feat(routing): pluggable local inference (MLX, llama.cpp, LM Studio) + openhuman:// deep-link + OpenSSL 3 PKCS12 fix (#750)

* feat(routing,local-ai): pluggable local inference + openhuman:// deep-link

Salvaged from recovery commit cfb1fd9f (Apr 19) — originally auto-backed up
by claude-mem recovery tool, not yet PR'd.

## Local AI routing becomes pluggable (factory.rs, health.rs)

Previously the IntelligentRoutingProvider hard-wired Ollama as the only local
inference backend. This commit lets operators point at any OpenAI-compatible
local server (llama.cpp / llama-server, LM Studio, MLX's mlx_lm.server /
mlx-omni-server, vLLM — anything exposing /v1/chat/completions + /v1/models).

Config surface:
- `OPENHUMAN_LOCAL_INFERENCE_URL` env var — full /v1 base URL. When set,
  health is probed via GET {base}/models (OpenAI-compat) instead of Ollama's
  /api/tags.
- `LocalAiConfig.provider` accepts "llamacpp" or "llama-server" as aliases
  that default to http://127.0.0.1:8080/v1.
- Default (no env, no override) stays Ollama — zero config change for
  existing users.

Motivation: Ollama's embedded llama.cpp cannot yet load Gemma 4 E2B and
other recent models; MLX is significantly faster on Apple Silicon.
Operators should be able to pick the backend; end users see nothing.

## Deep-link scheme (Info.plist)

Adds CFBundleURLTypes entry registering `openhuman://` for the macOS bundle
so browser-issued deep links (e.g. openhuman://auth?token=...) launch the
installed app. See .claude/rules/14-deep-link-platform-guide.md for the
flow — requires a .app bundle (not `tauri dev`).

## OpenSSL 3.x PKCS12 fix (setup-dev-codesign.sh)

Adds `-legacy` to the `openssl pkcs12 -export` call. Required on OpenSSL 3.x
because the modern SHA-256 MAC / AES-256-CBC defaults are not yet supported
by macOS `security` tool, which silently fails to import the cert.

Notes:
- Dropped the CLAUDE.md.new noise and the mic-description copy downgrade
  from the original recovery commit; kept only the load-bearing changes.
- Cargo.lock files reset to upstream/main to avoid dependency drift — the
  new code does not introduce new deps.

Co-Authored-By: WOZCODE <contact@withwoz.com>

* fix(routing,codesign): address CodeRabbit review on PR #750

- setup-dev-codesign.sh: probe for openssl `-legacy` support before
  using it so older OpenSSL/LibreSSL installs don't silently fail;
  drop the stderr suppression on pkcs12 so import errors surface.
- routing/factory.rs: trim provider before alias matching, lower the
  local-inference diagnostic to `debug` and drop raw URLs from the
  log fields, and honor `OPENHUMAN_OLLAMA_BASE_URL` via the existing
  `ollama_base_url()` helper in the default Ollama branch.
- local_ai/mod.rs: re-export `ollama_base_url` for the routing factory.

* style(app/src-tauri): apply rustfmt to merged main

`cargo fmt --check` in the merge workspace flagged two hunks inherited
from upstream main:

- app/src-tauri/src/lib.rs: move `mod notification_settings;` between
  the cfg(cef) imessage_scanner and slack_scanner declarations so the
  ordering matches rustfmt's reordering output.
- app/src-tauri/src/webview_accounts/mod.rs: wrap the long
  `try_state::<NotificationSettingsState>()` binding.

No behavior change.

* format

---------

Co-authored-by: Jwalin Shah <jshah1331@gmail.com>
Co-authored-by: WOZCODE <contact@withwoz.com>
Co-authored-by: Steven Enamakel <enamakel@tinyhumans.ai>

app/src-tauri/Cargo.lock

@@ -4,7 +4,7 @@ version = 4
 
 [[package]]
 name = "OpenHuman"
-version = "0.52.27"
+version = "0.52.28"
 dependencies = [
  "anyhow",
  "async-trait",

