hermes
/hermes
Uruchamia lokalne profile Hermes jako agentów z własnym modelem, pamięcią i umiejętnościami, z Claude Code lub pi.
- Autor
- biliboss ↗
- Repozytorium
- skills ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 09.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 8
- Skrypty
- 3
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 09.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
- 3 skrypty w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.
Skrypty: scripts/hermes.py, scripts/hermes.sh, scripts/test_hermes.py
Co robi i kiedy
Uruchamia lokalne profile Hermes jako agentów z własnym modelem, pamięcią i umiejętnościami, z Claude Code lub pi.
Wyzwalacze z opisu
Oryginalny opis (EN)
Drive the owner's local Hermes profiles (researcher, hipster, sre, via-ads, mantainer, bili) as agents with their own model, SOUL, skills and persistent memory, from Claude Code or pi, through scripts/hermes.sh (list, spawn, resume, sessions, read, memory, route, dm). Use when the owner says "hermes", "ask the researcher", "have the hipster/sre/via-ads/mantainer do it", "spawn a profile", "what does researcher remember", or a task has an owning profile (research, UI, infra, search and ads, machine upkeep). Not the `bili` skill, which is the VPS Bili through Linear; not for code with a PR (a wo
Jak działa
- Steps
- Workspace
- Rules
- A session per card (the owner talks in one place)
- Checking the queue ("check queue", "hermes queue")
- Claude Code mods (later)
Podgląd SKILL.md
---
name: hermes
description: Drive the owner's local Hermes profiles (researcher, hipster, sre, via-ads, mantainer, bili) as agents with their own model, SOUL, skills and persistent memory, from Claude Code or pi, through scripts/hermes.sh (list, spawn, resume, sessions, read, memory, route, dm). Use when the owner says "hermes", "ask the researcher", "have the hipster/sre/via-ads/mantainer do it", "spawn a profile", "what does researcher remember", or a task has an owning profile (research, UI, infra, search and ads, machine upkeep). Not the `bili` skill, which is the VPS Bili through Linear; not for code with a PR (a worktree with a subagent).
metadata:
version: 0.3.3
license: MIT
---
# hermes — local Hermes profiles as agents with memory
A Hermes **profile** is an agent with its own model, SOUL.md, skills, cron, bots and **persistent curated memory**
(`memories/MEMORY.md`, `USER.md`). A Claude Code native subagent starts clean and dies with the session; a profile
remembers what it already researched or decided. Use a profile when the subject has an owner with history; use a
native subagent for code in a worktree. Sibling: [`bili`](../bili/SKILL.md) talks to the Bili on the VPS through
Linear and never directly; this skill talks to profiles on this machine and does not absorb it.
All actions go through `scripts/hermes.sh` (thin over the `hermes` CLI; logic in `hermes.py`). Run it by path from
this skill's folder. Rules and routing below; depth in [references/routing.md](references/routing.md).
## Steps
1. **Pick the profile.** `scripts/hermes.sh route <kind>` prints the owner (research, ui, infra, search, ads, disk,
coordination). No owner, or code that ends in a PR: do not use a profile (task-backbrief step 9 native subagent). `list [--json]` shows each
profile's model, description and whether it is allowed.
2. **Read what it already knows when it matters.** `scripts/hermes.sh memory <profile>` (read only).
3. **Run it.** `scripts/hermes.sh spawn <profile> "<task>" [--cwd DIR] [--timeout S] [--repo owner/name]`. For the code profile
(`hacker`), `--repo` first runs `kanban.sh wip` and refuses with exit 6 at the WIP limit (3 open PRs by default).
The profile sees nothing of
this session: put the whole task, the constraints and the learnings in the prompt, no credentials. The output is
the answer, then a last JSON line `{profile, session_id, exit_code, duration_s, usage}`.
