Prompt deduplication (~20% reduction, 4,743 chars saved):
- evil_miku_lore.txt: remove intra-file duplication (height rule 2x,
cruelty-has-substance 2x, music secret 2x, adoration secret 2x),
trim verbose restatements, cut speech examples from 10 to 6
- evil_miku_prompt.txt: remove entire PERSONALITY section (in lore),
remove entire RESPONSE STYLE section (now only in preamble),
soften height from prohibition to knowledge
- miku_lore.txt: remove RELATIONSHIPS section (duplicates FRIENDS)
- miku_prompt.txt: remove duplicate intro, 4 personality traits
already in lore, FAMOUS SONGS section (in lore), fix response
length inconsistency (1-2 vs 2-3 -> consistent 2-3)
Preamble updates (evil_mode.py, evil_miku_personality.py, llm.py,
miku_personality.py):
- Response rules now exist in ONE place only (preamble)
- Height rule softened: model knows 15.8m, can say it if asked,
but won't default to quoting it when taunting
- Response length: 2-4 sentences (was 1-3), removed action template
list that model was copying literally (*scoffs*, *rolls eyes*)
- Added: always include actual words, never action-only responses
- Normal Miku: trim CHARACTER CONTEXT, fix 1-3 -> 2-3 sentences
Directory reorganization:
- Move 6 persona files to bot/persona/{evil,miku}/ subdirectories
- Update all open() paths in evil_mode.py, context_manager.py,
voice_manager.py, both Cat plugins
- Dockerfile: 6 COPY lines -> 1 (COPY persona /app/persona)
- docker-compose: 6 file mounts -> 2 directory mounts
(bot/persona/evil -> cat/data/evil, bot/persona/miku -> cat/data/miku)
Evil Miku system (previously unstaged):
- Full evil mood management: 2h rotation timer, mood persistence,
10 mood-specific autonomous template pools, mood-aware DMs
- Evil mode toggle with role color/nickname/pfp management
- get_evil_system_prompt() with mood integration
Add test_evil_moods.py: 10-mood x 3-message comprehensive test
39 KiB
39 KiB