app/src-tauri/Info.plist

@@ -6,5 +6,16 @@
 	<string>OpenHuman uses the microphone for voice dictation and for voice/video calls and huddles inside embedded apps (Google Meet, Discord, Slack).</string>
 	<key>NSCameraUsageDescription</key>
 	<string>OpenHuman uses the camera for video calls and huddles inside embedded apps (Google Meet, Discord, Slack).</string>
+	<key>CFBundleURLTypes</key>
+	<array>
+		<dict>
+			<key>CFBundleURLName</key>
+			<string>com.openhuman.app</string>
+			<key>CFBundleURLSchemes</key>
+			<array>
+				<string>openhuman</string>
+			</array>
+		</dict>
+	</array>
 </dict>
 </plist>

app/src-tauri/src/lib.rs

@@ -3,11 +3,11 @@ compile_error!("src-tauri host is desktop-only. Non-desktop targets are not supp
 
 mod core_process;
 mod core_update;
-mod notification_settings;
 #[cfg(feature = "cef")]
 mod discord_scanner;
 #[cfg(feature = "cef")]
 mod imessage_scanner;
+mod notification_settings;
 #[cfg(feature = "cef")]
 mod slack_scanner;
 #[cfg(feature = "cef")]

app/src-tauri/src/webview_accounts/mod.rs

@@ -267,7 +267,8 @@ fn forward_native_notification<R: Runtime>(
     payload: &tauri_runtime_cef::notification::NotificationPayload,
 ) {
     // Feature flag — bail early when the user hasn't opted in.
-    if let Some(settings) = app.try_state::<crate::notification_settings::NotificationSettingsState>()
+    if let Some(settings) =
+        app.try_state::<crate::notification_settings::NotificationSettingsState>()
     {
         if !settings.enabled() {
             log::debug!(

app/src/pages/Conversations.tsx

@@ -39,10 +39,7 @@ import {
   openhumanVoiceTts,
 } from '../utils/tauriCommands';
 import { formatTimelineEntry } from '../utils/toolTimelineFormatting';
-import {
-  AgentMessageBubble,
-  BubbleMarkdown,
-} from './conversations/components/AgentMessageBubble';
+import { AgentMessageBubble, BubbleMarkdown } from './conversations/components/AgentMessageBubble';
 import { LimitPill } from './conversations/components/LimitPill';
 import { ToolTimelineBlock } from './conversations/components/ToolTimelineBlock';
 import {
@@ -66,7 +63,6 @@ type ReplyMode = 'text' | 'voice';
 const AUTOCOMPLETE_POLL_DEBOUNCE_MS = 320;
 const AUTOCOMPLETE_MIN_CONTEXT_CHARS = 3;
 
-
 interface ConversationsProps {
   /**
    * `page` (default) renders the centered max-w-2xl card layout used as

scripts/setup-dev-codesign.sh

@@ -65,13 +65,22 @@ openssl req \
   2>/dev/null
 
 # ── Bundle to PKCS12 ─────────────────────────────────────────────────────────
+# `-legacy` keeps PKCS12 MAC/encryption compatible with macOS `security` tool
+# which does not yet support OpenSSL 3.x defaults (SHA256 MAC / AES-256-CBC).
+# Older OpenSSL/LibreSSL (including the macOS-bundled LibreSSL) do not know
+# about `-legacy`, so probe for support before adding it.
+PKCS12_LEGACY_ARGS=()
+if openssl pkcs12 -help 2>&1 | grep -q -- '-legacy'; then
+  PKCS12_LEGACY_ARGS=(-legacy)
+fi
+
 openssl pkcs12 \
   -export \
+  "${PKCS12_LEGACY_ARGS[@]}" \
   -out "$P12" \
   -inkey "$KEY" \
   -in "$CERT" \
-  -passout "pass:$P12_PASS" \
-  2>/dev/null
+  -passout "pass:$P12_PASS"
 
 # ── Import into login Keychain ───────────────────────────────────────────────
 security import "$P12" \

src/openhuman/local_ai/mod.rs

@@ -15,7 +15,7 @@ pub mod sentiment;
 mod install;
 pub(crate) mod model_ids;
 mod ollama_api;
-pub(crate) use ollama_api::OLLAMA_BASE_URL;
+pub(crate) use ollama_api::{ollama_base_url, OLLAMA_BASE_URL};
 mod parse;
 pub(crate) mod paths;
 mod service;

src/openhuman/routing/factory.rs

@@ -1,13 +1,18 @@
 use std::sync::Arc;
+use std::time::Duration;
 
 use crate::openhuman::config::LocalAiConfig;
-use crate::openhuman::local_ai::OLLAMA_BASE_URL;
+use crate::openhuman::local_ai::ollama_base_url;
 use crate::openhuman::providers::compatible::{AuthStyle, OpenAiCompatibleProvider};
 use crate::openhuman::providers::Provider;
 
 use super::health::LocalHealthChecker;
 use super::provider::IntelligentRoutingProvider;
 
