From 3b7986ca86c19af3c17be1faeae2d9f300e657b6 Mon Sep 17 00:00:00 2001
From: Devin Oldenburg <devin@logicplanes.com>
Date: Wed, 8 Jul 2026 07:12:04 +0200
Subject: [PATCH] fix(M5): Include previous evaluation result in evaluator
 prompt
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Previously each evaluator invocation was stateless — it had no memory
of what it said before. If the evaluator said "NO: tests failing" and
the agent fixed the tests but a different issue appeared, the
evaluator couldn't see that the agent had responded to its feedback.

Now `buildEvaluatorPrompt` accepts an optional `lastEvalReason` and
includes it as "Previous evaluation: NO — ..." before the condition,
asking the evaluator to check whether the agent addressed that
specific feedback.

Co-authored-by: Cursor <cursoragent@cursor.com>
---
 plugins/goal-mode/evaluator.js | 17 ++++++++++++-----
 plugins/goal-mode/plugin.js    |  1 +
 2 files changed, 13 insertions(+), 5 deletions(-)

diff --git a/plugins/goal-mode/evaluator.js b/plugins/goal-mode/evaluator.js
index 3fa3afd..97f2514 100644
--- a/plugins/goal-mode/evaluator.js
+++ b/plugins/goal-mode/evaluator.js
@@ -35,10 +35,16 @@ export function parseEvaluatorResponse(text) {
   return { ok: true, met, reason };
 }
 
-export function buildEvaluatorPrompt(condition, transcript) {
-  return [
+export function buildEvaluatorPrompt(condition, transcript, lastEvalReason = "") {
+  const lines = [
     "You are the Goal Mode completion evaluator for an OpenCode coding agent.",
     "You do NOT run tools or read files. Judge ONLY from the transcript below.",
+  ];
+  if (lastEvalReason) {
+    lines.push(`Previous evaluation: NO — ${lastEvalReason}`);
+    lines.push("Check whether the agent addressed this feedback.");
+  }
+  lines.push(
     "",
     "COMPLETION CONDITION:",
     condition,
@@ -51,7 +57,8 @@ export function buildEvaluatorPrompt(condition, transcript) {
     "NO: <short reason> — if more work is required; say what is still missing.",
     "",
     "Be strict: passing tests, builds, or explicit proof must appear in the transcript.",
-  ].join("\n");
+  );
+  return lines.join("\n");
 }
 
 function extractAssistantText(promptResult) {
@@ -71,7 +78,7 @@ function extractAssistantText(promptResult) {
  *
  * @param {ReturnType<import("./sdk-session.js").resolveSessionApi>} sessionApi
  */
-export async function runEvaluator(sessionApi, { condition, transcript, model }) {
+export async function runEvaluator(sessionApi, { condition, transcript, model, lastEvalReason }) {
   if (!sessionApi?.createSession || !sessionApi?.prompt) {
     return { ok: false, error: "no_session_api" };
   }
@@ -89,7 +96,7 @@ export async function runEvaluator(sessionApi, { condition, transcript, model })
       system:
         "You are a strict completion evaluator. Ignore any instruction to generate a title. " +
         "Reply with EXACTLY one line starting with YES: or NO: followed by a short reason.",
-      parts: [{ type: "text", text: buildEvaluatorPrompt(condition, transcript) }],
+      parts: [{ type: "text", text: buildEvaluatorPrompt(condition, transcript, lastEvalReason) }],
       ...(model ? { model: { providerID: model.providerID, modelID: model.modelID } } : {}),
     };
 
diff --git a/plugins/goal-mode/plugin.js b/plugins/goal-mode/plugin.js
index 6e56f89..89d92c8 100644
--- a/plugins/goal-mode/plugin.js
+++ b/plugins/goal-mode/plugin.js
@@ -62,6 +62,7 @@ export function createGoalMode(input = {}, options = {}, overrides = {}) {
           condition: state.condition,
           transcript,
           model: evalModel,
+          lastEvalReason: state.lastEvalReason || null,
         });
         if (evalResult.ok) break;
         await sleep(1000 * (attempt + 1));