@netnodeag/kraftwerk 0.11.0 → 0.12.0

package/dist/inspector/chat/sessions.js

@@ -25,6 +25,7 @@ async function loadState(id) {
         busy: false,
         pendingPermissions: new Map(),
         writeChain: Promise.resolve(),
+        needsContext: true,
     };
     // Two racing loads: keep whichever registered first.
     return states.get(id) ?? (states.set(id, state), state);
@@ -211,15 +212,25 @@ async function baseScopeContext(scope, agent) {
     }
     return "";
 }
+/**
+ * Every chat renders in the inspector UI, so agents should know two things:
+ * replies are markdown, and run artifacts are addressable over the
+ * inspector's own file endpoint (relative URLs keep working wherever the
+ * inspector is reachable — localhost, LAN, tunnel).
+ */
+const RENDERING_BLOCK = `## Chat rendering\n` +
+    `Your replies render as markdown in the inspector chat: headings, lists, tables, code blocks, ` +
+    `links and images all work. Files inside a workflow run's output folder are served by the ` +
+    `inspector itself at /api/runs/<run-id>/file?name=<relative-path>&raw=1 — embed an image inline ` +
+    `with ![alt](/api/runs/<run-id>/file?name=picture.png&raw=1), or link any artifact the same way. ` +
+    `Prefer these relative /api/... URLs over file paths when showing results: they render directly ` +
+    `in this chat and keep working wherever the inspector is reachable.`;
 async function scopeContext(scope, agent) {
     const [base, skills] = await Promise.all([
         baseScopeContext(scope, agent),
         availableSkills(scope),
     ]);
-    const block = skillsBlock(skills);
-    if (!block)
-        return base;
-    return base ? `${base}\n\n${block}` : block;
+    return [base, RENDERING_BLOCK, skillsBlock(skills)].filter(Boolean).join("\n\n");
 }
 /* ---------- backend lifecycle ---------- */
 /** Team chats carry the member's model/effort; resolved live so edits apply to new backends. */
@@ -280,6 +291,9 @@ async function ensureBackend(state) {
         agent === "pi"
             ? startPiBackend(cwd, hooks, tuning)
             : await startAcpBackend(agent, cwd, hooks, tuning);
+    // Fresh process = fresh context window: re-send scope context with the
+    // next prompt (matters for chats resumed after an inspector restart).
+    state.needsContext = true;
     return state.backend;
 }
 /* ---------- public API (used by server.ts) ---------- */
@@ -304,6 +318,7 @@ export async function createChat(opts) {
         busy: false,
         pendingPermissions: new Map(),
         writeChain: Promise.resolve(),
+        needsContext: true,
     });
     return meta;
 }
@@ -338,7 +353,10 @@ export async function postMessage(id, text) {
     void (async () => {
         try {
             const backend = await ensureBackend(state);
-            const context = isFirst ? await scopeContext(state.meta.scope, state.meta.agent) : "";
+            const context = state.needsContext
+                ? await scopeContext(state.meta.scope, state.meta.agent)
+                : "";
+            state.needsContext = false;
             // "/<skill> args" expands to the skill's instructions for the agent;
             // the thread keeps the short form the user typed.
             const body = await expandSkillInvocation(state.meta.scope, text);