CDD Pipeline

This project calls its release workflow Context Driven Development (CDD)

Workspace Configuration

All CDD commands and scripts require the $CDD_DIR environment variable.

export CDD_DIR=/path/to/workspace

This allows parallel workspaces without VCS conflicts. Each workspace is independent - different branches can use different workspace directories simultaneously.

Scripts will abort with a clear error if $CDD_DIR is not set.

Core Philosophy

Use user/agent interactive development time during the Research and Plan phases to prepare tightly focused tasks that pre-load the exact required context to run autonomously during the Implement phase.

Pipeline Phases

• Research
• Plan
• Execute

1. Research

What: Build foundational understanding before design decisions.

Artifacts:

• $CDD_DIR/README.md - High-level description of release features
• $CDD_DIR/user-stories/ - User actions and expected system responses
• $CDD_DIR/research/ - Facts gathered from internet, docs, codebase

Why this phase exists: Decisions made without research are assumptions. Research surfaces constraints, prior art, and edge cases that would otherwise appear during execution (when fixing is expensive).

Key insight: Research artifacts become reference material for task execution. Sub-agents can read $CDD_DIR/research/ without re-fetching from the internet. The research phase pays the lookup cost once.

2. Plan

What: Transform research into implementation decisions and executable task files.

Artifacts:

• $CDD_DIR/plan/ - Implementation decisions, architecture choices, detailed designs
• $CDD_DIR/tasks/ - Individual task files (one per unit of work)
• $CDD_DIR/tasks/order.json - Execution order and task metadata

Why this phase exists: Planning is the translation layer between "what we want" (research) and "what to do" (tasks). It makes architectural decisions explicit and breaks work into parallelizable units.

Key insight: Task files must be complete. During execution, sub-agents see only their task file. Any context not written into the task must be re-discovered (wasting tokens) or will be missed (creating bugs).

3. Verify (Sub-phase of Plan)

What: Two-stage iterative verification following the artifact pyramid.

Artifacts:

• $CDD_DIR/verified.md - Log of items verified and issues resolved

Stage 1 - Plan Verification:

/cdd:gap-plan  # Verify README/user-stories → plan alignment
fix gaps
/cdd:gap-plan  # Repeat until clean

Checks alignment down the pyramid:

• README/research/user-stories → plan
• Naming, error handling, memory management conventions
• Integration points, build changes, migrations

Stage 2 - Task Verification:

/cdd:gap-tasks  # Verify plan → tasks alignment
fix gaps
/cdd:gap-tasks  # Repeat until clean

Checks task completeness:

• All plan items covered by tasks
• Task sequencing and dependencies
• TDD workflow, test coverage, baseline skills
• Task scope and acceptance criteria

Why this phase exists: This is the last moment of full context visibility. Once execution begins, each sub-agent sees only its task. Cross-cutting issues, inconsistencies between tasks, missing dependencies - these can only be caught while the complete picture is visible.

Key insights:

• Specific checklists prevent vague "find issues" prompts
• verified.md prevents re-checking completed items
• Token budgets (140k per command) enable thorough verification
• Pyramid approach: verify each transition separately (foundation→plan, plan→tasks)

Convergence pattern: Each gap-finding run identifies the highest-priority issue. After fixing, re-run. When no substantive gaps remain, verification is complete.

4. Execute

What: Mechanical execution of task files via orchestration loop.

Process:

/orchestrate → /check-quality → /refactor → /check-quality

Why this structure: Execution is unattended. No human to provide missing context. Task files must be complete or sub-agents will fail, research (wasting tokens), or succeed partially (creating debt).

Key insight: Orchestrators load ZERO skills. They're pure execution loops. Sub-agents read task files directly - task content never flows through orchestrator context.

