feat(tracedetail): format openllmetry legacy prompt and completion keys

pkg/types/aiobservabilitytypes/keys.go

@@ -17,8 +17,10 @@ const (
 	GenAIUsageCacheCreationInputTokens = "gen_ai.usage.cache_creation.input_tokens"
 	GenAIUsageReasoningOutputTokens    = "gen_ai.usage.reasoning.output_tokens"
 
-	GenAIInputMessages  = "gen_ai.input.messages"
-	GenAIOutputMessages = "gen_ai.output.messages"
+	GenAIInputMessages         = "gen_ai.input.messages"
+	GenAIOutputMessages        = "gen_ai.output.messages"
+	GenAISystemInstructions    = "gen_ai.system_instructions"
+	GenAIResponseFinishReasons = "gen_ai.response.finish_reasons"
 
 	GenAIToolCallID        = "gen_ai.tool.call.id"
 	GenAIToolCallArguments = "gen_ai.tool.call.arguments"

pkg/types/aiobservabilitytypes/message.go

@@ -1,5 +1,6 @@
 // The message types mirror the OpenTelemetry GenAI message schemas, one type per schema definition:
-// open-telemetry/semantic-conventions-genai@06ec68e, model/gen-ai/gen-ai-{input,output}-messages.json.
+// https://github.com/open-telemetry/semantic-conventions-genai/blob/06ec68e722c45a7218e23ea1bc1339fe4e21ecae/model/gen-ai/gen-ai-input-messages.json
+// https://github.com/open-telemetry/semantic-conventions-genai/blob/06ec68e722c45a7218e23ea1bc1339fe4e21ecae/model/gen-ai/gen-ai-output-messages.json
 // Keys the schema does not name are dropped; the raw attribute on the span still has them.
 
 package aiobservabilitytypes

pkg/types/aiobservabilitytypes/message_format.go

@@ -4,23 +4,31 @@ import (
 	"encoding/json"
 	"fmt"
 	"regexp"
+	"sort"
+	"strconv"
 	"strings"
 
 	"github.com/SigNoz/signoz/pkg/valuer"
 )
 
 const (
-	FormatterSemconv    = "semconv"
-	FormatterOpenAIChat = "openai.chat"
-	FormatterText       = "text"
-	FormatterGeneric    = "generic"
+	FormatterSemconv       = "semconv"
+	FormatterOpenInference = "openinference"
+	FormatterOpenAIChat    = "openai.chat"
+	FormatterText          = "text"
+	FormatterGeneric       = "generic"
 )
 
 var dataURL = regexp.MustCompile(`^data:([^;,]+);base64,(.+)$`)
 
-// Provider spellings of the finish reasons; the enum values themselves pass through.
+// Provider spellings of the finish reasons; the enum values themselves pass through. tool_calls and
+// function_call are OpenAI's, end_turn, stop_sequence, max_tokens and tool_use Anthropic's, tool-calls the
+// Vercel AI SDK's:
+// https://github.com/openai/openai-openapi/blob/aeb6a041dc56c011ca9f12127ec898f911100aa0/openapi.yaml#L44782
+// https://platform.claude.com/docs/en/api/messages
+// https://github.com/vercel/ai/blob/a5e92fe3c692f8e152fe1eb9ef50a4e734ad0ffb/packages/provider/src/language-model/v2/language-model-v2-finish-reason.ts
 var finishReasonAliases = map[string]FinishReason{
-	"tool_calls": FinishReasonToolCall, "function_call": FinishReasonToolCall,
+	"tool_calls": FinishReasonToolCall, "function_call": FinishReasonToolCall, "tool-calls": FinishReasonToolCall, "tool_use": FinishReasonToolCall,
 	"end_turn": FinishReasonStop, "stop_sequence": FinishReasonStop,
 	"max_tokens": FinishReasonLength,
 }
@@ -35,16 +43,23 @@ type FormattedMessages struct {
 
 // FormatMessages builds the OTel GenAI view of a span's messages from the formats SDKs write. attributes
 // is the span's JSON attribute document, nested by key segment, so an object value is whole. Each
-// message value is a JSON string or a structured value. A tool call and its result take precedence over
-// messages beside them.
+// message value is a JSON string or a structured value. OpenInference's flattened messages come first,
+// then a tool call and its result, then the message attributes, then openllmetry's flattened prompts
+// and completions. gen_ai.system_instructions leads the
+// input and gen_ai.response.finish_reasons overrides the finish reasons the converter found, whichever
+// converter ran.
 func FormatMessages(attributes map[string]any) FormattedMessages {
 	f := &formatting{attributes: attributes, out: FormattedMessages{Input: InputMessages{}, Output: OutputMessages{}, Warnings: []string{}}}
 	switch {
+	case f.openInference():
 	case attribute(f.attributes, GenAIToolCallArguments) != nil || attribute(f.attributes, GenAIToolCallResult) != nil:
 		f.tool()
 	case attribute(f.attributes, GenAIInputMessages) != nil || attribute(f.attributes, GenAIOutputMessages) != nil:
 		f.messages()
+	case f.traceloop():
 	}
+	f.prependSystemInstructions()
+	f.fillFinishReasons()
 	return f.out
 }
 
@@ -63,18 +78,60 @@ func (f *formatting) messages() {
 	if input := attribute(f.attributes, GenAIInputMessages); input != nil {
 		var msgs OutputMessages
 		msgs, inputLabel = f.side(decode(input), RoleUser)
-		for _, m := range msgs {
-			f.out.Input = append(f.out.Input, ChatMessage{Role: m.Role, Parts: m.Parts, Name: m.Name})
-		}
+		f.out.Input = inputMessages(msgs)
 	}
 	if output := attribute(f.attributes, GenAIOutputMessages); output != nil {
 		f.out.Output, outputLabel = f.side(decode(output), RoleAssistant)
+		f.completionToolCalls()
 	}
+	f.warnMixedFormats(inputLabel, outputLabel)
+}
+
+func (f *formatting) warnMixedFormats(inputLabel, outputLabel string) {
 	if inputLabel != "" && outputLabel != "" && outputLabel != inputLabel {
 		f.warn("output formatted as %s, input as %s", outputLabel, inputLabel)
 	}
 }
 
+// prependSystemInstructions adds gen_ai.system_instructions, a list of parts, as a leading system message,
+// unless the input already starts with a system message of the same text: the OpenInference strands
+// processor keeps the attribute and inserts it as message 0 as well. A span without messages, such as
+// create_agent, stays without them.
+// https://github.com/open-telemetry/semantic-conventions-genai/blob/06ec68e722c45a7218e23ea1bc1339fe4e21ecae/model/gen-ai/gen-ai-system-instructions.json
+// https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/python/instrumentation/openinference-instrumentation-strands-agents/src/openinference/instrumentation/strands_agents/processor.py#L152
+// https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/python/instrumentation/openinference-instrumentation-strands-agents/src/openinference/instrumentation/strands_agents/processor.py#L293
+func (f *formatting) prependSystemInstructions() {
+	raw := attribute(f.attributes, GenAISystemInstructions)
+	if raw == nil || (len(f.out.Input) == 0 && len(f.out.Output) == 0) {
+		return
+	}
+	parts := systemParts(decode(raw))
+	if len(parts) == 0 || (len(f.out.Input) > 0 && f.out.Input[0].Role == RoleSystem && textOf(f.out.Input[0].Parts) == textOf(parts)) {
+		return
+	}
+	f.out.Input = append(InputMessages{{Role: RoleSystem, Parts: parts}}, f.out.Input...)
+}
+
+// fillFinishReasons applies gen_ai.response.finish_reasons, one entry per output message, over
+// whatever the converter found in the payload.
+// https://github.com/open-telemetry/semantic-conventions-genai/blob/06ec68e722c45a7218e23ea1bc1339fe4e21ecae/model/gen-ai/registry.yaml#L210
+func (f *formatting) fillFinishReasons() {
+	raw := attribute(f.attributes, GenAIResponseFinishReasons)
+	if raw == nil || len(f.out.Output) == 0 {
+		return
+	}
+	reasons := stringList(decode(raw))
+	if len(reasons) != len(f.out.Output) {
+		f.warn("gen_ai.response.finish_reasons has %d entries for %d output messages", len(reasons), len(f.out.Output))
+		return
+	}
+	for i, reason := range reasons {
+		if fr := finishReason(reason); fr != nil {
+			f.out.Output[i].FinishReason = fr
+		}
+	}
+}
+
 // side converts one attribute value by its shape; role is what a bare string on this side is
 // spoken as.
 func (f *formatting) side(value any, role Role) (OutputMessages, string) {
@@ -121,6 +178,44 @@ func attribute(doc map[string]any, key string) any {
 	return current
 }
 
+// indexed reads a flattened list, "<key>.<N>.*" keys stored as an object keyed by N, in numeric order.
+// A list value is returned as is.
+func indexed(doc map[string]any, key string) []map[string]any {
+	var items []any
+	switch v := attribute(doc, key).(type) {
+	case []any:
+		items = v
+	case map[string]any:
+		indexes := make([]int, 0, len(v))
+		byIndex := make(map[int]any, len(v))
+		for k, item := range v {
+			n, err := strconv.Atoi(k)
+			if err != nil {
+				continue
+			}
+			indexes = append(indexes, n)
+			byIndex[n] = item
+		}
+		sort.Ints(indexes)
+		for _, n := range indexes {
+			items = append(items, byIndex[n])
+		}
+	}
+	out := make([]map[string]any, 0, len(items))
+	for _, item := range items {
+		if m, ok := item.(map[string]any); ok {
+			out = append(out, m)
+		}
+	}
+	return out
+}
+
+// indexedAt reads one entry of a flattened list by its index.
+func indexedAt(doc map[string]any, key string, i int) map[string]any {
+	m, _ := attribute(doc, key+"."+strconv.Itoa(i)).(map[string]any)
+	return m
+}
+
 func firstObject(list []any) (map[string]any, bool) {
 	if len(list) == 0 {
 		return nil, false
@@ -133,6 +228,31 @@ func genericMessage(value any) OutputMessage {
 	return OutputMessage{Parts: Parts{part(GenericPart{"type": FormatterGeneric, "content": stringOf(value)})}}
 }
 
