deli-autoresearch

Deli AutoResearch

A set of conventions for running an agent loop unattended for hours or days. There's no code here — just a skill that tells the agent how to keep its state in files, try a different approach when it gets stuck, and recover when the loop dies.

The problem

Leave an agent running on its own long enough and it tends to fail in one of three ways:

It keeps trying the same approach and getting less out of it each time.
It finishes a piece of work, writes a summary, and then just sits there waiting.
The session closes or the context gets compacted, and the loop quietly stops with no error.

None of these are about the model being weak. They happen because nobody wired up the plumbing to keep the loop honest and alive. That plumbing is what this skill describes.

What's inside

One skill — deli-autoresearch. It covers:

The rule that, once started, the agent runs without stopping to ask questions
Which files hold state and logs (progress.json, findings.jsonl, directions_tried.json, and so on)
Prompts for the orchestrator, the workers, and the heartbeat watchdog
How to tell the loop is stuck by counting results, not by asking the model how it's doing
How to make each new attempt structurally different from the last
How to schedule subagents, validate their output, and escalate when blocked

Install

In Claude Code, add this repo as a plugin marketplace, then install the plugin:

/plugin marketplace add nanawanzii/deli-autoresearch-skill-protocol
/plugin install deli-autoresearch@deli-autoresearch

Or, if you'd rather skip the plugin system, drop the skill file straight into your project:

skills/deli-autoresearch/SKILL.md

Use

Invoke the skill when starting an unattended long-horizon task:

/skill deli-autoresearch

Then set up a task directory, write task_spec.md, start the orchestrator loop, and start a separate heartbeat watchdog alongside it. See skills/deli-autoresearch/SKILL.md for the full prompts, file formats, stall thresholds, and pivot rules.

Minimum viable startup

{task}/
  state/
    task_spec.md            # goal, scope, success criteria, stop conditions
    progress.json           # iteration counter + stale_count
    findings.jsonl          # append-only evidence
    directions_tried.json   # enforce direction diversity
    iteration_log.jsonl
  logs/
    work.jsonl
    orchestrator.jsonl
    heartbeat.jsonl

The orchestrator reads these files, picks a direction that's structurally different from what's been tried, launches a fresh worker, verifies the output, and updates state. An independent heartbeat watchdog restarts stalled loops without touching worker findings.

Skill layout

SKILL.md is a short entry point — the trigger conditions, the behavioral contract, a startup checklist, and a map of subfiles. Templates and references sit alongside so an agent only loads what its role needs, keeping context small.

skills/deli-autoresearch/
  SKILL.md                     # entry point + behavioral contract + startup checklist
  templates/
    orchestrator.md            # scheduler role: per-callback checklist + prompt
    worker.md                  # execution role: requirements + prompt
    heartbeat.md               # watchdog role: allowed/forbidden actions + prompt
  references/
    state-schema.md            # directory layout, JSON/JSONL formats, log record shape
    stall-and-pivot.md         # stall thresholds, structural vs tactical pivots
    patterns.md                # subagent patterns, validation, escalation, common mistakes

Load-when-needed: an orchestrator reads orchestrator.md + stall-and-pivot.md; a worker reads worker.md + state-schema.md; the heartbeat reads heartbeat.md only. patterns.md comes in at validation and escalation moments.

License

MIT