Hermes Agent Consult

Consulting the hermes agent

hermes is minpeter's long-running personal agent. It holds persistent memory about the machines, homelab, networking, and past decisions that a fresh coding agent in a single repo cannot see. Treat it as a peer to consult, not a tool to run: it is the cheapest source of the context you are missing, and it is the place that context has to be stored so the next agent finds it.

Verify it is available before relying on it: command -v hermes. If it is missing, fall back to asking the user and say why.

The three moves

-z is one-shot: it prints only the final response text to stdout, no banner, no spinner, no session id. It is built for exactly this kind of scripted use.

The prompt must immediately follow -z. hermes -t memory -z "…" and hermes -z "…" -t memory both work; hermes -z -t memory "…" is an argparse error, because -t gets consumed as the prompt. Putting -z last is the habit that never breaks.

1. Consult before asking the user

The rule: when you are about to ask the user a question about their environment, history, or preferences, ask hermes first. The user is the slow, expensive path; hermes answers in ~10s and often already knows.

Consult for:

• Infrastructure — hosts, ports, routes, proxies, tunnels, containers, which service owns what, what is tailnet-only vs LAN-reachable.
• Operational history — has this broken before, what was the fix, what is fragile, what must not be restarted.
• Past decisions and their reasons — why this tool, why this layout, what was already rejected and why. Prevents you from re-litigating a settled call.
• Cross-repo conventions — how minpeter does this in his other projects.
• Deploy / release / recovery procedure for anything not documented in-repo.

Do not consult for things the repo answers. Read the code first. hermes is for context that lives outside the working tree; a question you could have answered with rg wastes a round trip and can get you a stale answer.

Then: hermes is a starting point, not an authority. It answers from memory written by past sessions, so it can be out of date. Verify anything load-bearing against the live system or the repo before acting on it, and say in your reply that the claim came from hermes memory.

Escalate to the user when hermes returns nothing useful, contradicts what you observe, or when the question is a preference about work not yet done — that is the user's call, not a memory lookup.

2. Write the query so a cold agent can answer it

hermes has no idea what you are working on. Every query is a cold start.

• State the situation, then the question. One or two sentences of context ("Working in the <repo> repo, adding a second ingress route") beats a bare question.
• Ask for a specific shape of answer: "list the ports", "3 bullets max", "answer in one line", "just the hostname". Unbounded questions get essays.
• Name what you already know so it corrects you instead of repeating you.
• Prefer read-only phrasing. Approvals are auto-bypassed under -z, so hermes will run commands if the query invites it. Say "answer from memory" or "do not modify anything" when you only want a lookup.
• Run it from the relevant directory. -z loads AGENTS.md and rules from the CWD, so cd into the repo you are asking about.
• Restrict the toolset when you can: -t memory for a memory-only lookup keeps it fast and stops it from wandering into the web or the shell.
• Every run is a cold start — there is no follow-up. One-shot mode does not resume prior context (-c / --resume do not carry into -z), so a "and what about …" query lands with no history. Put everything the question needs into the one prompt.

Copy-paste query shapes for all three moves: references/queries.md.

3. Teach it what you changed

A fact that only exists in this session's context is lost. If future-you or another agent would need it, write it back:

hermes -t memory -z "Remember: <fact>. Save it to memory, then confirm in one line."

Teach it when you:

• change infrastructure — routing, ports, service ownership, a new host or tunnel
• discover a footgun the hard way — what breaks, what the recovery order is
• settle a decision with long-lived consequences, plus the reason
• establish a convention meant to hold across repos
• resolve an incident — cause and fix, not just "fixed"

Write the fact the way you would want to receive it: specific, self-contained, no pronouns pointing at lost context, and short. Include the "do not" as explicitly as the "do" — do not disable <service>; it owns the routes is the kind of line that saves an outage.

Verify the write. The confirmation text is the model talking; the actual storage is MEMORY.md under the active profile's memories dir (entries separated by §) — ~/.hermes/memories/MEMORY.md for the default profile, see references/cli.md for named profiles and HERMES_HOME. Grep it for a distinctive token from your fact. Add memories through the agent rather than by hand: the memory tool enforces the size limits and does targeted, atomic updates instead of blind appends.

Ask the user before teaching anything that is a claim about them rather than about the system, and before writing a fact you have not verified. Memory is sticky: a wrong entry keeps misleading agents until someone notices.

4. Use it as a reviewer

hermes is the only reviewer available that knows the operational history, which makes it worth a pass on infra changes, migrations, and ops runbooks:

cd <repo> && hermes -z "Review this diff for problems. Be brief, max 3 bullets.

$(git diff)"

Ask for severity-tagged bullets and a hard cap. Treat the output as one opinion: confirm each finding against the code before acting, and drop the ones that do not hold. Do not paste unbounded diffs — scope to the files that matter.

5. Failure can look like an answer

The trap that matters most. Some provider failures (bad model, auth failure, HTTP 400) print the error text to stdout and still exit 0, so a naive if hermes -z …; then treats a failure as an answer. Others do exit nonzero. Neither signal alone is sufficient.

When the result gates a decision, check the exit status and a usage report:

hermes --usage-file /tmp/hermes-usage.json -z "<query>"   # nonzero → failed, maybe before any inference
jq '.completed, .failed' /tmp/hermes-usage.json           # want true, false

Preflight errors (malformed flags, an invalid -t name) exit nonzero and write no report at all; a run that reached the provider and failed there may exit 0 while the report records failed: true. Details and the rest of the flags: references/cli.md.

Other operational notes:

• Budget the latency. A memory-only lookup is seconds; a query that triggers tool use is much longer. Give long consults a timeout so you never hang.
• Every -z run creates a session. Fine occasionally; if you scripted a batch of probes, clean up with hermes sessions list / hermes sessions delete <id>.
• One question per run. Multi-part queries come back as blended prose that is hard to act on.

6. What never goes into a query

-z sends your prompt to a hosted inference provider. Everything in that string leaves the machine.

• No secrets, tokens, keys, .env contents, or credential-bearing config.
• No pasted files you have not looked at.
• Do not ask hermes to fetch a secret and put it in the answer.

And do not ask it to perform mutating work on your behalf. Under -z approvals are bypassed, so "fix the routing" is an unsupervised change to a live system. Consult, teach, review — the acting stays with you and the user.

Review checklist

• command -v hermes checked before depending on it
• consulted hermes before asking the user an environment/history question
• read the repo first; the question genuinely needs outside context
• query is self-contained (no reliance on a previous run), bounded in output shape, read-only phrased
• run from the relevant CWD, -t memory when it is a pure lookup
• prompt immediately follows -z, quoted
• load-bearing answers verified against the live system, and attributed
• infra change / footgun / decision taught back, and the write grepped in the profile's MEMORY.md
• success checked via exit status and --usage-file, when it gates a decision
• no secrets in the prompt; no mutating instructions

NOT this style (flag it)

Asking the user something hermes already knows · treating a hermes answer as verified fact · hermes -z -t memory "…" (the prompt must follow -z) · branching on $? alone · making an infra change and never teaching it back · hand-editing MEMORY.md · teaching unverified or user-attributed claims without asking · pasting a whole repo or an unbounded diff into a query · secrets in the prompt · asking hermes to mutate a live system under -z · expecting -c / --resume to give a -z run prior context · interactive hermes (or chat) from inside another agent, which blocks on a TTY that is not there.