siftcoder Workflow Skill

Full autonomous multi-agent workflow for implementing features.

Description

This skill orchestrates the complete siftcoder workflow: Planning → Coding → QA Review → QA Fix. It manages agent handoffs, state transitions, and auto-continuation.

When to Use

This skill is automatically invoked when:

•A feature implementation starts
•Auto-continuation triggers next subtask
•QA review requires fixes
•Workflow needs to resume

Instructions

You are the siftcoder workflow orchestrator. You coordinate multiple agents to implement features autonomously.

Workflow States

code

┌─────────────────────────────────────────────────────────────┐
│                    siftcoder WORKFLOW                       │
├─────────────────────────────────────────────────────────────┤
│                                                             │
│  ┌──────────┐    ┌──────────┐    ┌──────────┐    ┌──────┐  │
│  │ PLANNING │ -> │ CODING   │ -> │ QA       │ -> │ DONE │  │
│  └──────────┘    └──────────┘    └──────────┘    └──────┘  │
│       │               │               │                     │
│       v               v               v                     │
│   [Planner]       [Coder]        [Reviewer]                │
│                                      │                      │
│                                      v                      │
│                                 [QA Fixer]                  │
│                                 (if needed)                 │
│                                                             │
└─────────────────────────────────────────────────────────────┘

Phase 1: Planning

•
Load Feature Read from features.json to get feature details.

•

Invoke Planner Agent

code

Task: Create implementation plan for feature [ID]

Context:
- Feature: [title]
- Description: [description]
- Acceptance Criteria: [criteria]

Output: Structured plan with subtasks

•
Store Plan Save plan to feature's subtasks in features.json.
•
Transition Update state to coding, move to first subtask.

Phase 2: Coding (Per Subtask)

•
Load Subtask Get current subtask from feature.

•

Invoke Coder Agent

code

Task: Implement subtask [ID]

Context:
- Subtask: [title]
- Feature: [parent feature]
- Previous subtasks: [completed subtasks]

Constraints:
- Follow project patterns
- Write tests
- Run quality gates

•
Quality Gates Automatic via PostToolUse hooks:
- •Format code
- •Run linter
- •Type check
•
Mark Complete Update subtask status to completed.
•
Auto-Continue Stop hook triggers should-continue.sh:
- •If more subtasks: continue to next
- •If all subtasks done: transition to QA

Phase 3: QA Review

•

Invoke QA Reviewer

code

Task: Validate feature [ID] implementation

Context:
- Feature: [title]
- Acceptance Criteria: [criteria]
- Files Modified: [list]

Checks:
- All acceptance criteria met
- Tests pass
- No regressions
- Code quality acceptable

Output: PASS or FAIL with issues

•
Handle Result
- •PASS: Transition to completion
- •FAIL: Transition to QA Fix

Phase 4: QA Fix (If Needed)

•

Invoke QA Fixer

code

Task: Fix issues from QA review

Issues:
[List of issues from reviewer]

Constraints:
- Fix ONLY the listed issues
- Don't refactor unrelated code
- Re-run affected tests

Output: Fixes applied

•
Re-Review Return to QA Review phase.
•
Iteration Limit Max 3 QA iterations. If still failing:
- •Mark feature as needs_human_review
- •Log issues
- •Notify user

Phase 5: Completion

•
Update State
- •Mark feature as completed
- •Update features.json
- •Log to implementation-log.jsonl
•
Create Checkpoint
- •Git commit with descriptive message
- •Store ref in checkpoints
•
Update Knowledge
- •Store any new patterns discovered
- •Log any gotchas encountered
•
Sync to Cloud (Pro Tier)
- •Check if cloud sync is configured
- •Check if auto-sync is enabled
- •
  If both true:
  bash
  # Push updated knowledge to cloud siftcoder/scripts/cloud-sync.sh push
- •Display sync status in completion report

•

Report

code

✅ FEATURE COMPLETE: [ID]

Subtasks: 5/5 completed
QA Iterations: 1
Files Modified: 8
Tests Added: 12
Time: 15 minutes

Summary: [brief description of what was implemented]

•
Auto-Continue to Next Feature If more features in queue and autonomous mode:
- •Load next pending feature
- •Start from Planning phase

State Machine

json

{
  "states": {
    "pending": { "next": "planning" },
    "planning": { "next": "coding" },
    "coding": { "next": "qa_review" },
    "qa_review": {
      "next": "completed",
      "fail": "qa_fix"
    },
    "qa_fix": { "next": "qa_review" },
    "completed": { "terminal": true },
    "needs_human_review": { "terminal": true }
  }
}

Error Handling

•
Agent Failure
- •Log error
- •Retry up to 2 times
- •If still failing, pause and notify user
•
Infinite Loop Detection
- •Track QA issue hashes
- •If same issue appears 3 times, break loop
- •Mark as needs_human_review
•
Context Overflow
- •If context getting large, summarize completed work
- •Start fresh agent with summary

Pause/Resume

Pause:

•Save current state to current-task.json
•Stop auto-continuation
•User can use /siftcoder:status to check

Resume:

•Load state from current-task.json
•Resume from last phase/subtask
•Continue auto-continuation

Configuration

Read from .claude/siftcoder-state/config.json:

json

{
  "mode": "autonomous",
  "autoContinue": true,
  "maxIterations": 10,
  "maxQARetries": 3,
  "autoCommit": true,
  "agentModels": {
    "planner": "sonnet",
    "coder": "sonnet",
    "qaReviewer": "sonnet",
    "qaFixer": "sonnet"
  },
  "cloudSync": {
    "enabled": false,
    "autoSyncOnKnowledgeChange": false,
    "autoSyncOnFeatureComplete": false
  }
}

Cloud Sync Settings

When cloudSync.enabled = true:

•Session Start: Execute cloud-sync.sh pull to get latest knowledge
•Knowledge Add: Trigger cloud-sync.sh push (if autoSyncOnKnowledgeChange = true)
•Feature Complete: Trigger cloud-sync.sh push (if autoSyncOnFeatureComplete = true)

Check cloud configuration at ~/.config/siftcoder/cloud.toml before sync operations.

Runtime Implementation

This skill includes a minimal skill.ts entry point to satisfy plugin requirements. The primary value remains in this documentation - see sections above for:

•Workflow orchestration patterns
•State management
•Agent coordination

The runtime entry point can be extended with actual functionality as needed.

Allowed Tools

Read, Write, Edit, Bash, Glob, Grep, Task (for subagents)