+/// Cache TTL for the non-ollama local health probe. Mirrors the default used
+/// by [`LocalHealthChecker::new`].
+const LOCAL_HEALTH_TTL: Duration = Duration::from_secs(30);
+
 /// Construct an [`IntelligentRoutingProvider`] from a remote backend provider
 /// and the local AI configuration.
 ///
@@ -22,16 +27,51 @@ pub fn new_provider(
     local_ai_config: &LocalAiConfig,
     remote_fallback_model: &str,
 ) -> IntelligentRoutingProvider {
-    let local_base = format!("{}/v1", OLLAMA_BASE_URL);
+    // Allow operators to point the local routing tier at an OpenAI-compatible
+    // server other than Ollama (e.g. llama-server for Gemma 4 E2B, which
+    // Ollama's embedded llama.cpp cannot load yet as of April 2026).
+    //
+    // `OPENHUMAN_LOCAL_INFERENCE_URL` — full `/v1` base URL of the local
+    // OpenAI-compat server. When set, health is probed via `GET {base}/models`
+    // instead of Ollama's `/api/tags`.
+    let override_base = std::env::var("OPENHUMAN_LOCAL_INFERENCE_URL")
+        .ok()
+        .map(|s| s.trim().trim_end_matches('/').to_string())
+        .filter(|s| !s.is_empty());
+
+    let provider_kind = local_ai_config.provider.trim().to_ascii_lowercase();
+    let use_openai_compat_local =
+        override_base.is_some() || matches!(provider_kind.as_str(), "llamacpp" | "llama-server");
+
+    let (provider_label, local_base, health) = if use_openai_compat_local {
+        let base = override_base.unwrap_or_else(|| "http://127.0.0.1:8080/v1".to_string());
+        let probe = format!("{base}/models");
+        tracing::debug!(
+            provider = %provider_kind,
+            "[routing] local inference configured via OpenAI-compat (non-ollama)"
+        );
+        (
+            "llamacpp",
+            base,
+            Arc::new(LocalHealthChecker::with_probe_url(probe, LOCAL_HEALTH_TTL)),
+        )
+    } else {
+        let ollama_base = ollama_base_url();
+        let local_v1 = format!("{ollama_base}/v1");
+        (
+            "ollama",
+            local_v1,
+            Arc::new(LocalHealthChecker::new(&ollama_base)),
+        )
+    };
+
     let local: Box<dyn Provider> = Box::new(OpenAiCompatibleProvider::new(
-        "ollama",
+        provider_label,
         &local_base,
-        None, // Ollama does not require authentication
+        None, // local servers do not require authentication
         AuthStyle::Bearer,
     ));
 
-    let health = Arc::new(LocalHealthChecker::new(OLLAMA_BASE_URL));
-
     IntelligentRoutingProvider::new(
         remote,
         local,

src/openhuman/routing/health.rs

@@ -45,6 +45,12 @@ impl LocalHealthChecker {
 
     /// Create a checker with a custom cache TTL (useful in tests).
     pub fn with_ttl(base_url: &str, ttl: Duration) -> Self {
+        Self::with_probe_url(format!("{base_url}/api/tags"), ttl)
+    }
+
+    /// Create a checker with an explicit full probe URL (for non-ollama local
+    /// backends such as llama-server, whose health endpoint is `/v1/models`).
+    pub fn with_probe_url(probe_url: String, ttl: Duration) -> Self {
         let client = reqwest::Client::builder()
             .timeout(PROBE_TIMEOUT)
             .build()
@@ -57,7 +63,7 @@ impl LocalHealthChecker {
             });
         Self {
             client,
-            probe_url: format!("{base_url}/api/tags"),
+            probe_url,
             cache: Mutex::new(None),
             ttl,
         }