feat(models): reach the fleet from off-network through the OpenWebUI door

Every provider in agent/models.yml is a LAN address, so a workstation that
leaves the house has no model at all. Add `openwebui (franksdank.cloud)`:
the already-public OpenWebUI vhost as an OpenAI-compatible provider that
proxies the same engines, with frank-dragoon and frank-zealot rows and the
WAN row as its own compaction target. LAN rows, modelRoles and the speech
runner stay untouched — the door is purely additive.

The key is command-backed (`apiKey: "!cat ~/.config/secrets/openwebui_api_key"`,
mode 600) instead of inline: a literal sk-... in a tracked file is a key in
every agent context on every box that pulls this repo, and the pre-commit
shape gate would block it anyway. Rotate = rewrite the file + restart omp,
because a `!cmd` value is cached for the process lifetime.

Compat blocks copy their LAN twins because the proxy was probed, not assumed
(2026-10-21): streaming SSE carries reasoning_content and tool_calls deltas
plus final usage, image_url content parts reach both vision towers, and
chat_template_kwargs is forwarded — proven on frank-zealot, where
enable_thinking false yields zero reasoning and true a real block.
frank-archon gets no WAN row: OpenWebUI answers 400 "Model not found"
because its picker never published it, so that path is a fleet change.

Verified end-to-end from a scratch HOME whose LAN baseUrls point at a dead
port, so any answer can only have come through the door (text and a real
tool round-trip); `omp models` still lists all four chat providers and the
tts runner, so the file validates. bats 137/137, kb-lint clean.
Docs: omp-docs/core/wan-model-access.md.

