Bot can now fetch and read the text content of URLs shared in chat.
Uses httpx for async HTTP and BeautifulSoup to strip HTML noise.
Output is capped at 8000 chars to avoid blowing the context window.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Persist per-channel rolling history to disk (channel_history.json) so
context survives bot restarts
- Add fetch_channel_history tool so the bot can pull Discord channel history
on demand when it needs more context (e.g. first mention in a channel)
- Add _ensure_user_tail() called proactively before every LLM call to
prevent Qwen "No user query found" template errors after tool results
- Set tool_choice="required" on the first agent call to guarantee the bot
always uses a tool before replying; subsequent calls use "auto"
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Replace <@USER_ID> Discord mention syntax with @DisplayName for all
non-bot users in incoming messages, so the LLM sees real names instead
of numeric IDs and responds using names naturally.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Structured logging with timestamps and per-task prefixes ([agent], [dossier], etc.)
- Background LLM semaphore to prevent concurrent background calls
- 120s cooldown after background task failures
- 90s timeout on background LLM calls to prevent deadlocks
- BACKGROUND_MODEL env var to use a separate model for dossier/compaction tasks
- Strip <think> tags from model responses (Qwen compat)
- Prompt: read the room for dynamic mood adaptation
- Prompt: blunt one-line self-correction, no apologies
- Prompt: more aggressive stance, never fold under pressure
- Updated personality.txt: less people-pleasing, more authentic
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Dossier extraction and compaction now use a separate BACKGROUND_MODEL
(dolphin-mistral) to avoid Qwen3.5 timeout issues on background tasks.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Switch model to qwen3.5-35b-a3b
- Add structured logging with timestamps and per-task prefixes
- Add background LLM semaphore to prevent concurrent background calls
- Add 120s cooldown after background task failures
- Add 90s timeout on background LLM calls to prevent semaphore deadlock
- Strip <think> tags from Qwen model responses
- Tune prompt: more aggressive stance, no folding under pressure
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Suppai_NekoandClaude Sonnet 4.6