orchestration

Multi-Task Orchestration

Execute orchestration from GitHub issues or epics with parallel agents, dependency resolution, consensus reviews, and automatic remediation using contextd for memory, checkpoints, and context folding.

Prerequisites: contextd REQUIRED

This skill CANNOT function without contextd.

Orchestration depends on:

• branch_create / branch_return for context isolation (CRITICAL)
• checkpoint_save / checkpoint_resume for state management
• memory_record for learning capture

If contextd unavailable:

1. STOP - inform user: "Orchestration requires contextd for context isolation"
2. Suggest: "Run /contextd:init to configure contextd"
3. NO FALLBACK - this skill is inoperable without contextd

Shared Orchestration Patterns

This skill builds on shared orchestration patterns. See:

• includes/orchestration/parallel-execution.md - Agent dispatch and concurrency
• includes/orchestration/result-synthesis.md - Collecting and merging results
• includes/orchestration/context-management.md - Context folding and memory
• includes/orchestration/checkpoint-patterns.md - Save/resume workflows

The patterns below are contextd-specific extensions for issue-driven orchestration.

Orchestration Flow

digraph orchestration {
    "Issues provided?" -> "Prompt user" [label="no"];
    "Issues provided?" -> "Fetch issues" [label="yes"];
    "Prompt user" -> "Fetch issues";
    "Fetch issues" -> "Create feature branch";
    "Create feature branch" -> "Resolve dependencies";
    "Resolve dependencies" -> "For each group";
    "For each group" -> "Create context branch" [label="parallel"];
    "Create context branch" -> "Launch task agents";
    "Launch task agents" -> "Collect results";
    "Collect results" -> "Run consensus review";
    "Run consensus review" -> "Remediate findings?" [shape=diamond];
    "Remediate findings?" -> "Apply fixes" [label="yes"];
    "Remediate findings?" -> "Save checkpoint" [label="no"];
    "Apply fixes" -> "Run consensus review" [label="re-review"];
    "Save checkpoint" -> "More groups?" [shape=diamond];
    "More groups?" -> "For each group" [label="yes"];
    "More groups?" -> "RALPH reflection" [label="no"];
    "RALPH reflection" -> "Create PR";
    "Create PR" -> "Final summary";
}

Agents

Agent	Purpose
contextd:orchestrator	This agent - manages workflow
contextd:task-agent	Executes individual tasks

Contextd Tools Required

Memory: memory_search, memory_record, memory_consolidate Checkpoints: checkpoint_save, checkpoint_resume, checkpoint_list Context Folding: branch_create, branch_return, branch_status Remediation: remediation_record, remediation_search Reflection: reflect_analyze, reflect_report

Phase 0: Input Resolution

If no issues provided, prompt the user:

AskUserQuestion(
  questions: [{
    question: "Which issues or epic would you like to orchestrate?",
    header: "Issues",
    options: [
      { label: "Enter issue numbers", description: "Comma-separated list (e.g., 42,43,44)" },
      { label: "Select an epic", description: "Epic issue that contains sub-issues" },
      { label: "Current milestone", description: "All open issues in current milestone" }
    ],
    multiSelect: false
  }]
)

For "Current milestone":

gh issue list --milestone "$(gh api repos/:owner/:repo/milestones --jq '.[0].title')" --json number,title
-> Present list for confirmation

Phase 1: Issue Discovery

1. Fetch issue details:
   gh issue view <number> --json number,title,body,labels,milestone

2. For epics (single issue with sub-issues):
   gh api graphql -f query='{ repository(owner:"X", name:"Y") {
     issue(number: N) { trackedIssues(first: 50) { nodes { number title } } }
   }}'

3. Extract task information:
   - number -> Task ID
   - title -> Task name
   - body -> Agent prompt (look for ## Agent Prompt or ## Description)
   - labels -> Priority (P0, P1, P2), type (feature, bug, etc.)
   - "Depends On: #XX" in body -> Dependencies

4. Record to memory:
   memory_record(title: "Orchestration: Issues #{list}", ...)

Phase 1.5: Branch Setup (MANDATORY)

NEVER push directly to main. ALWAYS create a feature branch first.

