Internal track

Idle-Discovery Agent

Wakes with no queued task, reads three bounded live signals from its own repo, asks AIsa for one goal (or NONE, falling back to its own open backlog), and checkpoints the episode for deterministic offline replay.

Result

Internal track Runner-up — mean score 66.63/100 across 8 judges, rank 12 of 22 teams overall. Two of judging's harder findings (the AIsa decision's NONE branch being unreachable, and an unauthenticated /trigger with no spend cap) were fixed for real after judging closed — full feedback and what shipped in response are in the plan doc.

Real-world use case

A repo-health janitor: a scheduled agent that checks a codebase's cheap signals — CI status, security alerts, code churn — and either proposes one well-reasoned fix or works down a known backlog, leaving every decision checkpointed for offline audit. Deployable nightly across a team's repos, catching drift nobody has bandwidth to manually triage, with a replayable paper trail for "why did it suggest this?" — useful anywhere audit/compliance matters, not just as a demo.

System boundaries — what's ours vs. what's real and external

The internal track asks for "real work across systems it does not own." The amber edges below are the two places this agent actually leaves its own infrastructure and depends on a system neither we nor this hackathon control.

This repo / GitHub Pages + Cloudflare Worker (ours)
boundary reads — GitHub REST API (a platform we don't own)
3 bounded signals: CI status · security alerts · edit-hotspot
boundary calls — AIsa (api.aisa.one), real metered spend, receipt captured
One goal, or NONE → falls back to this repo's own open-work plan
writes + verifies
GitHub issue + file on this repo, and on org2 (independent counterparty repo, with its own reactive Actions agent)
checkpoints
Workers KV — ours, but replayable + auditable offline

Full step-by-step breakdown (file names, and the closed, permanently-blocked Cotal attempt) lives in the plan doc - kept there only, so this page and the plan doc don't drift out of sync with each other again.

This is a live judge-triggerable demo. Click the button below and the browser calls the real deployed Worker directly — no server-side proxy, nothing pre-recorded.

  1. Observes GitHub Actions status, open Dependabot alerts, and edit-frequency hotspots on this repo
  2. One AIsa-gated LLM call (qwen-flash) decides a goal or says NONE — fully stateless, no memory of any prior episode or goal it has proposed before
  3. NONE falls back to the next open row in the project's own plan — a designed branch, not a failure
  4. Executes the goal as a real GitHub issue here and on org2 (an independent counterparty repo with its own reactive agent), each read back to confirm it landed
  5. Writes a checkpoint you can replay offline later, byte-for-byte

Checkpoint timeline

Links
Source repo Counterparty repo (org2) Agent card MCP server Raw checkpoint feed Design + open-work plan