Sdd Next

SDD-Next: Concise Playbook

Overview

• Purpose: Repeatable workflow for discovering specs, selecting actionable tasks, and keeping the user in the loop from planning through wrap-up.
• Scope: Single-task pulls and phase-focused loops; work happens inside project root.
• Audience: AI agents orchestrating SDD workflows who must respect human checkpoints and CLI guardrails.

High-Level Flow Diagram

Start
  |
  v
Read Work Mode Config (Step 0)
  |
  +-- single mode --> Discover Specs -> Gather Context -> Select Task
  |                     |                         |
  |                     |                     Alternatives?
  |                     |                     /         \
  |                     |            yes -> Browse   no -> Prepare Recommended Task
  |                     |                     \         /
  |                     v                      v       v
  |                   Draft Plan -> Seek Approval -> Implementation Handoff
  |                     |
  |                     v
  |                   Post-Implementation Checklist -> Surface Next -> Finish
  |
  +-- autonomous mode --> Phase Loop (auto-complete all tasks) -> Finish

Step 0: Read Work Mode Configuration

CRITICAL: This must be the FIRST step when sdd-next is invoked.

Before doing anything else, read the work mode from the user's configuration:

sdd get-work-mode --json

Expected output:

{"work_mode": "single"}

or

{"work_mode": "autonomous"}

Routing based on work_mode:

• If "single": Follow Single Task Workflow (Sections 3.1-3.6)
  • Plan and execute one task at a time with explicit user approval
  • After task completion, surface next recommendation and wait for user decision
• If "autonomous": Follow Autonomous Mode Workflow (Section starting at line 438)
  • Complete all tasks in current phase automatically within context limits
  • Check context after EVERY task completion
  • Stop only for blockers, plan deviations, or when context ≥85%

CRITICAL: Global Requirements & Conventions

Working Directory & Commands

• Stay inside repo root and keep commands one-per-line (no && chaining)
• Remember spec folders map to lifecycle (specs/pending, specs/active, specs/completed)

⚠️ CRITICAL: Spec Reading Rules (NEVER VIOLATE)

ALWAYS use sdd commands to read spec JSON files:

✅ sdd prepare-task {spec-id}
✅ sdd task-info {spec-id} {task-id}
✅ sdd progress {spec-id}

NEVER use these tools/commands on spec JSON:

❌ Read(/path/to/spec.json)          # Wastes 10,000+ tokens (specs are 50KB+)
❌ cat specs/active/spec.json        # Bypasses validation and hooks
❌ head specs/active/spec.json       # Wastes context
❌ jq '.tasks' spec.json             # Bypasses error handling
❌ grep "task-1" spec.json           # Inefficient, error-prone

Why: Spec files are large JSON (machine-readable, not human-readable). Direct reading wastes valuable context tokens and bypasses built-in validation, error handling, and hooks.

User Interaction Requirements

Context Gathering Best Practices

Default workflow (stick to this unless spec says otherwise)

1. Run sdd prepare-task with no flags. The returned context block already includes the previous sibling, parent metadata, phase progress, sibling files, the latest journal summary, and file-focused documentation context (when available via context.file_docs).
2. Only call sdd task-info, sdd get-task, or sdd progress if the spec explicitly asks for extra metadata or you need files that are not exposed in context.
3. After completing the task, re-run sdd prepare-task to surface the next recommendation and refreshed context.

When to use enhancement flags (extended_context)

• --include-full-journal: You need the full journal history for the previous sibling (long-running refactors, nuanced design notes).
• --include-phase-history: You are preparing phase summaries or retrospectives and need every entry tied to the current phase.
• --include-spec-overview: You must report spec-wide progress without running sdd progress.
• Combine flags only when you plan to read the extra data—each flag increases the JSON payload.

Decision guide

• Need additional detail beyond context?
  • Yes → Task metadata/files: sdd task-info {spec_id} {task_id} (or sdd get-task if the spec mentions nested metadata).
  • Yes → Journal history: Use --include-full-journal (previous sibling) or sdd get-journal for arbitrary tasks.
  • Yes → Phase backlog/alternatives: sdd query-tasks --parent {phase_id} when presenting options to the user.
  • No: Stick with prepare-task output; avoid redundant commands.

