fix: append diagram-state context after tool_result blocks, not before

src/renderer/src/ai/runner.ts

@@ -52,14 +52,18 @@ export interface RunAIOptions {
   relevantTypeIds?: Set<string>
 }
 
-/** Prepends a fresh text block to a turn's content ahead of sending it —
- *  used to carry the live diagram-state snapshot on the current round's new
+/** Appends a fresh text block to a turn's content ahead of sending it — used
+ *  to carry the live diagram-state snapshot on the current round's new
  *  content (the initial prompt on round 1, the tool-results message on
  *  round 2+) instead of a separate system message that would change every
- *  round and break the cacheable prefix in front of it (see below). */
-function withLeadingText(content: string | ChatContentBlock[], text: string): ChatContentBlock[] {
+ *  round and break the cacheable prefix in front of it (see below). Must be
+ *  APPENDED, not prepended: Anthropic requires `tool_result` blocks to be
+ *  the first content in the user turn immediately following the `tool_use`
+ *  they answer — a leading block ahead of them gets rejected with "tool_use
+ *  ids were found without tool_result blocks immediately after". */
+function withTrailingText(content: string | ChatContentBlock[], text: string): ChatContentBlock[] {
   const blocks: ChatContentBlock[] = typeof content === 'string' ? [{ type: 'text', text: content }] : content
-  return [{ type: 'text', text }, ...blocks]
+  return [...blocks, { type: 'text', text }]
 }
 
 /** One human-readable line for a tool call — resolves tempIds/labels through
@@ -153,7 +157,7 @@ export async function runAIPrompt(opts: RunAIOptions): Promise<RunAIResult> {
     const systemMessages = buildSystemMessages(metamodel, diagram.getNodes(), relevantTypeIds)
 
     // Re-read live state every round so the model sees whatever the previous
-    // round's tool calls just changed — attached as a leading text block on
+    // round's tool calls just changed — attached as a trailing text block on
     // THIS round's new turn (rather than a separate, ever-changing system
     // message) so everything before it stays a stable, appendable prefix.
     // This bakes a (soon stale) snapshot permanently into `turns[last]`,
@@ -167,7 +171,7 @@ export async function runAIPrompt(opts: RunAIOptions): Promise<RunAIResult> {
       diagram.getViews?.(),
     )
     const current = turns[turns.length - 1]
-    current.content = withLeadingText(current.content, contextText)
+    current.content = withTrailingText(current.content, contextText)
 
     if (round > 1) {
       // Clear the previous marker (else it'd accumulate past the 4-block

tests/aiRunner.test.ts

@@ -202,17 +202,20 @@ describe('runAIPrompt — end-to-end with mocked Anthropic tool-calling', () =>
       settings: anthropicSettings(),
       diagram: makeFacade(),
     })
-    // The initial user turn now carries a leading diagram-state snapshot
-    // block ahead of the original prompt text (see ai/runner.ts's
-    // `withLeadingText`) instead of the live state going out as a separate,
-    // ever-changing system message — that's what makes the system prefix in
-    // front of it byte-stable and cacheable across rounds/stages.
+    // The initial user turn now carries a trailing diagram-state snapshot
+    // block after the original prompt text (see ai/runner.ts's
+    // `withTrailingText`) instead of the live state going out as a
+    // separate, ever-changing system message — that's what makes the
+    // system prefix in front of it byte-stable and cacheable across
+    // rounds/stages. It must be TRAILING, not leading: Anthropic requires
+    // `tool_result` blocks to be the first content in a user turn that
+    // follows a `tool_use` — a block ahead of them is rejected.
     expect(result.history[0].role).toBe('user')
     const firstContent = result.history[0].content
     expect(Array.isArray(firstContent)).toBe(true)
     expect(firstContent).toEqual([
-      { type: 'text', text: expect.stringMatching(/Current diagram state/) },
       { type: 'text', text: 'add an API system' },
+      { type: 'text', text: expect.stringMatching(/Current diagram state/) },
     ])
     expect(result.history.some((m) => m.role === 'assistant' && Array.isArray(m.content))).toBe(true)
     expect(result.history.at(-1)).toEqual({ role: 'assistant', content: [{ type: 'text', text: 'Done.' }] })
@@ -240,10 +243,13 @@ describe('runAIPrompt — end-to-end with mocked Anthropic tool-calling', () =>
     expect(cacheControlBlocks(reqs[1])).toBe(2)
     expect(cacheControlBlocks(reqs[2])).toBe(2)
 
-    // Every round's outgoing final message carries a fresh leading
-    // diagram-state snapshot ahead of the tool_result blocks.
+    // Every round's outgoing final message carries a fresh diagram-state
+    // snapshot APPENDED after its tool_result blocks — never before them,
+    // since Anthropic requires tool_result to be the first content
+    // immediately following the tool_use it answers.
     const lastMsg2 = reqs[1].messages.at(-1)
-    expect(lastMsg2.content[0].text).toMatch(/Current diagram state/)
+    expect(lastMsg2.content[0].type).toBe('tool_result')
+    expect(lastMsg2.content.at(-1).text).toMatch(/Current diagram state/)
   })
 })