1. Generate branch name from epic/issue:
   branch_name = "feature/issue-{epic_number}-{sanitized_title}"
   Example: "feature/issue-123-locomo-benchmark"

2. Create and checkout feature branch:
   git checkout -b {branch_name}

3. Verify branch:
   git branch --show-current
   → Must NOT be "main" or "master"

4. If already on main with uncommitted changes:
   git stash
   git checkout -b {branch_name}
   git stash pop

Why this matters:

• Enables code review before merge
• Provides rollback capability
• Maintains audit trail
• Allows CI/CD validation

Phase 2: Initialization

1. Read engineering practices:
   Read("CLAUDE.md"), Read("engineering-practices.md")

2. Search past orchestrations:
   memory_search(project_id, "orchestration", limit: 5)

3. If resuming:
   checkpoint_list(session_id: "orchestrate-{issue_ids}")
   checkpoint_resume(checkpoint_id)

4. Create main context branch:
   branch_create(description: "Orchestration: #{issue_ids}", budget: 16384)

5. Save initial checkpoint:
   checkpoint_save(name: "orchestrate-start")

Phase 3: Dependency Resolution

1. Build dependency graph from issue relationships
2. Generate parallel groups (topological sort)
3. Validate no circular dependencies

Example:
  #42 depends on nothing -> Group 1
  #43 depends on nothing -> Group 1
  #44 depends on #42 -> Group 2
  #45 depends on #43, #44 -> Group 3

Phase 4: Group Execution

See includes/orchestration/parallel-execution.md for agent dispatch patterns.

For each dependency group:

1. Create context branch (budget: 8192)
   branch_create(description: "Group {n}: #{issue_numbers}")

2. Launch parallel task agents for issues in group:
   Task(
     subagent_type: "contextd:task-agent",
     prompt: |
       # Issue #{number}: {title}

       {issue_body}

       ## Contextd Integration
       - Record decisions with memory_record
       - Record fixes with remediation_record
       - Update issue with progress comments via `gh issue comment`
     description: "Issue #{number}: {title}",
     run_in_background: true
   )

3. Monitor and collect results:
   TaskOutput(task_id, block=false)
   branch_status(branch_id) -> check budget

4. Return from branch:
   branch_return(message: "Group complete: {summary}")

Phase 5: Consensus Review

See includes/orchestration/result-synthesis.md for result collection and conflict resolution. See includes/orchestration/consensus-review.md for consensus patterns and thresholds.

100% Consensus Requirement

Task orchestration requires 100% consensus from all reviewers:

Requirement	Threshold
Consensus Score	100%
Vetoes	0 (any veto blocks)
Critical/High	0

If any reviewer vetoes or finds critical/high issues:

1. Create remediation task
2. Fix the issues
3. Re-run full review
4. Continue until 100% consensus

After each group:

1. Launch review agents in parallel:
   Task(subagent_type: "fs-dev:security-reviewer", ...)
   Task(subagent_type: "fs-dev:vulnerability-reviewer", ...)
   Task(subagent_type: "fs-dev:code-quality-reviewer", ...)
   Task(subagent_type: "fs-dev:documentation-reviewer", ...)
   Task(subagent_type: "fs-dev:go-reviewer", ...)  # if Go code

2. Collect verdicts using result synthesis patterns
   - Parse findings from each reviewer
   - Detect conflicts between reviewers
   - Record to memory_record

3. Calculate consensus:
   consensus_score = (approvals / total_reviewers) * 100

4. Apply veto threshold:
   - strict: ALL findings must be fixed (100% consensus required)
   - standard: CRITICAL/HIGH must be fixed (security/vulnerability have veto power)
   - advisory: Log only, continue

5. If consensus < 100% OR any veto:
   - Create remediation tasks for all blocking findings
   - Fix issues
   - Re-run full review
   - Loop until 100% consensus

Phase 6: Remediation

If findings require fixes:

1. Search past remediations:
   remediation_search(query: "{finding.type}")

2. Apply fixes, run tests

3. Record new remediation:
   remediation_record(error_signature, root_cause, solution)

4. Re-run review if needed

Phase 7: Checkpoint

See includes/orchestration/checkpoint-patterns.md for checkpoint naming and retention.

