Daybreak AutoResearch

An agent-agnostic autonomous research protocol for long-horizon, grounded, multi-perspective investigation, packaged as a Codex-compatible skill.

Daybreak AutoResearch is a single-file agent protocol in SKILL.md. It is designed for runs that must continue for hours, days, or weeks without constant human steering while still keeping research honest: persistent state, source and claim ledgers, perspective coverage, contradiction and gap mining, heartbeat watchdogs, and sufficiency-based completion.

Lineage

This project is an adapted and extended version of Deli AutoResearch / Deli_AutoResearch, originally created by Deli Chen (DeepSeek).

Original references:

• Deli Chen / Victor Chen on X: https://x.com/victor207755822
• Original framework page: https://victorchen96.github.io/auto_research/framework.html

This adaptation preserves the reliability spine of the original framework:

• zero-interaction long-running work;
• state persisted to files rather than chat memory;
• stall detection and structural pivots;
• heartbeat watchdogs;
• guardian / worker separation;
• bounded subagent patterns;
• verification between iterations.

The adaptation adds a stronger discovery and grounding spine:

• STORM-style multi-perspective question asking;
• explicit perspective coverage via perspectives.json;
• retrieval-grounded source and claim ledgers;
• gap, contradiction, and hypothesis mining;
• a persistent backbone.md argument spine;
• source-quality and claim-confidence labels;
• sufficiency-based completion;
• the Discovery Tail Pass, which treats the first coherent synthesis as a baseline hypothesis and searches for live anomalies before accepting a result.

See ATTRIBUTION.md and MODIFICATIONS.md for more detail.

License

The Daybreak AutoResearch adaptation and original additions in this repository are released under the Apache License 2.0.

This repository also includes a NOTICE file with lineage and attribution notes. Deli AutoResearch has been publicly announced as open source; however, no explicit upstream license type was located while packaging this repository. If an upstream license is later published, review this repository for compatibility.

Agent-Agnostic Protocol

Daybreak is not tied to Codex. Codex support is the packaging layer in this repository, not the boundary of the idea. The protocol can be adapted to any agent environment that can:

• persist state to files or another durable store;
• read curated state at the start of each iteration;
• run bounded work loops without constant human steering;
• call retrieval, search, source, artifact, or experiment tools;
• cover multiple perspectives through subagents, personas, source classes, or structured question roles;
• validate outputs against external checks;
• maintain a heartbeat, watchdog, scheduler, or equivalent liveness signal.

Codex users can install SKILL.md and agents/openai.yaml directly. Other agent systems can treat SKILL.md as the canonical protocol and translate the state files, watchdogs, perspective coverage, and verification gates into their own runtime.

What It Is For

Use Daybreak AutoResearch when a task needs more than a one-pass answer:

• autonomous or long-horizon research;
• paper-like or report-like synthesis;
• source-backed investigation;
• multi-iteration writing, building, or experiment loops;
• discovering new project-specific relationships, contradictions, or next paths;
• recovering from stalls without waiting for the user to restart the work;
• keeping a durable audit trail of what was tried, found, checked, and left open.

Do not use it for tiny questions, one-off summaries, trivial code edits, or work where the user explicitly wants to steer every step interactively.

Core Loop

The skill keeps two axes alive at once:

alive enough to continue
honest enough to trust
wide enough to discover
disciplined enough to finish

At a high level:

1. Frame the discovery target.
2. Discover the perspective space.
3. Run grounded interrogation for each perspective.
4. Run the Discovery Tail Pass against the first coherent synthesis.
5. Mine gaps, contradictions, and hypotheses.
6. Reconcile results into one backbone.md.
7. Write, build, or decide from the backbone rather than raw agent output.

