fix: persist user input items to state.messages across callModel invocations

User input items (role: 'user') were never written to state.messages.
The two existing saveStateSafely write sites only persist response.output
and toolResults. Within a single callModel invocation this is masked
because makeFollowupRequest reconstructs input from the in-memory
this.resolvedRequest.input. But when a new callModel resumes from
persisted state, the loaded state.messages contains zero user items —
prior user turns are silently dropped.

This causes two problems:
1. cache_control prompt caching is defeated at every user-message
   boundary (cachedTokens = 0 on first request of each new message)
2. Conversation fidelity loss — the model never sees prior user turns

Fix: In initStream(), after request resolution, persist fresh user input
items to state.messages via saveStateSafely before the API call. Both
code paths (loaded history and no history) now capture freshItemsForState
and write it to state.

Co-Authored-By: Ben Heidorn <ben.heidorn@openrouter.ai>

Author: devin-ai-integration[bot]

.changeset/fix-user-input-persistence.md

@@ -0,0 +1,14 @@
+---
+'@openrouter/agent': patch
+---
+
+fix: persist user input items to state.messages across callModel invocations
+
+User input items (role: 'user') were never written to state.messages. The two
+existing saveStateSafely write sites only persist response.output and
+toolResults. When a new callModel resumes from persisted state, the loaded
+state.messages contains zero user items — prior user turns are silently dropped.
+
+This causes two problems:
+1. cache_control prompt caching is defeated at every user-message boundary
+2. Conversation fidelity loss — the model never sees prior user turns

packages/agent/src/lib/model-result.ts

@@ -78,6 +78,7 @@ import {
   isServerTool,
   isToolCallOutputEvent,
 } from './tool-types.js';
+import { normalizeInputToArray } from './turn-context.js';
 
 /**
  * Typeguard for plain-object records (non-null, non-array).
@@ -1498,6 +1499,13 @@ export class ModelResult<
         Array.isArray(this.currentState.messages) &&
         this.currentState.messages.length > 0;
 
+      // Track fresh user input items that need to be persisted to state.
+      // Without this, state.messages only contains response outputs and tool
+      // results — prior user turns are lost when a new callModel resumes from
+      // persisted state, breaking conversation fidelity and cache_control
+      // prompt caching at user-message boundaries.
+      let freshItemsForState: models.BaseInputsUnion[] | undefined;
+
       if (hasLoadedHistory && this.currentState) {
         // `currentState.messages` is InputsUnion — keep it as that union so
         // appendToMessages (which expects InputsUnion) accepts it directly.
@@ -1527,6 +1535,8 @@ export class ModelResult<
           ? await this.applyHooksToFreshItems(freshItems, historicalMessages, initialContext)
           : undefined;
 
+        freshItemsForState = hookedFresh;
+
         baseRequest = {
           ...baseRequest,
           input: hookedFresh
@@ -1548,6 +1558,23 @@ export class ModelResult<
           ...baseRequest,
           input: hookedInput,
         };
+        // All input is fresh when there's no loaded history — normalize to
+        // array form for state persistence.
+        freshItemsForState = normalizeInputToArray(hookedInput) as models.BaseInputsUnion[];
+      }
+
+      // Persist fresh user input items to state so they survive across
+      // callModel invocations. This ensures the resume path
+      // (state.messages + new input) reconstructs the full conversation.
+      if (
+        this.stateAccessor &&
+        this.currentState &&
+        freshItemsForState &&
+        freshItemsForState.length > 0
+      ) {
+        await this.saveStateSafely({
+          messages: appendToMessages(this.currentState.messages, freshItemsForState),
+        });
       }
 
       // Store resolved request with stream mode