v0.2.12 7a7bf0b

rpc.go RawHistoryFolder

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/packages/agent/extensions"
16	"github.com/patriceckhart/zot/packages/agent/extproto"
17	"github.com/patriceckhart/zot/packages/agent/modes"
18	"github.com/patriceckhart/zot/packages/core"
19	"github.com/patriceckhart/zot/packages/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) SubmitSlash(string)                                   {} // ignored in rpc mode
147func (h *rpcExtHooks) Insert(string)                                        {} // ignored in rpc mode
148func (h *rpcExtHooks) OpenPanel(string, extproto.PanelSpec)                 {}
149func (h *rpcExtHooks) UpdatePanel(string, string, string, []string, string) {}
150func (h *rpcExtHooks) ClosePanel(string, string)                            {}
151
152type rpcServer struct {
153	ctx      context.Context
154	args     Args
155	agent    *core.Agent
156	provider string
157	model    string
158	out      io.Writer
159	version  string
160
161	writeMu      sync.Mutex
162	turnMu       sync.Mutex // serialises one prompt at a time
163	activeCancel context.CancelFunc
164	authed       bool
165
166	// inFlight tracks long-running command goroutines so run() can
167	// wait for them before returning when stdin closes. Without this,
168	// piping a single 'prompt' command into 'zot rpc' would race the
169	// process exit against the agent loop and the prompt would never
170	// produce output.
171	inFlight sync.WaitGroup
172}
173
174// run reads NDJSON commands from in and dispatches them. Returns when
175// in is closed AND every in-flight long-running command (prompt /
176// compact) has finished, so a quick `echo cmd | zot rpc` invocation
177// still produces full output before the process exits.
178func (s *rpcServer) run(in io.Reader) error {
179	requireToken := os.Getenv("ZOTCORE_RPC_TOKEN") != ""
180	s.authed = !requireToken
181
182	sc := bufio.NewScanner(in)
183	sc.Buffer(make([]byte, 0, 64*1024), 16*1024*1024)
184	for sc.Scan() {
185		line := strings.TrimSpace(sc.Text())
186		if line == "" {
187			continue
188		}
189		var head struct {
190			Type string `json:"type"`
191			ID   string `json:"id,omitempty"`
192		}
193		if err := json.Unmarshal([]byte(line), &head); err != nil {
194			s.writeError("", "", fmt.Sprintf("malformed json: %v", err))
195			continue
196		}
197		if !s.authed {
198			if head.Type != "hello" {
199				s.writeError(head.ID, head.Type, "auth required: send hello with token first")
200				continue
201			}
202			var hello struct {
203				Token string `json:"token"`
204			}
205			_ = json.Unmarshal([]byte(line), &hello)
206			if hello.Token != os.Getenv("ZOTCORE_RPC_TOKEN") {
207				s.writeError(head.ID, head.Type, "invalid token")
208				return fmt.Errorf("rpc: bad auth token")
209			}
210			s.authed = true
211			s.writeResponse(head.ID, head.Type, map[string]any{
212				"protocol_version": 1,
213				"version":          s.version,
214				"provider":         s.provider,
215				"model":            s.model,
216			})
217			continue
218		}
219		s.dispatch(head.Type, head.ID, []byte(line))
220	}
221	err := sc.Err()
222	s.inFlight.Wait()
223	return err
224}
225
226// dispatch routes a command. Long-running commands (prompt, compact)
227// run on their own goroutine so the read loop stays responsive.
228func (s *rpcServer) dispatch(cmd, id string, raw []byte) {
229	switch cmd {
230	case "hello":
231		s.writeResponse(id, cmd, map[string]any{
232			"protocol_version": 1,
233			"version":          s.version,
234			"provider":         s.provider,
235			"model":            s.model,
236		})
237	case "prompt":
238		var req struct {
239			Message string `json:"message"`
240			Images  []struct {
241				MimeType string `json:"mime_type"`
242				Data     []byte `json:"data"`
