opensource/coder
3
0
Fork 0
mirror of https://github.com/coder/coder.git synced 2026-09-24 15:04:27 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix(coderd): strip markdown code fences from Anthropic task name responses (#23024)

Browse Source 
- Adds `extractJSON()` to strip markdown code fences before JSON parsing and wire into the `json.Unmarshal` call in `generateFromAnthropic`.
- Accepts variadic `RequestOption` in `generateFromAnthropic` so tests can inject a mock Anthropic server via `WithBaseURL`.
- Adds table-driven cases covering bare JSON, fenced with/without language tag, surrounding whitespace, and multiline JSON.
- Adds end-to-end cases using `httptest.NewServer` to serve fake Anthropic SSE streams with bare and fenced responses.
This commit is contained in:
Cian Johnston
2026-03-13 17:35:26 +00:00
committed by GitHub
parent efe114119f
commit 03d0fc4f4c
2 changed files with 177 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
coderd/taskname/taskname.go
+27 -3
View File
@@ -96,6 +96,27 @@ var (
ErrNoNameGenerated = xerrors.New("no task name generated")
)
// extractJSON strips optional markdown code fences (```json or
// ```) that LLMs sometimes wrap around JSON output, returning
// only the inner JSON string. Only well-formed fences with a
// newline after the opening backticks are stripped; malformed
// fences are left untouched so that json.Unmarshal fails
// cleanly and the caller can fall back to other strategies.
func extractJSON(s string) string {
s = strings.TrimSpace(s)
if strings.HasPrefix(s, "```") {
// Only strip when there is a newline separating the
// fence line from the body. Without one we cannot
// reliably tell the fence from the content.
if idx := strings.Index(s, "\n"); idx != -1 {
s = s[idx+1:]
s = strings.TrimSuffix(s, "```")
s = strings.TrimSpace(s)
}
}
return s
}
type TaskName struct {
Name string `json:"task_name"`
DisplayName string `json:"display_name"`
@@ -188,7 +209,7 @@ func generateFromPrompt(prompt string) (TaskName, error) {
// generateFromAnthropic uses Claude (Anthropic) to generate semantic task and display names from a user prompt.
// It sends the prompt to Claude with a structured system prompt requesting JSON output containing both names.
// Returns an error if the API call fails, the response is invalid, or Claude returns an "unnamed" placeholder.
func generateFromAnthropic(ctx context.Context, prompt string, apiKey string, model anthropic.Model) (TaskName, error) {
func generateFromAnthropic(ctx context.Context, prompt string, apiKey string, model anthropic.Model, opts ...anthropicoption.RequestOption) (TaskName, error) {
anthropicModel := model
if anthropicModel == "" {
anthropicModel = defaultModel
@@ -216,6 +237,7 @@ func generateFromAnthropic(ctx context.Context, prompt string, apiKey string, mo
anthropicOptions := anthropic.DefaultClientOptions()
anthropicOptions = append(anthropicOptions, anthropicoption.WithAPIKey(apiKey))
anthropicOptions = append(anthropicOptions, opts...)
anthropicClient := anthropic.NewClient(anthropicOptions...)
stream, err := anthropicDataStream(ctx, anthropicClient, anthropicModel, conversation)
@@ -234,9 +256,11 @@ func generateFromAnthropic(ctx context.Context, prompt string, apiKey string, mo
return TaskName{}, ErrNoNameGenerated
}
// Parse the JSON response
// Parse the JSON response. LLMs sometimes wrap JSON in
// markdown code fences (```json ... ```), so we strip
// those before unmarshalling.
var taskNameResponse TaskName
if err := json.Unmarshal([]byte(acc.Messages()[0].Content), &taskNameResponse); err != nil {
if err := json.Unmarshal([]byte(extractJSON(acc.Messages()[0].Content)), &taskNameResponse); err != nil {
return TaskName{}, xerrors.Errorf("failed to parse anthropic response: %w", err)
}