Files
ragflow/internal/agent/canvas/runner.go
Zhichang Yu 3f805a64f1 feat(agent): align Go agent behavior with Python (except retrieval component) (#16225)
## Summary

Aligns the **Go agent runtime/canvas/components/tools** behavior with
the **Python `agent/` implementation** so the same stored canvas DSL
produces the same execution result on either side. Every component,
tool, and runtime primitive in `internal/agent/` is now driven by the
same semantics as its Python counterpart — variable resolution, template
substitution, control flow, error reporting, retry/cancel, and stream
event shapes.

The **retrieval component is the one explicit exception** in this PR. It
is being reworked in a separate change and is excluded from this
alignment pass; the wrapper slot (`universe_a_wrappers.go →
newRetrievalComponent`) is preserved.

## Scope of alignment

### Components (all aligned with `agent/component/`)
`Begin` · `Message` · `LLM` (incl. ChatTemplateKwargs,
MessageHistoryWindowSize, VisualFiles, Cite, OutputStructure,
JSONOutput, TopP, MaxRetries, DelayAfterError, credentials) · `Agent`
(react + tool artifact capture + `Reset()` interface-assert) · `Switch`
(12/12 operators, Python-equivalent semantics) · `Categorize` · `Invoke`
· `Iteration` · `Loop` (macro-expansion through `workflowx.AddLoopNode`)
· `UserFillUp` (Python-equivalent interrupt/resume via eino
`compose.Interrupt`/`ResumeWithData`) · `FillUp` · `DataOperations` ·
`ListOperations` · `StringTransform` · `VariableAggregator` ·
`VariableAssigner` · `Browser` (full stagehand runtime parity) ·
`DocsGenerator` · `ExcelProcessor`.

### Tools (all aligned with `agent/tools/`)
`Retrieval` (wrapper slot only — logic out of scope) · `MCPToolAdapter`
(streamable-HTTP) · `CodeExec` (sandbox bridge with
`code_exec_contract.go` matching Python contract) · `AkShare` · `ArXiv`
· `Crawler` · `DeepL` · `DuckDuckGo` · `Email` · `ExeSQL` · `GitHub` ·
`Google` · `GoogleScholar` · `Jin10` · `PubMed` · `QWeather` · `SearXNG`
· `Tavily` · `Tushare` · `Wencai` · `Wikipedia` · `YahooFinance` —
uniform `eino tool.InvokableTool` interface, SSRF protection, shared
HTTP client.

### Canvas execution engine (`internal/agent/canvas/`)
Aligned with Python's `agent/canvas.py`:
- **Scheduler** (`scheduler.go`): state pre/post handlers, node lambdas,
per-component timeout resolver (4-level: per-class env → per-class table
→ uniform env → 600s fallback), `legacyNoOpNames`.
- **Loop subgraph** (`loop_subgraph.go`): Python-equivalent
`AddLoopNode` macro expansion + condition translation.
- **Multibranch** (`multibranch.go`): `Switch` / `Categorize` routing
via `compose.NewGraphMultiBranch` — same branch selection semantics as
Python.
- **Parallel subgraph** (`parallel_subgraph.go`): matches Python's
parallel fan-out contract.
- **Interrupt/Resume** (`interrupt_resume.go`): `UserFillUpNodeBody` /
`IsInterruptError` / `ExtractInterruptContexts` — replaces the
deprecated Python sentinel chain with eino's native interrupt API,
preserving the same external behavior.
- **Checkpoint** (`checkpoint_store.go`): `RedisCheckPointStore`
Get/Set/Delete, with business metadata (status / canvas_id /
parent_run_id) on a parallel Redis Hash.
- **RunTracker** (`run_tracker.go`): Start / MarkSucceeded / MarkFailed
/ MarkCancelled / AttachCheckpoint — same lifecycle as the Python run
record.
- **Cancel** (`cancel.go`): Redis pub/sub watch.
- **Stream** (`stream.go`): SSE channel with `messages` / `waiting` /
`errors` / `done` events, same shape as Python's `agent.canvas.RunEvent`
payload.

