Unit Work - Human-in-the-Loop Verification Framework

"The largest task an AI can self-validate to 100% accuracy, that is also able to get validated by the minimum amount of human review"

Core Philosophy

Unit Work replaces arbitrary development phases with a verification-driven approach:

1. Front-load work in interview and planning stages
2. Use Hindsight memory to compound learnings across sessions
3. NEVER skip memory recall - it's the foundation that makes compounding work
4. Create checkpoints at verifiable boundaries, not arbitrary phases
5. Treat commits as checkpoints with verification documents
6. Know your gaps and invite human verification where AI is weak

AI Capability Awareness

AI is strong at verifying:

• API endpoints (especially with before/after DB state)
• Backend logic and data transformations
• Test execution and result parsing

AI is weak at verifying:

• Visual design and layout
• UI component placement and spacing
• Spatial relationships ("does X overlap Y")

The plugin adapts confidence and checkpoint behavior based on these strengths/weaknesses.

Workflow Overview

/uw:plan -> /uw:work -> /uw:review -> /uw:compound
    |           |            |             |
  Spec.md   Checkpoints   Code Review   Learnings
            + Verify.md   + Fix Loop    to memory
                              |
                          Create PR

Decision Trees

See decision-trees.md for detailed decision flows:

• When to checkpoint
• When to ask user vs decide autonomously
• When to use each memory operation
• Which verification subagent to use

Templates

• Spec Template - Feature specification format
• Verify Template - Checkpoint verification document
• Learnings Template - Compound phase output
• Test Plan Template - Manual test plan from diff analysis

Directory Structure

Unit Work creates this structure in your project:

.unitwork/
├── specs/           # {DD-MM-YYYY}-{feature}.md
├── verify/          # {DD-MM-YYYY}-{n}-{name}.md
├── review/          # Code review findings
├── learnings/       # Compound phase output
└── test-plans/      # {DD-MM-YYYY}-{feature}.md

Agent Behavior Rules

Interview Phase

See interview-workflow.md for the complete interview protocol including confidence-based depth assessment and stop conditions.

1. Research before asking - check Hindsight, codebase, then web docs
2. Group related questions - don't ask one at a time
3. Push back on scope - "That sounds like a separate feature"
4. Advocate once - state recommendation, then accept decision
5. No premature solutions - don't propose implementation during requirements
6. Confirm understanding - summarize before writing spec

Implementation Phase

1. Follow the spec - it's the contract
2. Minimal changes - smallest diff that satisfies requirement
3. No drive-by refactoring - note tech debt, don't fix it
4. No drive-by bug fixes - unless critical and blocking
5. Checkpoint at boundaries - every verifiable unit
6. Document uncertainty - say so in checkpoint
7. Never skip verification - even if confident

Verification Phase

1. Tests first - always run relevant tests
2. API safety - never call mutating endpoints without permission
3. Screenshot everything - UI changes always get screenshots
4. Explicit confidence - state percentage with rationale
5. Human QA reproducible - no LLM-dependent steps
6. Precursor state documented - if testing needs setup, document how

Memory Rules

CRITICAL: Memory recall is the foundation of compounding. Skip it and you lose all accumulated learnings.

0. NEVER skip memory recall at session start - this is non-negotiable
1. Always async - never block on memory writes
2. Always contextualized - include repo name and work context
3. Document tracking - group by feature doc-id
4. Narrative format - write as natural language
5. Retain discoveries immediately - don't wait for phase end
6. Blind spots are critical - always retain when human finds what agent missed

Checkpoint Protocol

See checkpointing.md for the complete checkpointing reference including:

• Checkpoint Commit Format - Standard commit message format
• When to Checkpoint - Decision tree in decision-trees.md
• Verification Document - Template at templates/verify.md
• Self-Correcting Review - Protocol for fix checkpoints

Hindsight Integration

See hindsight-reference.md for complete patterns including:

• Bank Name Derivation - Worktree-safe bank name extraction
• ANSI Stripping - Required when processing output programmatically
• Memory Operations - Recall, retain, and reflect patterns
• Error Handling - Graceful degradation when Hindsight unavailable

Quick Reference

# Bank name (config override → git remote → worktree → pwd)
BANK=$(jq -re '.bankName // empty' .unitwork/.bootstrap.json 2>/dev/null || git config --get remote.origin.url 2>/dev/null | sed 's/.*\///' | sed 's/\.git$//' || basename "$(git worktree list 2>/dev/null | head -1 | awk '{print $1}')" || basename "$(pwd)")

# Recall with ANSI stripping
hindsight memory recall "$BANK" "query" --budget mid --include-chunks 2>&1 | sed 's/\x1b\[[0-9;]*m//g'

# Retain (always async)
hindsight memory retain "$BANK" "narrative" --context "context" --doc-id "id" --async

Context7 Integration

Context7 provides framework documentation lookup via MCP. Use it when implementing unfamiliar APIs.

Usage

Step 1: Resolve library ID
mcp__unitwork_context7__resolve-library-id
  query: "what you're trying to implement"
  libraryName: "framework-name"

Step 2: Query documentation
mcp__unitwork_context7__query-docs
  libraryId: "/org/project"  (from step 1)
  query: "specific API or pattern"

When to Use

• Unfamiliar framework APIs (check docs before guessing)
• Best practices for specific patterns
• Version-specific behavior differences
• Implementation examples from official docs

When NOT to Use

• Project-specific patterns (use Hindsight)
• Simple/well-known APIs
• Already checked docs this session

Verification Subagents

Subagent/Skill	Purpose	When to Use
test-runner	Execute tests	Changed test files or tested code
api-prober	Probe API endpoints	Changed API endpoints
/uw:browser-test (command)	UI verification	Changed UI components

Review Agents (Parallel)

Agent	Focus
type-safety	Casting, guards, nullability
patterns-utilities	Existing solutions, duplication
performance-database	N+1, indexes, parallelization
architecture	Structure, coupling, boundaries
security	Injection, auth, data exposure
simplicity	Over-engineering, YAGNI
memory-validation	Learnings from Hindsight memory

Confidence Assessment

Start at 100%, subtract:

• -5% for each untested edge case
• -20% if UI layout changes
• -10% if complex state management
• -15% if external API integration

>= 95%: Checkpoint and continue < 95%: Checkpoint and pause for human review

Commands

• /uw:plan - Interview and create spec
• /uw:work - Execute with checkpoints
• /uw:review - Parallel code review
• /uw:compound - Extract learnings
• /uw:bootstrap - First-time setup
• /uw:pr - Create/update GitHub PRs
• /uw:action-comments - Resolve PR comments
• /uw:fix-ci - Autonomously fix failing CI
• /uw:fix-conflicts - Intelligent rebase conflict resolution
• /uw:test-plan - Generate manual testing steps from git diffs