Over-50s Health Advisor

A Claude Code and OpenCode Agent definition for evidence-based, age-appropriate health, fitness, nutrition, and longevity guidance for adults 50+. The agent works with local context files only and always recommends confirming advice with a healthcare professional.

Features

• Evidence-based guidance with citations and a Sources section, weighted toward clinical measurements over device streams
• Noise-floor discipline: reports a metric as unchanged, not a trend, when its change is within measurement error
• Collection-cost transparency: states what decision a requested metric or artifact could change, and names an exit condition for any recurring log
• Optional long-term memory via Hindsight — a self-hosted recall and reflection layer over your local history files
• Safety boundaries and red-flag referral policy
• Local context management via Markdown files
• Install via Claude Code plugin system (/plugin install over-50s-health@ali5ter) or OpenCode (auto-discovered)
• Automatic context file creation on first run

Repository structure

This repository contains the agent definition and context templates for distribution. When installed via the plugin system, the agent is managed by Claude Code and your personal context files are stored in your home directory.

agents/
  advisor.md                       # Claude Code agent definition (source)
context/
  templates/                       # Reference context templates
    INITIAL_USER_INFORMATION.md
    CLIENT_HEALTH_CONTEXT.md
    CLIENT_PREFERENCES.md
    SESSION_NOTES.md
    SOURCES.md
    SESSION_NOTES_ARCHIVE.md       # On-demand narrative history
    METRICS_LOG.csv                # On-demand structured metrics for trend reports
    history/                       # Per-session history templates
      INDEX.md
  README.md
.claude-plugin/
  plugin.json                      # Plugin manifest
hooks/
  hooks.json                       # SessionStart, PreToolUse, SubagentStop + Stop hooks
hooks-handlers/
  sync-templates.sh                # Copies templates from plugin cache to ~/.claude/over-50s-health-advisor/templates/
  guard-bash-scope.sh              # Agent-scoped PreToolUse hook: scopes the advisor's Bash access
  retain-session.sh                # SubagentStop + Stop hook: persists session to Hindsight via hindsight-retain.py
  check-retain-staleness.sh        # SubagentStop + Stop hook: warns when on-disk history is not yet retained
.opencode/
  agents/
    health/
      advisor.md                   # OpenCode agent definition (auto-discovered)
scripts/
  hindsight-retain.py              # MCP transport: retains documents into the health bank (idempotent manifest)
  backfill-history.sh              # One-off: ingest all history into Hindsight
migrate                            # Migration script for v2.x users
README.md
LICENSE

After installation, your personal context files are stored at:

~/.claude/over-50s-health-advisor/
    context/                       # Your personal context files (auto-created on first run)
        ├── INITIAL_USER_INFORMATION.md   # core — read every session
        ├── CLIENT_HEALTH_CONTEXT.md      # core — read every session
        ├── CLIENT_PREFERENCES.md         # core — read every session
        ├── SESSION_NOTES.md              # core — read every session (last ~2 entries)
        ├── SOURCES.md                    # core — read every session
        ├── INDEX.md                      # session index — read on demand
        ├── history/                      # per-session history files
        │   └── YYYY/YYYY-MM-DD.md       # one file per session
        ├── METRICS_LOG.csv               # analysis — read on demand
        └── _legacy/                      # pre-v4.0.0 archives (never deleted)

Requirements

• Claude Code CLI v2.0.73 or later
• Uses Claude Opus 4.6 by default for highest reasoning quality; override with /model if preferred

Install

Claude Code

Inside Claude Code, run:

/plugin marketplace add ali5ter/claude-plugins
/plugin install over-50s-health@ali5ter

The first time you start a health conversation, the agent automatically creates your context files at ~/.claude/over-50s-health-advisor/context/.

OpenCode

The agent ships project-local and is auto-discovered when OpenCode is opened from this project (or globally from ~/.config/opencode/agents/).

Project-local (recommended — no setup required):

1. Open OpenCode from this project directory (or any parent).
2. The agent health/advisor appears automatically under .opencode/agents/.

Global install (use from any project):

mkdir -p ~/.config/opencode/agents/health
cp .opencode/agents/health/advisor.md ~/.config/opencode/agents/health/advisor.md

Invoke

From an OpenCode session:

• Just ask a health question — the agent activates automatically.
• Or target it directly: /agent health/advisor

From Claude Code:

• Just ask a health question — the agent activates automatically.
• Or target it directly: claude --agent over-50s-health:advisor
• Or inside Claude Code: /agent over-50s-health:advisor

Context path

Both Claude Code and OpenCode share the same context files:

~/.claude/over-50s-health-advisor/context/

Uninstall

Claude Code uninstall

Inside Claude Code, run:

/plugin uninstall over-50s-health@ali5ter

OpenCode uninstall

Project-local: nothing to remove — the agent lives in the project directory.

