Evaluate the profanity used working with Codex and Claude Code

Marcel Petrick released Agentic Swear Jar, a private, local, zero-AI dashboard that analyzes profanity in Claude Code and Codex CLI histories, with no prompts uploaded and no runtime dependencies beyond Linux and Python 3.10. The tool parses history files, generates a standalone report.html, and provides metrics such as matches per 100 prompts and per 1,000 words, intensity counts, and trends, while excluding raw prompts and identifiers.

A private, local, zero-AI dashboard for the spicy parts of your Claude Code and Codex CLI history. one command · two coding agents · zero prompts uploaded Building software with an agent is still building software. Tests fail, tools loop, the same bug returns wearing a small fake moustache, and occasionally the most precise technical response is a four-letter word. That is not a conduct problem. A coding harness is a machine. It has no feelings to hurt. Your language is simply another trace of the work: a tiny pressure gauge for friction, intensity, late nights, stubborn projects, and the glorious moment the fix finally lands. Agentic Swear Jar turns that trace into something worth smiling at. It compares Claude Code with Codex CLI, finds trends, ranks vocabulary, and produces a polished standalone report. It does all of this locally with ordinary parsing, Unicode-aware tokenization, dictionary lookups, and arithmetic. No model judges the language. No API receives the prompts. No raw prompt text goes into the report. Author: Marcel Petrick mail@marcelpetrick.it License: GPLv3 or later. See LICENSE. Note: projected is generated with AI. Requirements: Linux and Python 3.10 or newer. There are no runtime dependencies. ./swear-stats --open That reads both default histories: | Tool | Default input | Prompt field | |---|---|---| | Claude Code | ~/.claude/history.jsonl | display | | Codex CLI | ~/.codex/history.jsonl | text | The output is report.html , a self-contained file that works offline. If a history is missing, the tool warns and continues with the one it can find. Use --strict-inputs when both are required. Want aggregate JSON as well? ./swear-stats --json report.json --output report.html Want a particular era of your life? ./swear-stats --since 2026-01-01 --until 2026-06-30 --open The dates are inclusive and interpreted in the machine's local timezone. The report has combined, Claude Code, and Codex CLI views. Each includes: - prompt count, word count, total matches, and prompts containing at least one match; - matches per 100 prompts and per 1,000 words, so tools with different usage volumes compare fairly; - daily and monthly trends, local hour-of-day activity, and weekday patterns; - canonical vocabulary rankings fucked , fucking , and fucks become fuck ; - mild, moderate, and strong intensity counts; - match rates by prompt-length bucket; - longest clean and spicy prompt streaks; - distinct session and Claude project counts, without exposing their identifiers; and - malformed-record diagnostics, because JSONL occasionally has a bad day too. The dashboard intentionally does not contain prompts, snippets, session IDs, project names, or project paths. The analyzer streams each JSONL file one line at a time. Every prompt is tokenized once with a compiled Unicode-aware regular expression. Normalized tokens are checked against an in-memory hash map, making lookup effectively constant-time. Only counters, date buckets, and tiny timestamp/event tuples survive analysis. For an input containing n characters and w words, the main scan is O n + w . Memory use is independent of prompt content and grows only with the number of dates, sessions, and aggregate labels. It does not shell out to grep or jq , so multiline JSON strings, Unicode, and malformed records are handled consistently in one pass. Matching is case-insensitive and uses whole tokens. This avoids classic substring mistakes such as matching ass inside class or hell inside shell . Underscores act as separators, which is useful for identifiers such as what the hell . There is no universal database of swear words. Meaning depends on geography, context, reclaimed language, and personal taste; aggressive blocklists also run into the Scunthorpe problem https://en.wikipedia.org/wiki/Scunthorpe problem . This project therefore ships a small, transparent English list aimed at ordinary expletives—not hate-speech moderation. It has four tab-separated columns: variant canonical category severity fucking fuck expletive 3 wtf fuck abbreviation 2 Edit swearstats/data/en.tsv /marcelpetrick/AgenticSwearJar/blob/master/swearstats/data/en.tsv , or layer personal terms on top without changing the repository: ./swear-stats --lexicon my-words.tsv Later files override earlier variants. To discard the bundled English list entirely: ./swear-stats --replace-lexicon --lexicon my-words.tsv Useful public datasets