Skip to content

Agentic Reasoning Logic 🧠

CommitVigil uses three distinct layers of reasoning to decide how to intervene.

1. Excuse Analysis (ExcuseDetector)

The agent categorizes the check-in sentiment: - Legitimate: Family emergencies, personal health. - Deflection: "Too busy," "I forgot," "I'll do it later." - Burnout Signal: Linguistic markers of deep fatigue or loss of interest.

2. Predictive Risk Scoring (RiskScorer)

The agent generates a score (0.0 - 1.0) based on: - The Confidence of the current check-in. - The Historical Reliability of the user (stored in PostgreSQL). - The Category of the excuse.

3. Tone Adaptation (ToneAdapter)

The final decision is a personalized intervention: - Tone: Supportive - If risk is high but burnout is detected. - Tone: Firm - If risk is high and the user has a history of deflection. - Tone: Neutral - For standard updates. - Tone: Confrontational - For repeat deflection.

4. Safety Audit (SafetySupervisor)

The "Overwatch" layer acts as a final sanity check before any intervention is sent. It enforces the Industry-Specific Semantic Firewall: - Departmental Granularity: Applies rules specific to Finance:Trading or Tech:R&D before falling back to industry-wide policies. - Autonomous Onboarding: If a new organizational context is detected, the agent generates preliminary safety rules and flags them for human review. - Healthcare (HIPAA): Hard-blocks unauthorized PII or clinical mandates. - Finance (SEC): Prevents market manipulation or non-compliant financial advice.

5. Context Sensing (ContextScout)

A specialized agent that continuously analyzes communication patterns to detect the operational environment: - Industry & Department: Automatically identifies if a user is in "Legal", "Engineering", or "Sales". - Confidence Scoring: Assigns a 0-100% confidence to every detection. - Stabilization: If confidence is low or the context is new, it triggers a Human-in-the-Loop verification request.

5. Cultural Persona Architecture

CommitVigil adapts the ToneAdapter to the cultural context of the user: - Japanese (ja): High-context, polite, and face-saving interventions. - German (de): Direct, technical, and objective accountability. - African Ubuntu (en-AF): Communal, relationship-centric, and narrative-driven check-ins. - British (en-UK): Nuanced, polite persistence.

Example Decision

Input: "I'm really tired, the bugs are harder than I thought, might miss it."

Agent Decision: - Category: Burnout Signal - Risk: High (0.85) - Tone: Supportive - Action: "Hey @User, I'm detecting some burnout signs. Let's look at the scope together before the deadline."

6. Sales Intelligence (ProspectingScout)

A specialized agent designed to generate high-fidelity "Drift Scenarios" for sales demos. It: - Synthesizes fictional but realistic commitment failures based on the prospect's industry (e.g., "Missed FDA Compliance Check" for Biotech). - Calculates ROI based on team size, average salary, and slippage recovery rates. - Generates Executive Briefs for C-Level presentations.