Files
nixos/docs/superpowers/plans/2026-09-08-hermes-agent-mibook.md
T
finn.markwitzandClaude Sonnet 5 0756eb94cd docs: correct spec+plan for self-vs-inputs in imports
The fix subagent found that inputs.hermes-agent.nixosModules.default
inside `imports` causes infinite recursion in this repo's wiring
(inputs comes from _module.args, which merges after imports resolve;
self comes from specialArgs, which is available at import-time).
Tested both ways. Update docs to show self.inputs, matching commit
31e1910.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Vbvwg3Qt2u9nbkEA3wL17S
2026-09-08 17:51:31 +02:00

385 lines
24 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Hermes Agent on mibook Implementation Plan
> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
>
> **Note on nature:** Task 1 is repo-only work verifiable with `nix build` — no runtime tests exist for declarative config, matching this repo's existing convention (see the Immich plan). Tasks 2–5 are an **operator-run runbook**, not agent-automatable: Task 2 needs an interactive `claude setup-token` OAuth login, an Anthropic billing console check, and a Telegram BotFather chat; Task 3 requires typing real secret material into the `$EDITOR` sops opens; Task 4 runs `nixos-rebuild` on mibook, the operator's live daily-driver desktop. **Never paste secret values (tokens, passwords) into an agent session or chat** — Task 3 has the operator type them directly into the editor sops opens. Present Tasks 2–5 for the operator to run and confirm; do not attempt to execute them from an agent session.
**Goal:** Run Hermes Agent as a hardened NixOS systemd service on mibook, reachable via Telegram and Email, authenticated to Anthropic via a Claude subscription (`claude setup-token`) rather than a pay-per-token API key.
**Architecture:** A standard `my.profiles.hermes-agent` module wraps the upstream `services.hermes-agent` module (from `NousResearch/hermes-agent`'s `nixosModules.default`), imported from inside the profile module itself so both machines can evaluate it — only mibook ever turns it on (see Task 1 Step 3). Credentials (Anthropic token, Telegram bot token, email creds, both `*_ALLOWED_USERS` access-control keys) live in one sops secret (`hermes-env`) consumed via `environmentFiles`. No network exposure — Telegram long-polling and Email IMAP/SMTP are both outbound-initiated from mibook.
**Tech Stack:** NixOS (flake-parts), sops-nix, upstream `hermes-agent` flake (Tier 2/best-effort Nix support), Anthropic Claude (subscription auth).
**Spec:** `docs/superpowers/specs/2026-09-08-hermes-agent-nixos-design.md`
## Global Constraints
- Machine: **mibook** only. Do not enable on jupiter.
- Module source: official `github:NousResearch/hermes-agent` (`nixosModules.default`). Not the community `0xrsydn/nix-hermes-agent` wrapper.
- Auth: `claude setup-token` output goes into `ANTHROPIC_TOKEN` in the `hermes-env` secret — **not** `ANTHROPIC_API_KEY`. Confirmed supported by the locked revision's source (Task 4 Step 2); no longer conditional.
- Billing dependency: the setup-token/OAuth path only draws from **Claude Max "extra usage credits"** — confirm this pool is funded before Task 2's token generation, or auth will fail even with a valid token.
- Module wiring: the upstream `hermes-agent.nixosModules.default` import lives **inside** `modules/environments/hermes-agent/default.nix` (via the `inputs` module arg), never in `machines/configuration.nix`'s per-machine module lists — wiring it there breaks jupiter's evaluation even with the profile disabled (see Task 1 Step 3's correction).
- Access control: `TELEGRAM_ALLOWED_USERS` and `EMAIL_ALLOWED_USERS` are required in the `hermes-env` secret, not optional — see Task 2 Steps 3–4.
- Secret persistence: credentials end up in plaintext at `/var/lib/hermes/.hermes/.env` (mode 0640), not sops-nix's `/run/secrets` tmpfs — rollback (Task 5 Step 3) accounts for this.
- Secret: one sops secret named `hermes-env`, a full env-file blob (matches `environmentFiles` — list of paths, not a single-value secret).
- No network exposure: do not touch `modules/services/webserver` or open any inbound port. Telegram + Email are both outbound-only.
