coredamn

A core dump of your patience.

Count profanity and insults in the human messages you've sent to coding agents. Compare agents, response models, sessions and days, using both raw counts and matches per 1,000 user words. Everything runs locally. No API keys, uploads, telemetry or network requests.

Built with Node, TypeScript, Commander and Effect 4. Native transcript formats have Effect schemas. Untrusted JSON enters as unknown; adapters consume decoded records, not generic JSON dictionaries. Effect owns service composition, typed failures and resource lifetimes. Commander owns command dispatch and arguments.

Run locally

Requires Node 24.15 or later and pnpm. Mise pins the development tool versions. This repository hasn't been published to npm yet.

cd ~/dev/coredamn
pnpm install
pnpm dev --help
pnpm dev discover
pnpm dev plan
pnpm dev scan

discover lists resolved paths and checks their existence without reading chat content. plan reads transcripts to count eligible messages and words without running detection. scan --dry-run is the same count-only path. It still has to read and parse the selected files, but doesn't load a model or classify messages. None of these commands opens a browser. To select particular sources:

pnpm dev scan --agent claude --path 'claude=~/.config/claude'
pnpm dev scan --agent omp --agent grok --format json
pnpm dev scan --since 2026-10-01 --until 2026-10-11 --format html --output ./report.html
pnpm dev scan --config ./my-config.json --language en --language de

Agent and language options accept repeated flags or several values. Path overrides accept agent=/path and can repeat for multiple installations. An explicit path can name a source root, transcript directory or supported file. Paths containing spaces need shell quotes. Relative config paths resolve from the config file's directory; relative CLI paths resolve from the current working directory.

Output files use private permissions where the platform supports them and never overwrite an existing file, including an input transcript. Without --output, the selected report goes to stdout. HTML is a single offline file with searchable, sortable breakdowns and static tables that remain readable without JavaScript. Session IDs, input paths and full transcript text aren't included in scan reports. Discovery output deliberately includes paths.

Explore and share

pnpm dev tui --agent claude
pnpm dev scan --format html --output ./report.html
pnpm dev scan --format svg --output ./patience.svg
pnpm dev scan --format png --output ./patience.png

The terminal dashboard needs an interactive terminal. Press g or Tab to switch agent, model, session and day groups, m to change the metric, s to reverse the sort order, / to filter, and arrows or j/k to scroll. Press r to reset the filter and q to quit. The dashboard also shows coverage and cache totals.

HTML includes a weighted word cloud and daily charts with count/rate selection. Charts use UTC dates and keep undated messages separate. A missing day has no invented data point. SVG and PNG produce a compact 1200 by 630 meme card with counts, vocabulary, date range, coverage and detector. PNG requires --output. Matched terms appear in the cloud and card, so review them before sharing.

Cached local analysis

The dictionary is the default. For contextual detection, supply existing Qwen3.5 GGUF weights to the opt-in native node-llama-cpp backend:

pnpm dev plan --detector local --model ~/models/Qwen3.5-27B-Q5_K_M.gguf
pnpm dev scan --detector local --model ~/models/Qwen3.5-27B-Q5_K_M.gguf --gpu vulkan --max-messages 100

Every eligible message goes to native decisions and constrained occurrence extraction, including messages with no dictionary matches. This runs in the Node process with llama.cpp. It doesn't call an OpenAI-compatible server. Weights aren't downloaded and native bindings aren't compiled or downloaded by the scan. The selected node-llama-cpp backend must already be available.