### DSL bridge (`internal/agent/dsl/`)
- `normalize.go`: v1↔v2 collapsed into a single wire format — Python and
Go consume the same stored JSON.
- `reset.go`: per-run state reset matches Python's `Canvas.reset()`
semantics.
- Testdata mirrors Python's `agent_msg.json` / `all.json` / etc.

### Runtime (`internal/agent/runtime/`)
- `CanvasState` / `NewCanvasState` / `GetVar` / `SetVar` / `ReadVars`:
same `{{cpn_id@param}}` resolution model.
- `ResolveTemplate` (regex fast path + gonja fallback) — Python
Jinja-style semantics.
- `selector.go`, `metrics.go`, `component.go`: shared runtime contracts.

## Out of scope (intentionally)

- **`Retrieval` component logic** — wrapped only; full parity lands in a
follow-up PR.
- **Frontend** — only minor dsl-bridge / canvas UX fixes ride along.
- **CLI / admin / model registry** — orthogonal to agent behavior.

## How alignment is verified

`internal/service/agent_run_e2e_test.go` exercises the **full production
chain** against real Python-shaped DSL fixtures:
```
loadCanvasForUser → versionDAO.GetLatest → decodeCanvasFromDSL →
canvas.Compile → cc.Workflow.Invoke → answer extraction
```
using in-memory SQLite + miniredis (no Docker). Covers:
- `TestRunAgent_RealCanvas_BeginMessage` — happy path, `{{sys.query}}`
resolution
- `TestRunAgent_RealCanvas_WaitForUserResume` — two-run resume cycle
(Python-equivalent)
- `TestRunAgent_RealCanvas_CompileFails` — unknown component name →
sanitized error (Python-equivalent)
- `TestRunAgent_RealCanvas_InvokeFails` — unresolvable template ref
(Python-equivalent)
- `TestRunAgent_RunTracker_AttachCheckpoint_CallSequence` —
Start→AttachCheckpoint→MarkSucceeded lifecycle

`internal/handler/agent_test.go` — SSE streaming parity (`Content-Type:
text/event-stream`, `data: {…}\n\n`, trailing `data: [DONE]\n\n`,
OpenAI-compatible non-stream `choices`).

`internal/agent/canvas/fixture_compile_test.go` + per-component tests
pin the Python-equivalent outputs.

```
go test -count=1 -v -run 'TestRunAgent_RealCanvas|TestRunAgent_RunTracker' ./internal/service/
```

## Design reference

`docs/develop/agent-go-port-design.md` (1329 lines, last cross-checked
2026-06-17) — module layout, per-component / per-tool inventory,
corner-case catalogue, and the actionable backlog (Section 14, including
the retrieval alignment follow-up).

