No description
Find a file
2026-07-22 14:25:42 +09:00
server_app Increase piper timeout to 30s for container cold-start 2026-07-20 18:41:40 +09:00
skills/mental-health-advisor Add persistent patient file + background analysis + anti-sycophancy 2026-07-16 22:18:57 +09:00
.dockerignore Containerize: Dockerfile + podman run script 2026-07-20 18:34:33 +09:00
.gitignore Fix --file ordering: prompt BEFORE --file + logs in project dir 2026-07-16 23:19:23 +09:00
AGENTS.md Add mental health research advisor project 2026-07-16 21:06:34 +09:00
Dockerfile Containerize: Dockerfile + podman run script 2026-07-20 18:34:33 +09:00
entrypoint.sh Containerize: Dockerfile + podman run script 2026-07-20 18:34:33 +09:00
launch_mh.sh Add mental health research advisor project 2026-07-16 21:06:34 +09:00
launch_voice.sh Switch to OpenRouter cloud STT (openai/gpt-audio-mini) 2026-07-16 22:07:26 +09:00
opencode.jsonc Add mental health research advisor project 2026-07-16 21:06:34 +09:00
readme.md Show session ID as visible blockquote in README 2026-07-22 14:25:42 +09:00
run.sh Containerize: Dockerfile + podman run script 2026-07-20 18:34:33 +09:00
voice_chat.py Fix --file ordering: prompt BEFORE --file + logs in project dir 2026-07-16 23:19:23 +09:00

Mental Health Research Advisor

OpenCode session: ses_09535ec94ffe4rTjXnPNCrMRjU Resume: opencode --session ses_09535ec94ffe4rTjXnPNCrMRjU

An opencode-powered mental health research assistant for laypeople in Japan. Helps find and evaluate mental health evidence without the TikTok pop-psychology, Instagram therapy advice, or AI-generated misinformation of generic web search.

This is NOT a substitute for a mental health professional. It is a research librarian.

Voice Server (Containerized)

cd /src/llm_mentalhealth
OPENROUTER_API_KEY=sk-or-v1-... DEEPSEEK_API_KEY=sk-7e31... ./run.sh up
# Access: http://localhost:8010 (or Tailscale IP http://100.93.132.70:8010)

Quick Start

cd /src/llm_mentalhealth && ./launch_mh.sh

Then ask anything about mental health. The mental-health-advisor skill auto-loads.

What It Does

You ask You get
"Does CBT work for depression?" Systematic review with PMIDs, effect sizes in plain language, NNT
"Is my antidepressant available in Japan?" PMDA approval check, NHI coverage, Japan-specific dosing
"ADHDの治療法は" Evidence-graded treatment review, JSPN guidelines, Japan drug formulary
"Is microdosing psilocybin effective?" Honest assessment of evidence quality (early-stage, small samples)
"What helps with hikikomori?" Japan-specific context, cultural factors, evidence-limited but framework provided

Architecture

User query
    │
    ├─ 1. KB search (DSM-5-TR, WHO mhGAP, Cochrane MH, NICE MH, Japan MH system)
    │       └─ Ground-truth frameworks before any literature search
    │
    ├─ 2. PubMed (search → fetch articles → fulltext → related articles)
    │       └─ Primary biomedical evidence, MeSH-indexed, graded
    │
    ├─ 3. Web search (domain-whitelisted only)
    │       ├─ International: nimh.nih.gov, psychiatry.org, nice.org.uk, ...
    │       └─ Japan: mhlw.go.jp, ncnp.go.jp, jspn.or.jp, ...
    │
    └─ Output: disclaimer (+ crisis resources) → PICO → evidence summary → Japan context → plain language

No TikTok. No Instagram therapy influencers. No "Big Pharma hiding the cure." No AI therapy app marketing.

Japan-Specific Coverage

  • PMDA psychiatric drug approval status (Japan's formulary differs significantly)
  • JSPN clinical practice guidelines (Japanese equivalent of APA/NICE)
  • NHI coverage for psychiatric care and medications
  • Pharmacogenomics — CYP2D6*10, CYP2C19 variants in Japanese populations affect psychiatric medication
  • Cultural context — hikikomori, taijin kyofusho, karoshi/karojisatsu, mental health stigma
  • Care pathways — hokensho, mental clinics, university hospital psychiatry, designated mental health physicians
  • Japanese dosing — often substantially lower than Western dosing for psychiatric medications

Evidence Grading

Tier Type Example
S Cochrane/NICE/WHO mhGAP/APA guideline Gold standard
A Systematic review & meta-analysis Pooled RCT analysis
B Large multi-site RCT (n>200, pre-registered) Active comparator preferred
C Cohort/case-control, open-label Correlation ≠ causation
D Case series, expert opinion Hypothesis only
E Pre-prints, animal models Not peer-reviewed
F TikTok, Instagram, influencer claims Noise

Every response gets a grade.

Files

/src/llm_mentalhealth/
  launch_mh.sh          — one-command launcher
  opencode.jsonc        — project config
  AGENTS.md             — project instructions (loaded by opencode)
  readme.md             — this file

~/.config/opencode/skills/mental-health-advisor/
  SKILL.md              — the skill (auto-loads on mental health keywords)

~/content-repo/books/mental-health-reference/
  dsm5-tr-criteria.md   — key diagnostic frameworks
  who-mhgap-guide.md    — WHO Mental Health Gap Action Programme
  cochrane-mental-health.md — systematic review methods for MH
  nice-mh-guidelines.md — UK evidence-based guidelines (widely referenced)
  japan-mental-health-system.md — care pathways, law, culture
  mental-health-quick-reference.md — consolidated quick-reference

Requirements

  • opencode binary
  • PubMed MCP (remote, already in global config)
  • Tavily MCP (remote, already in global config)
  • kb-mcp-server (local, CONTENT_REPO=~/content-repo)

Run ./launch_mh.sh --check to verify without launching.

Crisis Resources (Japan)

If you or someone you know is in crisis:

Resource Contact Language
Tokyo English Lifeline (TELL) 03-5774-0992 English
Inochi no Denwa 0120-783-556 Japanese
Yorisoi Hotline 0120-279-338 Japanese / foreign languages
Emergency 119