@hypabolic/crossbar 0.1.0 → 0.2.0

package/CAPABILITY-MATRIX.md

@@ -7,7 +7,7 @@ adapter registers under (`oai` = `openai-completions`, `ant` = `anthropic-messag
 | Backend | port | pi api | listModels | introspectLoaded | switchModel | loadUnload | auth | health | perModelCaps | streaming | discovery fingerprint |
 |---|---|---|---|---|---|---|---|---|---|---|---|
 | **Ollama** | 11434 | oai | ✅ `/api/tags`,`/v1/models` | ✅ `/api/ps` | ✅ implicit (request id) | ✅ `keep_alive:0` | ◐ none local | ✅ `GET /` text | ✅ `/api/show` caps + ctx | ✅ | `GET /` → `Ollama is running` |
-| **LM Studio** | 1234 | oai | ✅ `/api/v0/models`,`/v1/models` | ✅ `state` field | ✅ JIT + `/api/v1/models/load` | ✅ load/unload + `lms` | ◐ Bearer, none default | ◐ infer 200 | ✅ type+`max_context_length` | ✅ | `/api/v0/models` w/ `state`,`compatibility_type` |
+| **LM Studio** | 1234 | oai | ✅ `/api/v1/models` (v0 fallback) | ✅ `state` field | ✅ JIT + `/api/v1/models/load` | ✅ load/unload + `lms` | ◐ Bearer, none default | ◐ infer 200 | ✅ type+`max_context_length` | ✅ | `/api/v1/models` (v0 fallback) w/ `state`,`compatibility_type` |
 | **llama-server** | 8080 | oai | ✅ `/v1/models` | ◐ `/props`,`/slots` (single) | ❌ (1/instance) | ❌ classic | ◐ none / `--api-key` | ✅ `/health` | ◐ ctx via `/props`,`meta` | ✅ | `/props` w/ `default_generation_settings`+`build_info` |
 | **llama-swap** | 8080 | oai/ant | ✅ `/v1/models` (all config) | ✅ `/running` | ✅ via `model` → restart upstream | ✅ `/api/models/unload`, ttl | ◐ optional multi-scheme | ✅ `/health`→OK | ◐ via upstream | ✅ | `/` → `/ui/`; `/running`,`/upstream/{model}` |
 | **vLLM** | 8000 | oai | ✅ `/v1/models` | ◐ `/is_sleeping` (dev) | ❌ base · ◐ LoRA | ◐ sleep/wake + LoRA | ◐ none / `--api-key` | ✅ `/health` | ◐ `max_model_len` only | ✅ | `/version` + `/metrics` `vllm:` + `owned_by:"vllm"` |
@@ -38,7 +38,7 @@ adapter registers under (`oai` = `openai-completions`, `ant` = `anthropic-messag
 
 1. `GET /` → `Ollama is running` ⇒ Ollama · redirect `/ui/` ⇒ llama-swap
 2. `GET /api/extra/version` → `{"result":"KoboldCpp"}` ⇒ KoboldCpp
-3. `GET /api/v0/models` 200 w/ `state`/`compatibility_type` ⇒ LM Studio
+3. `GET /api/v1/models` (v0 fallback) 200 w/ `state`/`compatibility_type` ⇒ LM Studio
 4. `GET /props` w/ `default_generation_settings`+`build_info` ⇒ llama-server / llamafile
 5. `GET /version` + `/metrics` `vllm:` ⇒ vLLM
 6. `GET /v1/models` shape: `owned_by:"vllm"`⇒vLLM · `meta.n_ctx_train`⇒llama.cpp ·

package/README.md

@@ -3,7 +3,7 @@
 [![CI](https://github.com/Hypabolic/Crossbar/actions/workflows/ci.yml/badge.svg)](https://github.com/Hypabolic/Crossbar/actions/workflows/ci.yml)
 [![npm](https://img.shields.io/npm/v/@hypabolic/crossbar)](https://www.npmjs.com/package/@hypabolic/crossbar)
 
-**The local/self-hosted inference connector Pi should have shipped with.**
+**Effortless local & self-hosted model backends for the Pi coding agent.**
 
 Crossbar is an extension for the [Pi coding agent](https://github.com/earendil-works/pi) that makes
 wiring Pi to *any* local or self-hosted model backend effortless — zero hand-edited JSON, all setup
@@ -12,6 +12,8 @@ in-place model switching.
 
 > Built by [Hypabolic](https://github.com/hypabolic).
 
+![Crossbar onboarding: discover a server, open the manage menu, switch the active model](docs/onboarding.gif)
+
 ---
 
 ## Why Crossbar
@@ -104,23 +106,10 @@ The `BackendAdapter` contract (`src/core/`) is the frozen boundary every adapter
 conformance suite (`tests/conformance/`) validates every adapter against it, and
 `tests/integration/` exercises the real discovery path over live sockets.
 