+func inputMessages(msgs OutputMessages) InputMessages {
+	out := make(InputMessages, 0, len(msgs))
+	for _, m := range msgs {
+		out = append(out, ChatMessage{Role: m.Role, Parts: m.Parts, Name: m.Name})
+	}
+	return out
+}
+
+// foldAssistant merges runs of assistant messages into one and drops messages without parts, for
+// SDKs that write one message per output item.
+func foldAssistant(msgs OutputMessages) OutputMessages {
+	out := make(OutputMessages, 0, len(msgs))
+	for _, msg := range msgs {
+		if len(out) > 0 && msg.Role == RoleAssistant && out[len(out)-1].Role == RoleAssistant {
+			out[len(out)-1].Parts = append(out[len(out)-1].Parts, msg.Parts...)
+			continue
+		}
+		if len(msg.Parts) == 0 {
+			continue
+		}
+		out = append(out, msg)
+	}
+	return out
+}
+
 // decode parses a JSON-encoded attribute value, twice when an SDK encoded the string again.
 // Anything that is not JSON is returned as is.
 func decode(value any) any {
@@ -179,6 +299,23 @@ func finishReason(value string) *FinishReason {
 	return &fr
 }
 
+// stringList reads a list of strings, or a bare string as a list of one.
+func stringList(value any) []string {
+	switch v := value.(type) {
+	case []any:
+		out := make([]string, 0, len(v))
+		for _, item := range v {
+			out = append(out, stringOf(item))
+		}
+		return out
+	case string:
+		if v != "" {
+			return []string{v}
+		}
+	}
+	return nil
+}
+
 func optionalString(value any) *string {
 	s := stringOf(value)
 	if s == "" {
@@ -195,8 +332,37 @@ func textPart(content string) TextPart {
 	return TextPart{Type: PartTypeText, Content: content}
 }
 
-// contentPart converts one OpenAI content block, which semconv parts may also carry; a block of
-// any other type is kept as sent.
+// textOf joins the text parts, for comparing a system message with gen_ai.system_instructions.
+func textOf(parts Parts) string {
+	var texts []string
+	for _, p := range parts {
+		if t, ok := p.Value.(TextPart); ok {
+			texts = append(texts, t.Content)
+		}
+	}
+	return strings.Join(texts, "\n")
+}
+
+// systemParts reads a system prompt given outside the history, a list of parts or a bare string.
+func systemParts(value any) Parts {
+	switch v := value.(type) {
+	case []any:
+		parts := make(Parts, 0, len(v))
+		for _, item := range v {
+			parts = append(parts, contentPart(item))
+		}
+		return parts
+	case string:
+		if v != "" {
+			return Parts{part(textPart(v))}
+		}
+	}
+	return nil
+}
+
+// contentPart converts one OpenAI content block, which semconv parts may also carry; a text part
+// in the semconv spelling has content where OpenAI has text. A block of any other type is kept
+// as sent.
 func contentPart(item any) Part {
 	p, ok := item.(map[string]any)
 	if !ok {
@@ -207,6 +373,9 @@ func contentPart(item any) Part {
 	}
 	switch stringOf(p["type"]) {
 	case "text":
+		if p["text"] == nil {
+			return part(textPart(stringOf(p["content"])))
+		}
 		return part(textPart(stringOf(p["text"])))
 	case "refusal":
 		return part(textPart(stringOf(p["refusal"])))

pkg/types/aiobservabilitytypes/message_format_openai_chat_test.go

@@ -167,7 +167,8 @@ func TestFormat_OpenAIChat(t *testing.T) {
 			]`,
 		},
 		{
-			// openinference capture, input.value of the ChatCompletion span
+			// ChatCompletionRequestMessageContentPartFile {file: {file_id}}, as OpenInference records the request in input.value:
+			// https://github.com/openai/openai-openapi/blob/aeb6a041dc56c011ca9f12127ec898f911100aa0/openapi.yaml#L42650
 			name:   "FileByID_FilePart",
 			input:  `{"model": "gpt-4o-mini", "messages": [{"role": "user", "content": [{"type": "text", "text": "Summarise this document in one sentence."}, {"type": "file", "file": {"file_id": "file-6F2ksmvXxt4VdoqmHRw6kL"}}]}], "max_tokens": 60}`,
 			output: `{"id":"chatcmpl-DktFileCapture0001","object":"chat.completion","model":"gpt-4o-mini-2024-07-18","choices":[{"index":0,"finish_reason":"stop","logprobs":null,"message":{"role":"assistant","content":"The document is a short PDF about dragons.","refusal":null,"annotations":[]}}]}`,
@@ -189,7 +190,8 @@ func TestFormat_OpenAIChat(t *testing.T) {
 			]`,
 		},
 		{
-			// openinference capture, input.value of the ChatCompletion span
+			// ChatCompletionRequestMessageContentPartFile {file: {filename, file_data}}, file_data a data URL:
+			// https://github.com/openai/openai-openapi/blob/aeb6a041dc56c011ca9f12127ec898f911100aa0/openapi.yaml#L42650
 			name:  "FileInline_BlobPart",
 			input: `{"model": "gpt-4o-mini", "messages": [{"role": "user", "content": [{"type": "text", "text": "Summarise this document in one sentence."}, {"type": "file", "file": {"filename": "draconomicon.pdf", "file_data": "data:application/pdf;base64,JVBERi0xLjQKMSAwIG9iajw8L1R5cGUvQ2F0YWxvZy9QYWdlcyAyIDAgUj4+ZW5kb2JqCjIgMCBvYmo8PC9UeXBlL1BhZ2VzL0tpZHNbXS9Db3VudCAwPj5lbmRvYmoKdHJhaWxlcjw8L1Jvb3QgMSAwIFI+PgolJUVPRgo="}}]}], "max_tokens": 60}`,
 			wantInput: `[

pkg/types/aiobservabilitytypes/message_format_openinference.go

@@ -0,0 +1,137 @@
+package aiobservabilitytypes
+
+import "github.com/SigNoz/signoz/pkg/valuer"
+
+// Format: the flattened llm.input_messages.N.message.* and llm.output_messages.N.message.* keys, the
+// provider-neutral copy of the conversation the OpenInference instrumentations write on LLM spans. A side
+// without them falls back to the raw provider payload the mapper moved onto gen_ai.*.messages; pipecat
+// streams the reply into output.value alone.
+// https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/spec/semantic_conventions.md#L77
+// https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/spec/tool_calling.md#L37
+// https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/spec/multimodal_attributes.md#L35
+// https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/spec/llm_spans.md#L181
+// Handled: message.content, the ordered message.contents items (text, image, audio, video, reasoning),
+// message.tool_calls, the legacy message.function_call_*, tool results, llm.finish_reason. A tool_use
+// content item repeats a call from message.tool_calls and is skipped.
+// Dropped: reasoning id, signature, data and encrypted_content; audio transcripts.
+
+const (
+	openInferenceInputMessages  = "llm.input_messages"
+	openInferenceOutputMessages = "llm.output_messages"
+	openInferenceFinishReason   = "llm.finish_reason"
+)
+
+func (f *formatting) openInference() bool {
+	input := indexed(f.attributes, openInferenceInputMessages)
+	output := indexed(f.attributes, openInferenceOutputMessages)
+	if len(input) == 0 && len(output) == 0 {
+		return false
+	}
+	inputLabel, outputLabel := FormatterOpenInference, FormatterOpenInference
+	if len(input) > 0 {
+		f.out.Input = inputMessages(openInferenceMessages(input))
+	} else if raw := attribute(f.attributes, GenAIInputMessages); raw != nil {
+		var msgs OutputMessages
+		msgs, inputLabel = f.side(decode(raw), RoleUser)
+		f.out.Input = inputMessages(msgs)
+	}
+	if len(output) > 0 {
+		f.out.Output = openInferenceMessages(output)
+		if fr := finishReason(stringOf(attribute(f.attributes, openInferenceFinishReason))); fr != nil {
+			for i := range f.out.Output {
+				f.out.Output[i].FinishReason = fr
+			}
+		}
+	} else if raw := attribute(f.attributes, GenAIOutputMessages); raw != nil {
+		f.out.Output, outputLabel = f.side(decode(raw), RoleAssistant)
+	}
+	f.warnMixedFormats(inputLabel, outputLabel)
+	return true
+}
+
+// openInferenceMessages converts the message buckets and folds runs of assistant messages: the Responses
+// path writes one message per output item, an empty one for a reasoning item without a summary.
+func openInferenceMessages(buckets []map[string]any) OutputMessages {
+	msgs := make(OutputMessages, 0, len(buckets))
+	for _, bucket := range buckets {
+		message, _ := bucket["message"].(map[string]any)
+		msgs = append(msgs, openInferenceMessage(message))
+	}
+	return foldAssistant(msgs)
+}
+
+// openInferenceMessage converts one message.* object. A message without a role is a tool response when
+// it has a tool_call_id, an assistant turn when it has tool calls, else a user turn, as the
+// collector-contrib normalizer decides it:
+// https://github.com/open-telemetry/opentelemetry-collector-contrib/blob/fc79b95297776dec6c47490cd6f6026b942bf6fb/processor/genainormalizerprocessor/internal/openinference/messages.go#L473
+func openInferenceMessage(m map[string]any) OutputMessage {
+	msg := OutputMessage{Role: role(stringOf(m["role"])), Parts: Parts{}, Name: optionalString(m["name"])}
+	id := stringOf(m["tool_call_id"])
+	calls := indexed(m, "tool_calls")
+	if msg.Role.IsZero() {
+		switch {
+		case id != "":
+			msg.Role = RoleTool
+		case len(calls) > 0:
+			msg.Role = RoleAssistant
+		default:
+			msg.Role = RoleUser
+		}
+	}
+	if msg.Role == RoleTool && id != "" {
+		msg.Parts = Parts{part(ToolCallResponsePart{Type: PartTypeToolCallResponse, ID: &id, Response: decode(m["content"])})}
+		return msg
+	}
+	if content := stringOf(m["content"]); content != "" {
+		msg.Parts = append(msg.Parts, part(textPart(content)))
+	}
+	msg.Parts = append(msg.Parts, openInferenceContents(m)...)
+	for _, call := range calls {
+		if c, ok := call["tool_call"].(map[string]any); ok {
+			msg.Parts = append(msg.Parts, openInferenceToolCall(c))
+		}
+	}
+	if name := stringOf(m["function_call_name"]); name != "" {
+		msg.Parts = append(msg.Parts, part(ToolCallRequestPart{Type: PartTypeToolCall, Name: name, Arguments: decode(m["function_call_arguments_json"])}))
+	}
+	return msg
+}
+
+// openInferenceContents converts the ordered message.contents items. A text item may come without
+// message_content.type, as the collector-contrib normalizer reads it:
+// https://github.com/open-telemetry/opentelemetry-collector-contrib/blob/fc79b95297776dec6c47490cd6f6026b942bf6fb/processor/genainormalizerprocessor/internal/openinference/messages.go#L285
+func openInferenceContents(m map[string]any) Parts {
+	items := indexed(m, "contents")
+	parts := make(Parts, 0, len(items))
+	for _, item := range items {
+		content, _ := item["message_content"].(map[string]any)
+		typ := stringOf(content["type"])
+		if typ == "" && content["text"] != nil {
+			typ = "text"
+		}
+		switch typ {
+		case "text":
+			parts = append(parts, part(textPart(stringOf(content["text"]))))
+		case "reasoning":
+			parts = append(parts, part(ReasoningPart{Type: PartTypeReasoning, Content: stringOf(content["text"])}))
+		case "image", "audio", "video":
+			if url := stringOf(attribute(content, typ+"."+typ+".url")); url != "" {
+				parts = append(parts, part(mediaPart(url, Modality{valuer.NewString(typ)})))
+			}
+		case "tool_use":
+		default:
+			parts = append(parts, part(GenericPart(content)))
+		}
+	}
+	return parts
+}
+
+// openInferenceToolCall reads a tool_call.* object, {id, function: {name, arguments}}.
+func openInferenceToolCall(call map[string]any) Part {
+	fn, _ := call["function"].(map[string]any)
+	p := ToolCallRequestPart{Type: PartTypeToolCall, Name: stringOf(fn["name"]), ID: optionalString(call["id"])}
+	if args, ok := fn["arguments"]; ok {
+		p.Arguments = decode(args)
+	}
+	return part(p)
+}