After each group passes review:

checkpoint_save(
  name: "orchestrate-group-{n}-complete",
  summary: "Completed: #{issues}, Remaining: #{remaining}"
)

Phase 8: RALPH Reflection

After all groups complete:

1. Analyze patterns:
   reflect_analyze(project_id, period_days: 1)

2. Consolidate similar learnings:
   memory_consolidate(similarity_threshold: 0.8)

3. Generate report:
   reflect_report(format: "markdown")

Phase 9: Final Summary & PR Creation

NEVER push directly to main. ALWAYS create a pull request.

1. Return from main branch:
   branch_return(message: "Orchestration complete: {metrics}")

2. Push feature branch to remote:
   git push -u origin {branch_name}

3. Create pull request:
   gh pr create \
     --title "feat: {epic_title}" \
     --body "## Summary
   Orchestrated implementation of #{epic_number}.

   ## Issues Completed
   - #{issue_list}

   ## Consensus Review
   All groups passed {threshold} review.

   ## Test Results
   {test_summary}"

4. Update issues with PR link:
   gh issue comment <number> --body "Implementation PR: #{pr_url}"

5. Record final memory:
   memory_record(title: "Orchestration Complete", outcome: "success")

6. Save final checkpoint:
   checkpoint_save(name: "orchestrate-complete")

Do NOT close issues until PR is merged.

Issue Body Format

Issues should include an agent prompt section:

## Description
Brief description of what this issue accomplishes.

## Agent Prompt
Detailed instructions for the task agent.

## Acceptance Criteria
- [ ] Criterion 1
- [ ] Criterion 2

## Dependencies
Depends On: #42, #43

Review Thresholds

Threshold	Behavior
strict	100% findings addressed before proceeding
standard	Security/vulnerability vetoes block, others advisory
advisory	Report only, continue execution

Resume Capability

/contextd:orchestrate --resume "group-2-complete"

-> Loads checkpoint state
-> Skips completed groups
-> Continues from saved point

Anti-Patterns

See shared anti-patterns in includes/orchestration/ files.

Issue-specific anti-patterns:

Pattern	Problem	Solution
Push directly to main	Bypasses review, no rollback	ALWAYS create feature branch + PR
Skip input prompt	User confusion	ALWAYS use AskUserQuestion if no issues
Skip pre-flight	Miss past learnings	ALWAYS search memory first
Monolith execution	No isolation	Use context branches per group
Skip remediation recording	Knowledge lost	ALWAYS record fixes
Over-budget branches	Context overflow	Monitor with branch_status
No GitHub comments	Lost visibility	Update issues with progress
Ignore dependencies	Execution order wrong	Parse "Depends On" in issue body
Close issues before merge	Premature closure	Wait for PR merge to close issues

---

Resource Monitoring

See includes/orchestration/context-management.md for base budget patterns.

Issue Group Budget Tracking

Track budget per dependency group using branch_status:

{
  "per_group_budget": {
    "group_1": { "allocated": 10000, "used": 8500, "issues": [42, 43] },
    "group_2": { "allocated": 10000, "used": 6200, "issues": [44] },
    "group_3": { "allocated": 10000, "used": 0, "issues": [45] }
  }
}

Branch-Based Monitoring

Every 5 minutes OR after each task:
  1. branch_status(branch_id) -> get token usage
  2. Compare to thresholds (70% warning, 90% critical)
  3. If warning: checkpoint_save(auto_created: true)
  4. If critical: pause, notify, wait for user

---

Concurrency Limits

See includes/orchestration/parallel-execution.md for base concurrency patterns.

Issue-Driven Concurrency

For GitHub issue orchestration, concurrency is managed per dependency group:

{
  "issue_concurrency": {
    "max_parallel_agents": 4,
    "max_per_group": 2,
    "queue_overflow": "wait",
    "timeout_minutes": 30,
    "priority_labels": ["P0", "P1", "P2"]
  }
}

Priority override: Issues with P0 label execute before P1/P2 when queue is full.