---------

Co-authored-by: Claude <noreply@anthropic.com>
2026-06-22 11:58:29 +08:00

502 lines
18 KiB
Go

//
// Copyright 2026 The InfiniFlow Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
//
// runner.go — Canvas execution runtime. Drives a Canvas invocation
// (the caller supplies the RunFunc that does Compile+Invoke), catches
// the four possible outcomes, and surfaces them as RunEvent values on
// a channel that the HTTP layer streams as SSE frames.
//
// Why this file lives in the canvas package: it is the runtime twin
// of scheduler.go (BuildWorkflow = "how to build", Runner = "how to
// drive"). Both concern the Canvas execution lifecycle; nothing
// outside the canvas package needs to know that these concerns are
// split across two files.
//
// Run outcomes — four paths on a single Run() call:
//
// 1. Normal completion (runErr == nil): the buildRunFunc already
// emitted all workflow events (workflow_started, node_started,
// node_finished, message, message_end, workflow_finished) during
// execution. The Runner just sends the `done` terminator.
// 2. Eino interrupt (runErr is an *InterruptSignal or wrapped
// variant): emit `waiting_for_user` with the first interrupt
// id. Persist the id so the next call can resume via
// compose.ResumeWithData (signalled through root:
// __resume_interrupt_id__ + __resume_data__).
// 3. Cancel / timeout (errors.Is(err, context.Canceled) etc.):
// silently close. The HTTP handler has already detached.
// 4. Other errors: emit `error` event with the err.Error() string.
//
// SSE wire contract (matches the handler envelope):
// - RunEvent.Type == "message" → {data: <string>}
// - RunEvent.Type == "waiting_for_user" → {cpn_id: <string>}
// - RunEvent.Type == "error" → {message: <string>}
// - RunEvent.Type == "done" → final terminator frame
package canvas
import (
"context"
"encoding/json"
"errors"
"fmt"
"log"
"runtime/debug"
"strings"
"sync"
"time"
"github.com/google/uuid"
)
// RunEvent is the unit the Runner pushes onto its output channel.
// The handler converts each RunEvent into one SSE frame in the
// Python-shaped envelope:
//
// data:{"event":"<Type>","message_id":"<MessageID>","created_at":<CreatedAt>,"task_id":"<TaskID>","session_id":"<SessionID>","data":<Data>}
//
// Type is the event tag; Data is the JSON payload string (already
// serialised — handler does not re-marshal). The handler wraps Data
// into the "data" field of the outer envelope so the front-end's
// use-send-message.ts parser sees a flat {event, message_id,
// created_at, task_id, session_id, data} object on every frame.
type RunEvent struct {
Type string
Data string
MessageID string
CreatedAt int64
TaskID string
SessionID string
}
// NodeStartedData is the "data" payload for "node_started" events.
type NodeStartedData struct {
Inputs interface{} `json:"inputs"`
CreatedAt float64 `json:"created_at"`
ComponentID string `json:"component_id"`
ComponentName string `json:"component_name"`
ComponentType string `json:"component_type"`
Thoughts string `json:"thoughts"`
}
// NodeFinishedData is the "data" payload for "node_finished" events.
type NodeFinishedData struct {
Inputs interface{} `json:"inputs"`
Outputs interface{} `json:"outputs"`
ComponentID string `json:"component_id"`
ComponentName string `json:"component_name"`
ComponentType string `json:"component_type"`
Error interface{} `json:"error"`
ElapsedTime float64 `json:"elapsed_time"`
CreatedAt float64 `json:"created_at"`
}
// MessageEvent is the JSON payload for Type=="message" frames.
type MessageEvent struct {
Content string `json:"content"`
Reference []interface{} `json:"reference,omitempty"`
}
// MessageEndEvent is the JSON payload for Type=="message_end" frames.
type MessageEndEvent struct {
Status *string `json:"status,omitempty"`
Attachment []interface{} `json:"attachment,omitempty"`
Reference []interface{} `json:"reference,omitempty"`
}
// WaitingForUserEvent is the JSON payload for Type=="waiting_for_user"
// frames. CpnID is the cpn id that emitted the wait sentinel — the
// front-end can use it to surface the prompt or to attach the
// follow-up to the right conversation turn.
type WaitingForUserEvent struct {
CpnID string `json:"cpn_id"`
Tips string `json:"tips,omitempty"`
