Skip to content

Add dualpass to Demo Harnesses#47

Open
Chris-Rebentisch wants to merge 1 commit into
ai-boost:mainfrom
Chris-Rebentisch:add-dualpass
Open

Add dualpass to Demo Harnesses#47
Chris-Rebentisch wants to merge 1 commit into
ai-boost:mainfrom
Chris-Rebentisch:add-dualpass

Conversation

@Chris-Rebentisch

Copy link
Copy Markdown

Summary

Adds dualpass to the Demo Harnesses section under Reference Implementations.

What it is

dualpass is a reliability-first agent harness whose headline primitive is cross-vendor independent review: the author agent runs on one CLI (e.g. claude), the reviewer agent runs on a different vendor's CLI (e.g. cursor-agent), and either is swappable without touching the harness. The pattern this productizes — a different-vendor model as the reviewer of every stage's output — is the strongest documented lever against self-evaluation bias in long-running coding agents.

Why it fits Demo Harnesses

Like Open SWE / OpenHands / Aider / SmallCode in this section, dualpass is a complete open-source harness with a specific design point worth studying:

  • Fixed-cycle pipeline (research → outline → spec → prompt → code → audit → handoff)
  • Dual-pass parallel reviewer with cross-vendor fallback on [resource_exhausted]
  • Build-marker YAML frontmatter — author-driven halt contract (only the author can stop the pipeline mid-stage; reviewers always auto-continue to next revision round)
  • Five built-in preflight gates (frontmatter, line-citations, single-flight, marker-frontmatter, acceptance-criteria-wording)
  • Circuit breaker on no-progress loops (SHA-256 artifact hash + 3-strike heuristic)
  • Background watchers (research/prompt/handoff) for overnight unattended runs
  • Cross-unit retro pattern aggregator (EventType counts + bigram frequencies across units)

Distilled from a real production-campaign of multi-day agent builds. Apache-2.0, Python 3.12+, on PyPI as of today.

Contribution checklist

  • Addresses a specific harness problem (cross-vendor reviewer as a primitive, preflight gates, author-driven halts).
  • Vendor-agnostic by principle — CLIs are templated; claude + cursor-agent are defaults but swap freely (codex, any CLI matching the agent-template contract).
  • Format: - [Name](URL) — 1–2 sentence note explaining why it's worth including.
  • Added under the appropriate section: Reference Implementations → Demo Harnesses.

Happy to adjust placement or wording per your preference.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant