1package agent
2
3import (
4 "bufio"
5 "context"
6 "encoding/json"
7 "errors"
8 "fmt"
9 "io"
10 "os"
11 "strings"
12 "sync"
13 "time"
14
15 "github.com/patriceckhart/zot/internal/agent/extensions"
16 "github.com/patriceckhart/zot/internal/agent/modes"
17 "github.com/patriceckhart/zot/internal/core"
18 "github.com/patriceckhart/zot/internal/extproto"
19 "github.com/patriceckhart/zot/internal/provider"
20)
21
22// runRPCMode implements the JSON-over-stdin/stdout RPC protocol.
23//
24// Wire format: one JSON object per line in both directions.
25//
26// Commands (stdin):
27//
28// {"id":"1","type":"prompt","message":"hello","images":[]}
29// {"id":"2","type":"abort"}
30// {"id":"3","type":"compact"}
31// {"id":"4","type":"get_state"}
32// {"id":"5","type":"set_model","model":"claude-opus-4-5"}
33// {"id":"6","type":"get_messages"}
34// {"id":"7","type":"clear"}
35// {"id":"8","type":"get_models"}
36//
37// Responses (stdout): {"type":"response","id":"1","command":"prompt","success":true}
38// Events (stdout): one JSON object per AgentEvent (same schema as --json mode).
39//
40// Auth: if $ZOTCORE_RPC_TOKEN is set, the first command must be
41// {"type":"hello","token":"..."} or the connection is closed.
42func runRPCMode(ctx context.Context, args Args, version string) error {
43 if args.NoYolo {
44 fmt.Fprintln(os.Stderr, "warning: --no-yolo has no effect in rpc mode (no interactive prompt available); tools will run without confirmation")
45 }
46 r, err := Resolve(args, true)
47 if err != nil {
48 return err
49 }
50
51 // Extensions: same lifecycle as interactive mode, minus the
52 // host-hooks integration. Notify/Display calls from extensions
53 // emit RPC events instead of TUI lines so any consumer can react.
54 extHooks := &rpcExtHooks{}
55 extMgr := extensions.New(ZotHome(), r.CWD, version, r.Provider, r.Model, extHooks)
56 for _, e := range extMgr.LoadExplicit(ctx, args.Exts) {
57 fmt.Fprintln(os.Stderr, "extension load:", e)
58 }
59 if !args.NoExt {
60 for _, e := range extMgr.Discover(ctx) {
61 fmt.Fprintln(os.Stderr, "extension load:", e)
62 }
63 }
64 extMgr.WaitForReady(3 * time.Second)
65 defer extMgr.Stop(2 * time.Second)
66 r.MergeExtensionTools(&extToolAdapter{mgr: extMgr})
67
68 ag := r.NewAgent()
69 ag.BeforeToolExecute = func(call provider.ToolCallBlock) (bool, string, json.RawMessage) {
70 r := extMgr.InterceptToolCall(ctx, call.ID, call.Name, call.Arguments)
71 if r.Block {
72 return false, r.Reason, nil
73 }
74 return true, "", r.ModifiedArgs
75 }
76 ag.BeforeTurn = func(step int) (bool, string) {
77 r := extMgr.InterceptTurnStart(ctx, step)
78 return !r.Block, r.Reason
79 }
80 ag.BeforeAssistantMessage = func(text string) (bool, string, string) {
81 r := extMgr.InterceptAssistantMessage(ctx, text)
82 if r.Block {
83 return false, r.Reason, ""
84 }
85 return true, "", r.ReplaceText
86 }
87 ag.OnEvent = func(ev core.AgentEvent) { fanoutAgentEvent(extMgr, ev) }
88
89 // /reload-ext hot-reload callback (also triggered via rpc
90 // `reload_ext` if/when added). Rebuilds the tool registry on the
91 // current agent so freshly-registered extension tools become
92 // callable without restarting the rpc process.
93 adapter := &extToolAdapter{mgr: extMgr}
94 extMgr.SetOnReload(func() {
95 resolved, err := Resolve(args, true)
96 if err != nil {
97 return
98 }
99 resolved.MergeExtensionTools(adapter)
100 ag.SetTools(resolved.ToolRegistry)
101 })
102
103 extMgr.EmitEvent(extproto.EventFromHost{Event: "session_start"})
104
105 server := &rpcServer{
106 ctx: ctx,
107 args: args,
108 agent: ag,
109 provider: r.Provider,
110 model: r.Model,
111 out: os.Stdout,
112 version: version,
113 }
114 extHooks.server = server
115 return server.run(os.Stdin)
116}
117
118// rpcExtHooks implements extensions.HostHooks for the headless RPC
119// loop. Notify and Display surface as `event` frames so any RPC
120// client can render them; Submit and Insert are no-ops because the
121// RPC loop has no editor and the prompt comes from the client.
122type rpcExtHooks struct {
123 server *rpcServer
124}
125
126func (h *rpcExtHooks) Notify(extName, level, message string) {
127 if h.server != nil {
128 h.server.writeEvent(map[string]any{
129 "type": "ext_notify",
130 "extension": extName,
131 "level": level,
132 "message": message,
133 })
134 }
135}
136func (h *rpcExtHooks) Display(extName, text string) {
137 if h.server != nil {
138 h.server.writeEvent(map[string]any{
139 "type": "ext_display",
140 "extension": extName,
141 "text": text,
142 })
143 }
144}
145func (h *rpcExtHooks) Submit(string) {} // ignored in rpc mode
146func (h *rpcExtHooks) Insert(string) {} // ignored in rpc mode
147func (h *rpcExtHooks) OpenPanel(string, extproto.PanelSpec) {}
148func (h *rpcExtHooks) UpdatePanel(string, string, string, []string, string) {}
149func (h *rpcExtHooks) ClosePanel(string, string) {}
150
151type rpcServer struct {
152 ctx context.Context
153 args Args
154 agent *core.Agent
155 provider string
156 model string
157 out io.Writer
158 version string
159
160 writeMu sync.Mutex
161 turnMu sync.Mutex // serialises one prompt at a time
162 activeCancel context.CancelFunc
163 authed bool
164
165 // inFlight tracks long-running command goroutines so run() can
166 // wait for them before returning when stdin closes. Without this,
167 // piping a single 'prompt' command into 'zot rpc' would race the
168 // process exit against the agent loop and the prompt would never
169 // produce output.
170 inFlight sync.WaitGroup
171}
172
173// run reads NDJSON commands from in and dispatches them. Returns when
174// in is closed AND every in-flight long-running command (prompt /
175// compact) has finished, so a quick `echo cmd | zot rpc` invocation
176// still produces full output before the process exits.
177func (s *rpcServer) run(in io.Reader) error {
178 requireToken := os.Getenv("ZOTCORE_RPC_TOKEN") != ""
179 s.authed = !requireToken
180
181 sc := bufio.NewScanner(in)
182 sc.Buffer(make([]byte, 0, 64*1024), 16*1024*1024)
183 for sc.Scan() {
184 line := strings.TrimSpace(sc.Text())
185 if line == "" {
186 continue
187 }
188 var head struct {
189 Type string `json:"type"`
190 ID string `json:"id,omitempty"`
191 }
192 if err := json.Unmarshal([]byte(line), &head); err != nil {
193 s.writeError("", "", fmt.Sprintf("malformed json: %v", err))
194 continue
195 }
196 if !s.authed {
197 if head.Type != "hello" {
198 s.writeError(head.ID, head.Type, "auth required: send hello with token first")
199 continue
200 }
201 var hello struct {
202 Token string `json:"token"`
203 }
204 _ = json.Unmarshal([]byte(line), &hello)
205 if hello.Token != os.Getenv("ZOTCORE_RPC_TOKEN") {
206 s.writeError(head.ID, head.Type, "invalid token")
207 return fmt.Errorf("rpc: bad auth token")
208 }
209 s.authed = true
210 s.writeResponse(head.ID, head.Type, map[string]any{
211 "protocol_version": 1,
212 "version": s.version,
213 "provider": s.provider,
214 "model": s.model,
215 })
216 continue
217 }
218 s.dispatch(head.Type, head.ID, []byte(line))
219 }
220 err := sc.Err()
221 s.inFlight.Wait()
222 return err
223}
224
225// dispatch routes a command. Long-running commands (prompt, compact)
226// run on their own goroutine so the read loop stays responsive.
227func (s *rpcServer) dispatch(cmd, id string, raw []byte) {
228 switch cmd {
229 case "hello":
230 s.writeResponse(id, cmd, map[string]any{
231 "protocol_version": 1,
232 "version": s.version,
233 "provider": s.provider,
234 "model": s.model,
235 })
236 case "prompt":
237 var req struct {
238 Message string `json:"message"`
239 Images []struct {
240 MimeType string `json:"mime_type"`
241 Data []byte `json:"data"`
242 } `json:"images"`
243 }
244 if err := json.Unmarshal(raw, &req); err != nil {
245 s.writeError(id, cmd, err.Error())
246 return
247 }
248 s.inFlight.Add(1)
249 go func() {
250 defer s.inFlight.Done()
251 s.runPrompt(id, req.Message, req.Images)
252 }()
253
254 case "abort":
255 if c := s.takeCancel(); c != nil {
256 c()
257 }
258 s.writeResponse(id, cmd, nil)
259
260 case "compact":
261 s.inFlight.Add(1)
262 go func() {
263 defer s.inFlight.Done()
264 s.runCompact(id)
265 }()
266
267 case "get_state":
268 s.writeResponse(id, cmd, s.snapshotState())
269
270 case "get_messages":
271 s.writeResponse(id, cmd, map[string]any{
272 "messages": messagesToJSON(s.agent.Messages()),
273 })
274
275 case "clear":
276 s.agent.SetMessages(nil)
277 s.writeResponse(id, cmd, nil)
278
279 case "set_model":
280 var req struct {
281 Model string `json:"model"`
282 }
283 if err := json.Unmarshal(raw, &req); err != nil {
284 s.writeError(id, cmd, err.Error())
285 return
286 }
287 if _, err := provider.FindModel(s.provider, req.Model); err != nil {
288 s.writeError(id, cmd, err.Error())
289 return
290 }
291 s.agent.Model = req.Model
292 s.model = req.Model
293 s.writeResponse(id, cmd, map[string]any{"model": req.Model})
294
295 case "get_models":
296 out := []map[string]any{}
297 for _, m := range provider.ModelsForProvider(s.provider) {
298 out = append(out, map[string]any{
299 "id": m.ID,
300 "provider": m.Provider,
301 "context_window": m.ContextWindow,
302 "max_output": m.MaxOutput,
303 "reasoning": m.Reasoning,
304 })
305 }
306 s.writeResponse(id, cmd, map[string]any{"models": out})
307
308 case "ping":
309 s.writeResponse(id, cmd, map[string]any{"pong": true})
310
311 default:
312 s.writeError(id, cmd, "unknown command")
313 }
314}
315
316// runPrompt executes a single prompt turn and streams events out.
317// Holds turnMu so a second concurrent prompt blocks until this one
318// finishes; the user can abort with the abort command.
319func (s *rpcServer) runPrompt(id, message string, images []struct {
320 MimeType string `json:"mime_type"`
321 Data []byte `json:"data"`
322}) {
323 s.turnMu.Lock()
324 defer s.turnMu.Unlock()
325
326 subCtx, cancel := context.WithCancel(s.ctx)
327 s.setCancel(cancel)
328 defer s.setCancel(nil)
329
330 s.writeResponse(id, "prompt", map[string]any{"started": true})
331
332 imgs := make([]provider.ImageBlock, 0, len(images))
333 for _, im := range images {
334 imgs = append(imgs, provider.ImageBlock{MimeType: im.MimeType, Data: im.Data})
335 }
336
337 err := s.agent.Prompt(subCtx, message, imgs, func(ev core.AgentEvent) {
338 // EvDone is emitted by the agent loop and we re-emit our own
339 // 'done' below; suppressing it here avoids duplicate frames.