-### CI / releasing
-
-- **CI** (`.github/workflows/ci.yml`) runs `tsc --noEmit` + the full test suite on every push and PR
-  (Node 22 & 24).
-- **Releases** (`.github/workflows/release.yml`) publish to npm via **GitHub→npm OIDC trusted
-  publishing** — no tokens or secrets. [Provenance](https://docs.npmjs.com/generating-provenance-statements)
-  is attached automatically. Two ways:
-  1. **Manual** — GitHub → *Actions → Release → Run workflow* → choose `patch` / `minor` / `major`.
-     It bumps `package.json`, commits, tags `vX.Y.Z`, and publishes.
-  2. **Tag push** — `npm version patch && git push --follow-tags` locally.
-
-  **One-time setup:** on npmjs.com, add a **Trusted Publisher** for `@hypabolic/crossbar`
-  (*Package settings → Trusted Publisher → GitHub Actions*) pointing at repo **`Hypabolic/Crossbar`**
-  and workflow **`release.yml`**. The workflow authenticates through the OIDC `id-token` it already
-  requests — no `NPM_TOKEN` needed.
+### CI
 
-<!-- TODO: add an onboarding demo GIF (docs/onboarding.gif) recorded against a live Ollama + LM Studio. -->
+CI (`.github/workflows/ci.yml`) runs `tsc --noEmit` + the full test suite on every push and PR
+(Node 22 & 24).
 
 ## License
 

package/package.json

@@ -1,11 +1,18 @@
 {
   "name": "@hypabolic/crossbar",
-  "version": "0.1.0",
-  "description": "The local/self-hosted inference connector Pi should have shipped with — multi-backend discovery, model switching, and zero-JSON in-TUI onboarding for the Pi coding agent.",
+  "version": "0.2.0",
+  "description": "The local/self-hosted inference connector for Pi — multi-backend discovery, model switching, and zero-JSON in-TUI onboarding for the Pi coding agent.",
   "type": "module",
   "license": "MIT",
   "author": "Hypabolic",
-  "homepage": "https://github.com/hypabolic/crossbar",
+  "homepage": "https://github.com/Hypabolic/Crossbar#readme",
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/Hypabolic/Crossbar.git"
+  },
+  "bugs": {
+    "url": "https://github.com/Hypabolic/Crossbar/issues"
+  },
   "keywords": [
     "pi-package",
     "pi-extension",
@@ -22,10 +29,12 @@
   "pi": {
     "extensions": [
       "./src/index.ts"
-    ]
+    ],
+    "image": "https://raw.githubusercontent.com/Hypabolic/Crossbar/main/docs/onboarding.gif"
   },
   "files": [
     "src",
+    "docs/onboarding.gif",
     "RESEARCH.md",
     "CAPABILITY-MATRIX.md",
     "ARCHITECTURE.md",
@@ -35,7 +44,9 @@
   "scripts": {
     "check": "tsc --noEmit",
     "test": "vitest run",
-    "test:watch": "vitest"
+    "test:watch": "vitest",
+    "demo:lmstudio": "node scripts/fake-lmstudio.mjs",
+    "demo:gif": "node scripts/gen-onboarding-gif.mjs"
   },
   "peerDependencies": {
     "@earendil-works/pi-coding-agent": "0.79.9",

package/src/adapters/generic.ts

@@ -140,9 +140,15 @@ class GenericAdapter implements BackendAdapter {
       name: model.name,
       reasoning: model.reasoning ?? false,
       input: model.input.length > 0 ? model.input : ["text"],
+      // Local inference is free → per-token costs are zero, but cache-hit token
+      // COUNTS still matter: Pi maps any `usage.prompt_tokens_details.cached_tokens` the
+      // backend reports to `Usage.cacheRead` and displays it regardless of cost. The
+      // flag only asks for usage in streaming (never fabricates), so it is safe even for
+      // unknown OpenAI-compatible servers that may not report cache hits.
       cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
       contextWindow: model.contextWindow ?? DEFAULT_CONTEXT_WINDOW,
       maxTokens: model.maxTokens ?? DEFAULT_MAX_TOKENS,
+      compat: { supportsUsageInStreaming: true },
     };
   }
 

package/src/adapters/llamacpp.ts

@@ -197,9 +197,14 @@ class LlamacppAdapter implements BackendAdapter {
       name: model.name,
       reasoning: model.reasoning ?? false,
       input: model.input.length > 0 ? model.input : ["text"],
+      // Local inference is free → per-token costs are zero, but cache-hit token
+      // COUNTS still matter: Pi maps the backend's `usage.prompt_tokens_details
+      // .cached_tokens` to `Usage.cacheRead` and displays it regardless of cost. Keep
+      // streaming usage reporting on so those prompt-cache hits are recorded.
       cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
       contextWindow: model.contextWindow ?? 8192,
       maxTokens: model.maxTokens ?? 4096,
+      compat: { supportsUsageInStreaming: true },
     };
   }
 