pkg/types/aiobservabilitytypes/message_format_openinference_test.go

@@ -0,0 +1,324 @@
+package aiobservabilitytypes
+
+import "testing"
+
+// Captured cases are the llm.input_messages.N.* and llm.output_messages.N.* keys as the collector
+// stored them for the SigNoz/scripts static-telemetry-generator captures.
+func TestFormat_OpenInference(t *testing.T) {
+	runFormatCases(t, []formatCase{
+		{
+			// openinference capture, ChatCompletion span: the flattened keys win over the raw request the
+			// mapper moved onto gen_ai.input.messages
+			name: "FlattenedMessages_RawRequestIgnored",
+			attributes: map[string]any{
+				"llm.input_messages.0.message.role":     "system",
+				"llm.input_messages.0.message.content":  "You are concise.",
+				"llm.input_messages.1.message.role":     "user",
+				"llm.input_messages.1.message.content":  "Say hello in five words.",
+				"llm.output_messages.0.message.role":    "assistant",
+				"llm.output_messages.0.message.content": "Hello! How are you today?",
+				"llm.finish_reason":                     "stop",
+				GenAIInputMessages:                      `{"messages": [{"role": "system", "content": "You are concise."}, {"role": "user", "content": "Say hello in five words."}], "model": "gpt-4o-mini", "max_tokens": 50, "temperature": 0.2}`,
+			},
+			wantInput: `[
+			  {"role": "system", "parts": [{"type": "text", "content": "You are concise."}]},
+			  {"role": "user", "parts": [{"type": "text", "content": "Say hello in five words."}]}
+			]`,
+			wantOutput: `[{"role": "assistant", "parts": [{"type": "text", "content": "Hello! How are you today?"}], "finish_reason": "stop"}]`,
+		},
+		{
+			// openinference capture, ChatCompletion span: message.tool_calls and llm.finish_reason
+			// https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/spec/tool_calling.md#L37
+			name: "ToolCalls_ToolCallParts_FinishReasonFromLLMKey",
+			attributes: map[string]any{
+				"llm.input_messages.0.message.role":                                       "user",
+				"llm.input_messages.0.message.content":                                    "What is the weather in Bengaluru in celsius?",
+				"llm.output_messages.0.message.role":                                      "assistant",
+				"llm.output_messages.0.message.tool_calls.0.tool_call.id":                 "call_MdYKTRrkshuBG9W5oBXYZaRc",
+				"llm.output_messages.0.message.tool_calls.0.tool_call.function.name":      "get_current_weather",
+				"llm.output_messages.0.message.tool_calls.0.tool_call.function.arguments": `{"city":"Bengaluru","unit":"c"}`,
+				"llm.finish_reason": "tool_calls",
+			},
+			wantInput:  `[{"role": "user", "parts": [{"type": "text", "content": "What is the weather in Bengaluru in celsius?"}]}]`,
+			wantOutput: `[{"role": "assistant", "parts": [{"type": "tool_call", "name": "get_current_weather", "id": "call_MdYKTRrkshuBG9W5oBXYZaRc", "arguments": {"city": "Bengaluru", "unit": "c"}}], "finish_reason": "tool_call"}]`,
+		},
+		{
+			// openinference capture, the follow-up ChatCompletion span: tool role messages carry the result
+			// in message.content with their tool_call_id
+			// https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/spec/llm_spans.md#L50
+			name: "ToolRoleMessages_ToolCallResponseParts",
+			attributes: map[string]any{
+				"llm.input_messages.0.message.role":                                      "user",
+				"llm.input_messages.0.message.content":                                   "What's the weather in Paris and in London? Use the tool for each.",
+				"llm.input_messages.1.message.role":                                      "assistant",
+				"llm.input_messages.1.message.tool_calls.0.tool_call.id":                 "call_SW4hst7nzysgvWSAz4OdyK9D",
+				"llm.input_messages.1.message.tool_calls.0.tool_call.function.name":      "get_current_weather",
+				"llm.input_messages.1.message.tool_calls.0.tool_call.function.arguments": `{"city": "Paris"}`,
+				"llm.input_messages.1.message.tool_calls.1.tool_call.id":                 "call_si8wiETYGpooMB7gMXSgMQpk",
+				"llm.input_messages.1.message.tool_calls.1.tool_call.function.name":      "get_current_weather",
+				"llm.input_messages.1.message.tool_calls.1.tool_call.function.arguments": `{"city": "London"}`,
+				"llm.input_messages.2.message.role":                                      "tool",
+				"llm.input_messages.2.message.tool_call_id":                              "call_SW4hst7nzysgvWSAz4OdyK9D",
+				"llm.input_messages.2.message.content":                                   `{"city": "Paris", "temp_c": 18, "summary": "Clear"}`,
+				"llm.input_messages.3.message.role":                                      "tool",
+				"llm.input_messages.3.message.tool_call_id":                              "call_si8wiETYGpooMB7gMXSgMQpk",
+				"llm.input_messages.3.message.content":                                   `{"city": "London", "temp_c": 18, "summary": "Clear"}`,
+				"llm.output_messages.0.message.role":                                     "assistant",
+				"llm.output_messages.0.message.content":                                  "The current weather is as follows:\n\n- **Paris**: 18°C, Clear\n- **London**: 18°C, Clear",
+				"llm.finish_reason":                                                      "stop",
+			},
+			wantInput: `[
+			  {"role": "user", "parts": [{"type": "text", "content": "What's the weather in Paris and in London? Use the tool for each."}]},
+			  {"role": "assistant", "parts": [
+			    {"type": "tool_call", "name": "get_current_weather", "id": "call_SW4hst7nzysgvWSAz4OdyK9D", "arguments": {"city": "Paris"}},
+			    {"type": "tool_call", "name": "get_current_weather", "id": "call_si8wiETYGpooMB7gMXSgMQpk", "arguments": {"city": "London"}}
+			  ]},
+			  {"role": "tool", "parts": [{"type": "tool_call_response", "id": "call_SW4hst7nzysgvWSAz4OdyK9D", "response": {"city": "Paris", "temp_c": 18, "summary": "Clear"}}]},
+			  {"role": "tool", "parts": [{"type": "tool_call_response", "id": "call_si8wiETYGpooMB7gMXSgMQpk", "response": {"city": "London", "temp_c": 18, "summary": "Clear"}}]}
+			]`,
+			wantOutput: `[{"role": "assistant", "parts": [{"type": "text", "content": "The current weather is as follows:\n\n- **Paris**: 18°C, Clear\n- **London**: 18°C, Clear"}], "finish_reason": "stop"}]`,
+		},
+		{
+			// openinference capture, vision call; base64 trimmed. An image as a data URL:
+			// https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/spec/multimodal_attributes.md#L47
+			name: "ContentsTextAndImage_TextAndBlobParts",
+			attributes: map[string]any{
+				"llm.input_messages.0.message.role":                                       "user",
+				"llm.input_messages.0.message.contents.0.message_content.type":            "text",
+				"llm.input_messages.0.message.contents.0.message_content.text":            "What animal is in this image? One sentence.",
+				"llm.input_messages.0.message.contents.1.message_content.type":            "image",
+				"llm.input_messages.0.message.contents.1.message_content.image.image.url": "data:image/jpeg;base64,/9j/4AAQSkZJRgABAQAAAQABAAD/2wBD",
+				"llm.output_messages.0.message.role":                                      "assistant",
+				"llm.output_messages.0.message.content":                                   "The animal in the image is a duck.",
+				"llm.finish_reason":                                                       "stop",
+			},
+			wantInput: `[{"role": "user", "parts": [
+			  {"type": "text", "content": "What animal is in this image? One sentence."},
+			  {"type": "blob", "modality": "image", "mime_type": "image/jpeg", "content": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBD"}
+			]}]`,
+			wantOutput: `[{"role": "assistant", "parts": [{"type": "text", "content": "The animal in the image is a duck."}], "finish_reason": "stop"}]`,
+		},
+		{
+			// openinference-openai-agents capture, Response span: one output message per item, an empty one
+			// for a reasoning item without a summary
+			name: "ResponsesOutputItems_AssistantMessagesFolded_EmptyDropped",
+			attributes: map[string]any{
+				"llm.input_messages.0.message.role":                                  "system",
+				"llm.input_messages.0.message.content":                               "You solve problems step by step and give a final answer.",
+				"llm.input_messages.1.message.role":                                  "user",
+				"llm.input_messages.1.message.content":                               "A bat and a ball cost $1.10. How much is the ball?",
+				"llm.output_messages.0.message.role":                                 "assistant",
+				"llm.output_messages.1.message.role":                                 "assistant",
+				"llm.output_messages.1.message.contents.0.message_content.type":      "text",
+				"llm.output_messages.1.message.contents.0.message_content.text":      "Final Answer: The ball costs $0.05.",
+				"llm.output_messages.2.message.role":                                 "assistant",
+				"llm.output_messages.2.message.tool_calls.0.tool_call.id":            "call_1",
+				"llm.output_messages.2.message.tool_calls.0.tool_call.function.name": "done",
+			},
+			wantInput: `[
+			  {"role": "system", "parts": [{"type": "text", "content": "You solve problems step by step and give a final answer."}]},
+			  {"role": "user", "parts": [{"type": "text", "content": "A bat and a ball cost $1.10. How much is the ball?"}]}
+			]`,
+			wantOutput: `[{"role": "assistant", "parts": [
+			  {"type": "text", "content": "Final Answer: The ball costs $0.05."},
+			  {"type": "tool_call", "name": "done", "id": "call_1"}
+			]}]`,
+		},
+		{
+			// openinference capture, the ChatCompletion span of a failed call
+			name: "InputOnly_NoOutput",
+			attributes: map[string]any{
+				"llm.input_messages.0.message.role":    "user",
+				"llm.input_messages.0.message.content": "this should fail",
+			},
+			wantInput:  `[{"role": "user", "parts": [{"type": "text", "content": "this should fail"}]}]`,
+			wantOutput: `[]`,
+		},
+		{
+			// the strands processor keeps the SDK's gen_ai.system_instructions on the span and inserts the same
+			// prompt as message 0
+			// https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/python/instrumentation/openinference-instrumentation-strands-agents/src/openinference/instrumentation/strands_agents/processor.py#L152
+			// https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/python/instrumentation/openinference-instrumentation-strands-agents/src/openinference/instrumentation/strands_agents/processor.py#L293
+			name: "StrandsSystemInstructionsRepeated_OneSystemMessage",
+			attributes: map[string]any{
+				"llm.input_messages.0.message.role":     "system",
+				"llm.input_messages.0.message.content":  "You are concise.",
+				"llm.input_messages.1.message.role":     "user",
+				"llm.input_messages.1.message.content":  "hi",
+				"llm.output_messages.0.message.role":    "assistant",
+				"llm.output_messages.0.message.content": "Hello.",
