diff --git a/plans/hermes-task-calendar.md b/plans/hermes-task-calendar.md
new file mode 100644
index 0000000..33643b7
--- /dev/null
+++ b/plans/hermes-task-calendar.md
@@ -0,0 +1,245 @@
+# Hermes co-managed tasks + calendar (SilverBullet)
+
+**Repo(s):** `nixconfig`. Service + Hermes skills land on **`204-agent`**; the
+traefik route registers on **`201-mono`** (same cross-host split as
+`hermes.nix`). The notes/tasks live in a SilverBullet *space* (markdown files) on
+`204-agent`; the calendar stays on the existing self-hosted CalDAV server.
+**Status:** draft.
+
+## Goal
+
+Manage my calendar and tasks **together with** Hermes — equal access, not
+delegation. I write detailed prose state for tasks/projects, so tasks must live
+**next to my notes** (note-adjacency). I also want **one big list of every open
+task** plus a **daily list** where either of us picks what fits the day. And it
+has to survive an agent and me writing concurrently, from any device including
+the phone.
+
+**Why not Obsidian (the pivot):** Obsidian is a single-user editor over flat
+files. Every hard thing in the earlier drafts — git merge layers, LiveSync,
+sync-conflict files, append-only conventions — existed only to make *two replicas
++ dumb sync* fake a shared datastore. That's the real defect: wrong storage
+model. Switching to another flat-file PKM (Logseq, Joplin, Foam) changes nothing.
+
+**The fix: a server-authoritative notes tool — SilverBullet.** It keeps
+note-adjacent markdown, but there is **one authoritative copy** (a `spaceDir` of
+`.md` files on one server) that my phone (PWA), my desktop, and Hermes all talk
+to — not N replicas that need merging. That single fact deletes the entire
+conflict-layer question. On top of that SilverBullet gives:
+
+- **Note-adjacency** — I write full project state in a page; tasks (`- [ ]`) live
+  inline in that page with the prose around them.
+- **A query/index engine** — it indexes tasks across the whole space, so "one big
+  list of every open task" and "today" are **live queries**, not files I hand-
+  maintain. My backlog+daily model is native, and note-adjacency stops fighting
+  the one-big-list want: tasks live where they belong; the queries assemble them.
+- **Trivial Hermes access** — the space is plain `.md` files *on the same host as
+  Hermes* (`204-agent`). Hermes reads/writes them with the terminal toolset it
+  already has — no MCP tool, no API, no replica. (`services.silverbullet` is in
+  our nixpkgs: `silverbullet-2.6.1`, module `services.silverbullet` with
+  `spaceDir`/`listenPort`/`listenAddress`/`envFile`.)
+
+## Approach
+
+### Notes + tasks (SilverBullet on `204-agent`)
+
+1. **Run SilverBullet on `204-agent`**, colocated with Hermes, following the
+   `hermes.nix` cross-host structure (not the generic `homelab-server` gate,
+   because `204-agent` isn't that tag):
+   - On **`reverse-proxy`** (201-mono): register `phonkds.modules.silverbullet`
+     with `ip = "192.168.3.204"`, the port, a dashboard tile, and a traefik route
+     (`silverbullet.w.phonkd.net`, `ipfilter = true`).
+   - On **`204-agent`**: `services.silverbullet.enable = true` with
+     `listenAddress = "0.0.0.0"`, a `listenPort`, `spaceDir =
+     /var/lib/silverbullet/space`, and `envFile` (basic-auth creds from a sops
+     secret). Add an nftables rule allowing that port from `192.168.3.201` only —
+     copy the hermes-dashboard firewall pattern.
+2. **Bridge the space to Hermes.** SilverBullet runs as its own user; Hermes runs
+   as `hermes`. Give both read/write on `spaceDir` — simplest is to set
+   `services.silverbullet.group = "hermes"` (or a shared group) and make the space
+   group-writable, so Hermes edits pages directly. SilverBullet watches the space
+   and re-indexes external file changes, so Hermes writing `.md` files straight to
+   disk is a supported path.