package/src/adapters/llamaswap.ts

@@ -256,9 +256,14 @@ class LlamaswapAdapter implements BackendAdapter {
       name: model.name,
       reasoning: model.reasoning ?? false,
       input: model.input.length > 0 ? model.input : ["text"],
+      // Local inference is free → per-token costs are zero, but cache-hit token
+      // COUNTS still matter: Pi maps the backend's `usage.prompt_tokens_details
+      // .cached_tokens` to `Usage.cacheRead` and displays it regardless of cost. Keep
+      // streaming usage reporting on so those prompt-cache hits are recorded.
       cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
       contextWindow: model.contextWindow ?? 8192,
       maxTokens: model.maxTokens ?? 4096,
+      compat: { supportsUsageInStreaming: true },
     };
   }
 

package/src/adapters/lmstudio.ts

@@ -2,16 +2,20 @@
  * LM Studio backend adapter.
  *
  * Implements the BackendAdapter contract for LM Studio's local server.
- * Uses the LM Studio-native /api/v0/* endpoints for discovery and management,
- * and delegates inference to the OpenAI-compatible /v1/* layer.
+ * Uses the LM Studio-native REST API for discovery and management, and delegates
+ * inference to the OpenAI-compatible /v1/* layer.
+ *
+ * LM Studio 0.4.0+ ships a native `/api/v1/*` REST API (recommended); the older
+ * `/api/v0/*` API carries the same rich model fields and is kept as a fallback for
+ * pre-0.4.0 servers. We prefer v1 and fall back to v0 only on a 404.
  *
  * Key API endpoints:
- *   GET  /api/v0/models            — model list with state, type, context lengths
- *   POST /api/v1/models/load       — load a model by id
- *   POST /api/v1/models/unload     — unload a model by id
+ *   GET  /api/v1/models  (→ /api/v0/models fallback)  — model list with state, type, context length
+ *   POST /api/v1/models/load                          — load a model by id
+ *   POST /api/v1/models/unload                        — unload a model by id
  *
  * Fingerprint discriminator: data[] entries have both `state` and
- * `compatibility_type` fields (unique to LM Studio's v0 API).
+ * `compatibility_type` fields (unique to LM Studio's native API).
  */
 
 import { Capability } from "../core/capability.ts";
@@ -24,9 +28,14 @@ import type {
   ModelDescriptor,
   PiModelEntry,
   Probe,
+  ProbeResult,
   ServerCredential,
 } from "../core/types.ts";
 
+/** Native model-list endpoints, in preference order (v1 first, v0 fallback for <0.4.0). */
+const MODELS_V1 = "/api/v1/models";
+const MODELS_V0 = "/api/v0/models";
+
 // ---------------------------------------------------------------------------
 // LM Studio API shapes (narrowed from unknown JSON)
 // ---------------------------------------------------------------------------
@@ -103,6 +112,7 @@ function hasLmsDiscriminator(json: unknown): boolean {
 function toDescriptor(m: LmsModelEntry): ModelDescriptor {
   const isEmbeddings = m.type === "embeddings";
   const isVlm = m.type === "vlm";
+  const isLoaded = m.state === "loaded";
 
   const input: ("text" | "image")[] = ["text"];
   if (isVlm) input.push("image");
@@ -112,11 +122,24 @@ function toDescriptor(m: LmsModelEntry): ModelDescriptor {
     name: m.id,
     input,
     embeddings: isEmbeddings,
-    loaded: m.state === "loaded",
+    loaded: isLoaded,
     raw: m,
   };
-  if (m.max_context_length !== undefined) {
-    desc.contextWindow = m.max_context_length;
+
+  // Context window: LM Studio reports both the model ceiling (`max_context_length`)
+  // and the window the model was actually loaded with (`loaded_context_length`),
+  // which is frequently configured well below the ceiling (e.g. a 128k model loaded
+  // at 4096). Register the OPERATIVE window so Pi budgets against what the server
+  // will really accept: prefer the loaded length when the model is resident (and
+  // non-zero), otherwise fall back to the model max. `loaded_context_length` is 0 or
+  // absent while the model is not loaded, so it never masks the ceiling in that case.
+  const loadedCtx =
+    isLoaded && typeof m.loaded_context_length === "number" && m.loaded_context_length > 0
+      ? m.loaded_context_length
+      : undefined;
+  const ctx = loadedCtx ?? m.max_context_length;
+  if (ctx !== undefined) {
+    desc.contextWindow = ctx;
   }
   return desc;
 }
@@ -140,10 +163,21 @@ class LmStudioAdapter implements BackendAdapter {
     Capability.Streaming,
   ]);
 