- Platforms: Telegram + Email only, first pass. No Discord/Slack/etc.
- Rebuild command: `sudo nixos-rebuild switch --flake '.#mibook'` (use `nixos-rebuild test` first, per Task 4).
- Build-check command: `nix build '.#nixosConfigurations.mibook.config.system.build.toplevel'`.
- Format Nix with `nixfmt-rfc-style` before committing.
- Secret material never enters an agent session, chat, or the git history in plaintext — only via `sops secrets/mibook.yaml` and `$EDITOR`.
- Base branch: `feat/sops-nix` (already pushed; do not merge or rebase it as part of this work).
---
## File Structure
- **Modify** `flake.nix` — add the `hermes-agent` flake input.
- **Not modified** `machines/configuration.nix` — Task 1 originally wired the upstream module import here (mibook-only); this broke jupiter's evaluation and was corrected to import from inside the profile module instead. See Task 1 Step 3.
- **Create** `modules/environments/hermes-agent/default.nix` — the `my.profiles.hermes-agent` module (single responsibility: declare Hermes Agent).
- **Modify** `modules/environments/default.nix` — add `./hermes-agent` to the import list.
- **Modify** `machines/mibook/environments.nix` — set `hermes-agent.enable = true`.
- **Modify** `machines/mibook/secrets.nix` — replace the `placeholder` secret with `hermes-env`.
- **Modify (encrypted, operator-only)** `secrets/mibook.yaml` — real credential values, never touched by an agent.
No other files change.
---
### Task 1: Wire up the module (flake input, profile module, secret declaration)
**Files:**
- Modify: `flake.nix`
- Create: `modules/environments/hermes-agent/default.nix`
- Modify: `modules/environments/default.nix`
- Modify: `machines/mibook/environments.nix`
- Modify: `machines/mibook/secrets.nix`
- **Not modified:** `machines/configuration.nix` — see Step 3's correction.
**Interfaces:**
- Produces: NixOS option `my.profiles.hermes-agent.enable` (bool). When true, configures `services.hermes-agent` with `environmentFiles = [ config.sops.secrets."hermes-env".path ]` and grants `finn` the `hermes` group. Declares `sops.secrets.hermes-env` (owner/group `hermes`, path resolves to `/run/secrets/hermes-env` at runtime — but note the *service* re-copies it to `/var/lib/hermes/.hermes/.env`, per the spec's persistence note).
- Consumes: `hermes-agent.nixosModules.default` from the new flake input, imported directly inside the profile module via `self.inputs` (the `self` module arg, from `specialArgs` — see Step 3 and Step 5's note on why not the `inputs` module arg) — not through `machines/configuration.nix`; existing `sops.secrets.<name>` mechanism from `machines/core/sops.nix` (already applied to every machine, unaffected by this task).
- [ ] **Step 1: Add the flake input**
In `flake.nix`, inside `inputs = { ... }`, add after the `sops-nix` block:
```nix
hermes-agent = {
url = "github:NousResearch/hermes-agent";
inputs.nixpkgs.follows = "nixpkgs";
};
```
- [ ] **Step 2: Lock the new input**
Run: `nix flake lock`
Expected: `flake.lock` gains a `hermes-agent` entry (and its transitive inputs). No other entries change.
- [ ] **Step 3: Do NOT wire the module import through `machines/configuration.nix`**
**Correction (found during final whole-plan review, 2026-09-08):** an
earlier version of this step added `hermes-agent` to the
`inherit (self.inputs) ...` destructure and `hermes-agent.nixosModules.default`
to mibook's `modules` list in `machines/configuration.nix`. That broke
`nixosConfigurations.jupiter` evaluation — confirmed by direct `nix eval`.
`modules/environments/default.nix` (Step 6) is machine-agnostic, so jupiter
loads the profile module too; NixOS's module system errors on a definition
for an undeclared option **regardless of `lib.mkIf`'s condition**. Since
jupiter never imported `hermes-agent.nixosModules.default`, it never
declared `services.hermes-agent` as an option at all, so even the disabled
(`mkIf false`) config definition in Step 5 crashed evaluation.