do exist. The MIT-licensed @dsojevic/profanity-list https://github.com/dsojevic/profanity-list adds severity, exceptions, and tags; rates terms by how likely they are to be profane; and https://github.com/words/cuss cuss LDNOOBW https://github.com/LDNOOBW/List-of-Dirty-Naughty-Obscene-and-Otherwise-Bad-Words covers many languages. They are good raw material, but importing a large list blindly will change the meaning of the statistics and increase false positives. A personal list you understand is usually more honest. usage: swear-stats -h --claude-history PATH --codex-history PATH --lexicon TSV --replace-lexicon --since YYYY-MM-DD --until YYYY-MM-DD -o PATH --json PATH --redact-terms --open --strict-inputs Examples: Nonstandard config directories ./swear-stats \ --claude-history /mnt/private/claude/history.jsonl \ --codex-history /mnt/private/codex/history.jsonl Machine-readable aggregates only in addition to HTML ./swear-stats --json build/stats.json --output build/dashboard.html Install an isolated CLI for development python3 -m venv .venv .venv/bin/pip install -e . .venv/bin/swear-stats --open This project analyzes the short, global prompt histories rather than full session transcripts, so each user input is counted once and tool outputs are excluded. - Anthropic documents ~/.claude/history.jsonl as every typed prompt with its timestamp and project path, retained until deletion. Full transcripts live below ~/.claude/projects/ and are subject to cleanupPeriodDays 30 days by default . See Explore the https://code.claude.com/docs/en/claude-directory . .claude directory - OpenAI's Codex source defines ~/.codex/history.jsonl as append-only JSONL records shaped like {"session id":"…","ts":1234567890,"text":"…"} . See the Codex message-history implementation https://github.com/openai/codex/blob/main/codex-rs/message-history/src/lib.rs . Both formats are implementation details that may evolve. Unknown or malformed records are skipped and reported rather than crashing the whole run. The script reads plaintext histories, so treat it with the same access controls as the coding tools themselves. Generated report.html and report.json are ignored by Git by default. Although they contain aggregates only, the canonical-term ranking is still personal information—share it because it is funny, not by accident. To stop future local prompt history, consult each tool's current settings. Anthropic documents CLAUDE CODE SKIP PROMPT HISTORY ; Codex supports history persistence = "none" in its config. Changing those settings also affects recall and resume behavior, so read the upstream documentation before flipping the switch. Run the same local quality gate used before every commit: ./localPipeline.sh The application has no runtime dependencies. The development pipeline expects shellcheck , ruff , and mypy on PATH ; it runs the complete test suite, shell and Python linting, formatting checks, strict type checking, and Git's whitespace validation, stopping at the first failure. The test suite covers both history shapes, canonical variants, Unicode-aware whole-word behavior, date filtering, malformed records, combined statistics, and the promise that raw prompt text and project paths never enter the HTML. The repository includes a reproducible public build based on the real Claude Code and Codex CLI histories: ./buildPublicArtifacts.sh It requires Chromium and img2pdf , then creates and updates: public/report.html — the interactive, redacted dashboard; public/report.json — aggregate data behind every published number; public/linkedin-carousel.html — four-slide carousel source; public/slides/ — four 1080×1080 RGB screenshots; and public/agentic-swear-jar-linkedin.pdf — the square, four-page LinkedIn document. The complete public/ directory is generated locally and ignored by Git. Rebuilding it never stages personal statistics or publication files by accident. The public build uses --redact-terms : observed vocabulary retains its first character for visual rhythm, while every remaining character becomes a solid marker block. The carousel builder refuses unredacted JSON. Raw prompts, project paths, session IDs, and uncensored observed terms are not published. - This is exact lexical matching, not contextual language understanding. - It will miss creative punctuation, deliberate obfuscation, and novel spellings. - It can count a quoted swear word or code identifier even when the prompt is discussing the word. - Severity is editorial metadata, not science. - Missing or disabled history cannot be reconstructed. - Counts answer “what matched this list?”—not “was this prompt offensive?” Those are very different questions, and this project only claims the first. May your tests be green and your vocabulary statistically significant. Copyright © 2026 Marcel Petrick. Licensed under the GNU General Public License v3.0 only /marcelpetrick/AgenticSwearJar/blob/master/LICENSE .