+  /**
+   * Fetch the native model list, preferring /api/v1/models and falling back to
+   * /api/v0/models for LM Studio < 0.4.0 (which only exposes the v0 REST API).
+   * Falls back ONLY on a 404 so auth (401) and unreachable (0) errors propagate.
+   */
+  private async modelsResponse(probe: Probe): Promise<ProbeResult> {
+    const v1 = await probe(MODELS_V1);
+    if (v1.status === 404) return probe(MODELS_V0);
+    return v1;
+  }
+
   // --- fingerprint ----------------------------------------------------------
 
   async fingerprint(baseUrl: string, probe: Probe): Promise<DiscoveredServer | null> {
-    const r = await probe("/api/v0/models");
+    const r = await this.modelsResponse(probe);
     if (!r.ok || r.status === 0) return null;
     if (!hasLmsDiscriminator(r.json)) return null;
 
@@ -163,7 +197,7 @@ class LmStudioAdapter implements BackendAdapter {
     _cred: ServerCredential,
     probe: Probe,
   ): Promise<HealthStatus> {
-    const r = await probe("/api/v0/models");
+    const r = await this.modelsResponse(probe);
     if (r.status === 0) return { state: "unreachable" };
     if (r.status === 401) return { state: "unauthorized" };
     if (!r.ok) return { state: "degraded" };
@@ -179,7 +213,7 @@ class LmStudioAdapter implements BackendAdapter {
     _cred: ServerCredential,
     probe: Probe,
   ): Promise<ModelDescriptor[]> {
-    const r = await probe("/api/v0/models");
+    const r = await this.modelsResponse(probe);
     if (!r.ok) {
       if (r.status === 401) throw new Error("401 Unauthorized");
       if (r.status === 0) throw new Error("listModels failed: server unreachable");
@@ -197,7 +231,7 @@ class LmStudioAdapter implements BackendAdapter {
     _cred: ServerCredential,
     probe: Probe,
   ): Promise<LoadedState> {
-    const r = await probe("/api/v0/models");
+    const r = await this.modelsResponse(probe);
     if (!r.ok) {
       if (r.status === 401) throw new Error("401 Unauthorized");
       if (r.status === 0) throw new Error("introspectLoaded failed: server unreachable");
@@ -242,7 +276,7 @@ class LmStudioAdapter implements BackendAdapter {
     }
 
     // Step 2: Confirm via model list that the target is now loaded
-    const r2 = await probe("/api/v0/models");
+    const r2 = await this.modelsResponse(probe);
     if (!r2.ok) {
       if (r2.status === 0) throw new Error("switchModel confirmation failed: server went down");
       if (r2.status === 401) throw new Error("401 Unauthorized");
@@ -287,9 +321,18 @@ class LmStudioAdapter implements BackendAdapter {
       name: model.name,
       reasoning: model.reasoning ?? false,
       input: model.input.length > 0 ? (model.input as ("text" | "image")[]) : ["text"],
+      // Local inference is free, so per-token COSTS are zero. The cache-hit token
+      // COUNTS still flow and are worth recording: LM Studio's OpenAI-compatible
+      // responses report `usage.prompt_tokens_details.cached_tokens`, which Pi maps to
+      // `Usage.cacheRead` and surfaces in the TUI regardless of cost. Keep usage
+      // reporting on during streaming so those automatic-prefix-cache hits are
+      // recorded. We intentionally do NOT set `cacheControlFormat`: LM Studio (llama.cpp
+      // engine) caches matching prefixes automatically, so injecting Anthropic-style
+      // `cache_control` markers would be wrong for this OpenAI-completions backend.
       cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
       contextWindow: model.contextWindow ?? 8192,
       maxTokens: model.maxTokens ?? 4096,
+      compat: { supportsUsageInStreaming: true },
     };
   }
 

package/src/adapters/ollama.ts

@@ -320,9 +320,14 @@ class OllamaAdapter implements BackendAdapter {
       name: model.name,
       reasoning: model.reasoning ?? false,
       input: model.input.length > 0 ? model.input : ["text"],
+      // Local inference is free → per-token costs are zero, but cache-hit token
+      // COUNTS still matter: Pi maps the backend's `usage.prompt_tokens_details
+      // .cached_tokens` to `Usage.cacheRead` and displays it regardless of cost. Keep
+      // streaming usage reporting on so those prompt-cache hits are recorded.
       cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
       contextWindow: model.contextWindow ?? DEFAULT_CONTEXT_WINDOW,
       maxTokens: model.maxTokens ?? DEFAULT_MAX_TOKENS,
+      compat: { supportsUsageInStreaming: true },
     };
   }
 