+				GenAISystemInstructions:                 `[{"type": "text", "content": "You are concise."}]`,
+			},
+			wantInput: `[
+			  {"role": "system", "parts": [{"type": "text", "content": "You are concise."}]},
+			  {"role": "user", "parts": [{"type": "text", "content": "hi"}]}
+			]`,
+			wantOutput: `[{"role": "assistant", "parts": [{"type": "text", "content": "Hello."}]}]`,
+		},
+		{
+			// pipecat writes the request as llm.input_messages and the streamed reply as output.value alone,
+			// which the mapper copies onto gen_ai.output.messages
+			// https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/python/instrumentation/openinference-instrumentation-pipecat/src/openinference/instrumentation/pipecat/_attributes.py#L430
+			// https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/python/instrumentation/openinference-instrumentation-pipecat/src/openinference/instrumentation/pipecat/_observer.py#L628
+			name: "OutputKeysAbsent_RawOutputFormatted_Warns",
+			attributes: map[string]any{
+				"llm.input_messages.0.message.role":    "user",
+				"llm.input_messages.0.message.content": "hi",
+				GenAIOutputMessages:                    "Hello there.",
+			},
+			warnings:   []string{"bare assistant text, role assumed", "output formatted as text, input as openinference"},
+			wantInput:  `[{"role": "user", "parts": [{"type": "text", "content": "hi"}]}]`,
+			wantOutput: `[{"role": "assistant", "parts": [{"type": "text", "content": "Hello there."}]}]`,
+		},
+		{
+			// message.role left unset by some instrumentors: tool_call_id, tool_calls or neither decide it,
+			// as the collector-contrib normalizer does
+			// https://github.com/open-telemetry/opentelemetry-collector-contrib/blob/fc79b95297776dec6c47490cd6f6026b942bf6fb/processor/genainormalizerprocessor/internal/openinference/messages.go#L473
+			name: "MissingRole_InferredFromToolFields",
+			attributes: map[string]any{
+				"llm.input_messages.0.message.content":                                   "Weather in Oslo?",
+				"llm.input_messages.1.message.tool_calls.0.tool_call.id":                 "call_9",
+				"llm.input_messages.1.message.tool_calls.0.tool_call.function.name":      "get_weather",
+				"llm.input_messages.1.message.tool_calls.0.tool_call.function.arguments": `{"city":"Oslo"}`,
+				"llm.input_messages.2.message.tool_call_id":                              "call_9",
+				"llm.input_messages.2.message.content":                                   "4C, cloudy",
+			},
+			wantInput: `[
+			  {"role": "user", "parts": [{"type": "text", "content": "Weather in Oslo?"}]},
+			  {"role": "assistant", "parts": [{"type": "tool_call", "name": "get_weather", "id": "call_9", "arguments": {"city": "Oslo"}}]},
+			  {"role": "tool", "parts": [{"type": "tool_call_response", "id": "call_9", "response": "4C, cloudy"}]}
+			]`,
+			wantOutput: `[]`,
+		},
+		{
+			// message_content.type left unset on text items by some instrumentors, as the collector-contrib
+			// normalizer reads them
+			// https://github.com/open-telemetry/opentelemetry-collector-contrib/blob/fc79b95297776dec6c47490cd6f6026b942bf6fb/processor/genainormalizerprocessor/internal/openinference/messages.go#L285
+			name: "UntypedContentWithText_TextPart",
+			attributes: map[string]any{
+				"llm.input_messages.0.message.role":                            "user",
+				"llm.input_messages.0.message.contents.0.message_content.text": "Summarise this.",
+				"llm.input_messages.0.message.contents.1.message_content.text": "Second block.",
+			},
+			wantInput:  `[{"role": "user", "parts": [{"type": "text", "content": "Summarise this."}, {"type": "text", "content": "Second block."}]}]`,
+			wantOutput: `[]`,
+		},
+		{
+			// an image given by URL
+			// https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/spec/multimodal_attributes.md#L42
+			name: "ImageURL_UriPart",
+			attributes: map[string]any{
+				"llm.input_messages.0.message.role":                                       "user",
+				"llm.input_messages.0.message.contents.0.message_content.type":            "image",
+				"llm.input_messages.0.message.contents.0.message_content.image.image.url": "https://example.com/duck.jpg",
+			},
+			wantInput:  `[{"role": "user", "parts": [{"type": "uri", "modality": "image", "uri": "https://example.com/duck.jpg"}]}]`,
+			wantOutput: `[]`,
+		},
+		{
+			// audio by URL with its transcript, or as the data URI built from a Chat Completions input_audio
+			// block; the transcript has no part to go to
+			// https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/spec/multimodal_attributes.md#L55
+			name: "Audio_UriAndBlobParts_TranscriptDropped",
+			attributes: map[string]any{
+				"llm.input_messages.0.message.role":                                              "user",
+				"llm.input_messages.0.message.contents.0.message_content.type":                   "audio",
+				"llm.input_messages.0.message.contents.0.message_content.audio.audio.url":        "https://example.com/hello.mp3",
+				"llm.input_messages.0.message.contents.0.message_content.audio.audio.transcript": "Hello, how are you?",
+				"llm.input_messages.0.message.contents.1.message_content.type":                   "audio",
+				"llm.input_messages.0.message.contents.1.message_content.audio.audio.url":        "data:audio/wav;base64,UklGRg==",
+			},
+			wantInput: `[{"role": "user", "parts": [
+			  {"type": "uri", "modality": "audio", "uri": "https://example.com/hello.mp3"},
+			  {"type": "blob", "modality": "audio", "mime_type": "audio/wav", "content": "UklGRg=="}
+			]}]`,
+			wantOutput: `[]`,
+		},
+		{
+			// reasoning items; id, signature, data and encrypted_content have no field on the reasoning part
+			// https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/spec/llm_spans.md#L181
+			name: "Reasoning_ReasoningParts_ProviderKeysDropped",
+			attributes: map[string]any{
+				"llm.output_messages.0.message.role":                                         "assistant",
+				"llm.output_messages.0.message.contents.0.message_content.type":              "reasoning",
+				"llm.output_messages.0.message.contents.0.message_content.text":              "The ball must cost less than a dollar.",
+				"llm.output_messages.0.message.contents.0.message_content.signature":         "EuYBCkQYAiJA",
+				"llm.output_messages.0.message.contents.1.message_content.type":              "reasoning",
+				"llm.output_messages.0.message.contents.1.message_content.data":              "EmwKAhgBEgy3",
+				"llm.output_messages.0.message.contents.2.message_content.type":              "reasoning",
+				"llm.output_messages.0.message.contents.2.message_content.id":                "rs_abc123",
+				"llm.output_messages.0.message.contents.2.message_content.encrypted_content": "gAAAAA==",
+				"llm.output_messages.0.message.contents.3.message_content.type":              "text",
+				"llm.output_messages.0.message.contents.3.message_content.text":              "The ball costs 5 cents.",
+			},
+			wantInput: `[]`,
+			wantOutput: `[{"role": "assistant", "parts": [
+			  {"type": "reasoning", "content": "The ball must cost less than a dollar."},
+			  {"type": "reasoning", "content": ""},
+			  {"type": "reasoning", "content": ""},
+			  {"type": "text", "content": "The ball costs 5 cents."}
+			]}]`,
+		},
+		{
+			// a tool call repeated as an ordered tool_use content item; message.tool_calls stays the one
+			// source. llm.finish_reason is Anthropic's stop_reason as the instrumentation passes it through
+			// https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/spec/tool_calling.md#L63
+			// https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/python/instrumentation/openinference-instrumentation-anthropic/src/openinference/instrumentation/anthropic/_wrappers.py#L938
+			name: "ToolUseBesideToolCalls_OneToolCallPart",
+			attributes: map[string]any{
+				"llm.output_messages.0.message.role":                                      "assistant",
+				"llm.output_messages.0.message.contents.0.message_content.type":           "text",
+				"llm.output_messages.0.message.contents.0.message_content.text":           "Let me check.",
+				"llm.output_messages.0.message.contents.1.message_content.type":           "tool_use",
+				"llm.output_messages.0.message.contents.1.tool_call.id":                   "toolu_01",
+				"llm.output_messages.0.message.contents.1.tool_call.function.name":        "get_weather",
+				"llm.output_messages.0.message.contents.1.tool_call.function.arguments":   `{"city": "Paris"}`,
+				"llm.output_messages.0.message.tool_calls.0.tool_call.id":                 "toolu_01",
+				"llm.output_messages.0.message.tool_calls.0.tool_call.function.name":      "get_weather",
+				"llm.output_messages.0.message.tool_calls.0.tool_call.function.arguments": `{"city": "Paris"}`,
+				"llm.finish_reason": "tool_use",
+			},
+			wantInput: `[]`,
+			wantOutput: `[{"role": "assistant", "parts": [
+			  {"type": "text", "content": "Let me check."},
+			  {"type": "tool_call", "name": "get_weather", "id": "toolu_01", "arguments": {"city": "Paris"}}
+			], "finish_reason": "tool_call"}]`,
+		},
+		{
+			// a content type the spec does not name is kept as sent
+			name: "UnknownContentType_GenericPart",
+			attributes: map[string]any{
+				"llm.input_messages.0.message.role":                                            "user",
+				"llm.input_messages.0.message.contents.0.message_content.type":                 "document",
+				"llm.input_messages.0.message.contents.0.message_content.document.document.id": "file_1",
+			},
+			wantInput:  `[{"role": "user", "parts": [{"type": "document", "document": {"document": {"id": "file_1"}}}]}]`,
+			wantOutput: `[]`,
+		},
+		{
+			// the legacy function_call keys
+			// https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/spec/semantic_conventions.md#L111
+			name: "LegacyFunctionCall_ToolCallPart",
+			attributes: map[string]any{
+				"llm.output_messages.0.message.role":                         "assistant",
+				"llm.output_messages.0.message.function_call_name":           "multiply",
+				"llm.output_messages.0.message.function_call_arguments_json": `{"x": 2}`,
+			},
+			wantInput:  `[]`,
+			wantOutput: `[{"role": "assistant", "parts": [{"type": "tool_call", "name": "multiply", "arguments": {"x": 2}}]}]`,
+		},
+		{
+			// a tool result with message.name naming the function
+			// https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/spec/llm_spans.md#L50
+			name: "ToolResultWithName_NameKept",
+			attributes: map[string]any{
+				"llm.input_messages.0.message.role":         "tool",
+				"llm.input_messages.0.message.name":         "multiply",
+				"llm.input_messages.0.message.tool_call_id": "call_62136355",
+				"llm.input_messages.0.message.content":      "2001",
+			},
+			wantInput:  `[{"role": "tool", "name": "multiply", "parts": [{"type": "tool_call_response", "id": "call_62136355", "response": "2001"}]}]`,
+			wantOutput: `[]`,
+		},
+	})
+}

pkg/types/aiobservabilitytypes/message_format_semconv.go

@@ -6,15 +6,20 @@ import (
 )
 
