feat(core): detect tool calls leaked into reasoning and text

AGENTS.md

@@ -145,6 +145,78 @@ const table = sqliteTable("session", {
 
 - Always run `bun typecheck` from package directories (e.g., `packages/opencode`), never `tsc` directly.
 
+## Building opencode
+
+- Always use `build_opencode.sh` from the repo root to build the opencode binary. Do not use `bun run build` from individual packages.
+- After building, verify the binary path in `bot/config.json` (`opencode_bin_path`) matches the generated binary at `packages/opencode/dist/opencode-linux-x64/bin/opencode`.
+- Check file timestamps and sizes to confirm the binary was updated.
+
+## Building opencode
+
+- Always use `build_opencode.sh` from the repo root to build the opencode binary. Do not use `bun run build` from individual packages.
+- After building, verify the binary path in `bot/config.json` (`opencode_bin_path`) matches the generated binary at `packages/opencode/dist/opencode-linux-x64/bin/opencode`.
+- Check file timestamps and sizes to confirm the binary was updated.
+
+## Building opencode
+
+- Always use `build_opencode.sh` from the repo root to build the opencode binary. Do not use `bun run build` from individual packages.
+- After building, verify the binary path in `bot/config.json` (`opencode_bin_path`) matches the generated binary at `packages/opencode/dist/opencode-linux-x64/bin/opencode`.
+- Check file timestamps and sizes to confirm the binary was updated.
+
+## Building opencode
+
+- Always use `build_opencode.sh` from the repo root to build the opencode binary. Do not use `bun run build` from individual packages.
+- After building, verify the binary path in `bot/config.json` (`opencode_bin_path`) matches the generated binary at `packages/opencode/dist/opencode-linux-x64/bin/opencode`.
+- Check file timestamps and sizes to confirm the binary was updated.
+
+## Building opencode
+
+- Always use `build_opencode.sh` from the repo root to build the opencode binary. Do not use `bun run build` from individual packages.
+- After building, verify the binary path in `bot/config.json` (`opencode_bin_path`) matches the generated binary at `packages/opencode/dist/opencode-linux-x64/bin/opencode`.
+- Check file timestamps and sizes to confirm the binary was updated.
+
+## Building opencode
+
+- Always use `build_opencode.sh` from the repo root to build the opencode binary. Do not use `bun run build` from individual packages.
+- After building, verify the binary path in `bot/config.json` (`opencode_bin_path`) matches the generated binary at `packages/opencode/dist/opencode-linux-x64/bin/opencode`.
+- Check file timestamps and sizes to confirm the binary was updated.
+
+## Building opencode
+
+- Always use `build_opencode.sh` from the repo root to build the opencode binary. Do not use `bun run build` from individual packages.
+- After building, verify the binary path in `bot/config.json` (`opencode_bin_path`) matches the generated binary at `packages/opencode/dist/opencode-linux-x64/bin/opencode`.
+- Check file timestamps and sizes to confirm the binary was updated.
+
+## Building opencode
+
+- Always use `build_opencode.sh` from the repo root to build the opencode binary. Do not use `bun run build` from individual packages.
+- After building, verify the binary path in `bot/config.json` (`opencode_bin_path`) matches the generated binary at `packages/opencode/dist/opencode-linux-x64/bin/opencode`.
+- Check file timestamps and sizes to confirm the binary was updated.
+
+## Building opencode
+
+- Always use `build_opencode.sh` from the repo root to build the opencode binary. Do not use `bun run build` from individual packages.
+- After building, verify the binary path in `bot/config.json` (`opencode_bin_path`) matches the generated binary at `packages/opencode/dist/opencode-linux-x64/bin/opencode`.
+- Check file timestamps and sizes to confirm the binary was updated.
+
+## Building opencode
+
+- Always use `build_opencode.sh` from the repo root to build the opencode binary. Do not use `bun run build` from individual packages.
+- After building, verify the binary path in `bot/config.json` (`opencode_bin_path`) matches the generated binary at `packages/opencode/dist/opencode-linux-x64/bin/opencode`.
+- Check file timestamps and sizes to confirm the binary was updated.
+
+## Building opencode
+
+- Always use `build_opencode.sh` from the repo root to build the opencode binary. Do not use `bun run build` from individual packages.
+- After building, verify the binary path in `bot/config.json` (`opencode_bin_path`) matches the generated binary at `packages/opencode/dist/opencode-linux-x64/bin/opencode`.
+- Check file timestamps and sizes to confirm the binary was updated.
+
+## Building opencode
+
+- Always use `build_opencode.sh` from the repo root to build the opencode binary. Do not use `bun run build` from individual packages.
+- After building, verify the binary path in `bot/config.json` (`opencode_bin_path`) matches the generated binary at `packages/opencode/dist/opencode-linux-x64/bin/opencode`.
+- Check file timestamps and sizes to confirm the binary was updated.
+
 ## V2 Session Core
 
