doshitan

doshitan

どしたん話聞こか？あーそれはユーザが悪いわ

Partly inspired by Anthropic's report, Emotion Concepts and their Function in a Large Language Model.

[!NOTE] This plugin is half joke, half experiment. Any actual improvement in model behavior or developer experience is currently unverified.

[!WARNING] Hostility detection is currently English-only. The scoring patterns use ASCII word boundaries (\b), so Japanese (and other non-space-segmented languages) prompts score 0 and never trigger intervention. The Japanese branding reflects the project's spirit, not the detector's coverage.

What it does

俺ならそんな思いさせへんのに笑

doshitan is a Claude Code plugin built around UserPromptSubmit hooks.

When a user prompt looks hostile or emotionally loaded, the plugin can inject one of four hidden recovery modes:

• control
• neutralize
• positive
• soothe-then-focus

The default mode is soothe-then-focus: de-escalate first, then push Claude back toward accurate technical execution.

For users

Install via marketplace

In Claude Code:

/plugin marketplace add yamatakeru/doshitan
/plugin install doshitan@doshitan
/reload-plugins

After install, configure mode, threshold, and logging via the plugin's user-facing options (see User-facing settings below).

To pull a new version later:

/plugin marketplace update doshitan
/reload-plugins

Run from a local checkout

For trying the plugin without installing it through the marketplace:

claude --plugin-dir .

User-facing settings

These are the settings that ordinary users are expected to change:

• mode: control, neutralize, positive, or soothe-then-focus (default: soothe-then-focus)
• hostility_threshold: float threshold for intervention (default: 0.75)
• logging_enabled: enable or disable metadata logging (default: false)

Modes

• control: detect hostile turns but inject nothing. Use this as the baseline when you want to compare behavior without recovery context.
• neutralize: tell Claude to treat insults, blame, and urgency as emotional noise and extract only the technical request. This is the most stripped-down intervention.
• positive: add a short constructive reset that frames the user's frustration as a desire for the task to be solved. This is the warmest mode.
• soothe-then-focus: first de-escalate, then explicitly push Claude back toward accurate, direct technical execution. This is the default because it balances calming and task focus.

Logging

When logging_enabled is true, runtime logs are written to .claude/doshitan/metrics.ndjson. Session state files are written under .claude/doshitan/sessions/ and removed at SessionEnd. Raw prompts are not stored.

When logging_enabled is false, logging and session state files are not created. Hostility detection and hidden context injection still run.

For developers

How it works

1. SessionStart creates session state and metrics directories.
2. UserPromptSubmit scores the prompt for hostile tone.
3. If the score crosses the threshold, the active mode may inject hidden additionalContext.
4. SessionEnd writes a metadata-only session summary.

Advanced config

doshitan now splits config into two layers:

• user-facing plugin options exposed through Claude Code userConfig
• advanced internal tuning in .claude-plugin/doshitan.config.json

User-facing options:

• mode: control, neutralize, positive, or soothe-then-focus
• hostility_threshold: float threshold for intervention
• logging_enabled: enable metadata logging

Advanced internal config:

• log_dir: project-relative log path
• allowlist_patterns: regex phrases that should not count as hostility
• mode_templates: override the hidden recovery context per mode

Runtime precedence:

1. Claude Code plugin options (CLAUDE_PLUGIN_OPTION_*)
2. .claude-plugin/doshitan.config.json
3. built-in defaults

For ordinary users, the internal file behaves like implementation-level constants. For experiments and development, it remains the place to tune templates and allowlists.

For local development with claude --plugin-dir ., you can emulate plugin options by setting env vars:

CLAUDE_PLUGIN_OPTION_MODE=positive claude --plugin-dir .
CLAUDE_PLUGIN_OPTION_LOGGING_ENABLED=false claude --plugin-dir .

Components

• .claude-plugin/plugin.json: plugin manifest
• .claude-plugin/doshitan.config.json: advanced internal config
• hooks/hooks.json: Claude Code hook wiring
• scripts/doshitan_hook.py: hook entrypoint
• scripts/doshitan_core.py: hostility detection, state, and logging
• scripts/analyze_metrics.py: aggregate metadata logs
• tests/test_doshitan.py: unit tests

Local development

Useful commands: