Documentation
¶
Index ¶
Constants ¶
const ( StatusSuccess = "success" StatusError = "error" EventTypeMessage = "message" EventTypeProgress = "progress" EventTypeTool = "tool" EventTypeError = "error" // EventTypeStderr carries a single line of subprocess stderr. Distinct // from Progress so the UI can render diagnostic chatter (npm install // progress, "API key required", protocol error dumps) differently from // real tool calls or agent thinking. EventTypeStderr = "stderr" )
Variables ¶
This section is empty.
Functions ¶
func AutoApproveOutcome ¶
func AutoApproveOutcome(req acp.RequestPermissionRequest) acp.RequestPermissionOutcome
AutoApproveOutcome picks the "best" allow option from a permission request and returns a Selected outcome for it. Preference order is allow_always > allow_once > any other option whose Kind starts with "allow_". If no allow option is present, the first option is selected so the agent can make forward progress; if the request carries no options at all, the request is canceled.
pi-go ACP subagents (claude / gemini / cursor) run inside a single user- initiated session that has already been authorized at the parent process boundary, so deferring every tool call back to the user is unnecessary noise. Routing all requests through this helper keeps that policy in one place.
func EnrichToolCallTitle ¶
EnrichToolCallTitle returns a title that combines the agent-provided display title with a short snippet pulled from RawInput. ACP agents often send generic titles like "Terminal", "TaskOutput", or "Read" with the real detail (the command, the file path, the task description) buried in RawInput, which produces a stream of unhelpful "TaskOutput / TaskOutput / TaskOutput" lines in the parent UI. Pulling the first useful field out of RawInput restores per-call context.
If RawInput is nil or carries no recognizable string field, the original title is returned unchanged.
Types ¶
type Event ¶
type Event struct {
Type string `json:"type"`
Content string `json:"content,omitempty"`
Error string `json:"error,omitempty"`
SessionID string `json:"session_id,omitempty"`
}
Event is the shared local streaming model for ACP updates.
type RunRequest ¶
type RunRequest struct {
Command []string
Prompt string
SessionID string
CWD string
Env []string
RPCTimeout time.Duration
}
RunRequest describes a local ACP turn request shared by client and server code.
func (RunRequest) Validate ¶
func (r RunRequest) Validate() error
Validate checks whether the request has the required fields for execution.
type RunResult ¶
type RunResult struct {
Status string `json:"status"`
Result string `json:"result,omitempty"`
Error string `json:"error,omitempty"`
SessionID string `json:"session_id,omitempty"`
Stderr string `json:"stderr,omitempty"` // Captured subprocess stderr for diagnostics
StopReason string `json:"stopReason,omitempty"` // ACP stopReason from PromptResponse
}
RunResult is the shared local result model for ACP executions.
type ToolCallTitleFilter ¶
type ToolCallTitleFilter struct {
// contains filtered or unexported fields
}
ToolCallTitleFilter dedupes tool-call title emissions per tool_call_id.
Most ACP agents send a generic ToolCall first (Title: "Terminal", "Read", "Bash") and then a ToolCallUpdate with a far more useful Title (the actual command, file path, etc.). Streaming both produces redundant lines like:
- Terminal
- `go test ./internal/acp/... 2>&1`
This filter buffers the initial title; if a ToolCallUpdate arrives for the same id with a non-empty title, the buffered title is dropped and only the update's title is emitted. If no update ever arrives (rare), Flush emits the buffered title at session end so the call is still visible.
func NewToolCallTitleFilter ¶
func NewToolCallTitleFilter(emit func(title string)) *ToolCallTitleFilter
NewToolCallTitleFilter wires the filter to the per-session emit callback. emit is invoked with the chosen title each time the filter decides to surface one — the caller is responsible for wrapping it in the appropriate event.
func (*ToolCallTitleFilter) Flush ¶
func (f *ToolCallTitleFilter) Flush()
Flush emits any pending titles whose tool calls produced no update before the session ended. Safe to call multiple times.
func (*ToolCallTitleFilter) OnToolCall ¶
func (f *ToolCallTitleFilter) OnToolCall(id, title string)
OnToolCall records an initial tool-call title without emitting it. Empty titles are ignored. If id is empty the title is emitted immediately, since without an id we cannot correlate a later update.
func (*ToolCallTitleFilter) OnToolCallUpdate ¶
func (f *ToolCallTitleFilter) OnToolCallUpdate(id, title string)
OnToolCallUpdate emits the update's title (when non-empty) and forgets any buffered initial title for the same id. If the update carries no title and nothing is pending, this is a no-op.
type ValidationError ¶
type ValidationError string
func (ValidationError) Error ¶
func (e ValidationError) Error() string
Source Files
Directories
¶
| Path | Synopsis |
|---|---|
|
claudecode
Package claudecode provides an ACP client for Claude Code via the @agentclientprotocol/claude-agent-acp subprocess adapter.
|
Package claudecode provides an ACP client for Claude Code via the @agentclientprotocol/claude-agent-acp subprocess adapter. |
|
cursor
Package cursor provides an ACP client for Cursor CLI via its `agent acp` subprocess (https://cursor.com/docs/cli/acp).
|
Package cursor provides an ACP client for Cursor CLI via its `agent acp` subprocess (https://cursor.com/docs/cli/acp). |
|
gemini
Package gemini provides an ACP client for Google Gemini CLI via the Gemini CLI subprocess adapter.
|
Package gemini provides an ACP client for Google Gemini CLI via the Gemini CLI subprocess adapter. |