diff --git a/packages/core/src/session/runner/to-llm-message.ts b/packages/core/src/session/runner/to-llm-message.ts index 3734d490f41..05d5c9522e4 100644 --- a/packages/core/src/session/runner/to-llm-message.ts +++ b/packages/core/src/session/runner/to-llm-message.ts @@ -139,28 +139,32 @@ const toolResult = (tool: SessionMessage.AssistantTool, providerMetadata: Provid } } -// A call cut off before it ran (by Esc, a crash or a failed stream) is not replayed. Input -// errors stay so the model can reissue the call. -const isAbandonedTool = (item: SessionMessage.AssistantContent) => - item.type === "tool" && - item.time.ran === undefined && - !(item.state.status === "error" && item.state.error.type.startsWith("tool.input")) +// Sessions migrated from V1 never recorded `time.ran`, so a completed call also counts as run. +const ranTool = (item: SessionMessage.AssistantContent) => + item.type === "tool" && (item.time.ran !== undefined || item.state.status === "completed") const assistant = (message: SessionMessage.Assistant, model: Model.Ref, providerMetadataKey: string) => { const sameProvider = message.model.providerID === model.providerID const sameModel = sameProvider && message.model.id === model.id // A failed stream is retried from what had already happened: the calls that ran and everything before // them. The rest was cut off, and dropping it keeps the request from ending in a partial reply. - const ran = message.content.findLastIndex((item) => item.type === "tool" && item.time.ran !== undefined) const kept = message.error === undefined || message.error.type === "aborted" ? message.content - : message.content.slice(0, ran + 1) - const items = kept.filter((item) => !isAbandonedTool(item)) + : message.content.slice(0, message.content.findLastIndex(ranTool) + 1) // A stream cut by Esc, a crash or a failure may end mid-block, so its last block's provider state is not - // replayed, and reasoning is only replayed once the model moved on to output. + // replayed, and reasoning is only replayed once the model moved on to output. Calls it cut off before + // they ran are dropped; input errors stay so the model can reissue the call. const cut = message.error !== undefined || message.finish === undefined const tail = cut ? message.content.at(-1) : undefined + const items = cut + ? kept.filter( + (item) => + item.type !== "tool" || + ranTool(item) || + (item.state.status === "error" && item.state.error.type.startsWith("tool.input")), + ) + : kept const content = items.flatMap((item, index): Array => { if (item.type === "text") return [ diff --git a/packages/core/test/session-runner-message.test.ts b/packages/core/test/session-runner-message.test.ts index 295d169b2dc..908e88eabec 100644 --- a/packages/core/test/session-runner-message.test.ts +++ b/packages/core/test/session-runner-message.test.ts @@ -806,6 +806,7 @@ Earlier work expect(messages[0]?.content).toEqual([ { type: "text", text: "Checking" }, { type: "reasoning", text: "Think", providerMetadata: { provider: { signature: "sig_1" } } }, + { type: "tool-call", id: "pending", name: "read", input: { path: "README.md" } }, { type: "tool-call", id: "running", name: "read", input: { path: "README.md" } }, { type: "tool-call", @@ -1507,6 +1508,43 @@ Earlier work expect(failedWithCalls[0]?.content[1]).toMatchObject({ type: "tool-call", id: "malformed" }) }) + test("replays tools migrated from V1, which never recorded when they ran", () => { + const migrated = (value: string, error?: SessionMessage.Assistant["error"]) => + SessionMessage.Assistant.make({ + id: id(value), + type: "assistant", + agent: build, + model: { id: Model.ID.make("same"), providerID: Provider.ID.make("provider") }, + content: [ + SessionMessage.AssistantTool.make({ + type: "tool", + id: `${value}-read`, + name: "read", + state: SessionMessage.ToolStateCompleted.make({ + status: "completed", + input: { path: "README.md" }, + content: [{ type: "text", text: "Hello" }], + }), + time: { created, completed: created }, + }), + ], + error, + finish: error === undefined ? "tool-calls" : "error", + time: { created, completed: created }, + }) + const messages = toLLMMessages( + [ + migrated("finished"), + migrated("aborted", { type: "aborted", message: "Step interrupted" }), + migrated("failed", { type: "provider.unknown", message: "Stream dropped" }), + ], + Model.Ref.make({ id: Model.ID.make("same"), providerID: Provider.ID.make("provider") }), + ) + expect( + messages.map((message) => message.content.map((part) => (part.type === "tool-call" ? part.id : part.type))), + ).toEqual([["finished-read"], ["tool-result"], ["aborted-read"], ["tool-result"], ["failed-read"], ["tool-result"]]) + }) + test("preserves assistant text provider state for the same model", () => { const messages = toLLMMessages( [