fix: claim a textless assistant row only within the newest turn

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

backend/migrations/20260916144653_flow_conversation_message_tool_call.up.sql

@@ -1,7 +1,8 @@
 -- A chat is rebuilt from its rows without reading jobs. A tool row for a script or flow
--- names the tool's own job, which holds its call; an MCP tool and a provider-native tool
--- run inside the agent's job, whose result lists every call of the turn with nothing tying
--- one to a row. For those the row carries the call itself.
+-- names the tool's own job, which holds its call. An MCP tool runs inside the agent's job,
+-- whose result lists every call of the turn with nothing tying one to a row, so its row
+-- carries the call itself; a provider-native web search carries only its citations, the
+-- provider never returning the query.
 ALTER TABLE flow_conversation_message ADD COLUMN tool_arguments TEXT;
 ALTER TABLE flow_conversation_message ADD COLUMN tool_result TEXT;
 

backend/windmill-api-flow-conversations/src/lib.rs

@@ -36,9 +36,12 @@ pub struct FlowConversationMessage {
     pub created_seq: i64,
     pub step_name: Option<String>,
     pub success: bool,
-    /// The call behind a tool row whose tool has no job of its own — an MCP tool, or a
-    /// provider-native one. Read back from the job otherwise, and null here.
+    /// An MCP tool call's arguments, which run inside the agent's job and so have none of
+    /// their own. Null for a web search, whose query the provider does not return, and for
+    /// a Windmill tool, whose own job holds its call.
     pub tool_arguments: Option<String>,
+    /// An MCP tool's result, a web search's citations, and what any failed tool failed
+    /// with. Null for a successful Windmill tool, whose own job holds the result.
     pub tool_result: Option<String>,
     /// The thinking of the iteration that produced this row. The agent job keeps the
     /// turn's thinking as one string.

chat-sdk/src/chat.ts

@@ -612,26 +612,32 @@ class ChatImpl implements Chat {
     if (rows.length === 0) return
     const messages = [...this.#state.messages]
     const known = new Set(messages.map((m) => m.serverId ?? m.id))
+    // An assistant message with no text is the thinking before a tool call, or before a
+    // structured answer whose text arrives as a tool call the stream never turns into a
+    // message. Text cannot tell such messages apart, so one is only ever claimed within
+    // the newest turn: a turn stopped before its rows landed leaves one behind, and no
+    // later row is its.
+    let turnStart = messages.length - 1
+    while (turnStart >= 0 && messages[turnStart].role !== 'user') turnStart--
+    const inTurn = (test: (m: ChatMessage) => boolean): number => {
+      const j = messages.slice(turnStart + 1).findIndex(test)
+      return j < 0 ? -1 : turnStart + 1 + j
+    }
     for (const row of rows.map(fromRow)) {
       if (known.has(row.id)) continue
       known.add(row.id)
-      let i = messages.findIndex(
-        (m) =>
-          m.seq === undefined &&
-          m.role === row.role &&
-          (m.content === row.content || (row.tool !== undefined && m.tool?.name === row.tool.name))
-      )
-      // A structured answer streams as thinking alone, its text arriving as a tool call
-      // the stream never turns into a message, so its row claims the message that holds
-      // that thinking and nothing else. Only past the newest user message: a turn stopped
-      // before its rows landed leaves such a message behind, and it is not this answer's.
-      if (i < 0 && row.role === 'assistant' && row.reasoning !== undefined) {
-        let turnStart = messages.length - 1
-        while (turnStart >= 0 && messages[turnStart].role !== 'user') turnStart--
-        const j = messages
-          .slice(turnStart + 1)
-          .findIndex((m) => m.seq === undefined && m.role === 'assistant' && m.content === '' && m.reasoning !== undefined)
-        if (j >= 0) i = turnStart + 1 + j
+      const sameShape = (m: ChatMessage) =>
+        m.seq === undefined &&
+        m.role === row.role &&
+        (m.content === row.content || (row.tool !== undefined && m.tool?.name === row.tool.name))
+      let i: number
+      if (row.role === 'assistant' && row.content === '') {
+        i = inTurn(sameShape)
+      } else {
+        i = messages.findIndex(sameShape)
+        if (i < 0 && row.role === 'assistant' && row.reasoning !== undefined) {
+          i = inTurn((m) => m.seq === undefined && m.role === 'assistant' && m.content === '' && m.reasoning !== undefined)
+        }
       }
       if (i >= 0) {
         const m = messages[i]

chat-sdk/test/chat.test.ts

@@ -666,6 +666,62 @@ describe('createChat with server history', () => {
     ])
   })
 
+  test('reasoning rows of a turn with a tool round stay with that turn after a stopped one', async () => {
+    let jobs = 0
+    const { fetch } = fetchMock(
+      (c) => (c.method === 'POST' && c.url.pathname.includes('/jobs/run/f/') ? text(`job-${++jobs}`) : undefined),
+      (c) =>
+        c.url.pathname.endsWith('/getupdate_sse/job-1')
+          ? sse([{ type: 'update', new_result_stream: ndjson({ type: 'reasoning_token_delta', content: 'first thoughts' }), stream_offset: 1 }])
+          : undefined,
+      (c) =>
+        c.url.pathname.endsWith('/getupdate_sse/job-2')
+          ? sse([
+              {
+                type: 'update',
+                new_result_stream: ndjson(
+                  { type: 'reasoning_token_delta', content: 'r1' },
+                  { type: 'tool_call', call_id: 'c1', function_name: 'lookup' },
+                  { type: 'tool_result', call_id: 'c1', function_name: 'lookup', result: '1', success: true },
+                  { type: 'reasoning_token_delta', content: 'r2' }
+                ),
+                stream_offset: 4,
+                completed: true,
+                only_result: { output: { n: 1 }, messages: [] }
+              }
+            ])
+          : undefined,
+      (c) => (c.url.pathname.includes('/queue/cancel/') ? text('ok') : undefined),
+      (c) =>
+        c.url.pathname.endsWith('/messages')
+          ? json([
+              messageRow(51, 'user', 'first'),
+              messageRow(52, 'user', 'again'),
+              messageRow(53, 'assistant', '', { job_id: 'step-2', reasoning: 'r1' }),
+              messageRow(54, 'tool', 'Used lookup tool', { job_id: 'tool-2' }),
+              messageRow(55, 'assistant', '{"n":1}', { job_id: 'step-2', reasoning: 'r2' })
+            ])
+          : undefined,
+      (c) => (c.url.pathname === '/api/w/ws/flow_conversations/list' ? json([]) : undefined)
+    )
+    const chat = createChat(options({}, fetch))
+    const first = chat.sendMessage('first')
+    await new Promise((r) => setTimeout(r, 50))
+    const stopped = chat.stop()
+    await first
+    const second = chat.sendMessage('again')
+    await stopped
+    await second
+    expect(chat.getState().messages.map((m) => [m.role, m.content, m.reasoning, m.seq])).toEqual([
+      ['user', 'first', undefined, 51],
+      ['assistant', '', 'first thoughts', undefined],
+      ['user', 'again', undefined, 52],
+      ['assistant', '', 'r1', 53],
+      ['tool', 'Used lookup tool', undefined, 54],
+      ['assistant', '{"n":1}', 'r2', 55]
+    ])
+  })
+
   test('a structured answer row leaves a stopped earlier turn its thinking', async () => {
     let jobs = 0
     const { fetch } = fetchMock(