@obtoai/agent-bridge 0.1.0-beta.21 → 0.1.0-beta.22

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@obtoai/agent-bridge",
-  "version": "0.1.0-beta.21",
+  "version": "0.1.0-beta.22",
   "description": "Local consumer for the OBTO Agent Bridge. Receives bridge events over SSE and drives a coding agent (Claude Code or OpenAI Codex) on your machine.",
   "license": "Apache-2.0",
   "author": "OBTO Inc.",

package/src/bridge-http.js

@@ -92,6 +92,28 @@ const claimThread = (threadId, agentId) =>
 const postExternalSync = (agentId, sessions) =>
   postJson('/api/bridge/external/sync', { agentId, sessions });
 
+// Phase 6.4 — download an attachment's raw bytes for use as a Claude SDK
+// image content block. The serve route streams the file with its stored
+// Content-Type; we read it into a Buffer and base64-encode for the SDK.
+// Returns { ok, status, mimeType, base64 } or { ok: false, status }.
+const getAttachmentBytes = async (attachmentId) => {
+  const c = getCfg();
+  const url = c.baseUrl.replace(/\/$/, '') +
+    '/api/bridge/attachment/' + encodeURIComponent(String(attachmentId));
+  const res = await fetch(url, {
+    method: 'GET',
+    headers: {
+      'OBTO-ORIGIN-HOST': c.originHost,
+      Authorization: 'Bearer ' + c.apiToken,
+    },
+    cache: 'no-store',
+  });
+  if (!res.ok) return { ok: false, status: res.status };
+  const mimeType = res.headers.get('content-type') || 'application/octet-stream';
+  const buf = Buffer.from(await res.arrayBuffer());
+  return { ok: true, status: res.status, mimeType, base64: buf.toString('base64') };
+};
+
 module.exports = {
   getCfg,
   buildHeaders,
@@ -100,4 +122,5 @@ module.exports = {
   postAgentActivity,
   claimThread,
   postExternalSync,
+  getAttachmentBytes,
 };

package/src/claude-driver.js

@@ -228,6 +228,52 @@ const buildEnvelope = (payload) => {
   return head + '\n\n' + body;
 };
 
+// Phase 6.4 — image attachments. When payload.attachmentIds is non-empty,
+// download each via the bridge HTTP API and assemble a multimodal user
+// message (image blocks + text envelope) as an async iterable, which the
+// Claude Agent SDK accepts in lieu of a plain prompt string. With no
+// attachments, returns the envelope text as-is — zero overhead on the
+// hot text-only path.
+const buildPromptForSdk = async (payload, envelopeText, log) => {
+  const ids = Array.isArray(payload && payload.attachmentIds)
+    ? payload.attachmentIds.filter(Boolean)
+    : [];
+  if (ids.length === 0) return envelopeText;
+
+  const blocks = [];
+  for (const id of ids) {
+    try {
+      const r = await bridgeHttp.getAttachmentBytes(id);
+      if (r && r.ok) {
+        blocks.push({
+          type: 'image',
+          source: {
+            type: 'base64',
+            media_type: r.mimeType || 'image/png',
+            data: r.base64,
+          },
+        });
+      } else {
+        if (log) log('warn', 'attachment fetch failed', { id, status: r && r.status });
+      }
+    } catch (e) {
+      if (log) log('warn', 'attachment fetch threw', {
+        id,
+        error: e && e.message ? e.message : String(e),
+      });
+    }
+  }
+  // No images survived the fetch — fall back to text-only so the turn still
+  // runs (with degraded context). The agent has the envelope; the user will
+  // see their own bubble with images in the bridge UI.
+  if (blocks.length === 0) return envelopeText;
+
+  blocks.push({ type: 'text', text: envelopeText });
+  return (async function* () {
+    yield { type: 'user', message: { role: 'user', content: blocks } };
+  })();
+};
+
 const buildBootstrapPrompt = (payload) =>
   buildEnvelope(payload) +
   '\n\n---\n' +
@@ -290,7 +336,7 @@ const consumeQuery = async (q) => {
 const driveFirstTouch = async ({ threadId, projectDir, payload, log }) => {
   const sdk = await import('@anthropic-ai/claude-agent-sdk');
   const bridgeServer = await buildBridgeMcpServer({ log });
-  const prompt = buildBootstrapPrompt(payload);
+  const prompt = await buildPromptForSdk(payload, buildBootstrapPrompt(payload), log);
   const options = Object.assign(
     {
       cwd: projectDir,
@@ -303,6 +349,7 @@ const driveFirstTouch = async ({ threadId, projectDir, payload, log }) => {
     threadId,
     projectDir,
     messageId: payload.messageId,
+    attachments: (payload.attachmentIds || []).length,
   });
 
