IntentFlow

Capture the reasoning behind your code decisions.

IntentFlow automatically saves your AI chat sessions from GitHub Copilot and Cursor, versioning the intent and decisions in your repository. Track what code was written by AI vs humans with Agent Trace and Human Trace.

Why IntentFlow?

You don't know why your Agents chose Drizzle over Prisma. Neither does the next agent in line to work on your codebase.

Intent captures the reasoning behind your code—from agent sessions to team discussions—versioned in your repo.

Features

Core Features

Automatic Capture: Sessions are saved automatically as you work
Git-Friendly: Markdown exports commit cleanly with your code
Privacy-First: PII and secrets are redacted before export
Local Search: Find past decisions without leaving VS Code
Opt-Out Control: Per-session ignore and global disable options
Rule Derivation: Generate coding rules from sessions using local LLMs (Ollama)
Multi-Format Export: Output rules to GitHub Copilot and Cursor formats
Auto-Tagging: Automatically categorize sessions using local AI analysis
Thinking Content: Capture AI extended reasoning (<think> tags) in collapsible sections

Experimental Features

Enable these in settings under intentflow.experimental.*:

NES (Next Edit Suggestions)

Track inline code completions from GitHub Copilot with metrics on AI-assisted typing.

Captures accepted AI suggestions with text content
Distinguishes AI completions from manual typing
Configurable detection thresholds
Language and file pattern filtering

Agent Trace

Track AI-generated code at the file and line level following the Agent Trace v0.1.0 spec.

Records which lines were written by AI assistants
Tracks file edits from Copilot tool calls (createFile, replaceString, multiReplaceString)
Content hashing for tracking code modifications
Links code to originating chat sessions

Human Trace

Calculate human-authored code by analyzing what ISN'T in Agent Trace.

Identifies code written by humans for compliance and audit
Git blame integration for contributor attribution
Pre-commit hooks showing human code coverage
Critical path monitoring for files requiring human review

Getting Started

Install the extension from the VS Code marketplace
Open a workspace folder
Start chatting with Copilot or Cursor
Sessions are automatically saved to .intentflow/history/

Directory Structure

.intentflow/
├── history/          # Markdown exports (git-friendly)
├── store/            # Structured JSON (gitignored)
├── traces/           # Agent Trace and Human Trace data
│   └── workspace-trace.json
├── backups/          # Rule file backups (gitignored)
└── .gitignore        # Auto-excludes store/ and backups/

Commands

Session Management

Command	Description
IntentFlow: Save Current AI Chat Session	Manually save the current session
IntentFlow: Search AI Chat History	Search past sessions
IntentFlow: Toggle Auto-Capture	Enable/disable automatic capture
IntentFlow: Ignore Current Session	Mark current session to not be saved
IntentFlow: Export Session to Markdown	Export session as markdown
IntentFlow: Open History Folder	Open the history directory
IntentFlow: Sync Workspace History	Import historical sessions from workspace

Rule Derivation

Command	Description
IntentFlow: Derive Rules from Sessions	Generate coding rules using Ollama

Human Trace (Experimental)

Command	Description
IntentFlow: Generate Human Trace	Calculate human-authored code in workspace
IntentFlow: Show Human Code Coverage	Display human vs AI code statistics
IntentFlow: Install Human Trace Git Hooks	Add pre-commit hook for coverage reporting
IntentFlow: Uninstall Human Trace Git Hooks	Remove the pre-commit hook

Configuration

Capture Settings

Setting	Default	Description
`intentflow.capture.enabled`	`true`	Enable automatic capture
`intentflow.capture.sources`	`["copilot", "cursor"]`	AI assistants to capture
`intentflow.storagePath`	`.intentflow`	Storage directory path
`intentflow.timezone.utc`	`true`	Use UTC timezone for timestamps
`intentflow.includeThinkingContent`	`true`	Include AI thinking/reasoning content
`intentflow.showDeveloperTools`	`false`	Show developer logging and debug tools

Redaction Settings

Setting	Default	Description
`intentflow.redaction.enabled`	`true`	Redact sensitive information
`intentflow.redaction.patterns`	`[]`	Additional regex patterns to redact

