      "display in your response. ";
+    
     function parseVerdictLine(text, label) {
       if (!text) return null;
       const re = new RegExp(`${label}:\\s*(\\w+)`, "i");
@@ -331,9 +347,9 @@ data:
     
     function judgePrompt(task, baseBranch, attempt, cwd) {
       return (
-        `Use the judge skill to review the diff against ${baseBranch}...HEAD for task ${task}. Write your verdict to ` +
-        `${path.join(cwd, `.task-result-${task}`)} as a single "VERDICT: PASS" or "VERDICT: FAIL" line plus one line of rationale, then ` +
-        `run: touch ${path.join(cwd, `.stage-done-${task}-judge-${attempt}`)}`
+        `${REQUIRE_REAL_TOOL_CALLS}Use the judge skill to review the diff against ${baseBranch}...HEAD for task ${task}. ` +
+        `Write your verdict to ${path.join(cwd, `.task-result-${task}`)} as a single "VERDICT: PASS" or "VERDICT: FAIL" ` +
+        `line plus one line of rationale, then run: touch ${path.join(cwd, `.stage-done-${task}-judge-${attempt}`)}`
       );
     }
     
@@ -694,7 +710,7 @@ data:
           cwd,
           repoId,
           "judge",
-          `Use the judge skill to review the full phase diff for phase ${phaseLabel} against ` +
+          `${REQUIRE_REAL_TOOL_CALLS}Use the judge skill to review the full phase diff for phase ${phaseLabel} against ` +
             `${baseBranch}...HEAD (covers every task in this phase: ${[...phaseTaskIds].join(", ")}). Every ` +
             `individual task already passed its own judge review -- your job here is different: confirm the ` +
             `tasks integrate correctly as one coherent narrative, and that real integration tests (not just ` +