feat: Complete dangling tool requests to avoid API errors (#840)

Co-authored-by: Hadley Wickham <[email protected]>

Author: gadenbuie

NEWS.md

@@ -21,6 +21,7 @@
 * `chat_openai_compatible()` replaces `chat_openai()` as the interface to use for OpenAI-compatible APIs, and `chat_openai()` is reserved for the official OpenAI API. Unlike previous versions of `chat_openai()`, the `base_url` parameter is now required (#801).
 * `chat_portkey()` now requires you to supply a model (#786).
 * `chat_portkey(virtual_key)` no longer needs to be supplied; instead Portkey recommends including the virtual key/provider in the `model` (#786).
+* `Chat$chat()`, `Chat$stream()`, and similar methods now add empty tool results when a the chat is interrupted during a tool call loop, allowing the conversation to be resumed without causing an API error (#840).
 * `Chat$chat_structured()` and friends now only warn if multiple JSON payloads found (instead of erroring) (@kbenoit, #732).
 * `Chat$get_tokens()` gives a brief description of the turn contents to make it easier to see which turn tokens are spent on (#618).
 * `Chat$get_tokens()` now also returns the cost, and returns one row for each assistant turn, better representing the underlying data received from LLM APIs. Similarly, the `print()` method now reports costs on each assistant turn, rather than trying to parse out individual costs (#824).

R/chat.R

@@ -193,7 +193,9 @@ Chat <- R6::R6Class(
     #'   `NULL`, then the value of `echo` set when the chat object was created
     #'   will be used.
     chat = function(..., echo = NULL) {
-      turn <- user_turn(...)
+      finish_tools <- private$complete_dangling_tool_requests()
+
+      turn <- user_turn(!!!finish_tools, ...)
       echo <- check_echo(echo %||% private$echo)
 
       # Returns a single turn (the final response from the assistant), even if
@@ -221,7 +223,9 @@ Chat <- R6::R6Class(
     #'   using the schema. For example, this will turn arrays of objects into
     #'   data frames and arrays of strings into a character vector.
     chat_structured = function(..., type, echo = "none", convert = TRUE) {
-      turn <- user_turn(..., .check_empty = FALSE)
+      finish_tools <- private$complete_dangling_tool_requests()
+
+      turn <- user_turn(!!!finish_tools, ..., .check_empty = FALSE)
       echo <- check_echo(echo %||% private$echo)
       check_bool(convert)
 
@@ -252,7 +256,9 @@ Chat <- R6::R6Class(
     #'   using the schema. For example, this will turn arrays of objects into
     #'   data frames and arrays of strings into a character vector.
     chat_structured_async = function(..., type, echo = "none", convert = TRUE) {
-      turn <- user_turn(..., .check_empty = FALSE)
+      finish_tools <- private$complete_dangling_tool_requests()
+
+      turn <- user_turn(!!!finish_tools, ..., .check_empty = FALSE)
       echo <- check_echo(echo %||% private$echo)
       check_bool(convert)
 
@@ -287,7 +293,9 @@ Chat <- R6::R6Class(
     #'   an interactive user interface. Concurrent mode is the default and is
     #'   best suited for automated scripts or non-interactive applications.
     chat_async = function(..., tool_mode = c("concurrent", "sequential")) {
-      turn <- user_turn(...)
+      finish_tools <- private$complete_dangling_tool_requests()
+
+      turn <- user_turn(!!!finish_tools, ...)
       tool_mode <- arg_match(tool_mode)
 
       # Returns a single turn (the final response from the assistant), even if
@@ -315,7 +323,9 @@ Chat <- R6::R6Class(
     #'   rich content types. When `stream = "content"`, `stream()` yields
     #'   [Content] objects.
     stream = function(..., stream = c("text", "content")) {
-      turn <- user_turn(...)
+      finish_tools <- private$complete_dangling_tool_requests()
+
+      turn <- user_turn(!!!finish_tools, ...)
       stream <- arg_match(stream)
       private$chat_impl(
         turn,
@@ -343,7 +353,9 @@ Chat <- R6::R6Class(
       tool_mode = c("concurrent", "sequential"),
       stream = c("text", "content")
     ) {
-      turn <- user_turn(...)
+      finish_tools <- private$complete_dangling_tool_requests()
+
+      turn <- user_turn(!!!finish_tools, ...)
       tool_mode <- arg_match(tool_mode)
       stream <- arg_match(stream)
       private$chat_impl_async(
@@ -735,6 +747,29 @@ Chat <- R6::R6Class(
 
     has_system_prompt = function() {
       length(private$.turns) > 0 && is_system_turn(private$.turns[[1]])
+    },
+
+    complete_dangling_tool_requests = function() {
+      if (length(private$.turns) == 0) {
+        return(NULL)
+      }
+
+      last_turn <- private$.turns[[length(private$.turns)]]
+      if (last_turn@role != "assistant") {
+        return(NULL)
+      }
+
+      tool_requests <- keep(last_turn@contents, is_tool_request)
+      if (length(tool_requests) == 0) {
+        return(NULL)
+      }
+
+      lapply(tool_requests, function(req) {
+        ContentToolResult(
+          error = "Chat ended before the tool could be invoked.",
+          request = req
+        )
+      })
     }
   )
 )
@@ -774,7 +809,7 @@ print.Chat <- function(x, ...) {
 turn_cost <- function(tokens, cost, prefix, suffix = "") {
   out <- paste0(prefix, "input=")
 
-  if (tokens[[3]] > 0) {
+  if (!is.na(tokens[[3]]) && tokens[[3]] > 0) {
     out <- paste0(out, tokens[[1]], "+", tokens[[3]])
   } else {
     out <- paste0(out, tokens[[1]])