 // Format: the OTel shape itself, [{role, parts}].
-// Written by: litellm, langchain, openllmetry, portkey, traceloop LLM spans.
-// Decoding is the genai types. Repairs: "text" used for "content" on text parts; OpenAI content
-// blocks left inside parts.
-// Not yet: "thinking" for "reasoning"; "result" or "output" for "response" on tool call responses.
+// Written by: litellm, langchain, openllmetry, portkey, traceloop LLM spans, pydantic-ai.
+// Decoding is the genai types. Repairs: "text" used for "content" on text parts; OpenAI content blocks
+// left inside parts; tool call arguments as a JSON string and "result" for "response", as pydantic-ai
+// writes them:
+// https://github.com/pydantic/pydantic-ai/blob/9ac853bb84ffed7ba0edccb026c5b4f3228c1de7/pydantic_ai_slim/pydantic_ai/messages.py#L3130
+// https://github.com/pydantic/pydantic-ai/blob/9ac853bb84ffed7ba0edccb026c5b4f3228c1de7/pydantic_ai_slim/pydantic_ai/messages.py#L1724
+// The choice finish_reason as the provider spells it, as litellm copies it:
+// https://github.com/BerriAI/litellm/blob/9d72bae336ec0d14c41651adcb0d2e1149721f46/litellm/integrations/opentelemetry.py#L2822
+// Not yet: "thinking" for "reasoning".
 
 // semconv decodes messages that already follow the schema through the genai types. A part of a
 // type the schema does not name is a provider content block, which contentPart converts.
 func (f *formatting) semconv(list []any) OutputMessages {
-	data, err := json.Marshal(semconvTextKey(list))
+	data, err := json.Marshal(semconvRepairs(list))
 	if err != nil {
 		f.warn("messages could not be encoded: %s", err)
 		return OutputMessages{genericMessage(list)}
@@ -28,6 +33,9 @@ func (f *formatting) semconv(list []any) OutputMessages {
 		if msgs[i].Parts == nil {
 			msgs[i].Parts = Parts{}
 		}
+		if msgs[i].FinishReason != nil {
+			msgs[i].FinishReason = finishReason(msgs[i].FinishReason.StringValue())
+		}
 		for j, p := range msgs[i].Parts {
 			if block, ok := p.Value.(GenericPart); ok {
 				msgs[i].Parts[j] = contentPart(map[string]any(block))
@@ -37,9 +45,9 @@ func (f *formatting) semconv(list []any) OutputMessages {
 	return msgs
 }
 
-// semconvTextKey copies the text and reasoning parts some SDKs write with "text" in place of the
-// schema's "content". The input is left untouched, it is also the span's raw attribute.
-func semconvTextKey(list []any) []any {
+// semconvRepairs copies the parts whose keys SDKs misspell into the schema's spelling. The input
+// is left untouched, it is also the span's raw attribute.
+func semconvRepairs(list []any) []any {
 	out := make([]any, 0, len(list))
 	for _, item := range list {
 		m, ok := item.(map[string]any)
@@ -50,13 +58,8 @@ func semconvTextKey(list []any) []any {
 		}
 		fixedParts := make([]any, 0, len(parts))
 		for _, p := range parts {
-			part, ok := p.(map[string]any)
-			typ := stringOf(part["type"])
-			if ok && (typ == "text" || typ == "reasoning") && part["content"] == nil && part["text"] != nil {
-				fixed := maps.Clone(part)
-				fixed["content"] = fixed["text"]
-				delete(fixed, "text")
-				p = fixed
+			if part, ok := p.(map[string]any); ok {
+				p = semconvRepairPart(part)
 			}
 			fixedParts = append(fixedParts, p)
 		}
@@ -66,3 +69,24 @@ func semconvTextKey(list []any) []any {
 	}
 	return out
 }
+
+func semconvRepairPart(part map[string]any) map[string]any {
+	fixed := maps.Clone(part)
+	switch stringOf(part["type"]) {
+	case "text", "reasoning":
+		if part["content"] == nil && part["text"] != nil {
+			fixed["content"] = fixed["text"]
+			delete(fixed, "text")
+		}
+	case "tool_call":
+		if args, ok := part["arguments"].(string); ok {
+			fixed["arguments"] = decode(args)
+		}
+	case "tool_call_response":
+		if _, ok := part["response"]; !ok && part["result"] != nil {
+			fixed["response"] = part["result"]
+			delete(fixed, "result")
+		}
+	}
+	return fixed
+}

pkg/types/aiobservabilitytypes/message_format_semconv_test.go

@@ -157,3 +157,34 @@ func TestFormat_Semconv(t *testing.T) {
 		},
 	})
 }
+
+func TestFormat_Semconv_Repairs(t *testing.T) {
+	runFormatCases(t, []formatCase{
+		{
+			// pydantic-ai keeps string tool call arguments as sent, writes the tool result under result, and the
+			// instructions as a gen_ai.system_instructions text part:
+			// https://github.com/pydantic/pydantic-ai/blob/9ac853bb84ffed7ba0edccb026c5b4f3228c1de7/pydantic_ai_slim/pydantic_ai/messages.py#L3130
+			// https://github.com/pydantic/pydantic-ai/blob/9ac853bb84ffed7ba0edccb026c5b4f3228c1de7/pydantic_ai_slim/pydantic_ai/messages.py#L1724
+			// https://github.com/pydantic/pydantic-ai/blob/9ac853bb84ffed7ba0edccb026c5b4f3228c1de7/pydantic_ai_slim/pydantic_ai/models/instrumented.py#L292
+			name:       "PydanticAI_ArgumentsDecoded_ResultBecomesResponse_InstructionsLead",
+			input:      `[{"role":"user","parts":[{"type":"text","content":"Weather in Paris?"}]},{"role":"assistant","parts":[{"type":"tool_call","id":"call_stub1","name":"get_weather","arguments":"{\"city\": \"Paris\"}"}],"finish_reason":"tool_call"},{"role":"user","parts":[{"type":"tool_call_response","id":"call_stub1","name":"get_weather","result":{"city":"Paris","temp_c":18}}]}]`,
+			output:     `[{"role":"assistant","parts":[{"type":"text","content":"Paris is clear at 18C."}],"finish_reason":"stop"}]`,
+			attributes: map[string]any{GenAISystemInstructions: `[{"type":"text","content":"Use the tool."}]`},
+			wantInput: `[
+			  {"role": "system", "parts": [{"type": "text", "content": "Use the tool."}]},
+			  {"role": "user", "parts": [{"type": "text", "content": "Weather in Paris?"}]},
+			  {"role": "assistant", "parts": [{"type": "tool_call", "name": "get_weather", "id": "call_stub1", "arguments": {"city": "Paris"}}]},
+			  {"role": "user", "parts": [{"type": "tool_call_response", "id": "call_stub1", "response": {"city": "Paris", "temp_c": 18}}]}
+			]`,
+			wantOutput: `[{"role": "assistant", "parts": [{"type": "text", "content": "Paris is clear at 18C."}], "finish_reason": "stop"}]`,
+		},
+		{
+			// litellm copies the choice finish_reason into the semconv message as OpenAI spells it:
+			// https://github.com/BerriAI/litellm/blob/9d72bae336ec0d14c41651adcb0d2e1149721f46/litellm/integrations/opentelemetry.py#L2822
+			name:       "LiteLLMFinishReasonSpelling_Normalised",
+			output:     `[{"role": "assistant", "parts": [], "finish_reason": "tool_calls"}]`,
+			wantInput:  `[]`,
+			wantOutput: `[{"role": "assistant", "parts": [], "finish_reason": "tool_call"}]`,
+		},
+	})
+}