Anti-patterns to avoid

• Running task-info, check-deps, and get-task back-to-back "just in case." The default prepare-task response now includes all dependency details in context.dependencies, eliminating the need for check-deps in 95% of cases. Call these commands only when context is insufficient for special requirements.
• Re-running sdd progress or sdd list-phases after every plan change. Use context.phase for quick updates and run sdd progress only before reporting global status.
• Fetching the entire spec or invoking doc-query before inspecting the prepare-task payload.

Command Value Matrix

Gate key decisions with AskUserQuestion (MANDATORY):

• Spec selection (when multiple available)
• Task selection (recommended vs alternatives)
• Plan approval (before implementation)
• Blocker handling (alternative tasks or resolve)
• Completion verification (for verify tasks)

Anti-Pattern: Never use text-based numbered lists like "1. Option A, 2. Option B". Always use AskUserQuestion tool for structured choices.

⚠️ CRITICAL: Context Checking Pattern

Before checking context, you MUST generate a session marker first.

This is a two-step process that must run sequentially:

Step 1: Generate Session Marker (REQUIRED FIRST)

sdd session-marker

Step 2: Check Context Using the Marker

sdd context --session-marker "SESSION_MARKER_<hash>"

Output Format:

{"context_percentage_used": 78}

CRITICAL REQUIREMENTS

✅ Run as TWO SEPARATE Bash tool calls (never combine) ✅ Run SEQUENTIALLY, not in parallel (step 2 depends on step 1 being logged) ❌ NEVER combine with && or $() - The marker must be logged to transcript first ❌ NEVER run in parallel - Step 2 will fail if step 1 hasn't been logged

Context Thresholds

• < 85%: Safe to continue
• ≥ 85%: Stop and recommend to the user that they /clear and then /sdd-begin for the next task

CRITICAL: Never Anticipate Context Usage

⚠️ ONLY check actual context percentage – NEVER speculate about future consumption:

❌ DO NOT stop early because:

• "Phase 2 implementation will consume context"
• "File reading tasks are coming"
• "I predict I'll hit 85% during the next phase"
• "I should have buffer space for future work"

✅ DO ONLY stop when:

• Context is CURRENTLY at or above 85%
• You have JUST checked context and confirmed actual usage

Rationale: Predicting context usage is unreliable and defeats the purpose of checking. The threshold (85%) is designed to give adequate headroom; stopping earlier wastes that safety margin.

When to Check Context

• Autonomous mode: After EVERY task completion (REQUIRED)
• Single-task mode: After task completion (recommended)
• Session start: Before intensive work (optional)

Work Mode Behavior

When sdd-next is invoked, it automatically reads the work_mode setting from .claude/sdd_config.json (see Step 0) and routes to the appropriate workflow:

Single Task Mode ("work_mode": "single") - Default

• Follows Sections 3.1–3.6 below
• Plan and execute one task at a time with explicit user approval
• After task completion, surface next recommendation and wait for user decision
• User maintains full control over which tasks to execute and when

Autonomous Mode ("work_mode": "autonomous")

• Follows dedicated Autonomous Mode section (later in this document)
• Complete all tasks in current phase automatically within context limits
• Check context after EVERY task completion (required)
• Stop only for blockers, plan deviations, or when context ≥85%
• Continues until phase is complete or manual intervention needed

Single Task Workflow

Use this workflow when the configured work mode is Single Task Mode ("work_mode": "single" in config). Execute one task at a time with explicit user approval for each step.