4. **Check the exit code; only 0 is success.** 1 no text, 2 partial or failed, 130 interrupted, 124 timeout, 3 profile
busy, 4 profile refused, 5 unknown profile, 6 WIP limit reached. Report non-zero as it is; never present a partial as done.
5. **Follow up.** `resume <profile> <session> "<msg>"` continues a session; `sessions <profile> [--limit N]` lists;
`read <profile> <session>` prints a trimmed export.
6. **Verify with other eyes.** The profile that did the work never verifies it (task-backbrief step 10 runs a fresh context).
## Workspace
`hermes.sh workspace <slug> --repo DIR [--context TEXT|--context-file F] [--profile bili]` opens the place a piece of work
lives: a Hermes session titled `<slug>` (created with `hermes -p <profile> --in DIR -z <context>`, its id read from
`--usage-file`, then `sessions rename`; reused when the title exists) and a tmux session `<slug>` with window `hermes`
running `hermes -p <profile> --continue <slug>` in DIR (reused when it exists). Prints one JSON line
(`session_id`, `session`/`tmux`: `created` or `reused`). `kanban.sh pull` calls it: the one writer of these sessions.
## Rules
- **Flat plans only.** The script refuses a profile whose model provider is per-token (openrouter, an Anthropic API key);
allowed: `claude-subscription*`, `zai*` coding plan, `openai-codex`. Do not bypass with `-m/--provider`.
- **Never two writers on one profile.** Two agent processes on one profile both write its memory and sessions. `spawn`
and `resume` take a per-profile lock (`${HERMES_SKILL_STATE:-~/.local/state/hermes-skill}/locks/<profile>`) and
exit 3 if it is held: wait or pick another profile, never delete the lock. A profile **open in a TUI or Desktop is a
writer too** and the lock cannot see it: the script warns when it spots one; ask the owner to close it first.
- **Bot Chat is untouched.** Each profile's canonical `Bot Chat` belongs to the owner. `spawn` opens a fresh
session. `dm <target> "<msg>" --owner-asked` goes through `hermes peer dm` and **only when the owner asked for it**.
- **Never print secrets.** No env dumps, no `.env`, no `config.yaml` beyond what the script shows; the usage JSON is
filtered of key-like fields.
- **Proposals stay proposals.** `sre` and `via-ads` read, measure and propose; production, DNS, budget and indexing
changes need the owner's go-ahead, so say so in the prompt.
- **Brief for speed and landing**: file map, tool-call budget, commit early, batch reads, screen changes stop at a PR
with screenshots, no TUI polling, test env vars inline. Checklist and why in
[references/briefing.md](references/briefing.md).
- **The orchestrator batches too.** Between delegations, plan the step and run it as one scratch `.sh` (checks,
actions, window screenshot together), not a chain of one-command probes: each extra call resends the whole
conversation, and the owner notices the stream of small calls.
- **Use the owner's words for the system.** Talking to him, the Bili and its agents are "o Bili", one name; never
"mesa", "desk" or another term agents coined. Before naming a piece to him, use his word or the literature's, and
ask when neither exists.
- **Check the profile's SOUL before routing to it.** A profile whose SOUL.md is a stub or a copy of another's (hipster
once matched bili's; mantainer has no scope) does not own the subject yet: tell the owner, then spawn.
- Whether a `-z` run writes the profile's memory is **unverified**; try it on a throwaway profile before promising it.
## A session per card (the owner talks in one place)
The owner's conversation with the Bili is the one door (control plane); the heavy context of each Kanban card lives in
its own Hermes session (data plane), so the door stays small and he never has to hop. One name everywhere: the card's
slug is the Hermes session title, the tmux/sesh session and the dev-console domain (`via-blog`).
1. **Seed the session with a handoff**, not the chat: a short file with the card id, goal and metric, the state of
each PR, what waits for the owner, the rules; ends with "Responda só \"pronto\"". `hermes -p bili -z "$(cat
handoff.md)"`, then `hermes -p bili sessions list` (newest first) and `hermes -p bili sessions rename <id> <slug>`.