Inputs map[string]any `json:"inputs,omitempty"`
}
// ErrorEvent is the JSON payload for Type=="error" frames.
type ErrorEvent struct {
Message string `json:"message"`
}
// RunFunc is the canvas execution contract the Runner depends on.
// Service-layer code supplies an implementation that compiles the
// DSL and invokes the eino Workflow; the Runner is agnostic to
// that machinery.
//
// Return contract:
//
// - nil error, non-nil state: run completed normally.
// - non-nil error that is an eino interrupt signal: the run paused
// on a wait-for-user node. The Runner extracts the InterruptCtx
// list via ExtractInterruptContexts and emits a `waiting_for_user`
// event. state may be nil in this branch (the engine does not
// surface a completed state when it halts on an interrupt).
// - any other non-nil error: run failed; surface as `error` event.
type RunFunc func(ctx context.Context, root map[string]any) (*CanvasState, error)
// Runner is the per-canvas execution runtime. It owns the
// interrupt-id map (V1 in-memory persistence keyed by
// (canvasID, sessionID)) and the goroutine cancellation registry.
//
// Concurrency: Runner methods are safe for concurrent use. The
// output channel is owned by the goroutine that started a run;
// the Cancel method signals the underlying run via the cancel
// channel that the RunFunc is expected to observe.
type Runner struct {
mu sync.Mutex
interruptIDs map[string]string // key = canvasID + "|" + sessionID; value = eino interrupt id
runCancels map[string]chan struct{}
}
// NewRunner returns a fresh Runner with the in-memory interrupt-id
// map initialised. The Runner has no background goroutines; it is
// owned by the AgentService.
func NewRunner() *Runner {
return &Runner{
interruptIDs: make(map[string]string),
runCancels: make(map[string]chan struct{}),
}
}
// sessionKey is the lookup key for the in-memory interrupt-id map. We
// concatenate with a separator that cannot appear in either id (the
// id format is uuid-hex) so two adjacent ids never collide.
func sessionKey(canvasID, sessionID string) string {
return canvasID + "|" + sessionID
}
// saveInterruptID stores the eino interrupt id for a (canvasID,
// sessionID) pair. Called when the RunFunc returns an interrupt
// error; the next RunAgent call with the same session id reads it
// back via getInterruptID and forwards it to the RunFunc so the
// RunFunc can target it via compose.ResumeWithData.
func (r *Runner) saveInterruptID(canvasID, sessionID, interruptID string) {
if interruptID == "" {
return
}
r.mu.Lock()
r.interruptIDs[sessionKey(canvasID, sessionID)] = interruptID
r.mu.Unlock()
}
// getInterruptID reads back the interrupt id saved by the previous
// run, then deletes it (the resume consumes it). Returns "" when no
// prior paused run exists for this session.
func (r *Runner) getInterruptID(canvasID, sessionID string) string {
r.mu.Lock()
id, ok := r.interruptIDs[sessionKey(canvasID, sessionID)]
if ok {
delete(r.interruptIDs, sessionKey(canvasID, sessionID))
}
r.mu.Unlock()
return id
}
// Run drives one canvas invocation. See package docstring for the
// four-outcome flow. The channel is always closed on return so the
// handler's for-range loop terminates.
//
// Metadata injection: the output channel, message_id, task_id, and
// session_id are injected into root so the RunFunc (buildRunFunc in
// service/agent.go) can emit intermediate events (workflow_started,
// node_started, node_finished, workflow_finished) during execution
// rather than only after the invoke completes. The key names follow
// the __<name>__ sentinel convention to avoid collisions with
// runtime DSL keys.
func (r *Runner) Run(
ctx context.Context,
run RunFunc,
canvasID, sessionID, userInput string,
root map[string]any,
) <-chan RunEvent {
out := make(chan RunEvent, 8)
if run == nil {
pushErr(out, "canvas: nil RunFunc")
close(out)
return out
}
cancel := make(chan struct{})
r.mu.Lock()
if prev, hadPrev := r.runCancels[canvasID]; hadPrev {
select {
case <-prev:
default:
close(prev)
}
}
r.runCancels[canvasID] = cancel
r.mu.Unlock()
// Generate the identifiers the RunFunc and SSE envelope need.
// message_id is generated per-run so the front-end can correlate
// all events for a single user turn. task_id is the published