package/src/adapters/vllm.ts

@@ -177,9 +177,15 @@ class VllmAdapter implements BackendAdapter {
       name: model.name,
       reasoning: model.reasoning ?? false,
       input: model.input.length > 0 ? model.input : ["text"],
+      // Local inference is free → per-token costs are zero, but cache-hit token
+      // COUNTS still matter: Pi maps the backend's `usage.prompt_tokens_details
+      // .cached_tokens` to `Usage.cacheRead` and displays it regardless of cost. vLLM
+      // reports cached tokens from its automatic prefix cache; keep streaming usage
+      // reporting on so those hits are recorded.
       cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
       contextWindow: model.contextWindow ?? DEFAULT_CONTEXT_WINDOW,
       maxTokens: model.maxTokens ?? DEFAULT_MAX_TOKENS,
+      compat: { supportsUsageInStreaming: true },
     };
   }
 

package/src/ui/onboarding.ts

@@ -24,45 +24,56 @@ import { Container, type SelectItem, SelectList, Text, matchesKey } from "@earen
 
 import type { BackendAdapter } from "../core/backend-adapter.ts";
 import { canIntrospect, canLoadUnload, canSwitch } from "../core/backend-adapter.ts";
-import type { DiscoveredServer, ModelDescriptor, ServerRecord } from "../core/types.ts";
+import type { DiscoveredServer, LoadedState, ModelDescriptor, ServerRecord } from "../core/types.ts";
 import type { ServerRegistry } from "../registry/registry.ts";
 import { serverId } from "../registry/ids.ts";
 import { adapterFor } from "../adapters/index.ts";
+import { unregisterServer } from "../shim/provider-shim.ts";
 import { createProbe } from "../discovery/probe.ts";
 
 // ─── Pure helpers ────────────────────────────────────────────────────────────
 
+/** Extract a `host:port` string from a base URL for compact labels. */
+function hostPortOf(baseUrl: string): string {
+  try {
+    const u = new URL(baseUrl);
+    return `${u.hostname}:${u.port || (u.protocol === "https:" ? "443" : "80")}`;
+  } catch {
+    return baseUrl.replace(/^https?:\/\//, "");
+  }
+}
+
+/** Capitalise a backend kind for display, e.g. "lmstudio" → "Lmstudio". */
+function kindLabelOf(kind: string): string {
+  return kind.charAt(0).toUpperCase() + kind.slice(1);
+}
+
 /**
- * Build a `SelectItem[]` representing the discovered servers for the top-level
- * onboarding list.  Already-registered servers are marked with a "(added)" suffix
- * so the user can see what is new vs. what Crossbar already knows about.
+ * Build a `SelectItem[]` representing the servers shown in the top-level onboarding
+ * list.  Three kinds of entry can appear:
+ *   - discovered servers (in discovery order) — already-registered ones get an
+ *     "(added)" suffix so the user can tell new from known;
+ *   - registered servers that are NOT currently discovered (e.g. offline), so they
+ *     can still be managed/removed;
+ *   - a sentinel "Add manually" entry, always last.
  *
- * Items are ordered: discovered servers first (in discovery order), then a
- * sentinel "Add manually" entry at the end.
+ * Selecting any already-registered entry opens the manage overlay; selecting a new
+ * discovered entry or the sentinel runs the add flow.
  */
 export function buildDiscoveredItems(
   discovered: DiscoveredServer[],
   existing: ServerRecord[],
 ): SelectItem[] {
   const existingIds = new Set(existing.map((r) => r.id));
+  const discoveredUrls = new Set(discovered.map((s) => s.baseUrl));
 
   const items: SelectItem[] = discovered.map((server): SelectItem => {
     const id = serverId(server.kind, server.baseUrl);
     const isAdded = existingIds.has(id);
 
-    // Extract host:port from baseUrl for the label suffix
-    let hostPort: string;
-    try {
-      const u = new URL(server.baseUrl);
-      hostPort = `${u.hostname}:${u.port || (u.protocol === "https:" ? "443" : "80")}`;
-    } catch {
-      hostPort = server.baseUrl.replace(/^https?:\/\//, "");
-    }
-
     // Compose a label: "[kind] host:port  ✓ healthy" or "(added)"
-    const kindLabel = server.kind.charAt(0).toUpperCase() + server.kind.slice(1);
     const healthMark = isAdded ? "(added)" : "✓ healthy";
-    const label = `${kindLabel} (${hostPort})`;
+    const label = `${kindLabelOf(server.kind)} (${hostPortOf(server.baseUrl)})`;
 
     return {
       value: server.baseUrl,
@@ -73,6 +84,18 @@ export function buildDiscoveredItems(
     };
   });
 