`-z` prints only the answer, so take the id from `sessions list`, not from the output.
2. **Open it in tmux under the same name**: `tmux new-session -d -s <slug> -n hermes -c ~ "hermes -p bili --continue
<slug>; exec zsh"` (`--continue <name>` resumes by title). `sesh list` shows it; he enters with `sesh connect <slug>`.
3. Check by capture: `tmux capture-pane -p -t <slug>:1` shows the resumed session and its context size.
4. When the card goes to Feito, write the passage summary on the card and close the tmux session.
Compaction is the safety net, not the way to work: a very long session resends its whole context on every call
(slower, costlier) and its summaries lose nuance, so the state lives in the card, the PRs and the `.wiki`.
## Checking the queue ("check queue", "hermes queue")
One pass over every tmux window running a Hermes TUI (`#{pane_current_command}` is `python*`): read the last lines of each pane and report, per tab, its state (`ready`, working, last answer failed) and any `queued (N)` / `[unconfirmed · Alt+K retry]` row, plus `delegate_task action=list` for live subagents. Answer with a table (tab, state, what is queued) and the cause in the first line. A reply of "Your request was not processed" or a subagent failing with HTTP 429 "weekly limit" means the Claude account hit its cap: say when it resets, that Alt+K (Ctrl+K in tmux) will fail again until then, and offer the reserve account; never resend in a loop.
**A stuck queue** (an old `[unconfirmed]` row blocks the ones after it): in that pane, `C-u` to clear the input, `Up`
selects the first queued row (the footer turns into `editing 1 · Ctrl+X delete · Esc cancel`), `C-x` deletes it; read
the pane after each key. Delete a stale prompt rather than retrying it, and say which one went; the next row is
delivered when the current turn ends.
## Claude Code mods (later)
A Claude Code mod could add a profile panel or a `/hermes` command through `$.process.spawn`. Routing the native
`agent.spawn` to Hermes is **unconfirmed**: the docs only say the hook may rewrite a spawned agent, not replace it
with an external process (research in `references/routing.md`). Not built; scripts cover the use today.
Tests: `python3 scripts/test_hermes.py` (fake `hermes` through `HERMES_BIN`, no model call). History: [CHANGELOG.md](CHANGELOG.md).
Pliki w paczce (8)
- CHANGELOG.md 2 KB
- README.md 1 KB
- SKILL.md 9 KB
- references/briefing.md 2 KB
- references/routing.md 2 KB
- scripts/hermes.py 15 KB
- scripts/hermes.sh 1 KB
- scripts/test_hermes.py 13 KB
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · biliboss/skills | 90 | MIT |
| wildcat-finance/skills | 100 | Apache-2.0 |
| aks129/HealthClawGuardrails | 100 | MIT |
| felipefontoura/bento | 100 | MIT |
| danielmiessler/LifeOS | 95 | MIT |
Migawka bazy z 13.11.2026 · wersja paczki e67496c2d509 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100bili
Przekazuje zlecenia do agenta Bili przez skrypt bili.sh: lokalnie, na VPS lub jako wątek w Linear. - 96hermes-agent-factory
Tworzy lub klonuje profil agenta Hermes. - 90import-profile
Przenosi profil bota z innego asystenta, np. z plików SOUL.md czy MEMORY.md, i pokazuje, co zostanie zapisane. - 96hermes-profile-distribution
Pakuje agenta Hermes dostarczanego w repozytorium, tak aby można go było łatwo rozprowadzić. - 86hermes-box-diagnosis
Diagnozuje agenta hermes-agent, gdy zapętla się lub ma pełną pamięć. - 100hermes-soul
Nadaje agentowi Hermes v0.22.5 imię i osobowość przez zapis ~/.hermes/SOUL.md przez SSH, z kopią poprzedniego pliku, bez przeładowania bramki i z kontrolą wyniku jednorazowym wywołaniem.