+3. **Task model in the space:**
+   - Tasks are inline `- [ ]` items inside project/note pages, with attributes for
+     scheduling (SilverBullet task attributes, e.g. a `due`/`scheduled` date or a
+     `#today` tag).
+   - A **`Backlog`** page holds a live query: every open task across the space.
+   - A **daily page** (`Journal/YYYY-MM-DD` or a `Today` page) holds a query for
+     **open tasks scheduled on-or-before today** — so anything planned but not
+     finished on its day *carries forward* and keeps showing until it's done,
+     never dropped. **Planning the day** — me or Hermes — is just *setting the date
+     attribute* on backlog tasks; the day view assembles itself.
+   - **Planning ≠ completion.** Scheduling a task for a day only sets metadata; the
+     task stays open until *I* check it off. A day is a lens over open tasks, not a
+     bucket that consumes them. The `cc-sync` deploy-stage tasks obey the same rule
+     — they close only when the change is actually live. This is the invariant the
+     `task-notes` skill enforces: Hermes may set/bump dates when planning, but
+     **never checks a task off to "clear" a day.**
+4. **Propose, don't mutate.** Hermes drops new/uncertain tasks onto an **`Inbox`**
+   page; I triage them into their real project pages. It may set date attributes /
+   check off only tasks I explicitly asked about. Read = the whole space; write is
+   governed by the skill rules below.
+
+### Calendar (CalDAV — unchanged from the earlier draft)
+
+1. `pkgs.khal` + `pkgs.vdirsyncer` in `services.hermes-agent.extraPackages`.
+2. `vdirsyncer`/`khal` config in the hermes state dir; creds from sops secret
+   `hermes-caldav` (`owner = "hermes"`, in `environmentFiles`).
+3. A `caldav` skill: `vdirsyncer sync` → `khal list` to read, `khal new` to
+   create, edit `.ics` + sync to push. Read unrestricted; **create/modify only for
+   events I asked for; never `khal delete`, never bulk-reschedule.**
+
+### Claude Code work-state sync (plans → PRs → deploys → tasks)
+
+The missing piece: **in-flight engineering work is itself a task, at every stage
+until it's actually live** — which is exactly the `done-means-deployed` rule
+encoded as task state. A single nixconfig change moves through:
+
+| Stage | Detected from | Task reads |
+|---|---|---|
+| planned, not executed | `plans/*.md` with `Status:` draft/approved/in-progress | "execute plan `<topic>`" |
+| executed, PR open, not merged | `gh pr list --state open` (incl. `hermes-autofix`'s own draft PRs) | "review & merge PR #N: `<title>`" |
+| merged, not deployed | host's deployed rev ≠ `origin/main` | "`deploy <host>`" |
+| deployed / verified | host rev == `origin/main` | task closes |
+
+This is a **projection**, not a hand-maintained list. A separate **`cc-sync`**
+skill on its own `hermes cron` job (every ~15m) polls those three sources and
+**rewrites a dedicated `Engineering` page** in the space — Hermes is the *sole
+writer* to that page, so it never conflicts with my edits, and I act on the
+PR/deploy, not the task line (which updates next tick). Tasks there are tagged
+`#cc` so the `Backlog` query can fold them in or filter them out.
+
+Detection notes:
+
+- **Plans / PRs** need no new plumbing — `Status:` is already parseable and Hermes
+  has `gh`/`GITHUB_TOKEN`.
+- **Merged-but-not-deployed** is the one gap: **set `system.configurationRevision`
+  in the flake** (currently unset) so every host reports its running rev via
+  `nixos-version`. Then Hermes learns each host's deployed rev **without ssh** by
+  querying Mimir — expose the rev as a node metric label (textfile collector) and
+  Hermes reads it the same way the `mimir-alerting`/`hermes-autofix` skills already
+  hit `10.9.0.1`. (ssh-per-host is the fallback if the metric route is more work
+  than it's worth.)