pkg/types/aiobservabilitytypes/message_format_test.go

@@ -120,3 +120,108 @@ So, the ball costs 5 cents.`,
 		},
 	})
 }
+
+func TestIndexed(t *testing.T) {
+	testCases := []struct {
+		name string
+		doc  map[string]any
+		want []map[string]any
+	}{
+		{
+			name: "NumericKeys_SortedNumerically",
+			doc:  document(map[string]any{"llm.input_messages.10.message.role": "user", "llm.input_messages.2.message.role": "system", "llm.input_messages.x.message.role": "skipped"}),
+			want: []map[string]any{{"message": map[string]any{"role": "system"}}, {"message": map[string]any{"role": "user"}}},
+		},
+		{
+			name: "StructuredList_KeptInOrder",
+			doc:  map[string]any{"llm": map[string]any{"input_messages": []any{map[string]any{"message": "b"}, map[string]any{"message": "a"}, "not an object"}}},
+			want: []map[string]any{{"message": "b"}, {"message": "a"}},
+		},
+		{name: "Absent_Empty", doc: map[string]any{}, want: []map[string]any{}},
+	}
+	for _, testCase := range testCases {
+		t.Run(testCase.name, func(t *testing.T) {
+			assert.Equal(t, testCase.want, indexed(testCase.doc, openInferenceInputMessages))
+		})
+	}
+}
+
+func TestFormat_SystemInstructions(t *testing.T) {
+	input := `[{"role":"user","parts":[{"type":"text","content":"hi"}]}]`
+	runFormatCases(t, []formatCase{
+		{
+			// the langchain instrumentation moves system messages out of the history into gen_ai.system_instructions,
+			// a JSON list of parts:
+			// https://github.com/traceloop/openllmetry/blob/93429cfb9767e6b51db6cb35f6b70f758117669a/packages/opentelemetry-instrumentation-langchain/opentelemetry/instrumentation/langchain/span_utils.py#L304
+			name:       "ListOfParts_LeadsInput",
+			input:      input,
+			attributes: map[string]any{GenAISystemInstructions: `[{"type": "text", "content": "You are concise."}]`},
+			wantInput: `[
+			  {"role": "system", "parts": [{"type": "text", "content": "You are concise."}]},
+			  {"role": "user", "parts": [{"type": "text", "content": "hi"}]}
+			]`,
+			wantOutput: `[]`,
+		},
+		{
+			// the OpenInference strands processor keeps gen_ai.system_instructions on the span and inserts the
+			// same prompt as message 0
+			// https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/python/instrumentation/openinference-instrumentation-strands-agents/src/openinference/instrumentation/strands_agents/processor.py#L152
+			// https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/python/instrumentation/openinference-instrumentation-strands-agents/src/openinference/instrumentation/strands_agents/processor.py#L293
+			name:       "SameTextLeadsInput_Skipped",
+			input:      `[{"role":"system","parts":[{"type":"text","content":"You are concise."}]},{"role":"user","parts":[{"type":"text","content":"hi"}]}]`,
+			attributes: map[string]any{GenAISystemInstructions: `[{"type": "text", "content": "You are concise."}]`},
+			wantInput: `[
+			  {"role": "system", "parts": [{"type": "text", "content": "You are concise."}]},
+			  {"role": "user", "parts": [{"type": "text", "content": "hi"}]}
+			]`,
+			wantOutput: `[]`,
+		},
+		{
+			name:       "BareString_TextPart",
+			input:      input,
+			attributes: map[string]any{GenAISystemInstructions: "be brief"},
+			wantInput: `[
+			  {"role": "system", "parts": [{"type": "text", "content": "be brief"}]},
+			  {"role": "user", "parts": [{"type": "text", "content": "hi"}]}
+			]`,
+			wantOutput: `[]`,
+		},
+		{
+			// a create_agent span carries instructions alone
+			name:       "NoMessages_Nothing",
+			attributes: map[string]any{GenAISystemInstructions: "be brief"},
+			wantInput:  `[]`,
+			wantOutput: `[]`,
+		},
+	})
+}
+
+func TestFormat_FinishReasons(t *testing.T) {
+	output := `[{"role":"assistant","parts":[{"type":"text","content":"a"}],"finish_reason":"stop"},{"role":"assistant","parts":[{"type":"text","content":"b"}]}]`
+	runFormatCases(t, []formatCase{
+		{
+			name:       "List_OverridesMessages",
+			output:     output,
+			attributes: map[string]any{GenAIResponseFinishReasons: []any{"tool_calls", "length"}},
+			wantInput:  `[]`,
+			wantOutput: `[{"role": "assistant", "parts": [{"type": "text", "content": "a"}], "finish_reason": "tool_call"}, {"role": "assistant", "parts": [{"type": "text", "content": "b"}], "finish_reason": "length"}]`,
+		},
+		{
+			name:       "LengthMismatch_LeavesMessages_Warns",
+			output:     output,
+			attributes: map[string]any{GenAIResponseFinishReasons: []any{"stop"}},
+			warnings:   []string{"gen_ai.response.finish_reasons has 1 entries for 2 output messages"},
+			wantInput:  `[]`,
+			wantOutput: `[{"role": "assistant", "parts": [{"type": "text", "content": "a"}], "finish_reason": "stop"}, {"role": "assistant", "parts": [{"type": "text", "content": "b"}]}]`,
+		},
+		{
+			// llm.finish_reason copied by the mapper as a bare string, in the Vercel AI SDK spelling:
+			// https://github.com/vercel/ai/blob/a5e92fe3c692f8e152fe1eb9ef50a4e734ad0ffb/packages/provider/src/language-model/v2/language-model-v2-finish-reason.ts
+			name:       "BareString_SingleChoice",
+			output:     `{"choices":[{"finish_reason":"stop","message":{"role":"assistant","content":"hi"}}]}`,
+			attributes: map[string]any{GenAIResponseFinishReasons: "tool-calls"},
+			wantInput:  `[]`,
+			wantOutput: `[{"role": "assistant", "parts": [{"type": "text", "content": "hi"}], "finish_reason": "tool_call"}]`,
+		},
+	})
+}

pkg/types/aiobservabilitytypes/message_format_tool.go

@@ -1,12 +1,15 @@
 package aiobservabilitytypes
 
 // Format: gen_ai.tool.call.arguments and gen_ai.tool.call.result on an execute_tool span, each an
-// object or its JSON string (open-telemetry/semantic-conventions-genai@06ec68e, model/gen-ai/registry.yaml).
+// object or its JSON string:
+// https://github.com/open-telemetry/semantic-conventions-genai/blob/06ec68e722c45a7218e23ea1bc1339fe4e21ecae/model/gen-ai/registry.yaml#L545
 // Written by: opentelemetry-instrumentation-langchain, traceloop @tool, OpenInference TOOL spans and
 // Vercel ai.toolCall.* through the span mapper.
-// Envelopes unwrapped: LangChain's callback payloads, {input_str, tags, metadata, inputs, kwargs} and
-// {output, kwargs} (openllmetry callback_handler.py on_tool_start/on_tool_end), and the traceloop
-// decorator's {args, kwargs} (traceloop-sdk decorators/base.py _handle_span_input).
+// Envelopes unwrapped: LangChain's on_tool_start and on_tool_end payloads, {input_str, tags, metadata,
+// inputs, kwargs} and {output, kwargs}:
+// https://github.com/traceloop/openllmetry/blob/93429cfb9767e6b51db6cb35f6b70f758117669a/packages/opentelemetry-instrumentation-langchain/opentelemetry/instrumentation/langchain/callback_handler.py#L837
+// The traceloop decorator's {args, kwargs}:
+// https://github.com/traceloop/openllmetry/blob/93429cfb9767e6b51db6cb35f6b70f758117669a/packages/traceloop-sdk/traceloop/sdk/decorators/base.py#L169
 
 // tool renders the call as an assistant tool_call and the result as a tool tool_call_response, both
 // carrying gen_ai.tool.call.id so the pair can be matched.

pkg/types/aiobservabilitytypes/message_format_tool_test.go

@@ -7,8 +7,9 @@ import "testing"
 func TestFormat_Tool(t *testing.T) {
 	runFormatCases(t, []formatCase{
 		{
-			// langchain capture, execute_tool get_weather span; opentelemetry-instrumentation-langchain
-			// callback_handler.py on_tool_start/on_tool_end envelopes
+			// gen_ai.tool.call.arguments and gen_ai.tool.call.result written by the langchain instrumentation itself,
+			// wrapped in its on_tool_start and on_tool_end payloads:
+			// https://github.com/traceloop/openllmetry/blob/93429cfb9767e6b51db6cb35f6b70f758117669a/packages/opentelemetry-instrumentation-langchain/opentelemetry/instrumentation/langchain/callback_handler.py#L837
 			name: "LangChainEnvelopes_InputsAndOutputUnwrapped",
 			attributes: map[string]any{
 				GenAIToolName:          "get_weather",
@@ -19,8 +20,9 @@ func TestFormat_Tool(t *testing.T) {
 			wantOutput: `[{"role": "tool", "parts": [{"type": "tool_call_response", "response": {"city": "Bengaluru", "temp_c": 18, "summary": "Clear"}}]}]`,
 		},
 		{
-			// openllmetry capture, get_weather.tool span; traceloop-sdk decorators/base.py
-			// _handle_span_input {args, kwargs} and the result encoded as a JSON string twice
+			// traceloop.entity.input and traceloop.entity.output from the @tool decorator, renamed by the mapper;
+			// the tool returned a JSON string, which json.dumps encoded again:
+			// https://github.com/traceloop/openllmetry/blob/93429cfb9767e6b51db6cb35f6b70f758117669a/packages/traceloop-sdk/traceloop/sdk/decorators/base.py#L169
 			name: "TraceloopDecorator_KwargsUnwrapped_ResultDecodedTwice",
 			attributes: map[string]any{
 				GenAIToolName:          "get_weather",
@@ -31,8 +33,8 @@ func TestFormat_Tool(t *testing.T) {
 			wantOutput: `[{"role": "tool", "parts": [{"type": "tool_call_response", "response": {"city": "Bengaluru", "temp_c": 28, "summary": "Partly cloudy"}}]}]`,
 		},
 		{
-			// a decorated function returning None is json.dumps(None); the openllmetry capture shows it on
-			// workflow spans, with this positional-args input
+			// traceloop.entity.output of a @tool returning None, json.dumps(None), renamed by the mapper:
+			// https://github.com/traceloop/openllmetry/blob/93429cfb9767e6b51db6cb35f6b70f758117669a/packages/traceloop-sdk/traceloop/sdk/decorators/base.py#L185
 			name: "TraceloopDecorator_PositionalArgsOnly_ListKept",
 			attributes: map[string]any{
 				GenAIToolName:          "lookup",
@@ -52,8 +54,8 @@ func TestFormat_Tool(t *testing.T) {
 			wantOutput: `[]`,
 		},
 		{
-			// openinference-openai-agents capture, get_weather TOOL span: input.value and output.value
-			// copied by the mapper (Arize-ai/openinference@12f2cc03, spec/traces.md)
+			// input.value and output.value of an OpenInference TOOL span, renamed by the mapper:
+			// https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/spec/traces.md
 			name: "OpenInferenceToolSpan_PlainJSONAndText",
 			attributes: map[string]any{
 				GenAIToolName:          "get_weather",
@@ -64,8 +66,8 @@ func TestFormat_Tool(t *testing.T) {
 			wantOutput: `[{"role": "tool", "parts": [{"type": "tool_call_response", "response": "27C, partly cloudy, light breeze from the south."}]}]`,
 		},
 		{
-			// vercel capture, ai.toolCall span: ai.toolCall.{id,name,args,result} copied by the mapper
-			// (https://ai-sdk.dev/docs/ai-sdk-core/telemetry)
+			// ai.toolCall.{id,name,args,result} of a Vercel ai.toolCall span, renamed by the mapper:
+			// https://ai-sdk.dev/docs/ai-sdk-core/telemetry
 			name: "VercelToolCall_IDOnBothParts",
 			attributes: map[string]any{
 				GenAIToolName:          "get_current_weather",