Skip this step — do not touch `machines/configuration.nix` at all. The fix
lives in Step 5: the upstream module is imported from inside the profile
module itself, which works correctly on both machines (declares the option
everywhere, only ever turns it on where `my.profiles.hermes-agent.enable = true`).
- [ ] **Step 4: Read a reference module to match repo style**
Read `modules/environments/immich/default.nix` (same shape: `cfg`, `lib.mkIf`, single-purpose options block). Match its formatting.
- [ ] **Step 5: Create the profile module**
Create `modules/environments/hermes-agent/default.nix`:
```nix
# Hermes Agent — Telegram + Email messaging gateway (Claude subscription auth)
{
config,
lib,
pkgs,
self,
...
}:
let
cfg = config.my.profiles.hermes-agent;
in
{
imports = [ self.inputs.hermes-agent.nixosModules.default ];
options.my.profiles.hermes-agent.enable =
lib.mkEnableOption "Hermes Agent (Telegram + Email gateway)";
config = lib.mkIf cfg.enable {
services.hermes-agent = {
enable = true;
# Deliberate pin, matches the approved design — bump on purpose, not by drift.
settings.model.default = "anthropic/claude-sonnet-4-5";
environmentFiles = [ config.sops.secrets."hermes-env".path ];
addToSystemPackages = true;
};
# addToSystemPackages only puts `hermes` on $PATH; using it needs group
# access to /var/lib/hermes/.hermes (mode 2770 hermes:hermes). finn is
# this machine's one human user and already knows hermes-env's contents
# (typed them into sops directly), so this is not new exposure.
users.users.finn.extraGroups = [ "hermes" ];
};
}
```
**Use `self`, not `inputs`, in the `imports` line.** `inputs` is supplied
via `_module.args.inputs = self.inputs;`, set inside a config block — but
NixOS resolves a module's `imports` before `_module.args` assignments
merge, so anything referenced inside `imports` must come from
`specialArgs` (fixed up front) instead. `self` is threaded via
`specialArgs = { inherit self; };` in `machines/configuration.nix` and is
available at import-time; `inputs` is not. Using `inputs` here reproduces
`infinite recursion encountered` — confirmed by testing it directly, not
theoretical. `self.inputs.hermes-agent` is the same flake input either way.
- [ ] **Step 6: Register the module in the environments import list**
In `modules/environments/default.nix`, add `./hermes-agent` after `./immich`:
```nix
./immich
./hermes-agent
];
}
```
- [ ] **Step 7: Enable it on mibook**
In `machines/mibook/environments.nix`, inside the `my.profiles = { ... }` block, add:
```nix
hermes-agent.enable = true;
```
- [ ] **Step 8: Replace the placeholder secret declaration**
In `machines/mibook/secrets.nix`, replace the whole file body:
```nix
# Secrets decrypted on mibook. sopsFile defaults to secrets/mibook.yaml
# via machines/core/sops.nix.
{ ... }:
{
sops.secrets.hermes-env = {
owner = "hermes";
group = "hermes";
mode = "0400";
restartUnits = [ "hermes-agent.service" ]; # confirmed against the locked revision's generated unit
};
}
```
This removes the `placeholder` secret declared here — it was an explicit throwaway (comment: "Safe to delete once a real secret is in place"). Its entry in the encrypted `secrets/mibook.yaml` content itself is cleaned up by the operator in Task 3, not here.
- [ ] **Step 9: Format**
Run: `nixfmt-rfc-style modules/environments/hermes-agent/default.nix machines/mibook/secrets.nix machines/mibook/environments.nix flake.nix`
- [ ] **Step 10: Build-check (this is the "test", with a known-expected failure until Task 3)**
Run: `nix build '.#nixosConfigurations.mibook.config.system.build.toplevel'`
**Correction (found during Task 1 implementation, 2026-09-08):** this does **not** build cleanly yet, and that's expected — `sops.validateSopsFiles = true` (set repo-wide in `machines/core/sops.nix`) makes secret-key existence a **build-time** dependency of `config.system.build.toplevel` (via a `manifest.json.drv` step), not an activation-time check as originally assumed here. Since `secrets/mibook.yaml` still only has the `placeholder` key at this point in the plan (Task 3 adds `hermes-env`), the build correctly fails with:
```
> .../sops-install-secrets: manifest is not valid: secret hermes-env in .../mibook.yaml is not valid: the key 'hermes-env' cannot be found
```
Do **not** treat this as a wiring defect, and do not weaken `sops.validateSopsFiles` to work around it — that's a repo-wide safety invariant, not this task's to change. Instead, confirm the wiring is correct by two other means: `nix build ... --dry-run` (should complete with no eval errors and the full derivation list) and Step 11 below. A fully green (non-dry-run) `nix build` of the toplevel is only reachable after Task 3 populates the real secret — treat that as the actual end-to-end checkpoint, revisited in Task 4 Step 1.
