dualpass 1.0.4

A reliability-first agent harness with cross-vendor independent review built in.

dualpass

One agent ships the work. A different vendor ships the review.

A reliability-first agent harness for staged agentic pipelines, with cross-vendor independent review built in as a primitive. Author with Claude. Audit with Cursor. Or swap either for whatever you trust — dualpass is CLI-agnostic about both.

Status: v1.0.1 — feature-complete. CLI surface, controller, dual-pass parallel reviewer, cross-vendor fallback, circuit breaker, watchers, retro tooling, real context-bundle + precedent-cache builders, build-marker frontmatter parsing, 5 built-in preflight gates, cross-unit retro pattern aggregation. Tests pass across Python 3.12/3.13 on Ubuntu and macOS.

---

What it is

dualpass is the orchestration layer around a staged agentic pipeline — the deterministic code that runs the loop, manages context, owns the gates, handles permissions, and (the headline feature) wires a different-vendor model as the independent reviewer of every stage's output.

The pattern this productizes:

1. The author agent (default: claude) produces a stage's artifact — research, outline, spec, code, audit, whatever your stages are.
2. The reviewer agent (default: cursor-agent) — running on a different vendor's model — judges the output, runs live-fire probes against the claims, and either approves or sends back findings.
3. The author iterates until approved, gates run, the next stage launches.
4. If either provider fails (resource_exhausted, transient outage), the harness falls back to a configured second reviewer (default: claude again) — never silently skips review.

Why cross-vendor matters: a single agent grading its own work is the most reliable way to ship plausible-but-wrong output. Anthropic engineering documented this pattern as a fix for self-evaluation bias. dualpass makes the pattern a first-class feature, not a thing you have to build yourself.

What it isn't

• Not a visual builder. No web UI, no DAG editor. CLI + YAML.
• Not a replacement for LangGraph or CrewAI. Different design point: opinionated reliability, not flexible composition.
• Not a hosted service. Local filesystem only. If someone wants SaaS, fork it.
• Not a model marketplace. Bring your own CLIs (claude, cursor-agent, codex, or anything matching the agent-template contract).

Who it's for

Two audiences, one v1:

• Solo devs and small teams building multi-stage coding agents — you want what a staged pipeline gives you (research → spec → code → audit → handoff), without inheriting somebody else's domain logic.
• Anyone running long-lived agentic workflows — research, content production, ops runbooks. The stage chain is configurable; coding is just the default example.

The nine components

dualpass is designed around the nine canonical components of a production agent harness. See docs/CONCEPTS.md for the full mapping.

Component	Where it lives
Model	config/agents.yaml — named agent roles → CLI invocation templates
Controller	src/dualpass/controller.py — the deterministic loop
Tools	Stage skills + drop-in gate plugins
Context window	src/dualpass/context.py — stage-context bundle + precedent-cache builders (compressed upstream-FINAL summaries + recent ratified precedents)
Memory	.dualpass-state/<unit>/ artifacts + project docs (PROJECT/DECISIONS/BACKLOG/DOC-MAP); memory.read_build_marker parses build-complete YAML frontmatter
Sub-agents	Documented affordance in stage skills; orchestration deferred to a future release (see ROADMAP if present, or CHANGELOG)
Planning	The first N stages of your pipeline (default: research → outline → spec)
Sandbox / permissions	config/permissions.yaml — tiered, opt-in autonomy
Observability	.dualpass-state/logs/ + structured event log + dualpass status

Quick start

pip install dualpass
dualpass init my-project
cd my-project && dualpass run --unit demo-001 --provider mock --ignore-breakpoints

Or, if you prefer uv:

uvx dualpass init my-project
cd my-project && uvx dualpass run --unit demo-001 --provider mock --ignore-breakpoints

--provider mock runs the full stage chain end-to-end without calling any real LLM — useful for verifying the harness works before pointing it at claude + cursor-agent for live runs.