+  // Append registered servers that weren't discovered this scan (offline / not
+  // reachable right now) so they remain manageable from the same list.
+  for (const record of existing) {
+    if (!record.enabled) continue;
+    if (discoveredUrls.has(record.baseUrl)) continue;
+    items.push({
+      value: record.baseUrl,
+      label: `${kindLabelOf(record.kind)} (${hostPortOf(record.baseUrl)})  (added)`,
+      description: "Registered · not currently discovered",
+    });
+  }
+
   // Always append the manual-add sentinel
   items.push({
     value: "__manual__",
@@ -151,6 +174,36 @@ export function capabilityActions(
   return actions;
 }
 
+/** One-line hints shown under each manage action. */
+const ACTION_DESCRIPTIONS: Record<string, string> = {
+  switch: "Make a model the active/served one",
+  load: "Load a model into memory",
+  unload: "Evict a loaded model from memory",
+  introspect: "Show which models are currently loaded",
+  remove: "Forget this server and delete its stored key",
+};
+
+/**
+ * Build the manage-overlay action list for an already-registered server: the
+ * adapter's capability-filtered actions (switch / load / unload / introspect) plus
+ * a "Remove server" action that is always available. Backends without any local
+ * capabilities (vLLM, OpenAI, Anthropic, generic) show only "Remove server".
+ */
+export function buildManageItems(adapter: BackendAdapter): SelectItem[] {
+  const items: SelectItem[] = capabilityActions(adapter).map((a) => {
+    const item: SelectItem = { value: a.value, label: a.label };
+    const desc = ACTION_DESCRIPTIONS[a.value];
+    if (desc !== undefined) item.description = desc;
+    return item;
+  });
+  items.push({
+    value: "remove",
+    label: "Remove server",
+    description: ACTION_DESCRIPTIONS["remove"]!,
+  });
+  return items;
+}
+
 /**
  * Coerce a user-supplied string (which may be bare "host:port", missing a scheme,
  * or already a valid URL) into a well-formed origin with no trailing slash.
@@ -177,6 +230,263 @@ export function normalizeManualUrl(input: string): string {
   return u.origin.replace(/\/+$/, "");
 }
 
