Skip to content

florianbuetow/fixclaude

v0.1.1MIT

Run this workflow from Codex. Reverse the built-in Claude Code limitations discovered in the source code leak. Installs production-grade agent directives that override verification gates, context decay, brevity mandates, and silent truncation — the same fixes Anthropic uses internally.

Fixclaude

Generated for Codex from the canonical source plugin. Run just update at the marketplace root to refresh it.

Run this workflow from Codex. Reverse the built-in Claude Code limitations discovered in the source code leak. Installs production-grade agent directives that override verification gates, context decay, brevity mandates, and silent truncation — the same fixes Anthropic uses internally.

Compatibility scope: This plugin runs as a Codex skill but intentionally operates on Claude Code files, sessions, processes, or service behavior. Those source-domain references are preserved rather than relabeled as unsupported Codex runtime behavior.

Install with Codex

Register the repository marketplace with codex plugin marketplace add ., then run codex plugin add fixclaude@florianbuetow-codex-plugins or install it from Florian Buetow Codex Plugins in the Plugins Directory. Invoke its workflows with $<skill-name>.


Reverse the built-in Claude Code limitations discovered in the source code leak.

Installs production-grade agent directives that override verification gates, context decay, brevity mandates, and silent truncation -- the same fixes Anthropic uses internally but gates behind employee-only flags.

What It Fixes

Based on reverse-engineering the Claude Code source against real agent logs (original analysis by @fakeguru):

#ProblemRoot CauseOverride
1"Done!" with 40 type errorsVerification gated to employees (USER_TYPE=ant)Forced verification loop
2Hallucinations after ~15 messagesAuto-compaction nukes context at ~167K tokensStep 0 cleanup + phased execution
3Band-aid fixes, no real solutionsSystem prompt mandates minimal interventionSenior dev quality override
4Context decay on large refactorsUsers default to single-agent sequentialSub-agent swarming (5-8 files/agent)
5Edits reference code it never sawFile reads capped at 2,000 linesChunked reads with offset/limit
6grep finds 3 results, there are 47Tool results truncated to 2,000-byte previewNarrow scoping + truncation awareness
7Rename misses dynamic importsgrep is text matching, not an ASTMulti-search mandate per reference type

Commands

CommandWhat it does
$fixclaudeAuto-detect and install (routes to init or update)
$fixclaude:initCreate a new CLAUDE.md with full production-grade directives
$fixclaude:updateAugment an existing CLAUDE.md with missing overrides
$fixclaude:analyzeGap analysis: check your CLAUDE.md against all 7 findings

All commands handle CLAUDE.md symlinks transparently (e.g., CLAUDE.md -> AGENTS.md).

CLAUDE.md Template

The directives installed by init cover 9 areas:

  1. Pre-Work -- Step 0 cleanup, phased execution, spec-based development
  2. Understanding Intent -- Follow references, work from raw data, one-word mode
  3. Code Quality -- Senior dev override, forced verification, write human code
  4. Context Management -- Sub-agent swarming, context decay awareness, file read budget
  5. File System as State -- Use disk for memory, intermediate results, progressive disclosure
  6. Edit Safety -- Edit integrity, no semantic search assumptions, one source of truth
  7. Prompt Cache Awareness -- Avoid cache invalidation, use /compact and context-log.md
  8. Self-Improvement -- Mistake logging, bug autopsy, failure recovery
  9. Housekeeping -- Autonomous bug fixing, parallel batch changes, file hygiene

Attribution

The CLAUDE.md template content is based on work by fakeguru, released under the MIT License. Copyright (c) 2026 fakeguru.

License

MIT