Install from source instead

git clone https://github.com/Chris-Rebentisch/dualpass && cd dualpass
pip install -e .   # Python 3.12 / 3.13. On 3.14 + macOS see CONTRIBUTING.md.

See docs/RUNBOOK.md for the full first-boot walkthrough, and docs/WALKTHROUGH.md for an annotated end-to-end run.

Why dualpass

Why a different-vendor reviewer? A single model grading its own work hits self-evaluation bias hard — it confidently praises mediocre output. Adversarial review by a different vendor's model (different training data, different RLHF signal, different failure modes) catches what a same-vendor reviewer misses. Anthropic's engineering team documented this as the strongest known lever against self-evaluation bias for long-running coding agents.

Why fixed stages instead of arbitrary DAGs? Fixed cycles are easier to reason about, easier to checkpoint, easier to recover. The DAG flexibility of frameworks like LangGraph buys you composability at the cost of debuggability. dualpass picks the other tradeoff. (If your use case genuinely needs a DAG, dualpass propose-dag will help you scope one — but v1 ships fixed-cycle only.)

Why ship watchers? Background daemons that auto-trigger downstream stages when an upstream artifact lands. The reliability lessons here (seed-before-going-live, single-flight guards, split-parent skip gates) are load-bearing — without them, a multi-day build is fragile to a single missed handoff. They're in v1 from day one.

Why CLI-template agent invocation? Because the LLM ecosystem is fragmenting faster than any SDK can keep up. As long as a model has a CLI that takes a prompt and returns a string, dualpass can drive it. No SDK lock-in.

How it compares

dualpass picks reliability over composability — fewer ways to wire things together, more guarantees about what happens when something goes wrong. The honest comparison below is meant to surface tradeoffs, not score points; where another project genuinely wins on a dimension, that's the point of putting it in the table.

Project	Cross-vendor reviewer	Fixed pipeline vs DAG	CLI-first	SDK lock-in	Watchers	Sandbox tiers
dualpass	Built-in	Fixed	Yes	None (CLI templates)	Yes (3)	Yes (4 postures)
LangGraph	DIY	DAG	No (Python API)	LangChain	No	DIY
CrewAI	DIY	Role-based	No (Python API)	CrewAI	No	DIY
HumanLayer 12-factor-agents	N/A (framework-agnostic spec)	N/A	N/A	N/A	N/A	N/A
smolagents	DIY	Single-turn	Yes	huggingface_hub	No	DIY
AutoGen	DIY	Multi-agent	No (Python API)	autogen	No	DIY

If you need DAG flexibility, LangGraph beats dualpass on that axis. If you want a single-turn, lightweight runner over Hugging Face Hub, smolagents is closer to that shape. dualpass is the right pick when you want a long-running staged pipeline with cross-vendor review and recoverable failure modes by default.

Documentation

• CONCEPTS.md — the nine components + dualpass-specific reliability patterns
• RUNBOOK.md — first-boot walkthrough and recovery procedures
• CONFIG-REFERENCE.md — every config file, every field
• CHANGELOG.md — release notes
• CONTRIBUTING.md — how to contribute

License

Apache 2.0. See LICENSE. Explicit patent grant included.

Acknowledgments

dualpass distills patterns from published guidance by Anthropic engineering (Claude Code, "Building Effective Agents," "Effective harnesses for long-running agents," "Harness design for long-running apps"), HumanLayer's 12-factor-agents (Dex Horthy), Cognition Labs ("Don't Build Multi-Agents"), and Simon Willison's writing on tool-use security, hardened against patterns learned running a production staged-pipeline build. Names credited where claims appear in the docs.

---

_{Python 3.14 + macOS editable installs (pip install -e .) silently fail to expose the package because Python 3.14's site.py skips .pth files that carry the auto-applied com.apple.provenance xattr. The published wheel works on 3.14, and the test suite works via the repo's conftest.py. See CONTRIBUTING.md for workarounds.}