@workermill/agent 0.7.14 → 0.7.16

package/dist/plan-validator.d.ts

@@ -39,7 +39,7 @@ export interface CriticResult {
         suggestedChanges?: string[];
     }>;
 }
-declare const AUTO_APPROVAL_THRESHOLD = 85;
+declare const AUTO_APPROVAL_THRESHOLD = 80;
 /**
  * Parse execution plan JSON from raw Claude CLI output.
  * Mirrors server-side parseExecutionPlan() in planning-agent-local.ts.

package/dist/plan-validator.js

@@ -16,7 +16,7 @@ import { generateText } from "./providers.js";
 // CONSTANTS
 // ============================================================================
 const MAX_TARGET_FILES = 5;
-const AUTO_APPROVAL_THRESHOLD = 85;
+const AUTO_APPROVAL_THRESHOLD = 80;
 // ============================================================================
 // PLAN PARSING
 // ============================================================================

package/dist/planner.d.ts

@@ -31,8 +31,9 @@ export interface PlanningTask {
  *   2. Run Claude CLI to generate plan
  *   3. Parse plan, apply file cap (max 5 files per story)
  *   4. Run critic validation via Claude CLI
- *   5. If critic approves (score >= 85): post validated plan to API
+ *   5. If critic approves (score >= 80): post validated plan to API
  *   6. If critic rejects: re-run planner with feedback (up to MAX_ITERATIONS)
- *   7. After MAX_ITERATIONS without approval: fail the task
+ *   7. After MAX_ITERATIONS without approval: post best plan if score >= 50 (fallback)
+ *   8. If no plan scored >= 50: fail the task
  */
 export declare function planTask(task: PlanningTask, config: AgentConfig, credentials?: ClaimCredentials): Promise<boolean>;

package/dist/planner.js

@@ -686,9 +686,10 @@ async function runTeamAnalysis(task, basePrompt, claudePath, model, env, repoPat
  *   2. Run Claude CLI to generate plan
  *   3. Parse plan, apply file cap (max 5 files per story)
  *   4. Run critic validation via Claude CLI
- *   5. If critic approves (score >= 85): post validated plan to API
+ *   5. If critic approves (score >= 80): post validated plan to API
  *   6. If critic rejects: re-run planner with feedback (up to MAX_ITERATIONS)
- *   7. After MAX_ITERATIONS without approval: fail the task
+ *   7. After MAX_ITERATIONS without approval: post best plan if score >= 50 (fallback)
+ *   8. If no plan scored >= 50: fail the task
  */
 export async function planTask(task, config, credentials) {
     const taskLabel = chalk.cyan(task.id.slice(0, 8));
@@ -922,7 +923,42 @@ export async function planTask(task, config, credentials) {
                 }
             }
         }
-        // All iterations exhausted — fail
+        // All iterations exhausted — try best-plan fallback before failing.
+        // If we have a plan that scored >= BEST_PLAN_FALLBACK_THRESHOLD, post it
+        // with a warning instead of discarding it entirely.
+        const BEST_PLAN_FALLBACK_THRESHOLD = 50;
+        if (bestPlan && bestScore >= BEST_PLAN_FALLBACK_THRESHOLD) {
+            const elapsed = Math.round((Date.now() - startTime) / 1000);
+            const msg = `${PREFIX} Best-plan fallback: posting plan with score ${bestScore}/100 (below ${AUTO_APPROVAL_THRESHOLD} threshold, above ${BEST_PLAN_FALLBACK_THRESHOLD} minimum)`;
+            console.log(`${ts()} ${taskLabel} ${chalk.yellow("⚠")} ${msg}`);
+            await postLog(task.id, msg);
+            const planningDurationMs = Date.now() - startTime;
+            const fallbackPosted = await postValidatedPlan(task.id, bestPlan, config.agentId, taskLabel, elapsed, bestScore, [`Best-plan fallback: critic rejected after ${MAX_ITERATIONS} iterations`], criticHistory, totalFileCapTruncations, planningDurationMs, MAX_ITERATIONS);
+            if (fallbackPosted) {
+                return true;
+            }
+            // Fallback post failed (404, 409, etc.) — fall through to plan-failed
+            // so the task doesn't stay stuck in "planning" status forever.
+            console.log(`${ts()} ${taskLabel} ${chalk.yellow("⚠")} ${PREFIX} Fallback post rejected by server, reporting plan-failed`);
+            await postLog(task.id, `${PREFIX} Fallback plan rejected by server — reporting failure`);
+        }
+        // No usable plan (or fallback rejected) — report failure to server so
+        // the task doesn't stay in "planning" status forever (infinite retry loop).
+        try {
+            const failReason = bestPlan && bestScore >= BEST_PLAN_FALLBACK_THRESHOLD
+                ? `Best-plan fallback rejected by server after ${MAX_ITERATIONS} iterations (best score: ${bestScore}/100)`
+                : `Critic rejected after ${MAX_ITERATIONS} iterations (best score: ${bestScore}/100, threshold: ${AUTO_APPROVAL_THRESHOLD}, fallback minimum: ${BEST_PLAN_FALLBACK_THRESHOLD})`;
+            await api.post("/api/agent/plan-failed", {
+                taskId: task.id,
+                agentId: config.agentId,
+                reason: failReason,
+                criticHistory,
+            });
+        }
+        catch {
+            // Best-effort — if the endpoint doesn't exist yet, the task will still
+            // be picked up again, but at least we tried.
+        }
         return false;
     }
     finally {
@@ -964,9 +1000,10 @@ async function postValidatedPlan(taskId, plan, agentId, taskLabel, elapsed, crit
     }
     catch (error) {
         const err = error;
-        const detail = err.response?.data?.detail || String(error);
-        console.error(`${ts()} ${taskLabel} ${chalk.red("✗")} Server validation failed: ${detail.substring(0, 100)}`);
-        await postLog(taskId, `${PREFIX} Server-side plan validation failed: ${detail.substring(0, 200)}`, "error", "error");
+        const detail = err.response?.data?.error || err.response?.data?.detail || String(error);
+        const statusCode = err.response?.status ? ` (${err.response.status})` : "";
+        console.error(`${ts()} ${taskLabel} ${chalk.red("✗")} Server validation failed${statusCode}: ${detail.substring(0, 100)}`);
+        await postLog(taskId, `${PREFIX} Server-side plan validation failed${statusCode}: ${detail.substring(0, 200)}`, "error", "error");
         return false;
     }
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@workermill/agent",
-  "version": "0.7.14",
+  "version": "0.7.16",
   "description": "WorkerMill Remote Agent - Run AI workers locally with your Claude Max subscription",
   "type": "module",
   "main": "./dist/index.js",