A Claude Code plugin marketplace whose coe plugin runs a Correction-of-Errors retrospective at the end of medium-to-big sessions and proposes systemic self-improvements — rules, memories, skills, hooks, settings.
- A
Stophook (coe-nudge.sh, pure shell, zero LLM tokens) detects a medium-to-big session and suggests running/coe— once per session. /coedelegates transcript analysis to the bundledcoe-analystsubagent, runs a confidence-gated confirm (or interviews you when the transcript is sparse), performs a blameless 5-whys, and writes a propose-only COE document todocs/coe/.- You review the proposed action items and apply the ones you want. The skill never edits your config.
Analyst findings are written to ${TMPDIR:-/tmp}/coe-findings/<slug>/ as per-transcript .json (+ optional .md) files and kept as an audit trail; the skill merges from those files, never from chat output.
Add this marketplace and install the coe plugin via Claude Code's plugin commands.
plugins/coe/skills/coe/SKILL.md— the skillplugins/coe/agents/coe-analyst.md— the analysis subagent (coe:coe-analyst)plugins/coe/scripts/coe-nudge.sh— the Stop-hook nudgeplugins/coe/commands/coe.md— the/coecommandplugins/coe/hooks/hooks.json— Stop-hook registration
[RULE] [MEMORY] [SKILL] [HOOK] [SETTING] [CLAUDE.md] [PROCESS] [TEST] [LINT] [CI]
Correction of Errors is Amazon's blameless, mechanisms-over-blame root-cause practice. Good starting points:
- Why you should develop a correction of error (COE) — AWS Cloud Operations Blog, the introductory post.
- Creating a correction of errors document — AWS Cloud Operations Blog, a walkthrough of the document sections (summary, timeline, 5 whys, action items).
- Amazon's approach to failing successfully — Amazon Builders' Library, on postmortems that drive real improvement.
The COE concepts here were inspired by two prior Claude skills, built fresh rather than copied:
- robisson/build-like-amazon-agent-skills (MIT) — blameless 5-whys and mechanisms-over-promises framing.
- rot13maxi/agent-post-mortem — skill packaging, the interview phase, and typed action items.
The novel twist this plugin adds: a session-end retrospective that runs even on success, targeting Claude Code's own configuration (rules, memories, skills, hooks, settings) rather than production incidents only.
Apache-2.0. See LICENSE.
