Open-source coding.sh measures where your coding agents waste time and money
AsianKidTV · reddit · 2026-11-16
A developer released coding.sh, a free MIT-licensed local tool (npx coding.sh) that measures cost and quality of Claude Code, Codex, Gemini CLI, OpenCode, Copilot CLI, Aider and more, logging prompts, tool calls, diffs, tests, tokens and cost to local SQLite.
Real findings from pointing it at his own runs:
- A/B test of a popular "be concise" plugin (10 tasks × 6 paired runs): replies 17% shorter, output tokens −8%, but cost per run +11% — the style guide injected into every request costs more than the shorter output saves.
- Across 57 workflow runs (666 subagents, $2,500 API-equivalent): broken prompt caches cost $106 (4% of spend); 41 cache breaks were agents idling at a barrier until expiry; 40 subagents failed outright.
- Mining 3,600 sessions found a find-and-replace typo in AGENTS.md sending Codex hunting a nonexistent folder in 92 sessions.
improve ranks what to fix first; an MCP server lets agents look up past sessions. No account, nothing leaves your machine.
More from coding & agent
- Two prompts that turn your AI assistant into Jarvis in about 10 minutes — VibeMarketer_ · 2026-11-16
- Cohere North 2 adds per-agent guardrails, with SOC 2, ISO 27001 and ISO 42001 certs — cohere · 2026-11-16
- FastMCP hits #1 on GitHub Trending with 28.5k stars as the go-to Python MCP framework — adnan_hashmi · 2026-11-16
- YC F26's FlashDrive launches a shared cloud drive for AI agents — damnGruz · 2026-11-16
- Reddit dev details 7 chronic failures of AI agents, from endless planning to fake progress — Nmoreland51 · 2026-11-16
- Dev ships MCP server for a trading journal so Claude can query PnL and win rate — Occ_inv · 2026-11-16