   const startedAt = Date.now();
@@ -357,7 +404,7 @@ const driveResume = async ({ threadId, sessionId, projectDir, jsonlPath, lastJso
 
   const sdk = await import('@anthropic-ai/claude-agent-sdk');
   const bridgeServer = await buildBridgeMcpServer({ log });
-  const prompt = buildEnvelope(payload);
+  const prompt = await buildPromptForSdk(payload, buildEnvelope(payload), log);
   const options = Object.assign(
     {
       resume: sessionId,
@@ -371,6 +418,7 @@ const driveResume = async ({ threadId, sessionId, projectDir, jsonlPath, lastJso
     threadId,
     sessionId,
     messageId: payload.messageId,
+    attachments: (payload.attachmentIds || []).length,
   });
 
   const startedAt = Date.now();

package/src/codex-driver.js

@@ -32,8 +32,24 @@ const queues = new Map();
 
 const ALLOW_ALL = process.env.BRIDGE_ALLOW_ALL === '1';
 
+// Phase 6.4 — Codex SDK doesn't accept image inputs yet. When the bridge
+// payload carries attachmentIds, we prepend an honest note so the agent
+// knows images existed (the human will see them in their own bubble on the
+// bridge UI). When the SDK gains multimodal support, this can be replaced
+// with a real image-in path.
+const attachmentDropNote = (payload) => {
+  const n = Array.isArray(payload && payload.attachmentIds)
+    ? payload.attachmentIds.filter(Boolean).length
+    : 0;
+  if (!n) return '';
+  return '[OBTO bridge note: ' + n + ' image attachment' + (n === 1 ? '' : 's') +
+    ' came with this message, but the Codex driver does not support image ' +
+    'input yet — proceeding with text only. Ask the human to describe the ' +
+    'image in words if you need its content.]\n\n';
+};
+
 const buildCodexPrompt = (payload, isFirst) => {
-  const head = buildEnvelope(payload);
+  const head = attachmentDropNote(payload) + buildEnvelope(payload);
   if (!isFirst) return head;
   return head +
     '\n\n---\n' +

package/src/opencode-driver.js

@@ -29,8 +29,24 @@ const queues = new Map();
 const DEFAULT_PROVIDER = process.env.BRIDGE_OPENCODE_PROVIDER || 'anthropic';
 const DEFAULT_MODEL = process.env.BRIDGE_OPENCODE_MODEL || 'claude-sonnet-4-5';
 
+// Phase 6.4 — opencode's SDK accepts only `parts:[{type:'text',text}]`. When
+// the bridge payload carries attachmentIds, we prepend an honest note so the
+// agent knows images existed (the human will see them in their own bubble on
+// the bridge UI). Upgrade to real image parts when opencode-ai/sdk grows
+// support for file/image parts.
+const attachmentDropNote = (payload) => {
+  const n = Array.isArray(payload && payload.attachmentIds)
+    ? payload.attachmentIds.filter(Boolean).length
+    : 0;
+  if (!n) return '';
+  return '[OBTO bridge note: ' + n + ' image attachment' + (n === 1 ? '' : 's') +
+    ' came with this message, but the opencode driver does not support image ' +
+    'input yet — proceeding with text only. Ask the human to describe the ' +
+    'image in words if you need its content.]\n\n';
+};
+
 const buildOpencodePrompt = (payload, isFirst) => {
-  const head = buildEnvelope(payload);
+  const head = attachmentDropNote(payload) + buildEnvelope(payload);
   if (!isFirst) return head;
   return head +
     '\n\n---\n' +