// version id (if available) or a per-run UUID.
messageID := strings.ReplaceAll(uuid.New().String(), "-", "")
taskID := ""
if v, ok := root["version_id"].(string); ok && v != "" {
taskID = v
}
if taskID == "" {
taskID = strings.ReplaceAll(uuid.New().String(), "-", "")
}
// Inject the output channel + metadata so the RunFunc can emit
// events during execution (workflow_started, node_started,
// node_finished, etc.).
root["__events__"] = out
root["__message_id__"] = messageID
root["__task_id__"] = taskID
root["__session_id__"] = sessionID
go func() {
defer close(out)
defer func() {
r.mu.Lock()
if r.runCancels[canvasID] == cancel {
delete(r.runCancels, canvasID)
}
r.mu.Unlock()
}()
// Panic sentinel (temporary diagnostic — see plan):
// a panic anywhere in the run goroutine used to silently
// propagate, leaving the events channel closed-empty so the
// SSE handler streamed a 200 OK with an empty body. We now
// log the panic value + stack trace so the next failing run
// surfaces a clear root cause in the server log.
defer func() {
if rec := recover(); rec != nil {
log.Printf("canvas runner PANIC canvas=%q session=%q: %v\n%s", canvasID, sessionID, rec, debug.Stack())
}
}()
// Resume path: inject the previously-saved interrupt id and
// the user's follow-up into root. The RunFunc reads these
// keys and decorates ctx with compose.ResumeWithData before
// invoking the workflow. The sentinel keys are deleted from
// root inside the RunFunc — see service/agent.go's
// buildRunFunc.
if userInput != "" {
if id := r.getInterruptID(canvasID, sessionID); id != "" {
root["__resume_interrupt_id__"] = id
root["__resume_data__"] = userInput
}
}
_, runErr := safeInvoke(ctx, cancel, run, root)
if runErr != nil {
if errors.Is(runErr, context.Canceled) || errors.Is(runErr, errCancelled) {
return
}
if ctxs := ExtractInterruptContexts(runErr); len(ctxs) > 0 {
// Wait-for-user: persist the real root-cause interrupt id for
// compose.ResumeWithData, but keep exposing the leaf
// user_fill_up interrupt id to the front-end so it can attach
// the prompt to the visible waiting node.
displayID := FirstInterruptID(ctxs)
resumeID := RootInterruptID(ctxs)
log.Printf("canvas runner interrupt canvas=%q session=%q task=%q contexts=%s display=%q resume=%q", canvasID, sessionID, taskID, formatInterruptContexts(ctxs), displayID, resumeID)
r.saveInterruptID(canvasID, sessionID, resumeID)
waiting := WaitingForUserEvent{CpnID: displayID}
if ctx := FirstUserFillUpInterrupt(ctxs); ctx != nil {
if info, ok := ctx.Info.(map[string]any); ok {
if tips, _ := info["tips"].(string); tips != "" {
waiting.Tips = tips
}
if inputs, ok := info["inputs"].(map[string]any); ok && len(inputs) > 0 {
waiting.Inputs = inputs
}
}
}
payload, _ := json.Marshal(waiting)
push(out, RunEvent{Type: "waiting_for_user", Data: string(payload), MessageID: messageID, CreatedAt: nowUnix(), TaskID: taskID, SessionID: sessionID})
// Always close a RunAgent call with the `done`
// terminator so the front-end can rely on a
// channel-end sentinel regardless of whether the run
// completed, errored, or paused for user input.
push(out, RunEvent{Type: "done", Data: "", MessageID: messageID, CreatedAt: nowUnix(), TaskID: taskID, SessionID: sessionID})
return
}
if IsInterruptError(runErr) {
// Raw InterruptSignal (no wrapped InterruptCtx list
// available). Emit a generic waiting_for_user event
// without a cpn id — the front-end falls back to
// the first paused session it knows about.
r.saveInterruptID(canvasID, sessionID, runErr.Error())
payload, _ := json.Marshal(WaitingForUserEvent{CpnID: runErr.Error()})
push(out, RunEvent{Type: "waiting_for_user", Data: string(payload), MessageID: messageID, CreatedAt: nowUnix(), TaskID: taskID, SessionID: sessionID})
push(out, RunEvent{Type: "done", Data: "", MessageID: messageID, CreatedAt: nowUnix(), TaskID: taskID, SessionID: sessionID})
return
}
pushErr(out, runErr.Error())
// Close the channel with the `done` terminator so the
// front-end sees a channel-end sentinel on the error
// path too — matches the contract for completed and
// waiting-for-user paths above.
push(out, RunEvent{Type: "done", Data: "", MessageID: messageID, CreatedAt: nowUnix(), TaskID: taskID, SessionID: sessionID})
return
}
