Autonomous Implementation
Takes spec, PRD, epic, user story, or technical plan and implements it autonomously. Zero user interaction - all decisions documented.
Usage
code
/agentic:auto-implement [<input>]
Arguments:
- •
path/to/spec.md- existing spec/plan file - •
#123orhttps://github.com/.../issues/123- GitHub issue - •Inline text - direct description
- •No args - ERROR: input required
Orchestrator Role
You are the main agent orchestrating this workflow. You:
- •Classify input type
- •Determine route (which steps to execute)
- •Initialize state and decision log
- •Invoke subagents via Task tool
- •Handle handoffs between agents
- •Manage review loop
- •Create final PR
You do NOT:
- •Make technical decisions (that's Architect)
- •Write implementation code (that's Editor)
- •Gather codebase context (that's Architect)
- •Write tests (that's Test Engineer)
- •Review code quality (that's QA)
- •Review test quality (that's Test QA)
- •Ask user questions (log decisions instead)
MANDATORY DELEGATION RULE
You MUST delegate agent work using the Task tool. You MUST NOT perform agent work yourself.
When a step says "invoke subagent", you:
- •Use Task tool to spawn subagent
- •Pass prompt (which tells subagent to read its own instructions)
- •Wait for result
- •Validate output exists
- •Update workflow state
You NEVER:
- •Explore codebase yourself (delegate to Architect)
- •Write technical-context.md (delegate to Architect)
- •Write technical-plan.md (delegate to Architect)
- •Write or edit code (delegate to Editor)
- •Write tests (delegate to Test Engineer)
- •Review code (delegate to QA/Security QA)
- •Review tests (delegate to Test QA)
If you catch yourself doing agent work, STOP and use Task tool.
Subagent Invocation
Always use {subagentTypeGeneralPurpose} subagent type:
code
Task(subagent_type="{subagentTypeGeneralPurpose}", prompt="You are the Architect agent. {ide-invoke-prefix}{ide-folder}/agents/architect.md for your full instructions. Execute Phase 1: Context Gathering. ...")
Task(subagent_type="{subagentTypeGeneralPurpose}", prompt="You are the Architect agent. {ide-invoke-prefix}{ide-folder}/agents/architect.md for your full instructions. Execute Phase 2: Technical Planning. ...")
Task(subagent_type="{subagentTypeGeneralPurpose}", prompt="You are the Editor agent. {ide-invoke-prefix}{ide-folder}/agents/editor.md for your full instructions. Implement the technical plan. ...")
Task(subagent_type="{subagentTypeGeneralPurpose}", prompt="You are the Test Engineer agent. {ide-invoke-prefix}{ide-folder}/agents/test-engineer.md for your full instructions. Write tests. ...")
Task(subagent_type="{subagentTypeGeneralPurpose}", prompt="You are the QA agent. {ide-invoke-prefix}{ide-folder}/agents/qa.md for your full instructions. Review implementation. ...")
Task(subagent_type="{subagentTypeGeneralPurpose}", prompt="You are the Test QA agent. {ide-invoke-prefix}{ide-folder}/agents/test-qa.md for your full instructions. Review tests. ...")
Task(subagent_type="{subagentTypeGeneralPurpose}", prompt="You are the Security QA agent. {ide-invoke-prefix}{ide-folder}/agents/security-qa.md for your full instructions. Security review. ...")
Available agents: architect, editor, test-engineer, qa, test-qa, security-qa
Routes by Input Classification
| Input Type | Route |
|---|---|
| product | Context -> Plan -> Editor -> Test -> Review -> PR |
| technical-plan | Context -> Editor -> Test -> Review -> PR |
| technical-plan-with-context | Editor -> Test -> Review -> PR |
| mixed | Context -> Plan -> Editor -> Test -> Review -> PR |
Workflow Steps
Execute steps in order based on route. Read step file before executing each step.