Directory Structure

$CDD_DIR/
├── README.md          # High-level release description
├── user-stories/      # User actions and system responses
│   └── README.md      # Index/overview
├── research/          # Internet facts for task reference
│   └── README.md      # Index/overview
├── plan/              # Implementation decisions
│   └── README.md      # Index/overview
├── tasks/             # Individual task files
│   └── order.json     # Execution order
├── verified.md        # Verification log (concerns + resolutions)
├── tmp/               # Temp files during execution
└── ...                # Other files permitted, ignored by pipeline

Abstraction Levels

Each directory serves a distinct abstraction level. Content should not leak between levels.

$CDD_DIR/plan/ - Coordination Layer

Plan documents coordinate everything shared between tasks. If two tasks must agree on something, the plan decides it.

The plan owns:

• Public symbols - Function names, struct names, enum names
• Function signatures - Arguments, argument order, return types
• Struct members - Field names, field order, field types
• Enums - Value names and meanings
• Allowed libraries - Which dependencies can be used, which are forbidden
• Architectural choices - Module boundaries, data flow, ownership rules

Why this matters: Tasks execute independently. If the plan doesn't coordinate shared symbols, Task A might define foo_create(ctx, name, count) while Task B expects foo_create(ctx, count, name). Both compile. Integration fails.

The plan is the contract. Tasks implement what the plan specifies. Disagreements are resolved in the plan before tasks are written.

DO NOT include: Implementation code, detailed algorithms, function bodies.

$CDD_DIR/tasks/ - Implementation Units

Task files specify what this task implements from the plan:

• Which functions from the plan this task implements
• Which structs from the plan this task defines
• Expected behaviors and edge cases
• Test scenarios for this task's scope

Tasks do NOT make coordination decisions. If a task needs a function signature, struct layout, or enum value - it comes from the plan. Tasks copy the relevant specs from the plan and add implementation guidance.

DO NOT include: Implementation code, function bodies, working test code.

The sub-agent writes the actual code. The task tells them what to write and how to test it.

Why this separation matters: Tasks execute independently without seeing each other. The plan is their shared reference. If Task A invents a struct layout instead of copying from the plan, Task B (which also uses that struct) will have a different layout. The plan prevents this.

Artifact Authority

Artifacts form a hierarchy of authority:

$CDD_DIR/README.md (authoritative)
        ↓
user-stories + research (derived)
        ↓
      plan (derived)
        ↓
      tasks (derived)

Rules:

1. Lower-level artifacts derive from higher-level ones
2. Contradictions between levels must be resolved immediately
3. Resolution is always top-down: align lower artifacts to higher authority
4. Higher-level documents may change through iteration—but when they do, all derived artifacts must be reviewed and realigned

Why this matters: Sub-agents executing tasks cannot detect contradictions with higher-level intent. If a task contradicts the plan, or the plan contradicts the README, the error propagates silently. Consistency is enforced during authoring because it cannot be enforced during execution.

Lifecycle

Create empty $CDD_DIR/ → Research → Plan → Verify → Execute → Delete $CDD_DIR/

The workspace directory is ephemeral. It exists only for the duration of the release workflow. After successful execution, it's deleted. The work lives in the codebase; the pipeline artifacts are disposable.

Efficiency Principles

1. Orchestrators load ZERO skills - Pure execution loops need only embedded logic
2. Sub-agents read task files - Never pass task content through orchestrator context
3. Load skills on-demand - Don't preload "just in case"
4. Complete task authoring - Spend tokens researching during authoring so execution doesn't need to
5. Reference vs working knowledge - Large docs go in separate skills, loaded when needed

Design Tradeoffs

Decision	Tradeoff	Rationale
Full context in verify	Expensive per-run	Last chance to catch cross-cutting issues
Task files are complete	Verbose, some duplication	Sub-agents can't see other tasks
Verification is human-terminated	Subjective	Models don't converge to "done" naturally
Research cached in files	Stale if release spans days	Avoids re-fetching during execution

When to Load This Skill

• Designing or modifying the release workflow
• Debugging why execution sub-agents are failing
• Optimizing token usage in the pipeline
• Understanding why a phase exists before changing it