CFN Loop Orchestration Skill

Metadata

• Skill ID: cfn-loop-orchestration
• Version: 3.2.0
• Category: Workflow Orchestration
• Maturity: Production
• Last Updated: 2026-07-03

Supported Entry Point (verified on disk)

The ONLY supported entry is:

$HOME/.claude/skills/cfn-loop-orchestration-v2/cli/orchestrate.sh

It routes to cli/resolve-provider-model.cjs via Node. Nothing else in this skill directory is a supported entry point.

In Task Mode (/cfn-loop-task) do NOT use Redis, helpers/, or dist/cli; agents return output directly. CLI/Redis sections apply only to /cfn-loop-cli mode.

The following entry points that earlier versions of this document referenced do NOT exist on disk and must not be invoked or cited:

• dist/cli/orchestrator-cli.js (missing)
• helpers/*.sh (gate-check, consensus, iteration-manager, deliverable-verifier, timeout-calculator: directory missing)
• .claude/skills/redis-coordination/ (skill removed; Redis helpers now live in .claude/skills/cfn-memory-persistence/lib/redis/)

Renamed, not missing: the old .claude/skills/cfn-loop-orchestration/test-cfn-orchestration.sh is now test-cfn-orchestration.sh in this directory.

For the mechanical test gate, use the verified script:

$HOME/.claude/skills/cfn-loop-orchestration-v2/cli/gate-check.sh \
  --out <test-output-file> --threshold <decimal>
# exit 0 = pass, 1 = fail, 2 = no tests detected (0/0 never passes)

Purpose

Orchestrates the Complete Fail Never (CFN) Loop workflow, managing the three-loop structure:

• Loop 3 (Primary Swarm - Implementation)
• Loop 2 (Consensus Validators - Review)
• Product Owner Decision (Strategic Approval)

Responsibilities

1. Coordinate multi-agent CFN Loop execution
2. Manage gate checks and consensus validation
3. Handle iteration cycles with feedback injection
4. Execute Product Owner decision flow
5. Enforce dependency ordering (Loop 3 -> Loop 2 -> PO)

CFN Loop Flow

1. Spawn Loop 3 Agents
   |
2. Collect Loop 3 results (Task Mode: direct output; CLI mode: provider runtime)
   |
3. Gate Check (cli/gate-check.sh on coordinator's test output)
   |- PASS -> Start Loop 2
   |- FAIL -> Iterate Loop 3 with feedback (goto step 1)
   |
4. Loop 2 Validates Loop 3 Work
   |
5. Collect Loop 2 consensus
   |
6. Product Owner Decision
   |- PROCEED -> Exit success
   |- ITERATE -> Iterate (goto step 1)
   |- ABORT -> Exit failure

Mode-Specific Thresholds

Single source of truth: .claude/skills/cfn-loop-orchestration-v2/THRESHOLDS.md. Do not restate values here; read that file. Summary of column meanings:

• testpassrate_gate: passing/total from the coordinator's authoritative test run (decimal). The only gate metric in Task Mode.
• confidence_gate: agent self-reported confidence, CLI mode only, computed by the mechanical rubric in lib/validation/SKILL.md.
• consensus: validators voting PASS / validators spawned.
• max_iter: iteration cap per mode.

Error Handling

Recoverable Failures

• Gate check failure: iterate Loop 3 with failing-test excerpts as feedback
• Consensus failure: iterate all agents
• Missing deliverables: force iteration with explicit feedback

Critical Failures

• Max iterations exceeded: report failure and exit
• Agent spawn failure: retry once, then escalate

Success Criteria

This skill is considered successful when:

1. CFN Loop slash commands (/cfn-loop-task, /cfn-loop-cli) work without modification
2. Gate decisions are made only by cli/gate-check.sh output, never by inline parsing
3. Thresholds are read from THRESHOLDS.md, never hardcoded