a3s_flow/engine/
mod.rs

1use chrono::{DateTime, Utc};
2use std::sync::Arc;
3use uuid::Uuid;
4
5use crate::error::{FlowError, Result};
6use crate::model::{
7    project_run, ActiveHookSnapshot, FlowEvent, FlowEventEnvelope, HookStatus, RuntimeCommand,
8    ScheduledWakeup, ScheduledWakeupKind, StepStatus, WaitStatus, WorkflowRunSnapshot,
9    WorkflowRunStatus, WorkflowRunSummary, WorkflowRunSuspension, WorkflowSpec,
10};
11use crate::observe::{FlowEventObserver, NoopFlowEventObserver};
12use crate::runtime::{FlowRuntime, WorkflowInvocation};
13use crate::store::{FlowEventStore, InMemoryEventStore};
14
15mod operations;
16mod steps;
17mod validation;
18use steps::StepExecutionContext;
19use validation::{
20    ensure_child_operation_matches, ensure_hook_command_matches, ensure_progress_matches,
21    ensure_same_start, ensure_step_batch_valid, ensure_step_command_matches,
22    ensure_wait_command_matches, is_event_conflict, validate_run_id,
23};
24
25/// Builder for a [`FlowEngine`].
26pub struct FlowEngineBuilder {
27    store: Arc<dyn FlowEventStore>,
28    runtime: Arc<dyn FlowRuntime>,
29    observer: Arc<dyn FlowEventObserver>,
30    max_replay_iterations: usize,
31}
32
33impl FlowEngineBuilder {
34    pub fn new(runtime: Arc<dyn FlowRuntime>) -> Self {
35        Self {
36            store: Arc::new(InMemoryEventStore::new()),
37            runtime,
38            observer: Arc::new(NoopFlowEventObserver),
39            max_replay_iterations: 1024,
40        }
41    }
42
43    pub fn with_store(mut self, store: Arc<dyn FlowEventStore>) -> Self {
44        self.store = store;
45        self
46    }
47
48    pub fn with_observer(mut self, observer: Arc<dyn FlowEventObserver>) -> Self {
49        self.observer = observer;
50        self
51    }
52
53    pub fn with_max_replay_iterations(mut self, max_replay_iterations: usize) -> Self {
54        self.max_replay_iterations = max_replay_iterations.max(1);
55        self
56    }
57
58    pub fn build(self) -> FlowEngine {
59        FlowEngine {
60            store: self.store,
61            runtime: self.runtime,
62            observer: self.observer,
63            max_replay_iterations: self.max_replay_iterations,
64        }
65    }
66}
67
68/// Event-sourced workflow engine.
69#[derive(Clone)]
70pub struct FlowEngine {
71    store: Arc<dyn FlowEventStore>,
72    runtime: Arc<dyn FlowRuntime>,
73    observer: Arc<dyn FlowEventObserver>,
74    max_replay_iterations: usize,
75}
76
77impl FlowEngine {
78    pub fn builder(runtime: Arc<dyn FlowRuntime>) -> FlowEngineBuilder {
79        FlowEngineBuilder::new(runtime)
80    }
81
82    pub fn new(store: Arc<dyn FlowEventStore>, runtime: Arc<dyn FlowRuntime>) -> Self {
83        Self {
84            store,
85            runtime,
86            observer: Arc::new(NoopFlowEventObserver),
87            max_replay_iterations: 1024,
88        }
89    }
90
91    pub fn in_memory(runtime: Arc<dyn FlowRuntime>) -> Self {
92        Self::new(Arc::new(InMemoryEventStore::new()), runtime)
93    }
94
95    pub fn store(&self) -> Arc<dyn FlowEventStore> {
96        Arc::clone(&self.store)
97    }
98
99    pub fn observer(&self) -> Arc<dyn FlowEventObserver> {
100        Arc::clone(&self.observer)
101    }
102
103    /// Start a workflow run and drive it until completion or suspension.
104    pub async fn start(&self, spec: WorkflowSpec, input: serde_json::Value) -> Result<String> {
105        let run_id = Uuid::new_v4().to_string();
106        self.start_with_id(run_id, spec, input).await
107    }
108
109    /// Start a workflow run using a caller-provided durable run id.
110    ///
111    /// Reusing the same `run_id` with the same workflow spec and input is
112    /// idempotent. Reusing it with different spec or input returns a conflict.