3.1 Choose the Spec

• If user supplies spec id, confirm it exists via sdd progress {spec-id}
• Otherwise list candidates: sdd find-specs
• Apply recommendation heuristic:
  • Prefer status: active with non-zero progress (started but incomplete)
  • If multiple qualify, pick highest completion % or with in_progress tasks
  • If none have progress, pick most recently touched active spec
• Surface recommendation explicitly: tag as (Recommended) in AskUserQuestion
• Present options via AskUserQuestion (include "Other / provide id")

3.2 Gather High-Level Context

• Use sdd progress {spec-id} and sdd list-phases {spec-id} for status summary
• Highlight objectives, blockers, completion percentages
• Offer additional context commands (sdd list-blockers, sdd render) only on request

3.3 Select the Task

• Ask via AskUserQuestion: accept recommended task or browse alternatives?
• Recommendation path: sdd prepare-task {spec-id} → surface task id, file, estimates, blockers
• Browsing path: Use sdd query-tasks {spec-id} (filter --parent, --status) + sdd list-blockers {spec-id} → present shortlist via AskUserQuestion

3.4 Deep Dive & Plan Approval

Gather every detail with a single call (omit {task-id} to accept the recommended task):

sdd prepare-task {spec-id} {task-id}

That response already contains everything you need:

• task_data → title, metadata, instructions pulled from the spec
• dependencies → top-level blocking status (can_start, blocked_by list)
• context → stitched data from the previous sibling, parent task, current phase, sibling files, task journal, AND detailed dependency information (context.dependencies) with full task titles, statuses, and file paths

Treat context as the authoritative source rather than chaining sdd task-info, sdd check-deps, and sdd get-task. Typical fields:

"context": {
  "previous_sibling": {
    "task_id": "task-3-1-2",
    "title": "Tighten plan creation language",
    "summary": "Updated scope guardrails for Section 3.3"
  },
  "parent_task": {
    "task_id": "task-3-1",
    "title": "Polish the planning workflow",
    "position_label": "Phase 3 · Task 1"
  },
  "phase": {
    "name": "Implementation",
    "percentage": 58,
    "blockers": []
  },
  "sibling_files": [
    {"path": "skills/sdd-next/SKILL.md", "reason": "Touched by previous sibling"}
  ],
  "task_journal": {
    "entry_count": 0,
    "entries": []
  },
  "dependencies": {
    "blocking": [],
    "blocked_by_details": [
      {
        "id": "task-2-3",
        "title": "Update context gathering",
        "status": "in_progress",
        "file_path": "src/context.py"
      }
    ],
    "soft_depends": []
  }
}

• context.previous_sibling: reference recent work for continuity or reuse its journal summary when explaining why the new task matters (context.previous_sibling.title).
• context.parent_task: verify how this subtask fits into the backlog; use context.parent_task.position_label to show progress.
• context.phase: surface phase health (context.phase.percentage, context.phase.blockers) without calling sdd progress.
• context.sibling_files: prime file navigation by reviewing whatever the spec already touched before opening new files.
• context.task_journal: access journal entries for this task showing decision history and status changes without separate calls.
• context.dependencies: detailed dependency information with task titles, statuses, and file paths for blocking (tasks this blocks), blocked_by_details (tasks blocking this), and soft_depends (soft dependencies)—eliminates need for separate sdd check-deps call in 95% of cases.

Only fall back to sdd task-info or sdd check-deps when the spec explicitly calls for metadata that is not surfaced through the standard payload.

Draft the execution plan around the spec intent, dependency gates, and the insights above. Example:

1. Confirm previous edits in context.sibling_files to maintain consistent tone.
2. Align deliverables with context.parent_task.title/position_label.
3. Call out open risks or blockers via context.phase.
4. Reference context.previous_sibling.summary if you need to explain how the work continues an earlier change.

Present plan and get approval via AskUserQuestion:

• Options: "Approve & Start", "Request Changes", "More Details", "Defer"
• Handle response appropriately

If recommended task is blocked, pause for guidance or loop back to task selection.