243			} `json:"images"`
244		}
245		if err := json.Unmarshal(raw, &req); err != nil {
246			s.writeError(id, cmd, err.Error())
247			return
248		}
249		s.inFlight.Add(1)
250		go func() {
251			defer s.inFlight.Done()
252			s.runPrompt(id, req.Message, req.Images)
253		}()
254
255	case "abort":
256		if c := s.takeCancel(); c != nil {
257			c()
258		}
259		s.writeResponse(id, cmd, nil)
260
261	case "compact":
262		s.inFlight.Add(1)
263		go func() {
264			defer s.inFlight.Done()
265			s.runCompact(id)
266		}()
267
268	case "get_state":
269		s.writeResponse(id, cmd, s.snapshotState())
270
271	case "get_messages":
272		s.writeResponse(id, cmd, map[string]any{
273			"messages": messagesToJSON(s.agent.Messages()),
274		})
275
276	case "clear":
277		s.agent.SetMessages(nil)
278		s.writeResponse(id, cmd, nil)
279
280	case "set_model":
281		var req struct {
282			Model string `json:"model"`
283		}
284		if err := json.Unmarshal(raw, &req); err != nil {
285			s.writeError(id, cmd, err.Error())
286			return
287		}
288		if _, err := provider.FindModel(s.provider, req.Model); err != nil {
289			s.writeError(id, cmd, err.Error())
290			return
291		}
292		s.agent.Model = req.Model
293		s.model = req.Model
294		s.writeResponse(id, cmd, map[string]any{"model": req.Model})
295
296	case "get_models":
297		out := []map[string]any{}
298		for _, m := range provider.ModelsForProvider(s.provider) {
299			out = append(out, map[string]any{
300				"id":             m.ID,
301				"provider":       m.Provider,
302				"context_window": m.ContextWindow,
303				"max_output":     m.MaxOutput,
304				"reasoning":      m.Reasoning,
305			})
306		}
307		s.writeResponse(id, cmd, map[string]any{"models": out})
308
309	case "ping":
310		s.writeResponse(id, cmd, map[string]any{"pong": true})
311
312	default:
313		s.writeError(id, cmd, "unknown command")
314	}
315}
316
317// runPrompt executes a single prompt turn and streams events out.
318// Holds turnMu so a second concurrent prompt blocks until this one
319// finishes; the user can abort with the abort command.
320func (s *rpcServer) runPrompt(id, message string, images []struct {
321	MimeType string `json:"mime_type"`
322	Data     []byte `json:"data"`
323}) {
324	s.turnMu.Lock()
325	defer s.turnMu.Unlock()
326
327	subCtx, cancel := context.WithCancel(s.ctx)
328	s.setCancel(cancel)
329	defer s.setCancel(nil)
330
331	s.writeResponse(id, "prompt", map[string]any{"started": true})
332
333	imgs := make([]provider.ImageBlock, 0, len(images))
334	for _, im := range images {
335		imgs = append(imgs, provider.ImageBlock{MimeType: im.MimeType, Data: im.Data})
336	}
337
338	err := s.agent.Prompt(subCtx, message, imgs, func(ev core.AgentEvent) {
339		// EvDone is emitted by the agent loop and we re-emit our own
340		// 'done' below; suppressing it here avoids duplicate frames.
341		if _, ok := ev.(core.EvDone); ok {
342			return
343		}
344		s.writeEvent(modes.EventToJSON(ev))
345	})
346	// Don't emit a stand-alone error event for cancellation; the prior
347	// turn_end with stop=aborted already carries that signal.
348	if err != nil && !errors.Is(err, context.Canceled) {
349		s.writeEvent(map[string]any{"type": "error", "message": err.Error()})
350	}
351	s.writeEvent(map[string]any{"type": "done"})
352}
353
354// runCompact mirrors runPrompt but for compaction.
355func (s *rpcServer) runCompact(id string) {
356	s.turnMu.Lock()
357	defer s.turnMu.Unlock()
358
359	subCtx, cancel := context.WithCancel(s.ctx)
360	s.setCancel(cancel)
361	defer s.setCancel(nil)
362
363	s.writeResponse(id, "compact", map[string]any{"started": true})
364	summary, err := s.agent.Compact(subCtx, 4, nil)
365	if err != nil {
366		if !errors.Is(err, context.Canceled) {
367			s.writeEvent(map[string]any{"type": "error", "message": err.Error()})