+
+### The Hermes skills
+
+Two skills, both installed via activation script like
+`mimir-alerting`/`hermes-autofix`:
+
+- **`task-notes`** — the SilverBullet space layout, task syntax + date attributes,
+  the `Backlog`/daily/`Inbox` pages, the "planning = set a date attribute"
+  workflow, and the read-all / stage-to-`Inbox` / write-only-what-I-asked rule.
+  Editing pages = plain file ops under `spaceDir`.
+- **`cc-sync`** — the projection above: poll plans/PRs/deploys and rewrite the
+  `Engineering` page. Sole writer to that page; read-only against the repo/hosts.
+
+These are the pieces I'll iterate on most — `SKILL.md` edits, not redeploys.
+
+### Keeping it in step
+
+Two `hermes cron` jobs (created once by hand like the autofix job): one `every
+15m` runs `vdirsyncer sync` so `khal`'s view is fresh and can surface today's
+overdue items; a second drives `cc-sync` to rewrite the `Engineering` page.
+`[SILENT]` when nothing's up.
+
+The morning tick is the **calendar-aware "plan my day"** flow, and it's the payoff
+of colocating tasks + calendar under one agent: Hermes reads the day's CalDAV
+events (`khal`) *and* the open `Backlog` (including carried-over and `#cc` deploy
+tasks), then proposes a realistic day by *setting date attributes* on the tasks
+that fit the gaps around my commitments — it doesn't invent time I don't have. I
+adjust; nothing gets closed by planning. Whatever I don't finish is still open
+tomorrow.
+
+## Steps
+
+1. **`nixconfig` / new `modules/homelab/apps/silverbullet.nix`** — the cross-host
+   module: traefik registration on `reverse-proxy`, `services.silverbullet` +
+   firewall rule on `204-agent`, group bridge to `hermes`. Model on `hermes.nix`.
+2. **`nixconfig` / `204-agent.nix`** — `pkgs.khal` + `pkgs.vdirsyncer` to
+   `extraPackages`; sops secrets `hermes-caldav` and `silverbullet-auth`
+   (`owner`/env as needed); append caldav secret to `environmentFiles`. Add both
+   secrets to `secret.yaml`.
+3. **`nixconfig`** — activation scripts writing the `vdirsyncer`/`khal` config +
+   the `caldav`, `task-notes`, and `cc-sync` `SKILL.md` files into the hermes
+   state dir.
+4. **`nixconfig` (for `cc-sync`)** — set `system.configurationRevision` from the
+   flake `self.rev`, and expose it as a node metric (textfile collector) so Hermes
+   can read each host's deployed rev from Mimir. (Small, generally-useful change:
+   you can then always see what rev a host runs.)
+5. **Deploy** `deploy 204-agent` (and `deploy 201-mono` for the traefik route);
+   redeploy hosts once so they report `configurationRevision`.
+6. **Runtime (one-time)** — `vdirsyncer discover`/first sync; seed the SilverBullet
+   space with `Backlog`, `Today`, `Inbox`, `Engineering` pages + their queries;
+   create the two `hermes cron` jobs; log into the SilverBullet PWA on
+   phone/desktop.
+7. **Verify** — `khal list` shows real events; I add a task in a note on the phone
+   and it appears in the `Backlog` query; Hermes sets a `due:today` on one and it
+   shows on the daily page; Hermes and I edit the same page seconds apart and the
+   server reconciles to one page (**no `.sync-conflict` anywhere**); this very
+   plan/PR shows up on the `Engineering` page as a task, and advances stage when it
+   merges and again when `204-agent` is deployed.