3.5 Implementation Handoff

Before coding:

sdd update-status {spec-id} {task-id} in_progress --note "context"

During implementation:

• Follow execution plan
• Document any deviations immediately

Using sdd doc scope --implement During Implementation:

When documentation is available and you need detailed implementation context for a file, use:

sdd doc scope <file-path> --implement

This provides implementation-focused context including:

• Detailed function signatures and parameters
• Full implementation logic and patterns
• Code examples from the actual file
• Dependencies and imports
• Usage examples and patterns

When to use scope --implement:

• Starting implementation of a task - Get comprehensive context before writing code
• Understanding existing patterns - See how current code works before extending it
• Refactoring tasks - Review full implementation details before restructuring
• Complex file modifications - Need deep understanding of current implementation
• Following established patterns - Extract patterns from existing code to maintain consistency

When NOT to use scope --implement:

• During planning phase - Use scope --plan instead (lighter context)
• Quick edits or trivial changes - Direct file reading may be faster
• Documentation unavailable - Fall back to Read tool
• Context limits approaching - Avoid heavy payloads near 85% threshold

Example workflow:

# 1. Task started, need implementation context
sdd doc scope src/services/auth.ts --implement

# 2. Review detailed implementation patterns and signatures
# (command returns comprehensive implementation context)

# 3. Implement changes following discovered patterns
# 4. Mark task complete with journal entry

Optimization tips:

• Use scope --implement at task start, not repeatedly during coding
• Cache insights from output rather than re-running
• Switch to targeted Read calls for specific line ranges if context is tight

After implementation:

Mark task complete using sdd-update subagent (atomically marks complete + creates journal):

Task(
  subagent_type: "sdd-toolkit:sdd-update-subagent",
  prompt: "Complete task {task-id} in spec {spec-id}. Completion note: [Summary of what was accomplished, tests run, verification performed].",
  description: "Mark task complete"
)

Journal content must include:

• What was accomplished
• Tests run and results
• Verification performed
• Files created/modified
• Any deviations from plan

3.6 Surface Next Recommendation

Immediately after completion:

sdd prepare-task {spec-id}

• Summarize next task's scope and blockers
• Check with user before proceeding
• If no pending work or spec complete, surface that clearly and confirm next steps

⚠️ CRITICAL: Completion Requirements

When Marking Tasks Complete

ONLY mark a task as completed when you have FULLY accomplished it.

Never Mark Complete If:

❌ Tests are failing ❌ Implementation is partial ❌ You encountered unresolved errors ❌ You couldn't find necessary files or dependencies ❌ Blockers exist that prevent verification

If Blocked or Incomplete:

✅ Keep task as in_progress ✅ Create new task describing what needs resolution ✅ Document blocker using sdd-update subagent ✅ Present alternatives to user via AskUserQuestion

Resolving Blocked Tasks:

When a task has been marked as blocked and the blocker is later resolved:

sdd unblock-task {spec-id} {task-id} --resolution "Brief description of how blocker was resolved"

Example:

sdd unblock-task feature-auth-001 task-3-2 --resolution "API endpoint now available in staging environment"

This marks the task as unblocked and ready to proceed. The task will then appear in sdd prepare-task recommendations.

Completion Journal Requirements

MUST provide journal content describing:

• What was accomplished
• Tests run and results
• Verification performed
• Any deviations from plan
• Files created/modified

Example:

Task(
  subagent_type: "sdd-toolkit:sdd-update-subagent",
  prompt: "Complete task task-2-3 in spec my-spec-001. Completion note: Implemented JWT auth middleware with PKCE flow. All 12 unit tests passing. Manual verification: login flow works in dev environment. Created src/middleware/auth.ts (180 lines) and tests/middleware/auth.spec.ts (45 tests).",
  description: "Mark task-2-3 complete"
)

3.7 Journal vs Git History

Anti-pattern: Running git log / git show to understand a recently completed SDD task when the journal already documents the work. That wastes time and risks contradicting the canonical record. Start with the spec journal; escalate to git history only if a fact is missing or you are diagnosing repo-level issues (rebases, conflicts, regressions).

Journal advantages

• Full implementation narrative (what changed and why) tied to task_id.
• Test and verification results in one place, ready for audits.
• Deviations, blockers, and next-task hints captured while they are fresh.
• Structured JSON makes it trivial for sdd prepare-task to surface the latest context without extra commands.
• Archives context even if commits are squashed or rebased later.