+// ─── Shared overlay + server-action helpers ─────────────────────────────────
+
+/** Reconstruct a minimal DiscoveredServer from a persisted record for adapter calls. */
+function serverFromRecord(record: ServerRecord): DiscoveredServer {
+  return {
+    kind: record.kind,
+    baseUrl: record.baseUrl,
+    auth: record.auth,
+    label: record.label,
+    confidence: 1,
+  };
+}
+
+/**
+ * Render a single-select overlay (titled SelectList in an accent border) and resolve
+ * to the chosen item value, or `null` on Esc/cancel. Shared by the model picker and
+ * the manage menus so they stay visually consistent.
+ */
+function selectOverlay(
+  ctx: ExtensionCommandContext,
+  title: string,
+  items: SelectItem[],
+  hint: string,
+): Promise<string | null> {
+  return ctx.ui.custom<string | null>(
+    (_tui, theme, _kb, done) => {
+      const container = new Container();
+      container.addChild(new DynamicBorder((s) => theme.fg("accent", s)));
+      container.addChild(new Text(theme.fg("accent", theme.bold(title))));
+
+      const list = new SelectList(items, Math.min(items.length, 12), getSelectListTheme());
+      list.onSelect = (item) => done(item.value);
+      list.onCancel = () => done(null);
+
+      container.addChild(list);
+      container.addChild(new Text(theme.fg("dim", hint)));
+      container.addChild(new DynamicBorder((s) => theme.fg("accent", s)));
+
+      return {
+        render: (width: number) => container.render(width),
+        invalidate: () => container.invalidate(),
+        handleInput: (data: string) => {
+          if (matchesKey(data, "escape")) {
+            done(null);
+            return;
+          }
+          list.handleInput(data);
+          _tui.requestRender();
+        },
+      };
+    },
+    { overlay: true, overlayOptions: { width: "60%" } },
+  );
+}
+
+const errMsg = (err: unknown): string => (err instanceof Error ? err.message : String(err));
+
+/** Fetch a server's models (live, falling back to last-known on failure). */
+async function fetchModels(
+  ctx: ExtensionCommandContext,
+  registry: ServerRegistry,
+  record: ServerRecord,
+): Promise<ModelDescriptor[] | null> {
+  const adapter = adapterFor(record.kind);
+  const cred = await registry.resolveCredential(record);
+  const probe = createProbe(record.baseUrl, { auth: cred, defaultTimeoutMs: 5000 });
+  try {
+    return await adapter.listModels(serverFromRecord(record), cred, probe);
+  } catch (err) {
+    if (record.lastKnownModels && record.lastKnownModels.length > 0) {
+      return record.lastKnownModels;
+    }
+    ctx.ui.notify(`Crossbar: could not list models — ${errMsg(err)}`, "error");
+    return null;
+  }
+}
+
+/** Switch the active model or load a model: pick from the list, then call the adapter. */
+async function performModelAction(
+  ctx: ExtensionCommandContext,
+  registry: ServerRegistry,
+  record: ServerRecord,
+  action: "switch" | "load",
+): Promise<void> {
+  const adapter = adapterFor(record.kind);
+  const models = await fetchModels(ctx, registry, record);
+  if (!models) return;
+  if (models.length === 0) {
+    ctx.ui.notify("Crossbar: server returned no models.", "warning");
+    return;
+  }
+
+  const title = action === "switch"
+    ? `Switch model — ${record.label}`
+    : `Load model — ${record.label}`;
+  const modelId = await selectOverlay(
+    ctx,
+    title,
+    buildModelItems(models.filter((m) => !m.embeddings)),
+    "↑↓ navigate · Enter select · Esc cancel",
+  );
+  if (!modelId) return;
+
+  const cred = await registry.resolveCredential(record);
+  // Loads can be slow (cold model into VRAM) — give them a generous budget.
+  const probe = createProbe(record.baseUrl, { auth: cred, defaultTimeoutMs: 60_000 });
+
+  ctx.ui.notify(
+    `Crossbar: ${action === "switch" ? "switching to" : "loading"} ${modelId}…`,
+    "info",
+  );
+  try {
+    if (action === "switch") {
+      if (!canSwitch(adapter)) return;
+      await adapter.switchModel(serverFromRecord(record), cred, modelId, probe);
+    } else {
+      if (!canLoadUnload(adapter)) return;
+      await adapter.loadUnload(serverFromRecord(record), cred, modelId, "load", probe);
+    }
+    ctx.ui.notify(
+      `Crossbar: ${modelId} ${action === "switch" ? "is now active" : "loaded"}.`,
+      "info",
+    );
+  } catch (err) {
+    ctx.ui.notify(`Crossbar: ${action} failed — ${errMsg(err)}`, "error");
+  }
+}
+
+/** Unload a currently-loaded model: resolve the loaded set, pick one, evict it. */
+async function performUnload(
+  ctx: ExtensionCommandContext,
+  registry: ServerRegistry,
+  record: ServerRecord,
+): Promise<void> {
+  const adapter = adapterFor(record.kind);
+  if (!canLoadUnload(adapter)) return;
+  const cred = await registry.resolveCredential(record);
+  const probe = createProbe(record.baseUrl, { auth: cred, defaultTimeoutMs: 5000 });
+
+  let loadedIds: string[] = record.lastKnownLoaded ?? [];
+  if (canIntrospect(adapter)) {
+    try {
+      const state = await adapter.introspectLoaded(serverFromRecord(record), cred, probe);
+      loadedIds = state.loadedModelIds;
+    } catch {
+      // Fall back to last-known on a failed introspection.
+    }
+  }
+  if (loadedIds.length === 0) {
+    ctx.ui.notify("Crossbar: no models are currently loaded.", "info");
+    return;
+  }
+
+  const modelId = await selectOverlay(
+    ctx,
+    `Unload model — ${record.label}`,
+    loadedIds.map((id) => ({ value: id, label: id })),
+    "↑↓ navigate · Enter select · Esc cancel",
+  );
+  if (!modelId) return;
+
+  ctx.ui.notify(`Crossbar: unloading ${modelId}…`, "info");
+  try {
+    await adapter.loadUnload(serverFromRecord(record), cred, modelId, "unload", probe);
+    ctx.ui.notify(`Crossbar: ${modelId} unloaded.`, "info");
+  } catch (err) {
+    ctx.ui.notify(`Crossbar: unload failed — ${errMsg(err)}`, "error");
+  }
+}
+
+/** Read and report the currently-loaded models for a server. */
+async function performIntrospect(
+  ctx: ExtensionCommandContext,
+  registry: ServerRegistry,
+  record: ServerRecord,
+): Promise<void> {
+  const adapter = adapterFor(record.kind);
+  if (!canIntrospect(adapter)) return;
+  const cred = await registry.resolveCredential(record);
+  const probe = createProbe(record.baseUrl, { auth: cred, defaultTimeoutMs: 5000 });
+
+  let state: LoadedState;
+  try {
+    state = await adapter.introspectLoaded(serverFromRecord(record), cred, probe);
+  } catch (err) {
+    ctx.ui.notify(`Crossbar: could not read loaded models — ${errMsg(err)}`, "error");
+    return;
+  }
+  if (state.loadedModelIds.length === 0) {
+    ctx.ui.notify(`Crossbar: ${record.label} has no models loaded.`, "info");
+    return;
+  }
+  const summary = state.loadedModelIds
+    .map((id) => {
+      const ctxLen = state.perModel?.[id]?.contextLength;
+      if (ctxLen === undefined) return id;
+      const ctxStr = ctxLen >= 1000 ? `${Math.round(ctxLen / 1000)}k` : `${ctxLen}`;
+      return `${id} (${ctxStr} ctx)`;
+    })
+    .join(", ");
+  ctx.ui.notify(`Crossbar: ${record.label} loaded — ${summary}`, "info");
+}
+
+/** Confirm and remove a server from the registry, auth.json, and Pi. */
+async function performRemove(
+  pi: ExtensionAPI,
+  ctx: ExtensionCommandContext,
+  registry: ServerRegistry,
+  record: ServerRecord,
+): Promise<void> {
+  const confirm = await ctx.ui.select(`Remove ${record.label}?`, ["Cancel", "Remove server"]);
+  if (confirm !== "Remove server") return;
+  unregisterServer(pi, record);
+  await registry.remove(record.id);
+  ctx.ui.notify(`Crossbar: removed ${record.label}.`, "info");
+}
+
+/**
+ * Open the manage overlay for an already-registered server: show the
+ * capability-filtered action menu and dispatch the chosen action.
+ */
+export async function openServerActions(
+  pi: ExtensionAPI,
+  ctx: ExtensionCommandContext,
+  deps: OnboardingDeps,
+  record: ServerRecord,
+): Promise<void> {
+  const { registry } = deps;
+  const adapter = adapterFor(record.kind);
+
+  const choice = await selectOverlay(
+    ctx,
+    `Manage — ${record.label}`,
+    buildManageItems(adapter),
+    "↑↓ navigate · Enter select · Esc close",
+  );
+  if (!choice) return;
+
+  switch (choice) {
+    case "switch":
+      await performModelAction(ctx, registry, record, "switch");
+      break;
+    case "load":
+      await performModelAction(ctx, registry, record, "load");
+      break;
+    case "unload":
+      await performUnload(ctx, registry, record);
+      break;
+    case "introspect":
+      await performIntrospect(ctx, registry, record);
+      break;
+    case "remove":
+      await performRemove(pi, ctx, registry, record);
+      break;
+  }
+}
+
 // ─── Overlay flow driver ────────────────────────────────────────────────────
 