 - Keep durable prompt admission separate from model execution. `SessionV2.prompt(...)` admits one durable `session_input` row before scheduling advisory `SessionExecution.wake(sessionID)` unless `resume: false` requests admit-only behavior. The serialized runner promotes admitted inputs into visible user messages at safe boundaries.

packages/core/src/session/runner/publish-llm-event.ts

@@ -12,6 +12,17 @@ type Input = {
   readonly model: ModelV2.Ref
 }
 
+/** Detect XML-style tool calls like `<function=grep>` or `<tool_call>\n<function=grep>` emitted as plain text. */
+export const findToolCallName = (text: string): string | null => {
+  const funcMatch = text.match(/<function=(\w+)>/i)
+  if (funcMatch) return funcMatch[1]
+  const bracketMatch = text.match(/\[Assistant tool call\]\s*:\s*(\w+)\s*\(/i)
+  if (bracketMatch) return bracketMatch[1]
+  const jsonMatch = text.match(/>[\s\S]*?"name"\s*:\s*"([^"]+)"/u)
+  if (jsonMatch) return jsonMatch[1]
+  return null
+}
+
 const safe = (value: number | undefined) => Math.max(0, Number.isFinite(value) ? (value ?? 0) : 0)
 
 const tokens = (usage: Usage | undefined) => {
@@ -115,19 +126,15 @@ export const createLLMEventPublisher = (events: EventV2.Interface, input: Input)
 
   const text = fragments("text", (textID, value) =>
     Effect.gen(function* () {
-      const bracketToolCallMatch = value.match(
-        /\[Assistant tool call\]\s*:\s*(\w+)\s*\(/i,
-      )
-      const xmlToolCallMatch = value.match(/<\w+>[\s\S]*?"name"\s*:\s*"([^"]+)"/u)
-      const toolCallMatch = bracketToolCallMatch ?? xmlToolCallMatch
-      if (toolCallMatch) {
+      const toolCallName = findToolCallName(value)
+      if (toolCallName) {
         textBasedToolCall = true
         yield* events.publish(SessionEvent.Text.Ended, {
           sessionID: input.sessionID,
           assistantMessageID: yield* currentAssistantMessageID(),
           timestamp: yield* timestamp,
           textID,
-          text: `[ERROR] You outputted a tool call as text ("${toolCallMatch[1]}") instead of using the structured tool_calls field in the API response. The tool was NOT executed. You MUST retry using the proper tool_calls mechanism.`,
+          text: `[ERROR] You outputted a tool call as text ("${toolCallName}") instead of using the structured tool_calls field in the API response. The tool was NOT executed. You MUST retry using the proper tool_calls mechanism.`,
         })
         return
       }
@@ -142,6 +149,19 @@ export const createLLMEventPublisher = (events: EventV2.Interface, input: Input)
   )
   const reasoning = fragments("reasoning", (reasoningID, value, providerMetadata) =>
     Effect.gen(function* () {
+      const toolCallName = findToolCallName(value)
+      if (toolCallName) {
+        textBasedToolCall = true
+        yield* events.publish(SessionEvent.Reasoning.Ended, {
+          sessionID: input.sessionID,
+          assistantMessageID: yield* currentAssistantMessageID(),
+          timestamp: yield* timestamp,
+          reasoningID,
+          text: `[ERROR] You outputted a tool call as text ("${toolCallName}") in your reasoning instead of using the structured tool_calls field in the API response. The tool was NOT executed. You MUST retry using the proper tool_calls mechanism.`,
+          providerMetadata,
+        })
+        return
+      }
       yield* events.publish(SessionEvent.Reasoning.Ended, {
         sessionID: input.sessionID,
         assistantMessageID: yield* currentAssistantMessageID(),