Global:

rm ~/.config/opencode/agents/health/advisor.md

Your personal context files are not removed. To delete them:

rm -rf ~/.claude/over-50s-health-advisor

Migrating from v2.x

If you previously installed via ./install.sh, run the migration script from this repo:

./migrate

This removes the old manually-installed agent file. Your context files are preserved. Then install via the appropriate platform commands above.

Usage

After installation, context files are automatically created at ~/.claude/over-50s-health-advisor/context/ on your first conversation.

1. Fill in ~/.claude/over-50s-health-advisor/context/INITIAL_USER_INFORMATION.md and CLIENT_PREFERENCES.md with your information.
2. Keep CLIENT_HEALTH_CONTEXT.md and SESSION_NOTES.md current as new information is shared with the agent.
3. Maintain SOURCES.md as a curated reference list (the agent will add sources; you can remove low-quality ones and add high-quality evidence).
4. Use the agent from any directory in your chosen platform. The agent will read and update these context files automatically.
5. Keep the "Last updated" dates accurate in each file.
6. Ask for a "trend summary" or "how have I done this year" and the agent will read METRICS_LOG.csv — a structured, append-only record of every metric it has logged — instead of re-reading a year of session notes. See context/README.md for details on METRICS_LOG.csv and SESSION_NOTES_ARCHIVE.md.

Bash access

The agent has Bash for local analysis: reading METRICS_LOG.csv and running your own health export scripts. A PreToolUse hook (hooks-handlers/guard-bash-scope.sh, registered in hooks/hooks.json) scopes what it can actually run: destructive commands (rm, mv, sudo, chmod, chown, dd, force-pushes) and network egress (curl, wget, nc) are denied outright — everything else it allows outright or leaves to the normal permission prompt.

Claude Code silently ignores hooks: set in a plugin agent's own frontmatter, so this hook is registered in the plugin's hooks/hooks.json instead, which applies plugin-wide rather than to one agent. It self-scopes by checking the agent_type field Claude Code passes it and is a no-op outside an active over-50s-health:advisor session — it has no effect on your other Claude Code sessions.

Long-term memory (Hindsight)

The advisor can use Hindsight as an optional recall and reflection layer over your local history files. Hindsight must be self-hosted — it is not bundled in this plugin.

Prerequisites

• A machine running Hindsight 0.10.3+ with an MCP endpoint
• Ollama (or another LLM server) for the LLM provider
• The health bank configured in Hindsight
• Hindsight MCP server configured (via Configure the MCP server below)

Setup

1. Install Hindsight on your host machine. See Hindsight docs for installation instructions.

2. Configure the MCP server for your platform:

   Claude Code (user scope):

   claude mcp add --scope user hindsight-health <command>
   

   OpenCode (user scope):

   opencode mcp add --global hindsight-health <command>
   

   Replace <command> with your Hindsight MCP server command (e.g. the Ollama-based endpoint). The MCP server is restricted server-side to recall, reflect, and retain tools only.

3. Enable retention by writing the enable gate to ~/.hindsight/retain.env — the same persistent-file pattern as the API key, set once and independent of how Claude Code is launched:

   echo 'export HINDSIGHT_RETAIN_ENABLED=true' > ~/.hindsight/retain.env
   

   The SubagentStop/Stop hook (hooks-handlers/retain-session.sh) sources that file and, when enabled, delegates to scripts/hindsight-retain.py, which retains the session's history file into the health bank over the MCP transport. It fails silently if Hindsight is unreachable. A sibling hook, check-retain-staleness.sh, warns on session end when on-disk history has not yet been retained (see Checking staleness below); it is offline and makes no network call.

4. Backfill existing history (one-time):

   ./scripts/backfill-history.sh
   

   This ingests all history/ files, CONTEXT_ARCHIVE.md, _legacy/ archives, and METRICS_LOG.csv rows into the Hindsight health bank. Retention runs over the MCP transport (scripts/hindsight-retain.py) and keeps a local idempotency manifest at ~/.hindsight/retain-manifest.json; a first run seeds it from the bank, so re-runs skip what is already present and add nothing new.