 export interface OnboardingDeps {
@@ -203,7 +513,7 @@ export interface OnboardingDeps {
  * @param deps - injected registry + discover function (for testability)
  */
 export async function openOnboarding(
-  _pi: ExtensionAPI,
+  pi: ExtensionAPI,
   ctx: ExtensionCommandContext,
   deps: OnboardingDeps,
 ): Promise<void> {
@@ -299,7 +609,15 @@ export async function openOnboarding(
 
     targetBaseUrl = normalizedUrl;
   } else {
-    // Discovered server path
+    // Already-registered server (discovered or offline) → open the manage overlay
+    // instead of re-running the add flow.
+    const existingRecord = registry.list().find((r) => r.baseUrl === chosenBaseUrl);
+    if (existingRecord) {
+      await openServerActions(pi, ctx, deps, existingRecord);
+      return;
+    }
+
+    // New discovered server path
     discoveredServer = discovered.find((s) => s.baseUrl === chosenBaseUrl);
     targetBaseUrl = chosenBaseUrl;
   }
@@ -368,47 +686,11 @@ export async function openOnboarding(
   }
 
   // ── Step 5: pick default model ─────────────────────────────────────────────
-  const modelItems = buildModelItems(models);
-
-  const chosenModelId = await ctx.ui.custom<string | null>(
-    (_tui, theme, _kb, done) => {
-      const container = new Container();
-
-      container.addChild(new DynamicBorder((s) => theme.fg("accent", s)));
-      container.addChild(
-        new Text(theme.fg("accent", theme.bold(`Pick default model — ${discoveredServer!.label}`))),
-      );
-
-      const list = new SelectList(
-        modelItems,
-        Math.min(modelItems.length, 12),
-        getSelectListTheme(),
-      );
-
-      list.onSelect = (item) => done(item.value);
-      list.onCancel = () => done(null);
-
-      container.addChild(list);
-      container.addChild(
-        new Text(theme.fg("dim", "↑↓ navigate · Enter select · Esc skip")),
-      );
-      container.addChild(new DynamicBorder((s) => theme.fg("accent", s)));
-
-      return {
-        render: (width: number) => container.render(width),
-        invalidate: () => container.invalidate(),
-        handleInput: (data: string) => {
-          // Allow Esc to skip model selection
-          if (matchesKey(data, "escape")) {
-            done(null);
-            return;
-          }
-          list.handleInput(data);
-          _tui.requestRender();
-        },
-      };
-    },
-    { overlay: true, overlayOptions: { width: "60%" } },
+  const chosenModelId = await selectOverlay(
+    ctx,
+    `Pick default model — ${discoveredServer.label}`,
+    buildModelItems(models),
+    "↑↓ navigate · Enter select · Esc skip",
   );
 
   // chosenModelId === null means the user skipped — still register the server