340 if _, ok := ev.(core.EvDone); ok {
341 return
342 }
343 s.writeEvent(modes.EventToJSON(ev))
344 })
345 // Don't emit a stand-alone error event for cancellation; the prior
346 // turn_end with stop=aborted already carries that signal.
347 if err != nil && !errors.Is(err, context.Canceled) {
348 s.writeEvent(map[string]any{"type": "error", "message": err.Error()})
349 }
350 s.writeEvent(map[string]any{"type": "done"})
351}
352
353// runCompact mirrors runPrompt but for compaction.
354func (s *rpcServer) runCompact(id string) {
355 s.turnMu.Lock()
356 defer s.turnMu.Unlock()
357
358 subCtx, cancel := context.WithCancel(s.ctx)
359 s.setCancel(cancel)
360 defer s.setCancel(nil)
361
362 s.writeResponse(id, "compact", map[string]any{"started": true})
363 summary, err := s.agent.Compact(subCtx, 4, nil)
364 if err != nil {
365 if !errors.Is(err, context.Canceled) {
366 s.writeEvent(map[string]any{"type": "error", "message": err.Error()})
367 }
368 return
369 }
370 s.writeEvent(map[string]any{
371 "type": "compact_done",
372 "summary": summary,
373 })
374}
375
376// snapshotState builds the get_state response.
377func (s *rpcServer) snapshotState() map[string]any {
378 cum := s.agent.Cost()
379 return map[string]any{
380 "provider": s.provider,
381 "model": s.model,
382 "cwd": s.args.CWD,
383 "message_count": len(s.agent.Messages()),
384 "busy": s.activeCancel != nil,
385 "usage": map[string]any{
386 "input": cum.InputTokens,
387 "output": cum.OutputTokens,
388 "cache_read": cum.CacheReadTokens,
389 "cache_write": cum.CacheWriteTokens,
390 "cost_usd": cum.CostUSD,
391 },
392 }
393}
394
395// ---- write helpers (single-line JSON, mutex-guarded) ----
396
397func (s *rpcServer) writeResponse(id, cmd string, data any) {
398 frame := map[string]any{
399 "type": "response",
400 "command": cmd,
401 "success": true,
402 }
403 if id != "" {
404 frame["id"] = id
405 }
406 if data != nil {
407 frame["data"] = data
408 }
409 s.write(frame)
410}
411
412func (s *rpcServer) writeError(id, cmd, msg string) {
413 frame := map[string]any{
414 "type": "response",
415 "command": cmd,
416 "success": false,
417 "error": msg,
418 }
419 if id != "" {
420 frame["id"] = id
421 }
422 s.write(frame)
423}
424
425func (s *rpcServer) writeEvent(payload map[string]any) {
426 s.write(payload)
427}
428
429func (s *rpcServer) write(v any) {
430 b, err := json.Marshal(v)
431 if err != nil {
432 return
433 }
434 s.writeMu.Lock()
435 defer s.writeMu.Unlock()
436 _, _ = s.out.Write(b)
437 _, _ = s.out.Write([]byte("\n"))
438}
439
440func (s *rpcServer) setCancel(c context.CancelFunc) {
441 s.writeMu.Lock()
442 s.activeCancel = c
443 s.writeMu.Unlock()
444}
445
446func (s *rpcServer) takeCancel() context.CancelFunc {
447 s.writeMu.Lock()
448 defer s.writeMu.Unlock()
449 c := s.activeCancel
450 s.activeCancel = nil
451 return c
452}
453
454// messagesToJSON serialises a transcript using the same schema as the
455// --json event mode for cross-format consistency.
456func messagesToJSON(msgs []provider.Message) []map[string]any {
457 out := make([]map[string]any, 0, len(msgs))
458 for _, m := range msgs {
459 out = append(out, map[string]any{
460 "role": string(m.Role),
461 "content": modes.ContentToJSON(m.Content),
462 "time": m.Time,
463 })
464 }
465 return out
466}