If `--dry-run` itself fails on an unknown `services.hermes-agent.*` option, reconcile against the actual module: the error will name the bad option; cross-check against `website/docs/getting-started/nix-setup.md` in the `NousResearch/hermes-agent` repo at the locked revision.
- [ ] **Step 11: Build-check jupiter too — this is the regression gate**
Run: `nix eval '.#nixosConfigurations.jupiter.config.system.build.toplevel.drvPath'`
Expected: succeeds, printing a `.drv` path. **Any change under `modules/` is a two-machine change even when it's "for one host"** — `modules/environments/default.nix` (Step 6) is loaded by every machine, so a mistake here can break jupiter's eval even though jupiter never enables this profile. This is exactly the failure Step 3's correction fixes; this step is what catches it if it ever regresses.
- [ ] **Step 12: Confirm the option evaluates on mibook**
Run: `nix eval '.#nixosConfigurations.mibook.config.services.hermes-agent.enable'`
Expected: `true`
- [ ] **Step 13: Commit**
```bash
git add flake.nix flake.lock modules/environments/hermes-agent/default.nix modules/environments/default.nix machines/mibook/environments.nix machines/mibook/secrets.nix
git commit -m "feat(mibook): add Hermes Agent profile module"
```
**Post-review note, 2026-09-08:** Task 1 was originally committed as
`731678e` with the jupiter-breaking bug Step 3 describes and fixes — the
upstream module was wired through `machines/configuration.nix`'s
mibook-only module list instead of imported inside the profile module. The
final whole-plan review caught it (jupiter's `nixosConfigurations` stopped
evaluating, breaking the repo's sole CI job even though jupiter never
enables this profile). It was corrected in a follow-up commit on the same
branch — run `git log --oneline -- modules/environments/hermes-agent/ machines/configuration.nix`
if you need both commit hashes. The steps above already reflect the
corrected, final approach — executing this plan from scratch would never
hit the bug.
---
### Task 2: Gather credentials (operator-run, not agent-automatable)
**Files:** none. All values are held by the operator for Task 3 — do not write them into any file an agent can read, and do not paste them into a chat session.
**Interfaces:**
- Produces: an Anthropic setup-token, a Telegram bot token, and email (IMAP/SMTP) credentials — ready to type into `$EDITOR` in Task 3.
- [ ] **Step 1: Confirm the billing pool is funded**
In the Anthropic console, confirm the Claude Max plan has **extra usage credits purchased**. The `claude setup-token` auth path bills against this pool only — it does not draw from the base Max allowance, and Claude Pro cannot use this path at all. If this isn't funded, stop here and decide whether to fund it or fall back to a plain `ANTHROPIC_API_KEY` (see Global Constraints).
- [ ] **Step 2: Generate the setup-token**
On mibook, run: `claude setup-token`
This opens an interactive OAuth flow. On completion it prints a long-lived token. Copy it somewhere only you control (e.g. your password manager) — you'll type it into `sops` in Task 3, not into this session.
- [ ] **Step 3: Create the Telegram bot, and get your own user ID**
In Telegram, message `@BotFather`, run `/newbot`, follow the prompts. Copy the resulting bot token (format `123456:ABC-...`). Also get your own numeric Telegram user ID (e.g. via `@userinfobot`) — Task 3 needs it for `TELEGRAM_ALLOWED_USERS`. Telegram defaults to open DM pairing (any stranger who messages the bot gets a pairing code and can request approval); without this allowlist, that's the default posture, not a hardened one.