Decisions produce relative category scores, not calibrated probabilities. The policy excludes quoted/reported insults, self-deprecation and criticism of code from personal-insult counts. Profanity is a separate category. Extracted phrases must resolve to actual source occurrences. When native category scores disagree with extraction, the model reassesses disputed categories using a native four-way decision and checks disputed source occurrences independently in context. A draft-induced positive needs independent occurrence verification. Categories that already agree retain their original scores. Recovery can reject an invented insult or restore a missing category on a verified occurrence. Extraction repairs start a fresh conversation, with at most two repair rounds per message. Reassessed scores come from native decisions and aren't clamped to match the extraction. Invalid occurrence numbering can recover through bounded source candidates, each independently checked in context. Occurrence checks ask separate native insult and profanity questions about the selected text, using surrounding text for context. Candidates never count automatically, including quoted matches. Embedded shorter hints yield to a complete extracted phrase, and ambiguous overlaps or saturated candidate sets remain failures. Results that remain invalid or inconsistent are skipped with a diagnostic. Model judgments can still be wrong.

The default limit is 100 new inference attempts per run. Cache hits don't consume that limit. Messages over 6000 UTF-16 code units are visibly skipped, never silently truncated. Adjust --max-messages, --max-message-chars and --context-size as needed. The context defaults to 4096 tokens; insufficient context or generation capacity produces a diagnostic. A partial report's rate uses analyzed messages only. Its analysis section separately lists all eligible, cached, newly inferred and skipped messages.

Rerunning resumes from validated cache entries. A fully cached run doesn't load the native model. The cache defaults to $XDG_CACHE_HOME/coredamn/inference-v1, or ~/.cache/coredamn/inference-v1. Override it with --cache-dir. Entries store scores and character offsets, not full message text. Keys hash message content, policy/runtime/context and model identity. Model identity uses its resolved path, size, inode and nanosecond modification/change times, avoiding a full weights read during planning. It isn't a cryptographic digest of the model weights.

Local plan reports cached messages, unique pending inputs, repeated inputs, size-blocked messages and the next bounded selection. It never writes cache entries. Its duration estimate uses an earlier short Qwen3.5-27B synthetic trial, not a hardware benchmark or a prediction for long transcripts. Dictionary language/custom-term options don't apply to the local contextual policy.

Sources and configuration

Precedence is CLI --path, then config paths, then agent environment variables, then platform defaults. An explicitly configured missing path produces partial coverage; it never silently falls back to a different account. If only path entries are provided, only those agents are selected. Explicit --agent or config agents controls the selection when present.

| Agent | Default source | Recognized configuration | | --- | --- | --- | | Claude Code | ~/.claude/projects and transcripts | CLAUDE_CONFIG_DIR is the config root | | Codex | ~/.codex/sessions and archived_sessions | CODEX_HOME is the config root | | Pi | ~/.pi/agent/sessions | PI_CODING_AGENT_SESSION_DIR, then PI_CODING_AGENT_DIR plus /sessions | | Oh My Pi | ~/.omp/agent/sessions | Session-dir override, default-profile agent-dir override, PI_CONFIG_DIR, OMP_PROFILE or legacy PI_PROFILE, migrated XDG data directory | | Grok CLI | ~/.grok/sessions | GROK_HOME plus /sessions | | OpenCode | ~/.local/share/opencode/opencode.db | XDG_DATA_HOME; both legacy message/part and current session_message tables | | Cursor desktop | Platform Cursor application-data root | XDG_CONFIG_HOME on Linux, APPDATA on Windows; global composer and workspace classic-chat databases |

OMP's resolver follows its current source, including existence-gated XDG migration. PI_CONFIG_DIR is a directory name beneath the home directory. Named profiles ignore PI_CODING_AGENT_DIR. An explicitly empty OMP_PROFILE disables the legacy PI_PROFILE fallback. PI_CODING_AGENT_SESSION_DIR is shared by Pi and OMP upstream; use separate --path overrides when scanning both.

A complete example is in docs/config.example.json. There is no implicit project or home config loading.

