The Improvement Cycle - Auto-Correction

"War Story": The Validator Who Cried "Wolf!"

Our first implementation of the GoalValidator was too sensitive.

Disaster Logbook (July 28th):

CRITICAL goal validation failures: 4 issues
⚠️ GOAL SHORTFALL: 0/50.0 contacts for contacts (100.0% gap, missing 50.0);
INFO: Creating corrective task: "URGENT: Collect 50.0 missing contacts"
... (5 minutes later)
CRITICAL goal validation failures: 4 issues
⚠️ GOAL SHORTFALL: 0/50.0 contacts for contacts (100.0% gap, missing 50.0)
INFO: Creating corrective task: "URGENT: Collect 50.0 missing contacts"

The system had entered a panic loop. It detected a gap, created a corrective task, but before the Executor could even assign and execute that task, the validator restarted, detected the same gap, and created another identical corrective task. Within hours, our task queue was flooded with hundreds of duplicate tasks.

The Lesson Learned: Self-Correction Needs "Patience" and "Awareness"

A proactive system without awareness of the state of its own corrective actions creates more problems than it solves. The solution required making our GoalValidator more intelligent and "patient".

1. Existing Corrective Task Check: Before creating a new corrective task, the validator now checks if theres already a pending or in_progress task trying to solve the same gap. If it exists, it does nothing.
2. Cooldown Period: After creating a corrective task, the system enters a "grace period" (e.g., 30 minutes) for that specific goal, during which no new corrective actions are generated, giving the agent team time to act.
3. AI-Driven Priority and Urgency: Instead of always creating "URGENT" tasks, we taught the AI to evaluate gap severity in relation to project timeline. A 10% gap at project start might generate a medium priority task; the same gap one day before deadline would generate a critical priority task.

The Prompt That Guides Correction

The heart of this system is the prompt that generates corrective tasks. It doesnt just say "solve the problem", but asks for a mini strategic analysis.

Reference code: _generate_corrective_task logic in goal_validator.py

prompt = f"""
You are an expert Project Manager in crisis management. A critical gap has been detected between the current project state and the preset objectives.

**Failed Objective:** {goal.description}
**Current State:** {current_progress}
**Detected Gap:** {failure_details}

**Lessons from the Past (from Memory):**
{relevant_failure_lessons}

**Required Analysis:**
1.  **Root Cause Analysis:** Based on past lessons and the gap, what is the most likely cause of this failure? (e.g., "Tasks were too theoretical", "Missing email verification tool").
2.  **Specific Corrective Action:** Define ONE SINGLE task, as specific and actionable as possible, to start bridging this gap. Dont be generic.
3.  **Optimal Assignment:** Which team role is best suited to solve this problem?

**Output Format (JSON only):**
{{
  "root_cause": "The main cause of the failure.",
  "corrective_task": {{
    "name": "Name of the corrective task (e.g., 'Verify Email of 50 Existing Contacts')",
    "description": "Detailed description of the task and expected result.",
    "assigned_to_role": "Specialized Role",
    "priority": "high"
  }}
}}
"""

This prompt doesnt just solve the problem, but does so intelligently, learning from the past and delegating to the right role, perfectly closing the feedback loop.