pkg/types/aiobservabilitytypes/message_format_traceloop.go

@@ -0,0 +1,87 @@
+package aiobservabilitytypes
+
+// Format: the flattened gen_ai.prompt.N.* and gen_ai.completion.N.* keys of openllmetry's instrumentations
+// from before their semconv rewrite, role, content, tool_call_id, tool_calls.M.{id,name,arguments} and
+// finish_reason, with list content as its JSON:
+// https://github.com/traceloop/openllmetry/blob/93429cfb9767e6b51db6cb35f6b70f758117669a/packages/opentelemetry-semantic-conventions-ai/opentelemetry/semconv_ai/__init__.py#L159
+// https://github.com/traceloop/openllmetry/blob/af487056a35625a2cff1c17ac83c2c0d90a02bca/packages/opentelemetry-instrumentation-openai/opentelemetry/instrumentation/openai/shared/chat_wrappers.py#L408
+// https://github.com/traceloop/openllmetry/blob/af487056a35625a2cff1c17ac83c2c0d90a02bca/packages/opentelemetry-instrumentation-openai/opentelemetry/instrumentation/openai/shared/chat_wrappers.py#L460
+// Written by: crewai through the openai instrumentation 0.45.6, mistralai 0.61.0 (https://github.com/traceloop/openllmetry/blob/4643b8826a6e6a4ce1968178d6ac9ca3ce607e2e/packages/opentelemetry-instrumentation-mistralai/opentelemetry/instrumentation/mistralai/__init__.py#L83),
+// litellm beside its semconv value.
+
+const (
+	traceloopPrompts     = "gen_ai.prompt"
+	traceloopCompletions = "gen_ai.completion"
+)
+
+func (f *formatting) traceloop() bool {
+	prompts := indexed(f.attributes, traceloopPrompts)
+	completions := indexed(f.attributes, traceloopCompletions)
+	if len(prompts) == 0 && len(completions) == 0 {
+		return false
+	}
+	f.out.Input = inputMessages(f.traceloopMessages(prompts))
+	f.out.Output = f.traceloopMessages(completions)
+	return true
+}
+
+// completionToolCalls fills an output message left without parts from the gen_ai.completion.N keys
+// litellm writes beside its semconv value, whose parts never carry the tool call:
+// https://github.com/BerriAI/litellm/blob/9d72bae336ec0d14c41651adcb0d2e1149721f46/litellm/integrations/opentelemetry.py#L2365
+// https://github.com/BerriAI/litellm/blob/9d72bae336ec0d14c41651adcb0d2e1149721f46/litellm/integrations/opentelemetry.py#L2770
+func (f *formatting) completionToolCalls() {
+	for i := range f.out.Output {
+		if len(f.out.Output[i].Parts) == 0 {
+			f.out.Output[i].Parts = append(f.out.Output[i].Parts, traceloopToolCalls(indexedAt(f.attributes, traceloopCompletions, i))...)
+		}
+	}
+}
+
+func (f *formatting) traceloopMessages(buckets []map[string]any) OutputMessages {
+	out := make(OutputMessages, 0, len(buckets))
+	for i, m := range buckets {
+		msg := OutputMessage{Role: role(stringOf(m["role"])), Parts: Parts{}, Name: optionalString(m["name"]), FinishReason: finishReason(stringOf(m["finish_reason"]))}
+		content, hasContent := m["content"]
+		if id := stringOf(m["tool_call_id"]); msg.Role == RoleTool && id != "" {
+			msg.Parts = Parts{part(ToolCallResponsePart{Type: PartTypeToolCallResponse, ID: &id, Response: decode(content)})}
+			out = append(out, msg)
+			continue
+		}
+		if hasContent {
+			msg.Parts = append(msg.Parts, traceloopContent(content)...)
+		}
+		msg.Parts = append(msg.Parts, traceloopToolCalls(m)...)
+		if len(msg.Parts) == 0 {
+			f.warn("message %d has no content", i)
+		}
+		out = append(out, msg)
+	}
+	return out
+}
+
+// traceloopContent reads a content value, a string or the JSON of an OpenAI content block list.
+func traceloopContent(content any) Parts {
+	if blocks, ok := decode(content).([]any); ok {
+		parts := make(Parts, 0, len(blocks))
+		for _, block := range blocks {
+			parts = append(parts, contentPart(block))
+		}
+		return parts
+	}
+	if text := stringOf(content); text != "" {
+		return Parts{part(textPart(text))}
+	}
+	return nil
+}
+
+// traceloopToolCalls reads tool_calls.M.{id,name,arguments} and litellm's function_call object.
+func traceloopToolCalls(m map[string]any) Parts {
+	var parts Parts
+	for _, call := range indexed(m, "tool_calls") {
+		parts = append(parts, part(ToolCallRequestPart{Type: PartTypeToolCall, Name: stringOf(call["name"]), ID: optionalString(call["id"]), Arguments: decode(call["arguments"])}))
+	}
+	if call, ok := m["function_call"].(map[string]any); ok {
+		parts = append(parts, part(ToolCallRequestPart{Type: PartTypeToolCall, Name: stringOf(call["name"]), Arguments: decode(call["arguments"])}))
+	}
+	return parts
+}

pkg/types/aiobservabilitytypes/message_format_traceloop_test.go

@@ -0,0 +1,106 @@
+package aiobservabilitytypes
+
+import "testing"
+
+// Captured cases are the gen_ai.prompt.N.* and gen_ai.completion.N.* keys as the collector stored
+// them for the SigNoz/scripts crewai, mistral and litellm captures.
+func TestFormat_Traceloop(t *testing.T) {
+	runFormatCases(t, []formatCase{
+		{
+			// crewai capture, openai.chat span of a tool round, as the openai instrumentation 0.45.6 wrote
+			// prompts and completions before its semconv rewrite
+			// https://github.com/traceloop/openllmetry/blob/af487056a35625a2cff1c17ac83c2c0d90a02bca/packages/opentelemetry-instrumentation-openai/opentelemetry/instrumentation/openai/shared/chat_wrappers.py#L408
+			name: "ToolRound_ToolCallAndToolMessages",
+			attributes: map[string]any{
+				"gen_ai.prompt.0.role":                   "system",
+				"gen_ai.prompt.0.content":                "You are Weather Reporter. Call get_weather and summarise the result.",
+				"gen_ai.prompt.1.role":                   "user",
+				"gen_ai.prompt.1.content":                "Current Task: What's the weather in Bengaluru?",
+				"gen_ai.prompt.2.role":                   "assistant",
+				"gen_ai.prompt.2.tool_calls.0.id":        "call_bbVgWPe8AMBol7JlAhDkA6R7",
+				"gen_ai.prompt.2.tool_calls.0.name":      "get_weather",
+				"gen_ai.prompt.2.tool_calls.0.arguments": `{"city":"Bengaluru"}`,
+				"gen_ai.prompt.3.role":                   "tool",
+				"gen_ai.prompt.3.tool_call_id":           "call_bbVgWPe8AMBol7JlAhDkA6R7",
+				"gen_ai.prompt.3.content":                `{"city": "Bengaluru", "temp_c": 18, "summary": "Clear"}`,
+				"gen_ai.completion.0.role":               "assistant",
+				"gen_ai.completion.0.finish_reason":      "stop",
+				"gen_ai.completion.0.content":            "The current weather in Bengaluru is clear with a temperature of 18°C.",
+			},
+			wantInput: `[
+			  {"role": "system", "parts": [{"type": "text", "content": "You are Weather Reporter. Call get_weather and summarise the result."}]},
+			  {"role": "user", "parts": [{"type": "text", "content": "Current Task: What's the weather in Bengaluru?"}]},
+			  {"role": "assistant", "parts": [{"type": "tool_call", "name": "get_weather", "id": "call_bbVgWPe8AMBol7JlAhDkA6R7", "arguments": {"city": "Bengaluru"}}]},
+			  {"role": "tool", "parts": [{"type": "tool_call_response", "id": "call_bbVgWPe8AMBol7JlAhDkA6R7", "response": {"city": "Bengaluru", "temp_c": 18, "summary": "Clear"}}]}
+			]`,
+			wantOutput: `[{"role": "assistant", "parts": [{"type": "text", "content": "The current weather in Bengaluru is clear with a temperature of 18°C."}], "finish_reason": "stop"}]`,
+		},
+		{
+			// crewai capture, openai.chat span answering with two tool calls
+			// https://github.com/traceloop/openllmetry/blob/af487056a35625a2cff1c17ac83c2c0d90a02bca/packages/opentelemetry-instrumentation-openai/opentelemetry/instrumentation/openai/shared/chat_wrappers.py#L460
+			name: "CompletionToolCalls_ToolCallParts",
+			attributes: map[string]any{
+				"gen_ai.prompt.0.role":                       "user",
+				"gen_ai.prompt.0.content":                    "What's the weather in Paris and in London?",
+				"gen_ai.completion.0.role":                   "assistant",
+				"gen_ai.completion.0.finish_reason":          "tool_calls",
+				"gen_ai.completion.0.tool_calls.0.id":        "call_1",
+				"gen_ai.completion.0.tool_calls.0.name":      "get_weather",
+				"gen_ai.completion.0.tool_calls.0.arguments": `{"city":"Paris"}`,
+				"gen_ai.completion.0.tool_calls.1.id":        "call_2",
+				"gen_ai.completion.0.tool_calls.1.name":      "get_weather",
+				"gen_ai.completion.0.tool_calls.1.arguments": `{"city":"London"}`,
+			},
+			wantInput: `[{"role": "user", "parts": [{"type": "text", "content": "What's the weather in Paris and in London?"}]}]`,
+			wantOutput: `[{"role": "assistant", "parts": [
+			  {"type": "tool_call", "name": "get_weather", "id": "call_1", "arguments": {"city": "Paris"}},
+			  {"type": "tool_call", "name": "get_weather", "id": "call_2", "arguments": {"city": "London"}}
+			], "finish_reason": "tool_call"}]`,
+		},
+		{
+			// mistral capture, mistralai.chat span: the instrumentation writes the role for every message but
+			// drops a content that is not a plain string
+			// https://github.com/traceloop/openllmetry/blob/4643b8826a6e6a4ce1968178d6ac9ca3ce607e2e/packages/opentelemetry-instrumentation-mistralai/opentelemetry/instrumentation/mistralai/__init__.py#L75
+			// https://github.com/traceloop/openllmetry/blob/4643b8826a6e6a4ce1968178d6ac9ca3ce607e2e/packages/opentelemetry-instrumentation-mistralai/opentelemetry/instrumentation/mistralai/__init__.py#L88
+			name: "RoleWithoutContent_EmptyMessageWarns",
+			attributes: map[string]any{
+				"gen_ai.prompt.0.role":              "user",
+				"gen_ai.completion.0.role":          "assistant",
+				"gen_ai.completion.0.finish_reason": "stop",
+				"gen_ai.completion.0.content":       "Observability is the ability to understand a system's internal state by analyzing its external outputs.",
+			},
+			warnings:   []string{"message 0 has no content"},
+			wantInput:  `[{"role": "user", "parts": []}]`,
+			wantOutput: `[{"role": "assistant", "parts": [{"type": "text", "content": "Observability is the ability to understand a system's internal state by analyzing its external outputs."}], "finish_reason": "stop"}]`,
+		},
+		{
+			// a vision prompt: list content is written as its JSON, the OpenAI content block list
+			// https://github.com/traceloop/openllmetry/blob/af487056a35625a2cff1c17ac83c2c0d90a02bca/packages/opentelemetry-instrumentation-openai/opentelemetry/instrumentation/openai/shared/chat_wrappers.py#L431
+			name: "ContentBlockList_TextAndBlobParts",
+			attributes: map[string]any{
+				"gen_ai.prompt.0.role":    "user",
+				"gen_ai.prompt.0.content": `[{"type":"text","text":"What animal is this?"},{"type":"image_url","image_url":{"url":"data:image/jpeg;base64,/9j/4AAQ"}}]`,
+			},
+			wantInput: `[{"role": "user", "parts": [
+			  {"type": "text", "content": "What animal is this?"},
+			  {"type": "blob", "modality": "image", "mime_type": "image/jpeg", "content": "/9j/4AAQ"}
+			]}]`,
+			wantOutput: `[]`,
+		},
+		{
+			// litellm capture, litellm_request span: the semconv output has no part for the tool call, the
+			// gen_ai.completion.0.function_call keys beside it still carry it
+			// https://github.com/BerriAI/litellm/blob/9d72bae336ec0d14c41651adcb0d2e1149721f46/litellm/integrations/opentelemetry.py#L2365
+			// https://github.com/BerriAI/litellm/blob/9d72bae336ec0d14c41651adcb0d2e1149721f46/litellm/integrations/opentelemetry.py#L2770
+			name: "SemconvEmptyParts_ToolCallFromCompletionKeys",
+			attributes: map[string]any{
+				GenAIInputMessages:                            `[{"role": "user", "parts": [{"type": "text", "content": "What is the weather in Bengaluru in celsius?"}]}]`,
+				GenAIOutputMessages:                           `[{"role": "assistant", "parts": [], "finish_reason": "tool_calls"}]`,
+				"gen_ai.completion.0.function_call.name":      "get_current_weather",
+				"gen_ai.completion.0.function_call.arguments": `{"city":"Bengaluru"}`,
+			},
+			wantInput:  `[{"role": "user", "parts": [{"type": "text", "content": "What is the weather in Bengaluru in celsius?"}]}]`,
+			wantOutput: `[{"role": "assistant", "parts": [{"type": "tool_call", "name": "get_current_weather", "arguments": {"city": "Bengaluru"}}], "finish_reason": "tool_call"}]`,
+		},
+	})
+}