Ollama Integration

Setting	Default	Description
`intentflow.ollama.enabled`	`true`	Enable Ollama integration
`intentflow.ollama.baseUrl`	`http://localhost:11434`	Ollama server URL
`intentflow.ollama.model`	`lfm2.5-thinking`	Model for analysis
`intentflow.ollama.timeout`	`120000`	Request timeout (ms)

Rule Derivation

Setting	Default	Description
`intentflow.rules.outputFormats`	`["copilot", "cursor"]`	Output targets
`intentflow.rules.minConfidence`	`0.7`	Minimum confidence threshold (0-1)
`intentflow.rules.createBackup`	`true`	Backup before overwrite

Auto-Tagging

Setting	Default	Description
`intentflow.tagging.enabled`	`true`	Enable auto-tagging of sessions
`intentflow.tagging.minConfidence`	`0.6`	Minimum confidence for tags (0-1)
`intentflow.tagging.maxMessages`	`10`	Max messages to analyze

Experimental: NES (Inline Completion Tracking)

Setting	Default	Description
`intentflow.experimental.nes.enabled`	`false`	Enable NES tracking
`intentflow.experimental.nes.retentionDays`	`30`	Days to retain metrics data
`intentflow.experimental.nes.minCharsForSingleLine`	`30`	Min chars for single-line completion
`intentflow.experimental.nes.minLinesForMultiLine`	`2`	Min lines for multi-line completion
`intentflow.experimental.nes.languages`	`[]`	Languages to track (empty = all)
`intentflow.experimental.nes.captureText`	`true`	Capture actual completion text
`intentflow.experimental.nes.maxTypingIntervalMs`	`200`	Max ms between keystrokes for typing detection
`intentflow.experimental.nes.maxCharsForCompletion`	`5000`	Max chars to track as completion
`intentflow.experimental.nes.excludePatterns`	`["/node_modules/", ...]`	File patterns to exclude

Experimental: Agent Trace

Setting	Default	Description
`intentflow.experimental.agentTrace.enabled`	`false`	Enable Agent Trace v0.1.0

Experimental: Human Trace

Setting	Default	Description
`intentflow.experimental.humanTrace.enabled`	`false`	Enable Human Trace
`intentflow.experimental.humanTrace.gitHooks.enabled`	`false`	Install git pre-commit hook
`intentflow.experimental.humanTrace.gitHooks.minHumanPercentage`	`10`	Min human % for passing status
`intentflow.experimental.humanTrace.gitHooks.criticalPaths`	`[]`	Paths requiring human review

How Agent Trace Works

Agent Trace records AI-generated code at edit-time by intercepting tool calls from AI assistants:

Tool Call Extraction: Captures createFile, replaceString, and multiReplaceString operations from Copilot sessions
Line Range Inference: Converts string-based edits (oldString → newString) to line numbers by searching the file content
Content Hashing: Generates MD5 hashes of code content to track modifications and code movement
Consolidation: Merges edits into a workspace-level trace stored at .intentflow/traces/workspace-trace.json

The trace follows the Agent Trace v0.1.0 specification with attribution ranges linking code to sessions.

How Human Trace Works

Human Trace is the complement of Agent Trace—it identifies code NOT written by AI:

Agent Trace Subtraction: Any line NOT in Agent Trace is potentially human-authored
Content Hash Verification: If a line exists in Agent Trace but the content_hash differs, it was modified by a human
Git Blame Integration: Uses git blame to identify contributors and timestamps for human code
Coverage Calculation: Groups human-authored ranges by contributor for audit and compliance

The pre-commit hook displays human coverage statistics without blocking commits.

Privacy

IntentFlow is designed with privacy in mind:

All data stays local by default
Sensitive information is automatically redacted:
- Emails, API keys, tokens (AWS, Azure, Slack, Google, GitHub)
- Public IP addresses (private IPs preserved)
- User paths (Unix and Windows)
- Database connection strings
- SSH keys, credit cards, SSNs
Custom redaction patterns can be added
Individual sessions can be ignored
Capture can be disabled globally

UI Components

IntentFlow adds an activity bar icon with two panels:

Control Panel: Overview and quick actions for session management
Rules Derivation: Interface for generating and managing coding rules

License

MIT