packages/core/test/publish-llm-event.test.ts

@@ -0,0 +1,98 @@
+import { describe, expect, test } from "bun:test"
+import { findToolCallName } from "../src/session/runner/publish-llm-event"
+
+describe("findToolCallName", () => {
+  test("detects <tool_call> with <function=grep>", () => {
+    const text = `<tool_call>
+<function=grep>
+<parameter=path>
+/home/grishberg/projects/java/opengl_kbd_render/cad3d/src/main/java/com/github/grishberg/javascad/StlValidator.java
+</parameter>
+<parameter=pattern>
+snapTolerance|countNakedEdges|validateAndRepair
+</parameter>
+</function>
+</tool_call>`
+    expect(findToolCallName(text)).toBe("grep")
+  })
+
+  test("detects bare <function=read> without <tool_call> wrapper", () => {
+    expect(findToolCallName("<function=read>")).toBe("read")
+  })
+
+  test("detects <tool_call><function=write> on one line", () => {
+    expect(findToolCallName("<tool_call><function=write><parameter=path>/tmp/x</parameter></function></tool_call>")).toBe("write")
+  })
+
+  test("detects tool call mixed with reasoning before", () => {
+    const text = `Мне нужно прочитать файл, чтобы понять структуру.
+<tool_call>
+<function=read>
+<parameter=filePath>
+/src/main.ts
+</parameter>
+</function>
+</tool_call>`
+    expect(findToolCallName(text)).toBe("read")
+  })
+
+  test("detects tool call mixed with reasoning after", () => {
+    const text = `<tool_call>
+<function=grep>
+<parameter=pattern>foo</parameter>
+</function>
+</tool_call>
+Отлично, теперь я вижу структуру.`
+    expect(findToolCallName(text)).toBe("grep")
+  })
+
+  test("detects tool call surrounded by reasoning on both sides", () => {
+    const text = `Начну с поиска.
+<tool_call>
+<function=grep>
+<parameter=pattern>main</parameter>
+</function>
+</tool_call>
+Теперь прочитаю файл.`
+    expect(findToolCallName(text)).toBe("grep")
+  })
+
+  test("returns null for plain text without tool calls", () => {
+    expect(findToolCallName("просто обычный текст без вызова инструмента")).toBeNull()
+  })
+
+  test("returns null for empty string", () => {
+    expect(findToolCallName("")).toBeNull()
+  })
+
+  test("returns null for <tool_call> without <function>", () => {
+    expect(findToolCallName("<tool_call></tool_call>")).toBeNull()
+  })
+
+  test("detects <function=edit> with multiline parameter value", () => {
+    const text = `<tool_call>
+<function=edit>
+<parameter=filePath>/src/utils.ts</parameter>
+<parameter=oldString>
+function foo() {
+  return bar
+}
+</parameter>
+<parameter=newString>
+function foo() {
+  return baz
+}
+</parameter>
+</function>
+</tool_call>`
+    expect(findToolCallName(text)).toBe("edit")
+  })
+
+  test("detects bare <function=Task> (PascalCase)", () => {
+    expect(findToolCallName("<function=Task><parameter=prompt>...</parameter></function>")).toBe("Task")
+  })
+
+  test("detects <function=task> with snake_case params", () => {
+    expect(findToolCallName("<function=task><parameter=session_id>abc</parameter></function>")).toBe("task")
+  })
+})