State Files

A typical Daybreak run writes inspectable state:

{task}/state/
  task_spec.md
  progress.json
  perspectives.json
  backbone.md
  sources.jsonl
  claims.jsonl
  findings.jsonl
  gaps.jsonl
  contradictions.jsonl
  hypotheses.jsonl
  directions_tried.json
  iteration_log.jsonl

{task}/logs/
  work.jsonl
  orchestrator.jsonl
  heartbeat.jsonl

These files are not decorative. They are the memory, evidence trail, recovery surface, and completion check for the run.

Installation / Adaptation

Codex

Copy this folder into your Codex skills directory:

New-Item -ItemType Directory -Force "$env:USERPROFILE\.codex\skills\daybreak-autoresearch"
Copy-Item -Force ".\SKILL.md" "$env:USERPROFILE\.codex\skills\daybreak-autoresearch\SKILL.md"
Copy-Item -Recurse -Force ".\agents" "$env:USERPROFILE\.codex\skills\daybreak-autoresearch\agents"

On Unix-like systems:

mkdir -p ~/.codex/skills/daybreak-autoresearch
cp SKILL.md ~/.codex/skills/daybreak-autoresearch/SKILL.md
cp -R agents ~/.codex/skills/daybreak-autoresearch/agents

Then invoke it explicitly:

Use $daybreak-autoresearch to run a grounded autonomous research pass on ...

Or let another routing skill call it when a task becomes research-grade, multi-iteration, source-backed, or not obvious enough for a local answer.

Other Agent Environments

Use SKILL.md as the reference protocol. Port the state folder shape, source and claim ledgers, perspective coverage, Discovery Tail Pass, watchdog behavior, and sufficiency checks into the target agent framework. The exact tool names can change; the contracts should not.

Suggested Prompt

Use $daybreak-autoresearch to investigate [topic]. Keep state in a task folder,
cover multiple perspectives, ground nontrivial claims in sources or artifacts,
mine gaps and contradictions, and finish only when the backbone is sufficient.

Limits

Daybreak AutoResearch is a protocol, not a guarantee of truth.

• A live heartbeat does not prove research quality.
• Simulated review does not replace external evaluation.
• Novelty pressure can create fiction if ungrounded.
• Retrieval bounds truth; missing corpora must be reported.
• Tail-search candidates are not discoveries until grounded, compressed, and attacked by a real check.
• Codex-compatible packaging does not make the protocol Codex-only.
• Deli AutoResearch was publicly announced as open source, but the exact upstream license type should still be checked if it is later published.

Repository Layout

.
|-- SKILL.md
|-- agents/
|   `-- openai.yaml
|-- README.md
|-- ATTRIBUTION.md
|-- MODIFICATIONS.md
|-- LICENSE
|-- NOTICE
|-- CITATION.cff
`-- .gitignore

Citation Request

If Daybreak AutoResearch helps your work, research, product, repo, article, or agent framework, please cite the repository and keep the upstream lineage visible. This is not only formal credit; it helps people trace where the ideas came from and how the protocol evolved.

Please credit:

• the original Deli AutoResearch / Deli_AutoResearch framework by Deli Chen (DeepSeek);
• the STORM research line when referring to STORM-style multi-perspective question asking;
• this Daybreak AutoResearch adaptation when using its grounding spine, backbone synthesis, sufficiency checks, or Discovery Tail Pass.

Suggested text:

Based on Deli AutoResearch / Deli_AutoResearch by Deli Chen (DeepSeek):
https://victorchen96.github.io/auto_research/framework.html

Daybreak AutoResearch adaptation: adds STORM-style multi-perspective grounding,
Discovery Tail Pass, evidence ledgers, backbone synthesis, and sufficiency-based
completion. Released under Apache-2.0 for the Daybreak adaptation and original
additions in this repository.

Also credit STORM when referring to the STORM-style research engine:

• Stanford OVAL STORM project: https://storm-project.stanford.edu/research/storm/
• STORM code: https://github.com/stanford-oval/storm
• Paper: https://arxiv.org/abs/2402.14207

This repository also includes CITATION.cff, so GitHub can show a "Cite this repository" option after publication.