feat(agent/multimodal): support [FILE:…] markers with text extraction (#2777)

Cargo.toml

@@ -186,7 +186,7 @@ curve25519-dalek = { version = "4", default-features = false, features = ["alloc
 matrix-sdk = { version = "0.16", optional = true, default-features = false, features = ["e2e-encryption", "rustls-tls", "markdown"] }
 fantoccini = { version = "0.22.0", optional = true, default-features = false, features = ["rustls-tls"] }
 serde-big-array = { version = "0.5", optional = true }
-pdf-extract = { version = "0.10", optional = true }
+pdf-extract = "0.10"
 # WhatsApp Web — upstream `whatsapp-rust` 0.5. Replaces the previous `wa-rs`
 # 0.2 fork: upstream now ships its own SqliteStore (so we no longer need the
 # 1.3K-line custom RusqliteStore) and dispatches `Event::Message` for
@@ -254,7 +254,6 @@ peripheral-rpi = ["dep:rppal"]
 browser-native = ["dep:fantoccini"]
 fantoccini = ["browser-native"]
 landlock = ["sandbox-landlock"]
-rag-pdf = ["dep:pdf-extract"]
 whatsapp-web = ["dep:whatsapp-rust", "dep:whatsapp-rust-tokio-transport", "dep:whatsapp-rust-ureq-http-client", "dep:wacore", "serde-big-array"]
 # Exposes the destructive `openhuman.test_reset` RPC. Off by default; the E2E
 # build (app/scripts/e2e-build.sh) flips it on. Shipped binaries never have

app/test/playwright/helpers/core-rpc.ts

@@ -215,6 +215,36 @@ export async function dismissWalkthroughIfPresent(page: Page): Promise<void> {
   }
 
   await markCompleted();
+  // Last-resort: a lingering #react-joyride-portal node will keep
+  // intercepting clicks on the page even after we've persisted the
+  // completion flag, AND React may re-mount one later (e.g. after a
+  // hash-route navigation that runs the walkthrough effect again).
+  // Strip every portal node now AND install a MutationObserver that
+  // keeps stripping any future mount for the rest of the page
+  // lifetime. The observer install is idempotent — re-runs of this
+  // helper on the same page no-op.
+  await page.evaluate(() => {
+    document.querySelectorAll('#react-joyride-portal').forEach(node => node.remove());
+    const win = window as unknown as { __openhumanJoyrideScrubInstalled?: boolean };
+    if (win.__openhumanJoyrideScrubInstalled) return;
+    win.__openhumanJoyrideScrubInstalled = true;
+    const scrub = (root: ParentNode) => {
+      root.querySelectorAll('#react-joyride-portal').forEach(node => node.remove());
+    };
+    const obs = new MutationObserver(mutations => {
+      for (const m of mutations) {
+        m.addedNodes.forEach(node => {
+          if (!(node instanceof Element)) return;
+          if (node.id === 'react-joyride-portal') {
+            node.remove();
+          } else {
+            scrub(node);
+          }
+        });
+      }
+    });
+    obs.observe(document.body, { childList: true, subtree: true });
+  });
 }
 
 async function waitForAuthenticatedSnapshot(page: Page): Promise<void> {

src/openhuman/agent/bus.rs

@@ -82,6 +82,10 @@ pub struct AgentTurnRequest {
     /// size caps).
     pub multimodal: MultimodalConfig,
 
+    /// File-attachment feature configuration (file marker count caps,
+    /// per-file size budget, extracted-text limits, MIME allowlist).
+    pub multimodal_files: crate::openhuman::config::MultimodalFileConfig,
+
     /// Maximum number of LLM↔tool round-trips before bailing out.
     /// Prevents infinite loops if a model gets "stuck" calling the same tool.
     pub max_tool_iterations: usize,
@@ -152,6 +156,7 @@ pub fn register_agent_handlers() {
                 silent,
                 channel_name,
                 multimodal,
+                multimodal_files,
                 max_tool_iterations,
                 on_delta,
                 target_agent_id,
@@ -245,6 +250,7 @@ pub fn register_agent_handlers() {
                     silent,
                     &channel_name,
                     &multimodal,
+                    &multimodal_files,
                     max_tool_iterations,
                     on_delta,
                     visible_tool_names.as_ref(),
@@ -402,6 +408,7 @@ mod tests {
             silent: true,
             channel_name: "test-channel".into(),
             multimodal: MultimodalConfig::default(),
+            multimodal_files: crate::openhuman::config::MultimodalFileConfig::default(),
             max_tool_iterations: 1,
             on_delta: None,
             target_agent_id: None,