+
+## Open decisions
+
+Defaults picked; say if you'd override:
+
+- **Notes tool = SilverBullet** (recommended) vs. Trilium/TriliumNext (server DB +
+  REST API, but not markdown and tasks aren't first-class) vs. staying on Obsidian
+  with LiveSync/CouchDB (keeps the editor, but Hermes access is clunkier and it's
+  a heavier service). SilverBullet uniquely gives markdown note-adjacency + a task
+  query engine + plain-file Hermes access on one authoritative store.
+- **Host placement = `204-agent`, colocated with Hermes** (recommended) — gives
+  Hermes direct filesystem access to the space. Alternative: run it on a
+  `homelab-server` host and have Hermes reach it over the HTTP API — cleaner tag
+  fit, but loses the plain-file access that makes the skill trivial.
+- **Hermes edit path = direct file ops on `spaceDir`** (recommended; SilverBullet
+  re-indexes external changes) vs. the SilverBullet **HTTP API**. Direct files is
+  simplest; switch to the API only if we ever see a race writing the same page.
+- **CalDAV access = `khal`/`vdirsyncer` CLI** (recommended) vs. raw `curl` vs. a
+  `caldav` MCP server — unchanged rationale.
+- **Deployed-rev detection = `configurationRevision` + Mimir metric**
+  (recommended) vs. Hermes ssh-ing each host. The metric route reuses the
+  Mimir-query capability Hermes already has and needs no ssh keys on `204-agent`;
+  ssh is the fallback.
+- **`Engineering` tasks: separate `#cc` page, one task per work-item advancing
+  through stages** (recommended) vs. a task per stage, and vs. folding them
+  straight into `Backlog`. One-item-with-stages keeps a change as a single line
+  from plan→PR→deploy; the `#cc` tag lets me include/exclude them from the big
+  list at will.
+- **Whole-space read exposure (accepted)** — Hermes can read every page, chosen so
+  it has project context. Write restraint rests on the skill rules + `Inbox`
+  staging.
+- **Autonomy envelope** (matches the `hermes-autofix` house rule): read
+  everything; write only what I asked; stage speculative items to `Inbox`; **no**
+  `khal delete`, **no** autonomous rescheduling, **no** deleting tasks.
+
+## Risks / rollout
+
+- **Concurrent edits** — the whole reason for SilverBullet. One authoritative
+  on-disk space + a re-indexing server means there's no two-replica merge and no
+  `.sync-conflict` files. Residual surface: Hermes and a client writing the *exact
+  same page* in the same instant is last-write on one store (not a fork); the
+  `Inbox`-staging + set-attributes-not-reorder convention keeps even that near
+  zero. Far smaller than the Obsidian model.
+- **Space is not in this repo.** The plan + nix wiring are versioned here; the
+  notes/tasks live in the SilverBullet space, the calendar in CalDAV. Both are
+  worth their own backup (a periodic `git`/`restic` snapshot of `spaceDir`) since
+  they're now a primary source of truth on `204-agent`, not just a replica.
+- **Whole-space read exposure** — see the decision above.
+- **`cc-sync` is read-derived** — it only *reads* plans/PRs/host-revs and
+  *writes* the one `Engineering` page it solely owns; it never merges, deploys, or
+  edits other pages. Worst case is a stale/wrong task line, corrected next tick —
+  no action taken on the repo or hosts. Requires `configurationRevision` to be set
+  or the deploy-stage tasks are blind.
+- **Permissions bridge** — silverbullet-user vs hermes-user access to `spaceDir`
+  is the one fiddly bit; get the shared group + group-writable dir right or Hermes
+  can't write.
+- **Rollout** is additive: `deploy 204-agent` + `deploy 201-mono`. Nothing else on
+  the reverse proxy changes beyond one route. Back out by removing the
+  silverbullet module, the two `extraPackages`, the secrets, and the activation
+  scripts, then redeploy; the space dir + cron job are runtime state removed by
+  hand.
+- **Credential blast radius** — `hermes-caldav` = full calendar write;
+  SilverBullet basic-auth guards the notes. Scope each as tightly as the servers
+  allow; the traefik `ipfilter` already gates network reach.