open-design-orchestrator

Orchestrates the Open Design pipeline from evidence collection through policy-gated merge.

Updated Aug 22, 2026
One-click install
npx skills add https://github.com/xegheplimo-web/hermes-ops --skill open-design-orchestrator-xegheplimo-web
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: open-design-orchestrator
Source: https://github.com/xegheplimo-web/hermes-ops/tree/main/skills/software-development/open-design-orchestrator
Command: npx skills add https://github.com/xegheplimo-web/hermes-ops --skill open-design-orchestrator-xegheplimo-web

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes scripts (resource) and references (resource) components.

What problem does it solve? Coordinating a governed software change pipeline—evidence collection, external review, task decomposition, agent dispatch, CI verification, and policy gating—requires many manual steps that are easy to skip or execute inconsistently. This Skill runs the full Open Design loop as a deterministic state machine so every change follows the same auditable path from user request to merge. ## Core Features & Use Cases - End-to-End State Machine: Runs stages from evidence collection through conflict detection, external review, task DAG decomposition, Devin dispatch, and a fail-closed policy gate returning PASS, REPAIR, ESCALATE, or BLOCK. - Bounded Repair Loop: Re-dispatches failed work to Devin with a cost- and attempt-bounded repair budget (max 3 attempts by default), escalating to a human when limits are hit. - Evidence-Bound Policy Gate: Builds a canonical EvidenceManifest v1 tied to the head SHA and CI status, so merges cannot pass on unverified or fabricated evidence. - Use Case: Ask Hermes to implement a feature on a real repository; the Skill collects repo evidence, gets an adversarial Codex review, decomposes tasks into the Ops DB, dispatches Devin, checks CI, and only allows merge after the policy gate passes. ## Quick Start Run the Open Design loop on my repository in dry-run mode with a mock reviewer and stop after the policy gate stage.

Frequently Asked Questions about open-design-orchestrator

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I run the Open Design orchestration loop on a repository?

Run scripts/open_design.py with --repo pointing at your repository, --out for the run artifact directory, --reviewer set to mock, codex, or openai, and --dispatch-mode dry-run for testing. Use --stop-after policy_gate to halt before dispatch and merge stages.

What does the policy gate do in this pipeline?

The hermes-policy-gate CLI evaluates a canonical EvidenceManifest v1 bound to the head SHA and CI status, returning PASS, REPAIR, ESCALATE, or BLOCK. It is fail-closed: missing or unverified CI evidence maps to failure, so merges cannot proceed on fabricated results.

What prerequisites are required before running this orchestrator?

You need Python 3.11+, git in PATH, PostgreSQL with Ops DB migrations applied and DATABASE_URL set, an authenticated Devin CLI for real dispatch, and Codex CLI or OPENAI_API_KEY for external review. A ~/.hermes/config.yaml must point skills.external_dirs at the repo skills directory.

Can I test the pipeline without real Devin dispatch or CI?

Yes, use --reviewer mock with --dispatch-mode dry-run to exercise the full state machine with deterministic mock review output. Unimplemented stages like real CI and Devin dispatch are stubbed but preserve the state contract for later completion.

When should I not use the Open Design orchestrator?

Skip it for single trivial edits that do not need governance overhead. It also cannot bypass the policy gate or human approval for CRITICAL changes, and it should not run without a configured DATABASE_URL when Ops DB dispatch is required.

How does the repair loop handle failed policy gate results?

On a REPAIR decision the orchestrator re-dispatches to Devin with the attempts counter incremented, bounded by a RepairBudget of max 3 attempts plus wall-clock and cost limits. ESCALATE requires human-in-the-loop approval, and BLOCK stops the run entirely.