Ringer README (NateBJones-Projects/ringer)

Tag: S-2026-07-26-ringer-readme Type: article (project README — vendor/author documentation) Author(s): Jon Edwards (Nate Jones Media LLC) Date of source: 2026-07-26 (main branch at capture; latest release ringside v0.1.1, 2026-07-03) Date ingested: 2026-07-26 Authority weight: medium — first-party tool documentation; accurate for design intent and mechanics, promotional in tone on cost/quality claims. Raw file: /_raw_sources/S-2026-07-26-ringer-readme.md. External URL: https://github.com/NateBJones-Projects/ringer

What it claims

Ringer is a single-file Python orchestrator for parallel AI-agent swarms in which an expensive model writes specs and reviews results while cheap CLI workers implement. Its central claim: worker self-reports are untrustworthy (“parallel agents lie”), so every task’s pass/fail is decided by executing a user-supplied check command against the artifact; failures retry once with the failure output injected. Runs are logged per attempt (verdict, tokens, duration, model, task_type) to a local JSONL; a scoreboard (models) reports pass rates per (model, task_type) with first_try_pass_rate as the routing signal and a promotion ladder (untested → probation → proven at 3+ tasks ≥0.67 first-try). Engines are pluggable config blocks; shipped lanes are Codex CLI (default), Grok Build CLI, and OpenCode+OpenRouter as a universal harness (default model z-ai/glm-5.2, cited at roughly $0.74/M in, $2.33/M out as of 2026-07). Ringside is a local-only web HUD showing live swarms and a versioned run library. install-agent installs a Claude Code skill plus two once-per-session, never-blocking hooks nudging agents toward Ringer for swarm-shaped work. Four invariants: stdin closed, explicit sandbox mode, executed verification, raw logs.

Notable quotes

“A check that cannot fail is trusting the worker with extra steps.” “The numbers are not portable between users.” (on the local model scoreboard)

What’s speculative vs. asserted

Asserted (mechanics, verifiable in code/use): executed checks, retry-once, logging, engine config, hooks. Promotional/estimated: “Fable-quality output without Fable-level burn”, per-task ~penny cost (depends on OpenRouter pricing at run time) [speculative — S-2026-07-26-ringer-readme]. Model pricing is dated 2026-07 and will drift.

Topics this feeds

• Ringer — technology page.

Open questions raised

• Windows support is via WSL only; non-macOS OpenCode use has no OS write-confinement.
• Native desktop HUD (Tauri) lags the web dashboard (v0.1.1 prototype).