fix: revert to stable baseline with minimal targeted improvements

agents/ollama.ts

@@ -148,6 +148,9 @@ async function runOllamaLoop(ctx: AgentRunContext): Promise<AgentResult> {
   let pendingModeNudge = false;
   let calledOutputTool = false;
   let addedContinueNudge = false;
+  let reportProgressCount = 0;
+  let editCommentCount = 0;
+  let selectedModeName = "";
 
   while (iterations < MAX_ITERATIONS) {
     iterations++;
@@ -222,13 +225,16 @@ async function runOllamaLoop(ctx: AgentRunContext): Promise<AgentResult> {
           "» model stopped before completing task — nudging to continue",
         );
         addedContinueNudge = true;
+        const isReview =
+          selectedModeName === "Review" || selectedModeName === "IncrementalReview";
         messages.push({
           role: "user",
           content:
             "Your task is not complete yet. Continue executing the workflow — " +
             "call the next required tool to finish. " +
-            "Do not stop until you have submitted a review (create_pull_request_review) " +
+            (isReview
-            "or called report_progress with a final summary.",
+              ? "Do not stop until you have submitted a review via create_pull_request_review."
+              : "Do not stop until you have submitted a review (create_pull_request_review) or called report_progress with a final summary."),
         });
         continue;
       }
@@ -279,6 +285,51 @@ async function runOllamaLoop(ctx: AgentRunContext): Promise<AgentResult> {
         role: "tool",
         content: result,
       });
+
+      // After checkout_pr returns, extract and pin the diffPath as a user
+      // message. Only tool messages are pruned — user messages survive context
+      // pressure — so the model retains the exact diff file path and knows how
+      // to call read_file with offset/limit to read diff ranges.
+      if (toolName === "checkout_pr") {
+        try {
+          const parsed = JSON.parse(result) as Record<string, unknown>;
+          if (typeof parsed?.diffPath === "string") {
+            const dp = parsed.diffPath;
+            messages.push({
+              role: "user",
+              content:
+                `The diff file is at: ${dp}\n\n` +
+                `Read each file's diff by calling read_file on this path with start_line/end_line from the TOC.\n` +
+                `Example — for "device.ts (163 lines, 1146-1308)": read_file(path="${dp}", start_line=1146, end_line=1308)\n` +
+                `IMPORTANT: Do NOT call read_file on source files (apps/..., packages/...). ` +
+                `Use read_file on the diff file path above with the TOC line ranges.`,
+            });
+            log.info(`» pinned diffPath to context: ${dp}`);
+          }
+        } catch {
+          // best-effort: if checkout_pr result isn't JSON, skip
+        }
+      }
+
+      // report_progress called repeatedly means the model is stuck.
+      if (toolName === "report_progress") {
+        reportProgressCount++;
+        if (reportProgressCount >= 2) {
+          log.info("» report_progress loop — stopping");
+          await unloadModel(ollama, model);
+          return { success: true };
+        }
+      }
+
+      // edit_issue_comment called repeatedly means the model is stuck.
+      if (toolName === "edit_issue_comment") {
+        editCommentCount++;
+        if (editCommentCount >= 2) {
+          log.info("» edit_issue_comment loop — stopping");
+          await unloadModel(ollama, model);
+          return { success: true };
+        }
+      }
     }
 
     // After the FIRST select_mode call, nudge the model to act on the guidance.
@@ -301,10 +352,18 @@ async function runOllamaLoop(ctx: AgentRunContext): Promise<AgentResult> {
         // best-effort
       }
 
-      const firstStep =
+      const isReviewMode =
-        selectedMode === "Review" || selectedMode === "IncrementalReview"
+        selectedMode === "Review" || selectedMode === "IncrementalReview";
-          ? "Your first tool call must be checkout_pr with the PR number from the event context."
+
-          : "Call the first tool required by the workflow now.";
+      selectedModeName = selectedMode;
+
+      const firstStep = isReviewMode
+        ? "Your first tool call must be checkout_pr. After that: (1) read diff ranges from the returned diffPath using the TOC line numbers — do NOT read full source files one-by-one, that will exhaust your context window; (2) every inline comment MUST identify a specific problem — never write praise or 'looks good' observations."
+        : "Call the first tool required by the workflow now.";
+
+      const endCondition = isReviewMode
+        ? "Your ONLY valid final action is create_pull_request_review. Do NOT call report_progress — the mode workflow explicitly forbids it."
+        : "Execute the complete workflow step by step until you call create_pull_request_review or report_progress.";
 
       messages.push({
         role: "user",
@@ -312,7 +371,7 @@ async function runOllamaLoop(ctx: AgentRunContext): Promise<AgentResult> {
           `Good. You have selected ${selectedMode || "a"} mode and received the workflow. ` +
           "Do NOT call select_mode again. " +
           `${firstStep} ` +
-          "Execute the complete workflow step by step until you call create_pull_request_review or report_progress.",
+          endCondition,
       });
     }
   }