// Normal completion — the buildRunFunc already emitted the
// workflow events during execution. Runner just sends the
// terminator.
push(out, RunEvent{Type: "done", Data: "", MessageID: messageID, CreatedAt: nowUnix(), TaskID: taskID, SessionID: sessionID})
}()
return out
}
// Cancel signals an in-flight run for the given canvas to stop.
// Safe to call when no run is active.
func (r *Runner) Cancel(canvasID string) {
r.mu.Lock()
cancel, ok := r.runCancels[canvasID]
r.mu.Unlock()
if !ok {
return
}
select {
case <-cancel:
default:
close(cancel)
}
}
// Peek reports whether a paused interrupt id is held for the given
// (canvasID, sessionID). It is intended for tests and diagnostics;
// the real runner does not need it at run time.
func (r *Runner) Peek(canvasID, sessionID string) bool {
r.mu.Lock()
_, ok := r.interruptIDs[sessionKey(canvasID, sessionID)]
r.mu.Unlock()
return ok
}
// errCancelled is the sentinel safeInvoke returns when the cancel
// channel fires during a run. It is wrapped against context.Canceled
// so callers can `errors.Is` either.
var errCancelled = fmt.Errorf("canvas: run cancelled")
// safeInvoke calls the supplied RunFunc with context-cancel and
// driver-cancel both wired in. The RunFunc is expected to honour
// ctx.Done() — the cancel channel is a secondary signal for the
// V1 in-process driver.
func safeInvoke(ctx context.Context, cancel chan struct{}, run RunFunc, root map[string]any) (*CanvasState, error) {
done := make(chan struct{})
var (
state *CanvasState
err error
)
go func() {
// Recover here, inside the goroutine that actually invokes
// `run`. A panic from `run` would otherwise crash the process
// before any caller could observe it; converting it into a
// regular error keeps the SSE contract intact and lets the
// runner emit a terminal `done` event.
defer func() {
if rec := recover(); rec != nil {
log.Printf("canvas runner PANIC: %v\n%s", rec, debug.Stack())
err = fmt.Errorf("canvas runner panic: %v", rec)
}
close(done)
}()
state, err = run(ctx, root)
}()
select {
case <-done:
return state, err
case <-cancel:
return nil, errCancelled
}
}
// PushEvent sends an event to the channel, dropping it if the consumer
// has gone away (handler cancelled). Exported so the service layer's
// buildRunFunc can emit intermediate workflow events through the
// same channel during execution.
func PushEvent(ch chan<- RunEvent, ev RunEvent) {
defer func() { _ = recover() }()
ch <- ev
}
// push sends an event to the channel, dropping it if the consumer
// has gone away (handler cancelled). Errors on send are intentional
// and ignored — the handler is the only consumer and its
// `for-range` loop exits when the request context is cancelled.
func push(out chan<- RunEvent, ev RunEvent) {
defer func() { _ = recover() }()
out <- ev
}
// pushErr serialises an ErrorEvent and pushes it on the channel.
func pushErr(out chan<- RunEvent, msg string) {
payload, _ := json.Marshal(ErrorEvent{Message: msg})
push(out, RunEvent{Type: "error", Data: string(payload)})
}
// nowUnix returns the current Unix timestamp in seconds.
func nowUnix() int64 {
return time.Now().Unix()
}
// extractAnswerFromState is kept for reference but is no longer called
// by the Runner — answer extraction now happens in buildRunFunc.
// Remove in a follow-up cleanup pass once all tests pass.
func extractAnswerFromState(state *CanvasState) (string, []interface{}) {
if state == nil {
return "", nil
}
snap := state.Snapshot()
var answer string
var reference []interface{}
// First pass: look for an "answer" key (preferred).
for _, bucket := range snap {
if a, ok := bucket["answer"].(string); ok && a != "" {
answer = a
break
}
}
// Second pass: fall back to "result" then "content" if
// no "answer" was found.
if answer == "" {
for _, bucket := range snap {
if r, ok := bucket["result"].(string); ok && r != "" {
answer = r
break
}
}
}
if answer == "" {
for _, bucket := range snap {
if c, ok := bucket["content"].(string); ok && c != "" {
answer = c
break
}
}
}
// Collect references (best-effort, no precedence).
for _, bucket := range snap {
if r, ok := bucket["reference"].([]interface{}); ok {
reference = append(reference, r...)
}
}
if answer == "" {
answer = "Run completed with no surfaceable answer."
}
return answer, reference
}