368		}
369		return
370	}
371	s.writeEvent(map[string]any{
372		"type":    "compact_done",
373		"summary": summary,
374	})
375}
376
377// snapshotState builds the get_state response.
378func (s *rpcServer) snapshotState() map[string]any {
379	cum := s.agent.Cost()
380	return map[string]any{
381		"provider":      s.provider,
382		"model":         s.model,
383		"cwd":           s.args.CWD,
384		"message_count": len(s.agent.Messages()),
385		"busy":          s.activeCancel != nil,
386		"usage": map[string]any{
387			"input":       cum.InputTokens,
388			"output":      cum.OutputTokens,
389			"cache_read":  cum.CacheReadTokens,
390			"cache_write": cum.CacheWriteTokens,
391			"cost_usd":    cum.CostUSD,
392		},
393	}
394}
395
396// ---- write helpers (single-line JSON, mutex-guarded) ----
397
398func (s *rpcServer) writeResponse(id, cmd string, data any) {
399	frame := map[string]any{
400		"type":    "response",
401		"command": cmd,
402		"success": true,
403	}
404	if id != "" {
405		frame["id"] = id
406	}
407	if data != nil {
408		frame["data"] = data
409	}
410	s.write(frame)
411}
412
413func (s *rpcServer) writeError(id, cmd, msg string) {
414	frame := map[string]any{
415		"type":    "response",
416		"command": cmd,
417		"success": false,
418		"error":   msg,
419	}
420	if id != "" {
421		frame["id"] = id
422	}
423	s.write(frame)
424}
425
426func (s *rpcServer) writeEvent(payload map[string]any) {
427	s.write(payload)
428}
429
430func (s *rpcServer) write(v any) {
431	b, err := json.Marshal(v)
432	if err != nil {
433		return
434	}
435	s.writeMu.Lock()
436	defer s.writeMu.Unlock()
437	_, _ = s.out.Write(b)
438	_, _ = s.out.Write([]byte("\n"))
439}
440
441func (s *rpcServer) setCancel(c context.CancelFunc) {
442	s.writeMu.Lock()
443	s.activeCancel = c
444	s.writeMu.Unlock()
445}
446
447func (s *rpcServer) takeCancel() context.CancelFunc {
448	s.writeMu.Lock()
449	defer s.writeMu.Unlock()
450	c := s.activeCancel
451	s.activeCancel = nil
452	return c
453}
454
455// messagesToJSON serialises a transcript using the same schema as the
456// --json event mode for cross-format consistency.
457func messagesToJSON(msgs []provider.Message) []map[string]any {
458	out := make([]map[string]any, 0, len(msgs))
459	for _, m := range msgs {
460		out = append(out, map[string]any{
461			"role":    string(m.Role),
462			"content": modes.ContentToJSON(m.Content),
463			"time":    m.Time,
464		})
465	}
466	return out
467}

Keyboard shortcuts

Single-key shortcuts are on. ctrl k and esc work whatever this says.

ctrl k :
go to a section, branch, tag, file or repository
?
this list
j k
move the row cursor down / up
enter
open the row under the cursor; on Files, a directory opens or closes in place
↑ ↓ ← → home end
in the file tree (focus on a row): previous / next row, close / open a directory or go to its parent / first entry, first / last row. They never open a file
l h
on Files: open a directory or a file / close a directory or go up to its parent, then to the parent page. Elsewhere: into the side panel (on a merge request, its files) / back to the content; there j k move between its controls
g g G
first / last row
/
filter the list; on Files, find a file anywhere in the repository (esc clears, enter opens the best match)
b
switch branch or tag
g s g f g c g b g m g i g r g a
tabs: summary, files, commits, branches & tags, merge requests, issues, releases, actions (when available)
g h
home
w
wide mode: use the whole width of the screen, or not (remembered)
y
copy a permalink pinned to the commit
d u x
diff: split / unified / collapse or expand all
e E
diff: the whole file in view / every file, with the lines the hunks hide
n
merge request: next note in the diff
esc
close a dialog, leave a field

Branches 1

Tags 476

↑↓ moveenter openesc close