agent/models.yml

@@ -306,3 +306,135 @@ providers:
         name: "frank-tts (dankclaw) — cloned voices"
         api: openai-speech
         input: [text]
+
+  # ─── WAN door ───────────────────────────────────────────────────────────
+  # The only block in this file that answers OFF the home network. Every
+  # provider above is a LAN IP, and `modelRoles` (config.yml) still points at
+  # those, so away from home you must NAME one of the rows below:
+  #   omp --model 'openwebui (franksdank.cloud)/frank-dragoon:medium'
+  # The door is the OpenWebUI vhost `ai.franksdank.cloud` — Caddy → OWU on
+  # Dankclaw:3005, Anubis-exempt because OWU enforces its own auth, public A
+  # record present (LAN clients resolve 192.168.0.143 via Pi-hole, measured
+  # 2026-10-21). It exposes an OpenAI-compatible surface at /api/v1 and
+  # proxies to the SAME engines as the LAN rows — no second inference stack,
+  # so every engine-side law below still holds through it.
+  # CREDENTIAL LAW: the key is not in this file. `apiKey` accepts a
+  # command-backed value (`!cmd`, resolved by src/config/resolve-config-value.ts:
+  # stdout trimmed, cached for the process lifetime; anything else is tried as
+  # an env-var name, then taken literally), so the secret lives in
+  # ~/.config/secrets/openwebui_api_key, mode 600, like every other credential
+  # on these boxes. A literal `sk-…` here would trip the pre-commit shape gate
+  # and then live in every agent context on every box that pulls this repo.
+  # Rotate: rewrite the secret file, then RESTART omp — the `!cmd` result is
+  # cached per process, so a live session keeps serving the old key.
+  # WHAT THE DOOR SHOWS (probed 2026-10-21 with this key): /api/v1/models
+  # answers frank-dragoon and frank-zealot only. frank-archon is NOT published
+  # in OpenWebUI's picker and the API answers 400 "Model not found" for it, so
+  # the fleet's biggest endpoint has no WAN path until it is published
+  # server-side on Dankclaw — a fleet change, not a harness change.
+  # CONTRACT PROBED THROUGH THE DOOR (2026-10-21): streaming SSE carries
+  # `reasoning_content` deltas and `tool_calls` deltas, finish_reason
+  # `tool_calls`, and a final `usage` under stream_options.include_usage;
+  # `image_url` content parts reach both vision towers (a 64 px red PNG comes
+  # back "Red" from each); and `chat_template_kwargs` IS forwarded — proven on
+  # frank-zealot, where enable_thinking false gives reasoning 0 chars and true
+  # gives a real thinking block on the same prompt. Every compat block below is
+  # therefore a copy of its LAN twin, never a simplification: dropping the
+  # kwargs here would quietly change thinking behaviour on the WAN path.
+  # Chat completions made with an API key create NO OpenWebUI chat row (the
+  # chat list is unchanged after probing), so an off-network session leaves no
+  # trail in the family's chat history.
+  # NOT BEHIND THIS DOOR: speech. OWU does route /api/v1/audio/speech and it
+  # does reach frank-tts:7772, but OWU substitutes its own voice enum first —
+  # a cloned pack name comes back 400 from the upstream. `modelRoles.speech`
+  # stays on `frank-tts (dankclaw)`; do not repoint it here.
+  "openwebui (franksdank.cloud)":
+    baseUrl: https://ai.franksdank.cloud/api/v1
+    api: openai-completions
+    # Quoted because an unquoted leading `!` is a YAML tag, not a command.
+    apiKey: "!cat ~/.config/secrets/openwebui_api_key"
+    models:
+      - id: frank-dragoon
+        # Same halogen flash-server as `llamacpp (dankclaw)` above, one hop
+        # further out; the picker row and the engine share this id, so there
+        # is no alias to keep in sync.
+        name: "frank-dragoon (WAN)"
+        # Off-network the compaction hub must be reachable too: the LAN
+        # frank-zealot row is a dead address from anywhere but the house.
+        # Same engine, same slot arithmetic, so the keepRecent/threshold
+        # math in the LAN rows transfers untouched.
+        compactionModel: openwebui (franksdank.cloud)/frank-zealot
+        # Inherited from the engine, not invented: 200,000 under the native
+        # 262,144 ceiling (the 400-zone law — never sit at the ceiling). The
+        # proxy neither widens the window nor truncates: an over-ceiling
+        # request is still a 400.
+        contextWindow: 200000
+        # Same clamp arithmetic as the LAN row (see frank_dragoon_unit.bats);
+        # HALOGEN_MAX_TOKENS_CAP on the engine is unchanged.
+        maxTokens: 45056
+        reasoning: true
+        input: [text, image]
+        compat:
+          # Verified through the door: reasoning arrives as
+          # `reasoning_content` in the deltas, not folded into content.
+          reasoningContentField: "reasoning_content"
+          # `:off` must be a real switch, and it is: this engine honours
+          # reasoning_effort "none"; the harness default would only send the
+          # lowest effort, which still thinks.
+          reasoningDisableMode: none-effort
+          # Identity map, same reason as the LAN row — flash-server lists
+          # minimal|low|medium|high|xhigh|max|none and 400s anything else.
+          reasoningEffortMap:
+            minimal: minimal
+            low: low
+            medium: medium
+            high: high
+            xhigh: xhigh
+            max: max
+          # Pinned so a server-side HALOGEN_TOP_K edit cannot change this
+          # session's recipe; the engine accepts all three keys.
+          extraBody:
+            temperature: 1.0
+            top_p: 0.95
+            top_k: 20
+      - id: frank-zealot
+        # The 3090 endpoint, same llama.cpp engine behind the proxy. Stays
+        # the widest slot in the file, so it is the WAN compaction target.
+        name: "frank-zealot (WAN)"
+        compactionModel: openwebui (franksdank.cloud)/frank-zealot
+        # PER-SLOT window: llama.cpp divides --ctx-size across slots and
+        # answers a longer prompt with a 400, never a truncation. Follows
+        # every reslice on the engine (3 × 200,704 today).
+        contextWindow: 200704
+        # Prompt + thinking budget inside the slot; compaction fires at 85 %
+        # of the window on this row set (see the LAN twin for the full
+        # worst-case arithmetic).
+        maxTokens: 16384
+        reasoning: true
+        input: [text, image]
+        compat:
+          reasoningContentField: "reasoning_content"
+          # LOAD-BEARING, exactly as on the LAN row: the Bonsai-2 template
+          # accepts ONLY xhigh|medium|low and raises on anything else, so an
+          # unmapped harness level is a hard failure on the role's first
+          # request. Probed 2026-10-21 that the mapped value, not the
+          # harness-level, is what reaches the template through this proxy.
+          reasoningEffortMap:
+            minimal: low
+            low: low
+            medium: medium
+            high: xhigh
+            max: xhigh
+          # The payload's own sampling recipe plus the thinking switches.
+          # chat_template_kwargs survives the proxy (verified), so
+          # preserve_thinking keeps prior assistants' thinking blocks in
+          # history for multi-turn tool traffic — same contract as at home.
+          extraBody:
+            temperature: 1.0
+            top_p: 0.95
+            top_k: 20
+            min_p: 0.0
+            presence_penalty: 0.0
+            chat_template_kwargs:
+              enable_thinking: true
+              preserve_thinking: true