tests/integration/tests/tracedetail/02_thread.py

@@ -21,17 +21,25 @@ def test_thread_returns_gen_ai_spans_in_order(
     seed_attribute_evolution("traces", ATTRIBUTE_JSON_ROLLOUT_TIME)
     now = datetime.now(tz=UTC).replace(microsecond=0)
     trace_id = TraceIdGenerator.trace_id()
-    root_id, first_llm_id, tool_id, agent_id, second_llm_id, third_llm_id = (TraceIdGenerator.span_id() for _ in range(6))
+    root_id, first_llm_id, tool_id, agent_id, second_llm_id, legacy_llm_id, third_llm_id, flattened_llm_id = (TraceIdGenerator.span_id() for _ in range(8))
     resources = {"service.name": "tracedetail-thread"}
     first_input = json.dumps([{"role": "user", "parts": [{"type": "text", "content": "weather in Bangalore?"}]}])
-    first_output = json.dumps([{"role": "assistant", "parts": [{"type": "tool_call", "id": "call_1", "name": "get_weather", "arguments": {"city": "Bangalore"}}], "finish_reason": "tool_call"}])
+    first_output = json.dumps([{"role": "assistant", "parts": [{"type": "tool_call", "id": "call_1", "name": "get_weather", "arguments": {"city": "Bangalore"}}]}])
+    first_instructions = json.dumps([{"type": "text", "content": "Answer with the tool."}])
     second_input = json.dumps([{"role": "tool", "content": "sunny", "tool_call_id": "call_1"}])
 
     insert_traces(
         [
             Traces(timestamp=now - timedelta(seconds=10), duration=timedelta(seconds=9), trace_id=trace_id, span_id=root_id, name="POST /chat", kind=TracesKind.SPAN_KIND_SERVER, resources=resources, attribute_write_mode="json_only"),
             Traces(
-                timestamp=now - timedelta(seconds=8), trace_id=trace_id, span_id=first_llm_id, parent_span_id=root_id, name="chat gpt-4o", resources=resources, attributes={"gen_ai.request.model": "gpt-4o", "gen_ai.input.messages": first_input, "gen_ai.output.messages": first_output}, attribute_write_mode="json_only"
+                timestamp=now - timedelta(seconds=8),
+                trace_id=trace_id,
+                span_id=first_llm_id,
+                parent_span_id=root_id,
+                name="chat gpt-4o",
+                resources=resources,
+                attributes={"gen_ai.request.model": "gpt-4o", "gen_ai.system_instructions": first_instructions, "gen_ai.input.messages": first_input, "gen_ai.output.messages": first_output, "gen_ai.response.finish_reasons": json.dumps(["tool_calls"])},
+                attribute_write_mode="json_only",
             ),
             Traces(
                 timestamp=now - timedelta(seconds=6),
@@ -45,7 +53,34 @@ def test_thread_returns_gen_ai_spans_in_order(
             ),
             Traces(timestamp=now - timedelta(seconds=5), trace_id=trace_id, span_id=agent_id, parent_span_id=root_id, name="invoke_agent planner", resources=resources, attributes={"gen_ai.agent.name": "planner"}, attribute_write_mode="json_only"),
             Traces(timestamp=now - timedelta(seconds=4), trace_id=trace_id, span_id=second_llm_id, parent_span_id=root_id, name="chat gpt-4o", resources=resources, attributes={"gen_ai.request.model": "gpt-4o", "gen_ai.input.messages": second_input}, attribute_write_mode="json_only"),
+            Traces(
+                timestamp=now - timedelta(seconds=3),
+                trace_id=trace_id,
+                span_id=legacy_llm_id,
+                parent_span_id=root_id,
+                name="openai.chat",
+                resources=resources,
+                attributes={"gen_ai.request.model": "gpt-4o", "gen_ai.prompt.0.role": "user", "gen_ai.prompt.0.content": "and tomorrow?", "gen_ai.completion.0.role": "assistant", "gen_ai.completion.0.content": "Also sunny.", "gen_ai.completion.0.finish_reason": "stop"},
+                attribute_write_mode="json_only",
+            ),
             Traces(timestamp=now - timedelta(seconds=2), trace_id=trace_id, span_id=third_llm_id, parent_span_id=root_id, name="chat gpt-4o", resources=resources, attributes={"gen_ai.request.model": "gpt-4o", "gen_ai.output.messages": "It is sunny in Bangalore."}, attribute_write_mode="json_only"),
+            Traces(
+                timestamp=now - timedelta(seconds=1),
+                trace_id=trace_id,
+                span_id=flattened_llm_id,
+                parent_span_id=root_id,
+                name="ChatCompletion",
+                resources=resources,
+                attributes={
+                    "gen_ai.request.model": "gpt-4o",
+                    "llm.input_messages.0.message.role": "user",
+                    "llm.input_messages.0.message.content": "thanks",
+                    "llm.output_messages.0.message.role": "assistant",
+                    "llm.output_messages.0.message.content": "You're welcome.",
+                    "llm.finish_reason": "stop",
+                },
+                attribute_write_mode="json_only",
+            ),
         ]
     )
 
@@ -54,10 +89,10 @@ def test_thread_returns_gen_ai_spans_in_order(
     assert response.status_code == HTTPStatus.OK, response.text
 
     thread = response.json()["data"]
-    assert [span["span_id"] for span in thread["spans"]] == [first_llm_id, tool_id, agent_id, second_llm_id, third_llm_id]
+    assert [span["span_id"] for span in thread["spans"]] == [first_llm_id, tool_id, agent_id, second_llm_id, legacy_llm_id, third_llm_id, flattened_llm_id]
     assert "nextCursor" not in thread
 
-    first, tool, agent, input_only, output_only = thread["spans"]
+    first, tool, agent, input_only, legacy, output_only, flattened = thread["spans"]
     assert agent["attributes"]["gen_ai.agent.name"] == "planner"
     assert agent["formattedInput"] == []
     assert agent["formattedOutput"] == []
@@ -73,7 +108,7 @@ def test_thread_returns_gen_ai_spans_in_order(
     assert output_only["attributes"]["gen_ai.output.messages"] == "It is sunny in Bangalore."
 
     tool_call = {"type": "tool_call", "name": "get_weather", "id": "call_1", "arguments": {"city": "Bangalore"}}
-    assert first["formattedInput"] == [{"role": "user", "parts": [{"type": "text", "content": "weather in Bangalore?"}]}]
+    assert first["formattedInput"] == [{"role": "system", "parts": [{"type": "text", "content": "Answer with the tool."}]}, {"role": "user", "parts": [{"type": "text", "content": "weather in Bangalore?"}]}]
     assert first["formattedOutput"] == [{"role": "assistant", "parts": [tool_call], "finish_reason": "tool_call"}]
     assert first["formatterWarnings"] == []
     assert tool["formattedInput"] == [{"role": "assistant", "parts": [tool_call]}]
@@ -83,6 +118,12 @@ def test_thread_returns_gen_ai_spans_in_order(
     assert output_only["formattedInput"] == []
     assert output_only["formattedOutput"] == [{"role": "assistant", "parts": [{"type": "text", "content": "It is sunny in Bangalore."}]}]
     assert output_only["formatterWarnings"] == ["bare assistant text, role assumed"]
+    assert legacy["attributes"]["gen_ai.prompt.0.content"] == "and tomorrow?"
+    assert legacy["formattedInput"] == [{"role": "user", "parts": [{"type": "text", "content": "and tomorrow?"}]}]
+    assert legacy["formattedOutput"] == [{"role": "assistant", "parts": [{"type": "text", "content": "Also sunny."}], "finish_reason": "stop"}]
+    assert flattened["attributes"]["llm.input_messages.0.message.content"] == "thanks"
+    assert flattened["formattedInput"] == [{"role": "user", "parts": [{"type": "text", "content": "thanks"}]}]
+    assert flattened["formattedOutput"] == [{"role": "assistant", "parts": [{"type": "text", "content": "You're welcome."}], "finish_reason": "stop"}]
 
 
 def test_thread_paginates_with_cursors(