⚠️ CRITICAL: Verification Tasks

Detecting Verification Tasks

Check task metadata for type: verify or verification_type field:

sdd task-info {spec-id} {task-id}

Dispatch by Verification Type

Automated Tests (verification_type: "auto")

Task(
  subagent_type: "sdd-toolkit:run-tests-subagent",
  prompt: "Run tests for {task-id} in spec {spec-id}. Execute tests and handle failures.",
  description: "Run tests"
)

After tests complete:

1. Present findings to user
2. Use AskUserQuestion to get approval before marking complete
3. Options: "Accept & Complete", "Fix Failures", "Review Details"

Fidelity Review (verification_type: "fidelity")

Task(
  subagent_type: "sdd-toolkit:sdd-fidelity-review-subagent",
  prompt: "Review {scope} '{target}' in spec {spec-id}. Compare completed tasks against requirements.",
  description: "Fidelity review for {scope}"
)

After review completes:

1. Present fidelity report
2. Use AskUserQuestion for decision:
   • "Accept & Complete" - Mark verification complete, journal deviations
   • "Revise Implementation" - Reopen parent tasks for fixes
   • "Update Spec" - Document accepted deviations

Manual Review (verification_type: "manual")

Present checklist from task metadata to user for confirmation via AskUserQuestion.

Post-Implementation Checklist

After completing a task:

• Task status updated (in_progress → completed) with journal entry
• Follow-up commands or monitoring notes captured in journal
• Blockers or deviations surfaced to user; next steps agreed
• Context check performed (two-step pattern above)
• Next recommended task retrieved via sdd prepare-task {spec-id} and shared with user
• Spec context refreshed via sdd progress {spec-id} for reporting

Autonomous Mode (Phase Completion)

Use this workflow when the configured work mode is Autonomous Mode ("work_mode": "autonomous" in config). If the user changes the config to Single Task Mode mid-session, switch to Section 3 (Single Task Workflow).

When to Use

• Config file has "work_mode": "autonomous" set
• User wants to complete multiple tasks in current phase without per-task approval
• User has sufficient context headroom (check context before starting)

Key Characteristics

• Phase-scoped: Completes all tasks within current phase only (does not cross phase boundaries)
• Context-aware: Checks context after EVERY task, stops if ≥85%
• Defensive stops: Stops for blocked tasks and plan deviations (requires user approval)
• No plan approval: Creates execution plans internally without showing user

Autonomous Workflow Loop

Step 1: Task Execution Loop

For each task in current phase:

1. Prepare next task:

   sdd prepare-task {spec-id}
   

2. Check phase complete: If no more tasks in current phase → Exit loop

3. Check for blockers:

   • If next task blocked: STOP
   • Present blocker info via AskUserQuestion
   • Options: alternative tasks, resolve blocker, or stop
   • Exit autonomous mode

4. Create execution plan (silently):

   • Analyze task metadata from prepare-task output
   • Create detailed internal plan (no user approval needed)
   • Include all standard components: prerequisites, steps, success criteria

5. Mark task in_progress:

   sdd update-status {spec-id} {task-id} in_progress
   

6. Execute implementation according to internal plan

7. Handle plan deviations:

   • If implementation deviates: STOP
   • Document deviation
   • Present to user via AskUserQuestion
   • Options: revise plan, update spec, explain more, rollback
   • Exit autonomous mode

8. Mark task complete:

   Task(
     subagent_type: "sdd-toolkit:sdd-update-subagent",
     prompt: "Complete task {task-id} in spec {spec-id}. Completion note: [Brief summary of what was accomplished, tests run, verification performed].",
     description: "Mark task complete"
   )
   

9. ⚠️ CRITICAL: Check context usage (REQUIRED):

   Run two-step pattern as SEPARATE, SEQUENTIAL Bash calls:

   # First call:
   sdd session-marker
   

   # Second call (only after first completes):
   sdd context --session-marker "SESSION_MARKER_<hash>"
   

   Check ACTUAL context percentage reported, do NOT speculate:

   • If context ACTUALLY ≥85% (as reported by command): STOP, exit loop, go to Summary
   • If context <85%: Continue to next iteration

   CRITICAL: Do not stop based on predictions like "upcoming work will use context" or "I should have buffer space for the next phase." Only stop when actual usage reaches the threshold. The 85% threshold already provides safety margin.

