Hermes Doctor

Linters check your code. Hermes Doctor checks your agent's mind.

A local, read-only annual physical for a long-lived personal Hermes Agent installation.

Hermes Doctor scans bloated memories, stale skills, reminder/cron drift, oversized sessions, broken local Markdown links, and runtime warnings — without modifying anything.

As of v0.4, it also includes a conservative Hermes update pre-flight check: upstream-behind count, local repository changes, known Telegram timeout-patch signatures, UTF-8 BOM risks, and skill platforms: frontmatter hygiene.

Hermes gets better as it remembers. But long-lived personal agents also accumulate entropy.

It started as a family physician's tool for treating a different kind of patient: my personal AI agent. Designed by someone who thinks in preventive checkups, not emergency surgery.

Early public preview. Useful, small, and intentionally conservative.

Non-goals

Hermes Doctor is deliberately small. The following are not going to happen:

• ❌ No --fix mode, ever. The doctor writes the prescription; you go to the pharmacy.
• ❌ No cloud service, dashboard, or telemetry. Reports are local-only.
• ❌ No automatic memory deduplication, skill rewriting, or cron reconciliation.
• ❌ No support for other agent frameworks (AutoGPT, LangChain, etc.) until Hermes itself is stable.
• ❌ No generic Markdown linting — there are better tools for that.
• ❌ No runtime dependencies in the core scanner. Stdlib only.

Who is this for?

Hermes Doctor is for people who run Hermes Agent as a long-lived personal agent and want to know when local state is getting messy:

• people using Hermes memories and skills heavily
• people relying on reminder / cron automation
• people keeping Markdown-based personal state around an agent
• people who want safe diagnosis before cleanup
• people interested in personal agent observability and hygiene

What it checks

• Markdown bloat and broken local links
• Memory / skill size, duplication, and mutable project-fact candidates
• REMINDERS.md vs hermes cron list consistency
• Session file size
• Recent runtime / gateway warning and error event counts
• Hermes update readiness: upstream-behind count, local changes, Telegram timeout patch signature, UTF-8 BOM risks
• Skill platforms: frontmatter drift, including the folded-description regression pattern
• Domain scores and overall health score

Safety model

Hermes Doctor v1 is observational and read-only.

It does not:

• edit files
• delete files
• deduplicate memories or skills
• reconcile reminders
• modify cron jobs
• scan external project folders unless explicitly requested
• send your local data anywhere

Reports redact paths, secret-like strings, and identifier-like strings on a best-effort basis. Do not publish reports from real personal deployments without reviewing them first.

Privacy

• No telemetry.
• No network calls, except local execution of installed hermes CLI commands.
• Reports are generated locally.
• Raw Hermes command output is excluded by default.
• --debug-raw is for local debugging only and should not be used for public reports.
• Redaction is best-effort, not a formal secret-scanning guarantee.

Install

Recommended (zero-install, ephemeral):

pipx run hermes-doctor --summary

Persistent install:

pipx install hermes-doctor

Or from source:

git clone https://github.com/samahn0601/hermes-doctor.git
cd hermes-doctor
python -m pip install -e .

Usage

# Full Markdown report
hermes-doctor

# Compact output for cron/watchdogs
hermes-doctor --summary

# Safe JSON output
hermes-doctor --json

# Write timestamped report and refresh <HERMES_HOME>/reports/health/latest.md
hermes-doctor --write-report

# Automation gate: exit 2 on critical findings
hermes-doctor --summary --fail-on critical

# Print the package version without scanning local state
hermes-doctor --version

# Diagnose stale installs / active-environment version mismatches
hermes-doctor --self-check

# Pre-flight before updating Hermes itself
hermes-doctor --summary --fail-on critical

# Stricter pre-flight gate for automation
hermes-doctor --summary --fail-on warning

By default, Hermes Doctor scans only the Hermes home directory:

hermes-doctor --hermes-home ~/.hermes

External Markdown paths are opt-in:

hermes-doctor --include ./my-notes
hermes-doctor --include-project-hub

Raw Hermes command output is excluded by default. For local debugging only:

hermes-doctor --json --debug-raw

Example summary

Hermes Health: 100/100 (healthy)
Findings: critical=0 warning=0 info=3
Domains: markdown=100, memory_skills=100, reminder_cron=100, session_context=100, runtime_gateway=100
Reminder/Cron: ids=['r_0001']
Runtime: errors=0 warnings=0
Actionable: none

Example finding

Actionable:
- [HD-MEM-002 warning] Memory/skill size warning: <HERMES_HOME>/memories/notes.md size=84KB
- [HD-RMD-005 critical] Active reminder missing cron job: r_0007

Each finding carries a stable ID (HD-MD-…, HD-MEM-…, HD-RMD-…, HD-SES-…, HD-RT-…) — safe to grep for, pin in CI, and reference in issues. The complete list lives in FINDING_IDS in src/hermes_doctor/cli.py.

Hermes Doctor does not fix these automatically. It tells you what to inspect before you change state.

Scoring model

Hermes Doctor uses heuristic domain scores, not a formal proof of system health.

• warning findings apply a small penalty
• critical findings apply a larger penalty
• the weakest domain is weighted heavily so one bad subsystem is not hidden by a good average
• info findings do not reduce the score

Treat the score as a screening result, not a diagnosis carved in stone.

Limitations

• Hermes Doctor is heuristic and may produce false positives or false negatives.
• Hermes CLI output formats may change over time.
• Redaction is best-effort; review real reports manually before sharing.
• v1 is intentionally conservative and does not repair state.
• This is not an official Hermes Agent project.

Development

python -m pip install -e .[dev]
python -m pytest
python -m ruff check .

Roadmap

• v0.1 — initial read-only scanner
• v0.2 — golden fixture corpus, score invariants, PyPI publication
• v0.3 — stable finding IDs, confidence ratings, adversarial redaction corpus, security/contribution docs
• v0.3.1 — install sanity checks: --version, --self-check, source/metadata version guard (current)
• v0.4 — dry-run "review candidates" suggestions. Never executable scripts. Still no silent mutation.
• v1.0+ — boring, trusted, intentionally feature-frozen. Wins by credibility and restraint, not feature volume.

Reporting issues safely

Before opening an issue, please read SECURITY.md — it explains what redaction is and is not guaranteed, and how to share a report without leaking secrets.

For private vulnerability reports (e.g. suspected redaction leak), open a GitHub security advisory instead of a public issue.

Contributing

See CONTRIBUTING.md. The project is intentionally narrow — please skim the Non-goals section first.

License

MIT