Verified Documentation Generator

Generate audit-grade documentation from a codebase feature with full traceability and zero hallucination.

Arguments

$ARGUMENTS should be:

• Feature name and description (e.g., "User Registration", "Bookmark System", "Reading Progress")

Prerequisites

• Clear feature name and description from user
• Access to the codebase root directory

Hard Rules

• ❌ No assumptions or inferred behavior
• ❌ No undocumented states or transitions
• ✅ Every claim must link to file:line
• ✅ Unknowns must be explicitly reported
• ✅ Halt if verification fails

Stage 1: Scope Resolution

Goal: Define feature boundaries and reject ambiguous scopes.

1. Parse the feature description to identify expected entry points (API routes, UI components, CLI commands)
2. Search codebase for matching routes, exported functions/classes, and test files
3. Validate at least one concrete code location exists

Gate:

• If no code locations found → HALT with failure report
• If scope is ambiguous → Ask user to narrow scope
• If confirmed → Proceed with entry_point_candidates list

Output: scope_definition, entry_point_candidates, scope_confidence

Stage 2: Codebase Discovery

Goal: Map all code artifacts related to the feature.

1. Verify each entry point exists and classify as: API | UI | JOB | EVENT | CLI
2. Trace dependencies (imports, injected dependencies, runtime-resolved modules)
3. Build dependency graph (DAG)
4. Classify artifacts: Controllers, Services, Repositories, Models, Utils, External
5. Extract data models with types, relationships, and validation rules

Gate:

• Every entry point must be verifiable (file_exists, line_exists, callable)
• If no verifiable entry points → HALT

Output: entry_points, dependency_graph, artifacts_by_type, data_models, external_integrations, discovery_gaps

Stage 3: Execution Flow Tracing

Goal: Trace complete execution paths from entry to exit.

1. Synchronous paths: Follow function call chains, record decision branches, track variable transformations
2. Async boundaries: Identify Promise/async-await, event emitters, message queues, webhooks
3. Side effects: Catalog DATABASE_WRITE, DATABASE_READ, EXTERNAL_CALL, FILE_SYSTEM, CACHE_OP, EVENT_EMIT, LOGGING
4. Flow enumeration: Build all paths, tag as HAPPY_PATH | ERROR_PATH | EDGE_CASE

Gate:

• Every flow step must have a code reference
• If coverage below threshold → Ask user to accept partial or HALT

Output: execution_flows, async_boundaries, side_effects, flow_coverage, untraced_branches

Stage 4: State & Transition Extraction

Goal: Extract explicit state machines; reject inferred states.

1. Search for status/state enums, FSM patterns, status columns in schemas
2. Map transitions: STATE_A --[condition]--> STATE_B @ file:line
3. Verify each state value exists in code (enum, constant, or literal)
4. Identify orphan states (no inbound) and terminal states (no outbound)

Gate:

• Every state must have a code reference
• Every transition must link verified states
• If undocumented states found → HALT with failure report

Output: states, transitions, state_machine, orphan_states, terminal_states

Stage 5: Failure & Edge Case Analysis

Goal: Identify error handling coverage and missing failure modes.

1. Locate try/catch blocks, error callbacks, error boundaries, middleware handlers
2. Trace exception propagation paths
3. Detect edge cases: NULL_INPUT, EMPTY_ARRAY, BOUNDARY_VALUES, PERMISSION, TIMEOUT, CONFLICT
4. Score risk: LOW | MEDIUM | HIGH | CRITICAL

No hard gate – outputs are informational.

Output: handled_failures, unhandled_failures, silent_failures, edge_cases, risk_areas

Stage 6: Diagram Generation

Goal: Produce diagrams strictly matching verified flows.

1. Generate State Diagram (Mermaid stateDiagram-v2) from state_machine
2. Generate Sequence Diagrams (Mermaid sequenceDiagram) from execution_flows
3. Generate Dependency Diagram (Mermaid flowchart) from dependency_graph
4. Cross-check every diagram element against source stage outputs

Gate:

• Remove any unverified elements
• Log warnings for removed elements

Output: state_diagram, sequence_diagrams, dependency_diagram, diagram_verification_log

Stage 7: Documentation Assembly

Goal: Produce structured documentation.

Generate markdown with these sections:

# [Feature Name] Documentation

## Overview

## Entry Points

## Data Models

## Execution Flows

## State Machine

## Error Handling

## Known Gaps & Limitations

## Appendix: Code References

1. Every claim gets [ref: file:line] annotation
2. Build clickable traceability index
3. Generate two views: Technical (full detail) and Non-Technical (simplified)

Output: documentation_technical, documentation_summary, traceability_index, confidence_report

Stage 8: Verification & Validation

Goal: Cross-check all claims; fail if verification cannot pass.

1. Extract all factual claims from documentation
2. Verify each claim against traceability_index:
   • File exists
   • Line content matches claim basis
3. Check diagram-flow consistency
4. Verify all gaps appear in Known Limitations
5. Hallucination detection: Flag claims without code references or contradicting code

Gate:

• If hallucination detected → HALT immediately
• If too many unverifiable claims → HALT
• If passed → Output final documentation

Output: verification_passed, verified_claims, failed_claims, hallucination_report, final_documentation or failure_report

Failure Behavior

When any stage encounters a blocking failure:

1. STOP execution immediately
2. PRESERVE all partial outputs
3. Generate failure report:

   {
     "stage_failed": "<stage_name>",
     "failure_type": "<failure_enum>",
     "failure_reason": "<explanation>",
     "partial_outputs": { ... },
     "recovery_suggestions": [...]
   }

4. AWAIT user intervention

Example Usage

User: Generate documentation for the "User Registration" feature

Stage 1: Found POST /auth/register, /signup page
Stage 2: Traced AuthController → AuthService → UserRepository, EmailService
Stage 3: Mapped registration flow with email verification async boundary
Stage 4: Extracted UserStatus enum { PENDING, ACTIVE, SUSPENDED }
Stage 5: Identified missing timeout handling on EmailService
Stage 6: Generated state diagram and sequence diagram
Stage 7: Assembled full technical documentation
Stage 8: Verified 24/24 claims, no hallucinations → SUCCESS