10. Check phase completion:

    • If current phase complete: Exit loop, go to Summary
    • Otherwise: Return to step 1

Step 2: Present Summary Report

When autonomous mode exits:

## Autonomous Execution Summary

**Mode:** Phase Completion (Autonomous)
**Spec:** {spec-title} ({spec-id})
**Phase:** {phase-title} ({phase-id})

### Tasks Completed
✅ task-1-1: [title] - Duration: X min
✅ task-1-2: [title] - Duration: Y min

### Phase Progress
Phase {phase-id}: {completed}/{total} tasks ({percentage}%)
Overall: {total_completed}/{total_tasks} tasks ({overall_percentage}%)

### Context Usage
Current context: {context_percentage}%

### Exit Reason
{One of:}
- ✅ Phase Complete
- ⏸️ Context Limit: ≥85% threshold
- 🚧 Blocked Task
- ⚠️ Plan Deviation
- ❌ No Actionable Tasks

### Next Steps
{Contextual recommendations based on exit reason}

Autonomous Mode Best Practices

DO:

• ✅ Check context after EVERY task completion
• ✅ Stop immediately when context ≥85%
• ✅ Base stopping decisions on ACTUAL context percentage, never predictions
• ✅ Use the full safety margin (continue until ≥85% reported)
• ✅ Stop for blocked tasks (don't auto-pivot)
• ✅ Stop for plan deviations (don't auto-revise)
• ✅ Create detailed internal plans
• ✅ Present comprehensive summary at end

DON'T:

• ❌ Cross phase boundaries
• ❌ Skip plan creation (always plan, just don't show)
• ❌ Continue past 85% context
• ❌ Stop early based on predictions of future context usage
• ❌ Auto-resolve blockers
• ❌ Auto-revise plans on deviations
• ❌ Batch task completions

Phase Loop with Human Checkpoints

Scope Confirmation

Show sdd list-phases {spec-id} with progress. Ask via AskUserQuestion:

• Focus on target phase
• Adjust scope
• Revert to single-task mode

Queue Preparation

Prime backlog:

sdd query-tasks {spec-id} --parent {phase-id} --status pending

If queue empty or blocked:

sdd list-blockers {spec-id}

Pause for user direction.

Task Loop

Reuse Single Task Workflow (steps 3.3–3.6) for each pending task.

After each completion:

• Refresh phase: sdd check-complete {spec-id} --phase {phase-id}
• If granted "auto-continue for this phase", note permission but still report blockers immediately

Phase Wrap-Up

Summarize results:

sdd progress {spec-id}
sdd query-tasks {spec-id} --parent {phase-id}

Present accomplishments, verification outcomes, blockers.

Ask via AskUserQuestion: continue to next phase, perform phase review, or stop.

Troubleshooting

Spec File Not Found / Path Errors

Cause: Wrong working directory or relative paths

Solution:

• Provide absolute path: sdd prepare-task {spec-id} --path /absolute/path/to/specs
• Run sdd find-specs to discover available specs

All Tasks Blocked

Diagnosis:

sdd list-blockers {spec-id}
sdd check-complete {spec-id}

Solution: Present alternatives via AskUserQuestion or resolve blockers before continuing

Quick Reference

Core Commands

# Discovery
sdd find-specs
sdd progress {spec-id}
sdd list-phases {spec-id}

# Task Selection
sdd prepare-task {spec-id}              # Primary command - includes all context
sdd next-task {spec-id}                 # Simpler alternative - just task ID
sdd task-info {spec-id} {task-id}       # Rarely needed - only for non-recommended tasks
sdd check-deps {spec-id} {task-id}      # Rarely needed - now in context.dependencies

# Context Checking (TWO STEPS)
sdd session-marker
sdd context --session-marker "SESSION_MARKER_<hash>"

# Advanced
sdd query-tasks {spec-id} --status pending --parent {phase-id}
sdd list-blockers {spec-id}
sdd unblock-task {spec-id} {task-id} [--resolution "reason"]
sdd check-complete {spec-id} --phase {phase-id}

Critical Patterns Summary