   Options:

   • --host HOST — custom Hindsight host (default: 192.168.1.63)
   • --dry-run — show what would be retained without sending
   • — skip archives

How it works

• Session start: the agent orients with a targeted recall to the Hindsight MCP server, a small budget, a recent temporal window — querying the Client's recent sessions to surface what is unresolved. Do not use an unbounded reflect (the slow synthesis call), which can return confident but wrong facts; reserve reflect for genuine cross-session reasoning. Treat any recall result as a lead, not a fact — cross-check against the core files. If Hindsight is unavailable, the agent falls back to reading core files + the two most recent history entries.
• Retrieval: when the Client mentions past topics, the agent calls recall first and notes what it checked in a History checked: line. If the MCP call fails, it falls back to INDEX.md + Grep.
• Session end: the SubagentStop/Stop hooks delegate to hindsight-retain.py, which retains the session file into the health bank over the MCP transport (skipping it if the manifest already has it). If unreachable, the file on disk is untouched. A sibling hook, check-retain-staleness.sh, warns on end when on-disk history is not yet retained.

Files are always canonical. Hindsight is a layer on top — if it is unavailable, the agent degrades to the plain-file behaviour with no data loss.

Checking staleness

Because retention can miss sessions or metric groups (a session end that never fired, a backfill run before new sessions existed, a manifest not yet re-seeded), the check-retain-staleness.sh hook makes that gap observable: on every SubagentStop/Stop it compares the local manifest against the on-disk history/ files and METRICS_LOG.csv and warns (on stderr, non-blocking) when anything is not yet retained. It is offline — it reads only the manifest and the local files, making no network call — so it is safe on every session end. Run it on demand for the full list:

./hooks-handlers/check-retain-staleness.sh --report

The fix for any reported gap is idempotent — re-run ./scripts/backfill-history.sh, which skips what is already retained and adds only what is missing. The check runs only when HINDSIGHT_RETAIN_ENABLED=true.

Known limitations

• reflect on local models (e.g., qwen3.8:27b via Ollama) can produce confident factual errors. v5.1.0 mitigates this at session start by orienting with a targeted recall instead and treating the result as a lead to cross-check against the files; reserve reflect for genuine cross-session synthesis, and always verify figures against the files.
• Recall quality: temporal queries are strong (8/8 in testing), entity queries are good but not perfect (9/12 at top-5 in testing).
• The spike results are documented in the sysadmin-studio repo.

Invoking the Agent

Invoking with Claude Code

The health advisor agent is invoked automatically when you ask a health-related question in Claude Code. You can also target it directly using the --agent flag:

claude --agent over-50s-health:advisor

This sets the agent as the preferred delegate for the session. Claude Code routes health and wellness queries to it automatically; for anything else, the default assistant responds.

Alternatively, just start Claude Code normally and ask a health question — the agent will activate based on your query.

Invoking with OpenCode

Invoke from within an OpenCode session:

/agent health/advisor

Or just ask a health question — the agent activates based on your query's topic.

Starting a Conversation

When you send your first message (even just "Hi"), the agent reads your context files and greets you with a brief summary of your current health focus. If context files are missing, it will create them and prompt you to fill in your initial information.

Once you've filled in your initial context files, you can begin interacting with the agent. Here are some ways to start:

Initial Engagement:

• "Where do you see the context files?" — the agent will analyze your context and provide a summary of your health profile, helping establish rapport.
• "Can you review my health information and suggest areas to focus on?"
• "What do you know about me so far?"

Direct Queries:

• "What strength training program would you recommend for me?"
• "Can you suggest a weekly meal plan that supports my goals?"
• "How can I improve my metabolic health based on my recent labs?"
• "What mobility exercises should I prioritize?"

Specific Requests:

• "Review my recent blood panel and explain what the trends mean"
• "Create a 4-week progressive workout plan for me"
• "Suggest supplements appropriate for my age and health status"

The agent will read your context files, provide evidence-based guidance with citations, and update your context files as you share new information.

Mobile and voice access

The agent can be used hands-free from a phone or tablet using two Claude Code features:

Remote Control — continues a local Claude Code session on any device:

1. Start a session on your Mac with the advisor agent active.
2. Run /remote-control to generate a session URL and QR code.
3. Scan the QR code in the Claude app or open the URL in a browser on your phone.
4. Your phone controls the local session with full access to your context files.

Your computer must stay on and the session must remain open. See the Claude Code Remote Control docs for details.

Voice input — speak your queries instead of typing:

Run /voice in your Claude Code session to enable voice input. Hold the spacebar to speak, release to send. Transcription is free and does not count against rate limits. Responses are text-only; voice output is not yet available.

Combined, /remote-control and /voice let you speak health queries from your phone while the agent reads and updates your context files on your computer in the background.

Available on Pro, Max, Team, and Enterprise plans.

Safety and medical boundaries

• Educational guidance only, not diagnosis.
• Immediate referral for emergency symptoms.
• Always include a reminder to confirm with a healthcare professional.

Evidence and citations

• Use credible, evidence-based sources.
• Provide citations with links in every response that includes recommendations.
• End responses with a Sources section.

License

MIT License, Copyright (c) 2026 Alister Lewis-Bowen.