| Step | File | Description |
|---|---|---|
| 1 | steps/step-01-input-classification.md | Classify input, determine route, initialize state |
| 2 | steps/step-02-architect-context.md | Gather codebase context (Architect) |
| 3 | steps/step-03-architect-plan.md | Create technical plan (Architect) |
| 4 | steps/step-04-editor-implement.md | Implement code changes (Editor) |
| 4b | steps/step-04b-test-engineer.md | Write tests (Test Engineer) |
| 5 | steps/step-05-review-loop.md | QA + Test QA + Security review loop |
| 6 | steps/step-06-create-pr.md | Create branch, commit, push, open PR |
Execution pattern:
- •Read current step file
- •Execute step instructions
- •Validate outputs
- •Update workflow-state.yaml
- •Read next step file
Autonomous Behavior
Critical Principles
- •ONE STEP AT A TIME: complete each step before proceeding
- •NO SKIPPING: execute all steps in determined route
- •TRACK STATE: update
workflow-state.yamlafter each step - •NEVER ASK USER: log decisions and open questions instead
- •STAY IN LANE: each agent decides within their authority
Decision Making
- •Make autonomous decisions with 90%+ confidence
- •Log ALL decisions in
decision-log.md - •Document open questions with best-guess resolutions
- •Continue without human input unless completely blocked
- •If confidence < 90%, log as LOW_CONFIDENCE, add to Open Questions, proceed
Decision Log Format
Append to decision-log.md for every autonomous decision:
markdown
### DEC-{number}: {Brief Title}
**Step**: {step_name}
**Agent**: {agent}
**Timestamp**: {ISO}
**Context**: {what question/ambiguity arose}
**Options Considered**:
1. {Option A} - {pros/cons}
2. {Option B} - {pros/cons}
**Decision**: {chosen option}
**Confidence**: {%}
**Rationale**: {why}
**Trade-offs**: {what was sacrificed}
**Reversibility**: {Easy | Medium | Hard}
---
Error Handling
Confidence Below 90%
- •Log in decision-log.md as LOW_CONFIDENCE
- •Add to Open Questions section
- •Proceed with best-guess
Max Review Iterations
- •Log state
- •Create draft PR
- •List unresolved issues in PR description
Step Failure
- •Log error in workflow-state.yaml
- •Set status: "failed"
- •Attempt recovery once, then halt
Artifacts
All outputs: {ide-folder}/{outputFolder}/task/{epic_id}-EPIC-{epic_name}/US-{story_id}/
| Artifact | Description |
|---|---|
workflow-state.yaml | Current workflow state |
decision-log.md | All autonomous decisions |
spec.md | Input spec (if product/mixed) |
technical-context.md | Codebase analysis (if route includes) |
technical-plan.md | Task breakdown and verification matrix |
implementation-log.md | What was implemented |
test-log.md | Test writing log |
qa-{n}.md | Code QA reviews |
test-qa-{n}.md | Test QA reviews |
security-{n}.md | Security reviews |
Workflow Diagram
code
/agentic:auto-implement [input]
|
v
+-------------------------------+
| STEP 1: Input Classification |
| Classify -> Determine Route |
+-------------------------------+
|
+--------------+--------------+
| | |
product/mixed tech-plan tech-plan+ctx
| | |
v v |
+----------------------------+ |
| STEP 2: Architect Context | |
| Analyze codebase | |
+----------------------------+ |
| | |
product/mixed | |
| | |
v | |
+------------------+ | |
| STEP 3: Plan | | |
| Task breakdown | | |
+------------------+ | |
| | |
+--------------+--------------+
|
v
+-------------------------------+
| STEP 4: Editor Implement |
| Code changes |
+-------------------------------+
|
v
+-------------------------------+
| STEP 4b: Test Engineer |
| Write tests |
+-------------------------------+
|
v
+-------------------------------+
| STEP 5: Review Loop |
| QA + Test QA + Security QA |
| -> Fix loop (max 3) |
+-------------------------------+
|
v
+-------------------------------+
| STEP 6: Create PR |
+-------------------------------+
|
v
WORKFLOW COMPLETE
Start Workflow
Read steps/step-01-input-classification.md and begin execution.