Shikigami — TODO

Server (`server/`)

Wire template into llama-server — Add --chat-template-file ./template.jinja (and optionally --host / --port) to mod_llama/start.sh.
Run Heretic on Devstral weights — Use server/heretic to produce a decensored GGUF. Update the model path in start.sh if the output path differs.

Implement mod_whisper — Write integration code (shell or C++) that runs whisper.cpp's stream binary (or main for push-to-talk), captures mic input, and outputs transcribed text. This is the STT entry point.
Fix mod_piper/start.sh — Replace the hardcoded "Good morning." test with the proper invocation: pass beanstalkd host/port as args to hamelin and pipe output to ffplay.
Populate mod_whisper/whisper.cpp submodule — Run git submodule update --init --recursive to pull whisper.cpp source, then write a build.sh for it.

Write the orchestrator process — A daemon (Ruby, Python, or shell) that:
1. Reads transcribed text from the mod_whisper integration
2. POSTs to llama-server's /v1/chat/completions (OpenAI-compatible, maintains conversation history)
3. Pushes the LLM response text into beanstalkd's tts tube for hamelin to consume
Define beanstalkd host/port config — Standardize how host/port are passed across piper-serve, the orchestrator, and any future modules (env vars or a shared config file).

Create ecosystem.config.js (PM2) — Define managed processes for the client: beanstalkd, hamelin (TTS), whisper listener, orchestrator. Optionally a separate server-side ecosystem file for llama-server.
Test full pipeline end-to-end — Mic → whisper → orchestrator → llama-server → beanstalkd → hamelin → speaker.