| name | auto-dev |
| description | Direct development with automated review. Plans implementation, executes via subagent, then auto-reviews. Use for work that isn't test-driven (UI, config, styling, infrastructure, scaffolding). |
| argument-hint | <feature description> [--review quick|std|max] [--max-iterations N] [--spec <path>] [--report] |
Auto Dev
Build a feature directly, then auto-review the result: plan → implement → review-fix loop.
Same structure as auto-tdd but without the TDD workflow — for work where test-first development isn't practical (UI design, configuration, styling, scaffolding, infrastructure, templates, etc.).
- Step 1 plans the implementation in the main context and gets user approval
- Step 2 executes the approved plan via subagent (context isolation)
- Step 3 runs
/decaf-quality:auto-code-review on the result
Argument Parsing
Parse $ARGUMENTS:
- Feature description (required): Everything that isn't a flag — describes what to build
- Review mode:
--review quick|std|max (default: std) — passed to auto-review
- Max review iterations:
--max-iterations N (default: 3) — passed to auto-review
- Spec path:
--spec <path> — passed to auto-review for spec compliance checking
--report: passed to auto-review, which produces a comparison-grade session report for skill tuning (@../../conventions/session-report.md). When set, this skill contributes the implementation-phase record: the implementation subagent's harness-reported usage (tokens / tool calls / duration, verbatim from the Agent tool result), changeset stats (files changed, +/− lines, new files), and a one-line scope description.
Execution Steps
Step 1: Plan (Main Context)
This step runs in the main context so the user can interact with the plan.
1a. Detect project context:
- Identify the project language(s), framework(s), and build system
- Identify the build/verify command if applicable (e.g.,
dotnet build, npm run build, go build ./..., cargo build)
- Explore the codebase to understand existing modules, patterns, conventions, and similar features
- Note project structure, naming conventions, and existing patterns to follow
1b. Design the implementation plan:
Based on the feature description and codebase exploration:
- Identify the files to create or modify
- List implementation steps — ordered so each step builds on the last and produces a working (or at least compilable) state
- Identify patterns from existing code that should be followed
- Identify the first tracer bullet (the one step that proves the approach works end-to-end)
- Note any dependencies, prerequisites, or setup needed
1c. Present plan and ask for approval:
## Auto Dev Plan
**Feature**: {feature description}
**Language**: {language} | **Framework**: {framework} | **Build command**: {command or "N/A"}
### Implementation Steps
1. 🎯 {tracer bullet step} ← tracer bullet
2. {next step}
3. {next step}
...
### Files
| Action | File | Purpose |
|--------|------|---------|
| Create | src/components/Foo.tsx | New component |
| Modify | src/routes.ts | Register new route |
| ... | ... | ... |
### Patterns to Follow
{Existing conventions and patterns identified in the codebase, or "None — new area"}
After implementation, `/auto-review` will run in **{reviewMode}** mode (max {maxIterations} iterations).
Ask via AskUserQuestion:
question: "Review the plan above. Approve, adjust steps, or change the approach?"
Wait for user response. Apply adjustments if any.
Step 2: Implementation (Subagent)
Launch a general-purpose subagent using the Agent tool to execute the approved plan.
Subagent prompt template:
Implement the following pre-approved plan. The user has already reviewed and approved it — do NOT ask for confirmation. Just implement.
Feature: {feature description}
Language: {language} | Framework: {framework} | Build command: {buildCommand or "N/A"}
Approved implementation steps:
- 🎯 {tracer bullet step} ← start here
- {step 2}
- {step 3}
...
Files to create/modify:
{files table from approved plan}
Patterns to follow:
{patterns from approved plan}
Execution rules:
- Implement each step in order
- After each step, verify the project still builds/compiles (run
{buildCommand} if available)
- If a step fails to compile, fix the issue before moving on
- Follow existing project conventions and the patterns listed above
- If the code you write is testable and the project has test infrastructure, feel free to add tests — but this is not required and you do not need to follow a red-green-refactor cycle
- Do NOT refactor unrelated code
When complete, report:
- Summary of steps completed (which succeeded, which had issues)
- List of all files created or modified
- Build verification result (pass/fail)
Wait for the subagent to complete. Record results. With --report, also record the subagent's harness-reported usage from the tool result and the changeset stats — this becomes the implementation-phase record for the session report.
Report to the user:
## Implementation Complete
**Steps completed**: {X}/{total}
**Files modified**: {list}
**Build**: {passing / N/A}
Starting auto-review...
If the subagent reports that it could not complete any steps (total failure), ask the user whether to proceed with auto-review anyway or stop.
Step 3: Auto Review
Run /decaf-quality:auto-code-review using the Skill tool, passing through the review arguments:
/decaf-quality:auto-code-review {reviewMode} --max-iterations {maxIterations} {--spec specPath if provided} {--report if set}
Auto-review will automatically detect the scope from uncommitted changes (which includes everything the implementation subagent produced). With --report, the implementation-phase record from Step 2 is in this context — auto-review's session report picks it up for its agent inventory and token accounting.
The auto-review skill manages its own subagent lifecycle — let it run to completion.
Step 4: Final Summary
After auto-review completes, present a combined summary:
## Auto Dev Complete
### Implementation Phase
**Steps**: {X}/{total} completed
**Files**: {list of files created/modified}
**Build**: {passing / N/A}
### Review Phase
{Summary from auto-review — iterations, findings fixed, deferred, skipped}
### Result
{One-line overall assessment: e.g., "Feature implemented in N steps, M review findings addressed."}
Notes
- The implementation subagent gets a fresh context window — this allows complex features without exhausting the main context
- The planning step stays in the main context so the user can interact naturally
- Auto-review runs in the main context because it manages its own subagent lifecycle
- If the user didn't specify
--review, default to std mode
- This skill does not require TDD — use
/auto-tdd when test-first development is appropriate
- Tests are welcome if the code is testable, but not mandatory — no red-green-refactor cycle is enforced
- The review phase runs on uncommitted changes and may briefly mutate the working tree (a reviewer's non-destructive revert-probe, or a fix that fails verification).
auto-code-review takes a recoverable snapshot (git stash create) before its fix/probe phase so nothing is lost; for a stronger guarantee, commit the implementation before review.