ZeroHandoff — Autonomous Software Delivery

One short build request in. A working, tested application and narrated demo out.

ZeroHandoff is an autonomous software-delivery pipeline for the OpenAI Build Week Developer Tools track. A person describes only what they want built, for whom, the desired outcome, and any constraints. Codex turns that request into a sequence of deliverables without human handoffs:

Build Request → Opportunity Model → Outcome Model → Capability Graph → Decision Graph → Scenario Model → Autonomous Change → Evidence + Learning → Demo

Seven two-agent lifecycle cells—SENSE, MODEL, COMPOSE, DECIDE, SIMULATE, EXECUTE, and OBSERVE—own those seven stages, respectively. Each pair forms independent judgments and passes one versioned artifact forward. These compact artifacts are designed for agent-to-agent execution instead of long-form specifications, planning ceremonies, or human handoff documents. Codex preserves project memory, enforces quality gates, and routes failed contract checks through bounded repair loops.

Every run produces an auditable delivery bundle:

• a runnable application with source and setup instructions;
• automated tests and quality-gate evidence;
• append-only JSON logs covering timestamps, runtime and model configuration, the immutable training baseline, run-start inference state, handoff rewards, shadow deltas, night commit, decisions, gates, repairs, artifacts, and the final outcome; and
• a generated narrated video demo.

Setup and running

Supported platforms. The full ZeroHandoff engine supports macOS and Linux with Python 3.11+, Node.js 22+, npm, Chrome/Chromium, FFmpeg, and Codex CLI 0.144.0+ authenticated with GPT-5.6 Sol access. The prebuilt EchoLedger sandbox works on any desktop with a modern browser and a local static-file server.

Fastest judge test — no install or rebuild

The repository includes the production build from the final autonomous run:

python3 -m http.server 8000 --directory submission/sandbox/echoledger/dist

Open http://127.0.0.1:8000. EchoLedger requires no account, backend, API key, network service, or rebuild. Select Complaint · 02:31, redact evidence EV-015, inspect the recurring signal, assign its action, export the local case brief, and reset the fictional experience.

Install the complete engine

python3 -m venv .venv
.venv/bin/pip install -e '.[dev]'
cd ui
npm ci
npm run build
cd ..
.venv/bin/python -m zerohandoff.cli --repo . doctor --live

Run without model calls

This deterministic path exercises training, orchestration, gates, repair, artifact handoffs, demo assembly, and bundling without spending model credits:

.venv/bin/python -m pytest -q
.venv/bin/python -m zerohandoff.cli --repo . puzzles validate
.venv/bin/python -m zerohandoff.cli --repo . train --adapter fixture --rounds 10
.venv/bin/python -m zerohandoff.cli --repo . run \
  --request tests/fixtures/build_request.json --adapter fixture

Run the real Codex pipeline

The checked-in trained baseline means judges do not need to retrain. From the cloned repository root, activate the installed environment and launch Codex:

source .venv/bin/activate
codex

Then, inside the Codex chat, enter:

$run-pipeline
Build a local React application for [audience] that [desired outcome].
Constraints: [non-negotiable boundaries].

Do not run the internal Python delivery command manually. The $run-pipeline skill collects one Build Request, validates the frozen baseline and current inference lineage, then invokes the engine and runs all seven stages, demo generation, and bundling with GPT-5.6 Sol. Use $pipeline-status inside Codex to inspect progress. To observe delivery in the local Control Room, open a second terminal from the repository and run:

.venv/bin/python -m zerohandoff.cli --repo . serve --port 8765

Then open http://127.0.0.1:8765.

Sample data, outputs, and verification

• tests/fixtures/build_request.json is the agent-free sample request.
• data/puzzles.jsonl and data/puzzle_stats.json are the deterministic ten-puzzle training corpus and validation summary.
• .zerohandoff/frozen/latest.json is the checked-in immutable relationship baseline; real delivery starts from it rather than retraining.
• submission/sandbox/echoledger/src/domain.ts and submission/sandbox/echoledger/public/audio/echoledger-el1042.wav contain EchoLedger's entirely fictional local transcript/domain fixtures and synthetic recording. No real customer information is included.
• Each run is written to .zerohandoff/runs/<run_id>/, including versioned artifacts, logs/*.jsonl, the runnable app, narrated demo, checksums, and delivery_bundle.nosync/. Versioned continual-learning commits are under .zerohandoff/learning/commits/.

Verify the tracked judge package, training/run completion, trust invariants, checksums, sandbox, and media streams with:

python3 scripts/submission_package.py verify

More internal commands and recovery behavior are in the judge guide and runbook.

Trust training supports the pipeline; it is not the product itself. Before delivery, a separate ten-round, solver-validated puzzle pilot trains asymmetric directed trust through reward prediction error and one single-prompt Night Curator call per agent after each round. The resulting ten-dimensional vectors form an immutable trained baseline for software-delivery inference: trust alone selects authority, while a deterministic policy compiler turns up to three strong non-trust dimensions into qualitative collaboration guidance without exposing raw scores. See the visual explainer or the source specification for the full design.

Inference copies that baseline into a separate evolving state. Trained pair edges begin at their learned values; new cross-team edges begin at 0.0. Each next team gives a binary handoff reward—accept 1, request revision 0—but trust stays stable during the run. One extra-high-reasoning Night Curator then commits reduced-plasticity trust (α=0.05, step cap ±0.1) and consolidated memory once the prototype completes. The nine non-trust dimensions and the original training JSON remain unchanged forever.

Codex was both the build environment and the hackathon reference runtime. It accelerated the typed contracts, orchestration engine, agent configurations, gates, repair/resume logic, browser proof, demos, and regression suite; GPT-5.6 Sol powers the fourteen specialists and Curators. The human chose the learning rules, immutable-training/evolving-inference boundary, seven agent-native artifacts, reduced-plasticity continual learning, zero post-authorization human handoffs, and Codex-only hackathon scope. The deeper build story and ready-to-use submission copy are in submission/.

Use $run-pipeline, $pipeline-status, and $train-trust from Codex chat. Internal commands and the Control Room are documented in the runbook.

ZeroHandoff is an independent original implementation conceptually inspired by the BMad Method. It does not copy or redistribute BMad source, prompts, agent definitions, names, or branded assets, and is not affiliated with or endorsed by BMad Code, LLC.

Explore: visual pipeline · system design · execution plan · trust architecture · judge guide · submission package · project idea · tracked constraints · MIT license