- [ ] **Step 4: Gather email credentials**
Decide which mailbox Hermes Agent should use, and get its IMAP/SMTP host, username, and password (or app-specific password if the provider requires one, e.g. Gmail). The exact env var names (confirmed against the locked `hermes-agent` revision, not guessed) are `EMAIL_ADDRESS`, `EMAIL_PASSWORD`, `EMAIL_IMAP_HOST`, `EMAIL_SMTP_HOST`, plus optional `EMAIL_IMAP_PORT` (default 993) / `EMAIL_SMTP_PORT` (default 587) if your provider differs — see `website/docs/user-guide/messaging/email.md` and `website/docs/reference/environment-variables.md` in the `NousResearch/hermes-agent` repo for more detail. Also decide `EMAIL_ALLOWED_USERS` (your own sending address) — without it, mail from unrecognized senders is **silently ignored** with no log line, which Task 4 Step 4 depends on knowing.
---
### Task 3: Populate the secret (operator-run, secret material)
**Files:**
- Modify (encrypted): `secrets/mibook.yaml`
**Interfaces:**
- Consumes: the three credentials gathered in Task 2, and the `hermes-env` declaration from Task 1.
- Produces: a decryptable `hermes-env` entry that Task 1's `environmentFiles` reference resolves at activation.
- [ ] **Step 1: Open the encrypted file**
Run: `nix develop -c sops secrets/mibook.yaml`
This decrypts and opens `$EDITOR` on the plaintext.
- [ ] **Step 2: Remove the placeholder key, add `hermes-env`**
Delete the `placeholder: ...` entry (Task 1 already removed its Nix declaration; this removes the matching content). Add:
```yaml
hermes-env: |
ANTHROPIC_TOKEN=<paste the claude setup-token output>
TELEGRAM_BOT_TOKEN=<paste the BotFather token>
TELEGRAM_ALLOWED_USERS=<your numeric Telegram user id, from Task 2 Step 3>
EMAIL_ADDRESS=<mailbox address>
EMAIL_PASSWORD=<password or app-password>
EMAIL_IMAP_HOST=<imap host>
EMAIL_SMTP_HOST=<smtp host>
EMAIL_ALLOWED_USERS=<your own sending address, from Task 2 Step 4>
```
(Exact YAML block-scalar syntax per whatever your `$EDITOR` shows — sops re-encrypts on save regardless of formatting, as long as it's valid YAML. Both `*_ALLOWED_USERS` keys are access control, not optional — see Task 2 Steps 3–4 for why.)
- [ ] **Step 3: Save and let sops re-encrypt**
Save and quit `$EDITOR`. Run: `git diff --stat secrets/mibook.yaml`
Expected: the file shows as changed (still fully encrypted — confirm by `git diff secrets/mibook.yaml | head` and seeing sops's `ENC[...]` ciphertext, never plaintext).
- [ ] **Step 4: Stage it**
Run: `git add secrets/mibook.yaml`
Required before the next `nix build` — untracked sops files are invisible to flake evaluation (repo rule, `docs/secrets.md`).
- [ ] **Step 5: Commit**
```bash
git commit -m "feat(mibook): populate Hermes Agent credentials"
```
---
### Task 4: Deploy and verify (operator-run, on mibook)
**Files:** none in repo. Run all commands on mibook.
**Interfaces:**
- Consumes: the built config from Task 1, the secret from Task 3.
- Produces: a running, verified `hermes-agent.service` reachable via Telegram and Email.
- [ ] **Step 1: Test-activate without persisting**
Run: `sudo nixos-rebuild test --flake '.#mibook'`
Expected: activates without reboot-persisting. If it fails at secret decryption ("hermes-env not found in sops file"), Task 3 wasn't saved/staged correctly — redo Task 3.
- [ ] **Step 2: Watch the service start**
`hermes-agent.service` is the confirmed unit name (verified against the locked revision's generated unit, not a guess). Run: `journalctl -u hermes-agent.service -f`
Expected: starts, reads the env file (lands at `/var/lib/hermes/.hermes/.env` — see spec Part 2), and does **not** crash-loop on auth. `ANTHROPIC_TOKEN` is confirmed supported by the module's credential resolver, so an auth failure here points at the token itself (expired, or the extra-credits pool from Task 2 Step 1 isn't actually funded) rather than the module rejecting the variable name.