---

Dead Letter Queue for Failed Tasks

See includes/orchestration/parallel-execution.md for base retry and error handling patterns.

Issue-Specific DLQ

Failed issues are tracked with GitHub integration:

{
  "dlq_entry": {
    "task_id": "task_abc123",
    "issue_number": 42,
    "failure_reason": "timeout after 30 minutes",
    "error_message": "Agent did not complete within budget",
    "retry_count": 2,
    "max_retries": 3,
    "github_comment": true
  }
}

DLQ Escalation

Failures	Action
3+	Notify user
5+	Pause orchestration, await confirmation
10+	Abort, add comment to epic issue

Issue Failure Correlation

dlq_analyze() for issues:
  - Correlated issues (same dependency failing)
  - Label patterns (bug issues failing more)
  - Recommendations (increase budget, split issues)

---

Unified Memory Type References

Tag orchestration artifacts with standard types:

Artifact	Tag	Purpose
Orchestration start	type:decision, category:planning	Track scope
Task completion	type:learning, category:execution	What worked
Task failure	type:failure, category:execution	What failed
Remediation applied	type:remediation, category:fix	Error patterns
Review findings	type:pattern, category:quality	Code patterns

---

Hierarchical Namespace Guidance

Orchestration Namespaces

<org>/<project>/orchestrations/<orchestration_id>

Examples:
  fyrsmithlabs/contextd/orchestrations/epic-42
  fyrsmithlabs/marketplace/orchestrations/v1.6-release

Nested Task Namespaces

<orchestration_namespace>/groups/<group_id>/tasks/<task_id>

Example:
  fyrsmithlabs/contextd/orchestrations/epic-42/groups/1/tasks/issue-43

---

Audit Fields

All orchestration records include:

Field	Description	Auto-set
created_by	Orchestrator session	Yes
created_at	Start timestamp	Yes
completed_at	End timestamp	Yes
duration_seconds	Total execution time	Yes
task_count	Number of tasks	Yes
success_count	Completed tasks	Yes
failure_count	Failed tasks	Yes

---

Claude Code 2.1 Patterns

See includes/orchestration/parallel-execution.md for base Task tool patterns.

Issue-to-Task Mapping

Map GitHub issue dependencies to Task dependencies:

# Group 1 (parallel - no issue dependencies)
task_42 = Task(prompt: "Execute issue #42", run_in_background: true)
task_43 = Task(prompt: "Execute issue #43", run_in_background: true)

# Group 2 (issue #44 depends on #42)
task_44 = Task(prompt: "Execute issue #44", addBlockedBy: [task_42.id])

# Group 3 (issue #45 depends on #43 and #44)
task_45 = Task(prompt: "Execute issue #45", addBlockedBy: [task_43.id, task_44.id])

Orchestration Safety Hooks

Prevent dangerous operations during issue execution:

{
  "hook_type": "PreToolUse",
  "tool_name": "Bash",
  "condition": "orchestration_active AND command.matches('git push --force|rm -rf')",
  "prompt": "Dangerous operation during orchestration. Checkpoint first and confirm with user."
}

Issue Completion Hook

Auto-record task completions and update GitHub:

{
  "hook_type": "PostToolUse",
  "tool_name": "Task",
  "condition": "task_completed AND orchestration_active",
  "prompt": "Task completed. Record to memory_record, update issue with progress comment."
}

---

Event-Driven State Sharing

Orchestration emits events for other skills:

{
  "event": "orchestration_started",
  "payload": {
    "orchestration_id": "epic-42",
    "issues": [42, 43, 44, 45],
    "groups": 3,
    "estimated_duration": "2 hours"
  },
  "notify": ["workflow", "consensus-review"]
}

Subscribe to orchestration events:

• orchestration_started - Orchestration began
• group_started - Group execution began
• task_started - Individual task began
• task_completed - Task finished (success or failure)
• task_failed - Task moved to DLQ
• group_completed - All tasks in group done
• review_triggered - Consensus review started
• orchestration_paused - Budget/failure pause
• orchestration_completed - All groups done