backlog-drift

AI-powered drift detection for Backlog.md. Detects when code changes make your backlog tasks stale, inaccurate, or silently resolved.

The Problem

Backlog.md keeps issues alongside your code, but it doesn't detect when tasks become outdated as code evolves:

• A task references src/auth/oauth.ts but that file was deleted in a refactor
• Acceptance criteria say "implement rate limiting" but it already exists
• A completed task's referenced files have been modified since completion
• All of a task's referenced files are gone, but the task is still open

For AI agents consuming the backlog as context, stale tasks produce stale plans.

How It Works

Layer 1 — Structural Drift (deterministic, fast, zero false positives)

Check	What it detects
Dead references	A ref file has been deleted or renamed
Dependency state	A dependency task has been completed or archived
Stale completion	A Done task's ref was modified after completion
Orphaned task	All ref files deleted, task still active

Layer 2 — Semantic Drift (AI-powered, opt-in)

Uses an LLM to analyze whether task descriptions and acceptance criteria still match the referenced code. Supports Anthropic, OpenAI, and Ollama (local models).

Quick Start

# Install
bun add -g backlog-drift       # or: npm i -g backlog-drift

# Initialize (creates config + installs pre-commit hook)
backlog-drift init

# Discover file references for existing tasks
backlog-drift scan --apply

# Run drift checks
backlog-drift check

Commands

backlog-drift check [options]     Run drift checks on backlog tasks
  -t, --task <id>                 Check a specific task
  -s, --since <ref>               Only check changes since a git ref
  --semantic                      Also run AI-powered semantic checks
  --json                          Output as JSON

backlog-drift fix [options]       Auto-fix structural drift issues
  -t, --task <id>                 Fix a specific task
  --dry-run                       Show proposed changes without applying
  --semantic                      Also apply AI-suggested updates
  -y, --yes                       Apply without confirmation

backlog-drift scan [options]      Discover file refs from descriptions & git history
  -t, --task <id>                 Scan a specific task
  --apply                         Auto-add discovered refs to tasks
  --json                          Output as JSON

backlog-drift report [--json]     Generate a drift health report
backlog-drift init                Create config and install pre-commit hook
backlog-drift config              Show current configuration
backlog-drift hook install        Install pre-commit hook
backlog-drift hook remove         Remove pre-commit hook
backlog-drift hook status         Show hook installation status

Pre-Commit Hook

When installed (automatically via backlog-drift init), the pre-commit hook runs on every commit:

1. Checks staged files against task refs for structural drift
2. Discovers untracked refs from staged files mentioned in task descriptions
3. Reports findings based on hook.mode:
   • warn (default): print warnings, allow commit
   • block: print warnings, abort commit if drift detected
   • silent: log only

Example output:

⚠ backlog-drift: 2 tasks affected by this commit

  DRIFT-42 "Add rate limiting"
    ⚠ ref src/middleware/rateLimit.ts was modified

  DRIFT-19 "OAuth integration"  [Done]
    ✗ ref src/auth/oauth.ts no longer exists

ℹ backlog-drift: discovered new refs from staged files

  DRIFT-31 "Add social login"
    + src/auth/jwt.ts (not yet tracked)

  Run `backlog-drift scan --apply` to add these references.

Configuration

backlog-drift init creates .backlog-drift.yml:

structural:
  enabled: true
  check_dead_refs: true
  check_dependency_state: true
  check_stale_completions: true
  check_orphaned_tasks: true
  stale_completion_days: 14

semantic:
  enabled: false
  provider: anthropic          # anthropic | openai | ollama
  model: claude-sonnet-4-20250514
  # API key via BACKLOG_DRIFT_API_KEY env var

hook:
  mode: warn                   # warn | block | silent
  structural_only: true

scope:
  statuses:
    - To Do
    - In Progress
    - Review
  check_completed: true
  check_archived: false
  ignore_tasks: []
  ignore_refs: []              # e.g., "*.test.ts"

How Tasks Get File References

backlog-drift uses Backlog.md's existing --ref field to know which files each task cares about:

backlog task create "Add auth" --ref src/auth.ts --ref src/middleware.ts

For tasks without refs, backlog-drift scan discovers them:

• Parses task descriptions for file paths
• Searches git history for commits mentioning the task ID
• Suggests adding them: backlog-drift scan --apply

Upstream Contribution Path

backlog-drift is designed to eventually merge into Backlog.md as a native backlog drift subcommand:

• Same stack: Bun + TypeScript + Commander.js + gray-matter + Biome
• Same code conventions and formatting
• Additive-only changes to task frontmatter (drift_status, drift_checked, drift_log)
• No new dependencies beyond what Backlog.md already uses

Development

# Install dependencies
bun install

# Run tests (121 tests, ~99% coverage)
bun test

# Type check
bun run check:types

# Lint & format
bun run check
bun run check:fix

# Run CLI locally
bun run src/index.ts check

License

MIT