- [ ] **Step 3: Verify Telegram connects**
Send your bot a message from Telegram. Expected: it responds (confirms both the bot token and the LLM auth work end-to-end).
- [ ] **Step 4: Verify Email connects**
Send a test email to the configured mailbox. Expected: Hermes Agent picks it up (check `journalctl -u hermes-agent.service` for an inbound-message log line) and can send a reply.
- [ ] **Step 5: Persist across reboot**
Run: `sudo nixos-rebuild switch --flake '.#mibook'`
Expected: same running state as Step 1–4, now persisted.
---
### Task 5: Sign-off (operator-run)
**Files:** none in repo beyond what Tasks 1 and 3 already committed.
**Interfaces:**
- Consumes: a verified running instance (Task 4).
- [ ] **Step 1: Open a PR**
`feat/sops-nix` is already pushed to `origin` but not merged; this work sits on top of it. Push this branch (or continue on `feat/sops-nix` directly, per your preference) and open a PR per the repo's usual Gitea workflow. Note in the PR description that it depends on `feat/sops-nix` landing first (or is stacked on it).
- [ ] **Step 2: Confidence window**
Use it for a few days — send yourself Telegram/Email messages periodically, including after mibook has slept and woken up, to see how the accepted availability gap actually behaves in practice.
- [ ] **Step 3: Rollback (if needed)**
`my.profiles.hermes-agent.enable = false` in `machines/mibook/environments.nix`, `sudo nixos-rebuild switch --flake '.#mibook'` stops the service — but the plaintext credential file survives on disk (see spec's Rollback section for why). Also run `sudo rm -rf /var/lib/hermes` and revoke the Telegram bot token via `@BotFather` → `/revoke`. No data migration is involved, so this is still cheap, just not a single command.
---
## Self-Review
**Spec coverage:**
- Host mibook, not jupiter → Global Constraints + File Structure. ✓
- NixOS native module, official upstream source → Task 1. ✓
- Home Manager rejected → not present anywhere (correctly absent, matches "deliberately deferred" reasoning — no task references it). ✓
- `claude setup-token` → `ANTHROPIC_TOKEN`, Max extra-credits caveat → Task 2 Steps 1–2, Task 4 Step 2, Global Constraints. ✓
- `hermes-env` single env-file secret → Task 1 Step 8, Task 3. ✓
- No sops prerequisite (mibook already registered) → confirmed by reusing existing `machines/mibook/secrets.nix` rather than any `.sops.yaml`/Phase-9-style bootstrapping. ✓
- No network exposure → Global Constraints; no task touches `modules/services/webserver`. ✓
- Telegram + Email only → Task 2 Steps 3–4, Task 4 Steps 3–4. ✓
- Rollback path → Task 5 Step 3. ✓
- Deferred (other platforms, availability-gap workarounds, public exposure) → correctly absent. ✓
**Placeholder scan:** No bare TBD/TODO. The bracketed values (`<paste the claude setup-token output>`, etc.) are operator-supplied secret material by design — Task 2 explicitly produces them, Task 3 explicitly consumes them; they're never meant to be concrete in this document since they're credentials, not code.
**Type consistency:** `my.profiles.hermes-agent.enable`, `services.hermes-agent.*`, and the `hermes-env` secret name are used identically across Tasks 1, 3, and 4. The `hermes-agent.service` unit name and `ANTHROPIC_TOKEN` var are confirmed (not guessed) consistently in both Task 1's declaration and Task 4's verification.
**Post-final-review addendum (2026-09-08):** the final whole-plan review found the module-wiring bug described in Task 1 Step 3, plus three access/persistence gaps now folded into Global Constraints, Task 2 Steps 3–4, Task 3's template, and Task 5 Step 3's rollback (TELEGRAM_ALLOWED_USERS/EMAIL_ALLOWED_USERS required; secrets persist to `/var/lib/hermes/.hermes/.env`, not `/run/secrets`; `hermes` CLI needs `finn` in the `hermes` group to be usable). All four are corrected in place above, not left as separate addenda, so the plan reads as a single coherent document rather than a stack of patches.