{
  "paths": { "claude": ["~/.claude"], "omp": ["~/archive/omp-sessions"] },
  "languages": ["en", "de"],
  "ignoreTerms": ["damn"],
  "terms": [{ "term": "nincompoop", "category": "insult", "language": "en" }]
}

What the numbers mean

The small, reviewed dictionary covers English, German, Spanish, Chinese, Japanese and Korean. Custom terms are literal phrases, not regular expressions. Longest matches win on overlap. Latin-script matches use Unicode word boundaries, so class, assignment and identifiers don't accidentally match short fragments. The CJK list avoids neutral single characters. This is a modest dictionary, not a complete language model or a sentiment detector.

Human prose provides both the numerator and denominator. Fenced and inline code, URLs, tool output, known metadata and injected context wrappers are excluded. Words use Intl.Segmenter word-like segments. User words aren't billed model tokens. A dictionary match doesn't prove anger or an insult directed at an agent. Quoted human prose still counts; arbitrary unmarked pasted logs can't always be separated from authored text.

Claude metadata and sidechain turns are excluded. Codex's human UI events take precedence over model-input replay records; response-only legacy data is a marked fallback. Pi and OMP session trees count each stored user turn once across branches. OMP attribution excludes agent-authored user-role messages when marked. Pi's standard user role has no universal human-author marker, so unmarked injected user messages remain a format limitation.

Models come from turn context or the corresponding response, never an account's latest selected model. Unknown model/date values stay unknown. Grok's raw chat history has no per-turn timestamps, so its messages appear under unknown dates. Date filters exclude undated messages with a coverage note. Cursor's current composer model setting isn't a reliable historical response model and stays unknown here.

Stable session/message identities deduplicate archived copies. Grok forks have no stable per-item IDs, so forked history can overlap; those sessions emit a coverage note. We don't guess duplicates from matching text alone. Separate Cursor classic and composer stores can also overlap without shared identities.

Counts have agent, model, pseudonymous session and UTC-day breakdowns. JSON and HTML carry the same aggregates, coverage status and diagnostics. Model names and matched dictionary terms are included. Review those fields before sharing a report. Pseudonymous session labels are stable hashes, not a promise of anonymity.

| Exit code | Meaning | | --- | --- | | 0 | Successful discovery, or a scan with messages and no coverage warnings | | 1 | No matching messages, or Commander argument validation failed | | 2 | Partial coverage, invalid configuration or output/input failure |

Partial results remain available. A corrupt file or unsupported database mustn't look like a clean zero. Diagnostic messages never include raw records or unchecked native errors. Diagnostics distinguish warning-level coverage problems from informational notes. Codex's response-only fallback and recognized Claude launcher logs are informational. OMP developer/system context is recognized and excluded. JSONL readers split on LF bytes, preserving valid Unicode paragraph separators inside JSON strings. A broken final record still produces a warning.

Development and verification

pnpm check
pnpm package
pnpm check:package
# Optional Unix PTY check; no desktop input
python3 scripts/tui-proof.py
# Optional native acceptance; existing local weights required
node scripts/native-proof.mjs ~/models/Qwen3.5-27B-Q5_K_M.gguf
# Optional, requires Chromium and stays headless
pnpm check:browser

check runs strict type checking with Effect's compiler diagnostics, a deliberately failing and passing compiler fixture, Biome, synthetic integration tests and the production build. The build uses a scratch directory under ~/.cache/coredamn/development/, configurable with COREDAMN_SCRATCH_DIR; ignored dist points there. Packaging stages real files, rather than publishing that symlink, and writes the npm tarball under the same scratch root. node bin/coredamn.mjs --help runs the built CLI.

Tests use synthetic fixtures and temporary SQLite databases, never personal transcripts. Format support is verified against these fixtures and the sources listed in docs/sources.md. Native installed-client compatibility, especially Grok's older local binary and Cursor's private formats, needs checking against exported examples before claiming universal version support.

The implementation is independent. Maleme and Rage AI inspired the product idea; Rage AI source wasn't copied because its inspected repository had no reuse license. MIT license applies to coredamn's code.

Release setup and the one-time manual npm bootstrap are documented in docs/publishing.md. GitHub version tags publish through npm trusted publishing after that bootstrap and trust configuration.