Ringer

Category: tool Maturity: adopted (in active use on Paul’s Mac since 2026-07-25) First seen: 2026-07-25 Last updated: 2026-07-26

What it is

Ringer is a single-file Python orchestrator that fans work out to parallel, cheap CLI worker agents (Codex, OpenCode/OpenRouter, Grok) and verifies every result by executing a user-written check command — pass/fail is decided by running the artifact, never by the worker’s self-report [S-2026-07-26-ringer-readme]. Ringside, its local web HUD at 127.0.0.1:8700, shows live swarms and a versioned run library; nothing leaves the machine [S-2026-07-26-ringer-readme].

How it’s used

• Paul’s install (2026-07-25/26): repo at ~/Projects/ringer, two lanes — Codex CLI on the flat-rate ChatGPT plan, and OpenCode+OpenRouter on GLM-5.2 at roughly a cent per task; Grok declined [S-2026-07-26-clm-dashboard-v2-rebuild].
• First production use: the four-task clm-data-layer swarm that built the CLM Taxonomy Dashboard v2 data layer (converter, reference extractor, integrity validator, CSV export), 4/4 PASS with two tasks rescued by the retry-with-failure-context lane; all checks independently re-verified [S-2026-07-26-clm-dashboard-v2-rebuild].
• install-agent is active: a Claude Code skill plus two once-per-session, never-blocking hooks steer agents toward Ringer when work is swarm-shaped [S-2026-07-26-ringer-readme][S-2026-07-26-clm-dashboard-v2-rebuild].
• The local scoreboard (./ringer.py models) routes on first_try_pass_rate; after the first runs GPT-5.6 Sol is proven on probe tasks, GLM-5.2 on probation with 100% first-try on data-pipeline work [S-2026-07-26-clm-dashboard-v2-rebuild].

Theoretical basis

Orchestrator–worker role split (expensive model plans/reviews, cheap models type) with executed verification as the trust mechanism, and local evidence-based model routing (untested → probation → proven ladder) [S-2026-07-26-ringer-readme]. No dedicated Concept pages yet.

Strengths / weaknesses

• Strength: executed checks caught two failing first attempts in the CLM run and the retry lane fixed both without frontier-model tokens [S-2026-07-26-clm-dashboard-v2-rebuild].
• Strength: per-attempt local logging makes routing decisions evidence-based and personal to the workload [S-2026-07-26-ringer-readme].
• Weakness: work that shares one large artifact is not swarm-shaped — parallel workers collide; the CLM UI rebuild was deliberately done directly for this reason [S-2026-07-26-clm-dashboard-v2-rebuild].
• Weakness: cost claims (“~a penny a task”) track OpenRouter pricing and will drift; plan-billed lanes report no token costs [speculative — S-2026-07-26-ringer-readme].

Tensions

No source disagreements surfaced yet.

Sources

• [S-2026-07-26-ringer-readme] → S-2026-07-26-ringer-readme
• [S-2026-07-26-clm-dashboard-v2-rebuild] → S-2026-07-26-clm-dashboard-v2-rebuild