113    pub async fn start_with_id(
114        &self,
115        run_id: impl Into<String>,
116        spec: WorkflowSpec,
117        input: serde_json::Value,
118    ) -> Result<String> {
119        spec.validate()?;
120        let run_id = run_id.into();
121        validate_run_id(&run_id)?;
122
123        for _ in 0..self.max_replay_iterations {
124            match self.store.list(&run_id).await {
125                Ok(history) => {
126                    let snapshot = project_run(&run_id, &history)?;
127                    ensure_same_start(&run_id, &snapshot, &spec, &input)?;
128                    if !history
129                        .iter()
130                        .any(|event| matches!(event.event, FlowEvent::RunStarted))
131                    {
132                        match self
133                            .record_event_at(&run_id, snapshot.last_sequence, FlowEvent::RunStarted)
134                            .await
135                        {
136                            Ok(_) => {}
137                            Err(err) if is_event_conflict(&err) => continue,
138                            Err(err) => return Err(err),
139                        }
140                    }
141                    match self.drive(&run_id).await {
142                        Ok(_) => return Ok(run_id),
143                        Err(err) if is_event_conflict(&err) => continue,
144                        Err(err) => return Err(err),
145                    }
146                }
147                Err(FlowError::RunNotFound(_)) => {
148                    let created = match self
149                        .record_event_at(
150                            &run_id,
151                            0,
152                            FlowEvent::RunCreated {
153                                spec: spec.clone(),
154                                input: input.clone(),
155                            },
156                        )
157                        .await
158                    {
159                        Ok(created) => created,
160                        Err(err) if is_event_conflict(&err) => continue,
161                        Err(err) => return Err(err),
162                    };
163                    match self
164                        .record_event_at(&run_id, created.sequence, FlowEvent::RunStarted)
165                        .await
166                    {
167                        Ok(_) => {}
168                        Err(err) if is_event_conflict(&err) => continue,
169                        Err(err) => return Err(err),
170                    }
171                    match self.drive(&run_id).await {
172                        Ok(_) => return Ok(run_id),
173                        Err(err) if is_event_conflict(&err) => continue,
174                        Err(err) => return Err(err),
175                    }
176                }
177                Err(err) => return Err(err),
178            }
179        }
180
181        Err(FlowError::ReplayLimitExceeded(self.max_replay_iterations))
182    }
183
184    /// Resume a wait once its timer has fired.
185    pub async fn resume_wait(&self, run_id: &str, wait_id: &str) -> Result<()> {
186        for _ in 0..self.max_replay_iterations {
187            let snapshot = self.snapshot(run_id).await?;
188            if snapshot.status.is_terminal() {
189                return Err(FlowError::RunTerminal(run_id.to_string()));
190            }
191            match snapshot.waits.get(wait_id) {
192                Some(wait) if wait.status == WaitStatus::Waiting => {
193                    match self
194                        .record_event_at(
195                            run_id,
196                            snapshot.last_sequence,
197                            FlowEvent::WaitCompleted {
198                                wait_id: wait_id.to_string(),
199                            },
200                        )
201                        .await
202                    {
203                        Ok(_) => {}
204                        Err(err) if is_event_conflict(&err) => continue,
205                        Err(err) => return Err(err),
206                    }
207                    match self.drive(run_id).await {
208                        Ok(_) => return Ok(()),
209                        Err(err) if is_event_conflict(&err) => continue,
210                        Err(err) => return Err(err),
211                    }
212                }
213                Some(_) => match self.drive(run_id).await {
214                    Ok(_) => return Ok(()),
215                    Err(err) if is_event_conflict(&err) => continue,
216                    Err(err) => return Err(err),
217                },
218                None => {
219                    return Err(FlowError::InvalidTransition(format!(
220                        "wait {wait_id} does not exist for run {run_id}"
221                    )))
222                }
223            }
224        }
225
226        Err(FlowError::ReplayLimitExceeded(self.max_replay_iterations))
227    }
228
229    /// Resume an active hook with external payload.
230    pub async fn resume_hook(
231        &self,
232        run_id: &str,
233        hook_id: &str,
234        payload: serde_json::Value,
235    ) -> Result<()> {
236        for _ in 0..self.max_replay_iterations {
237            let snapshot = self.snapshot(run_id).await?;
238            if snapshot.status.is_terminal() {
239                return Err(FlowError::RunTerminal(run_id.to_string()));
240            }
241            match snapshot.hooks.get(hook_id) {
242                Some(hook) if hook.status == HookStatus::Active => {
243                    match self
244                        .record_event_at(
245                            run_id,
246                            snapshot.last_sequence,
247                            FlowEvent::HookReceived {
248                                hook_id: hook_id.to_string(),
249                                payload: payload.clone(),
250                            },
251                        )
252                        .await
253                    {
254                        Ok(_) => {}
255                        Err(err) if is_event_conflict(&err) => continue,
256                        Err(err) => return Err(err),
257                    }
258                    match self.drive(run_id).await {
259                        Ok(_) => return Ok(()),
260                        Err(err) if is_event_conflict(&err) => continue,
261                        Err(err) => return Err(err),
262                    }
263                }
264                Some(_) => match self.drive(run_id).await {
265                    Ok(_) => return Ok(()),
266                    Err(err) if is_event_conflict(&err) => continue,
267                    Err(err) => return Err(err),
268                },
269                None => {
270                    return Err(FlowError::InvalidTransition(format!(
271                        "hook {hook_id} does not exist for run {run_id}"
272                    )))
273                }
274            }
275        }
276
277        Err(FlowError::ReplayLimitExceeded(self.max_replay_iterations))
278    }
279
280    /// Dispose an active hook without accepting a callback payload.
281    ///
282    /// This is useful when a host withdraws an approval request, expires a
283    /// webhook token, or closes an external callback route. The workflow is
284    /// driven after the disposal event so replay code can observe
285    /// [`WorkflowContext::hook_disposed`](crate::context::WorkflowContext::hook_disposed)
286    /// and complete, fail, or schedule an alternate path.
287    pub async fn dispose_hook(&self, run_id: &str, hook_id: &str) -> Result<()> {
288        for _ in 0..self.max_replay_iterations {
289            let snapshot = self.snapshot(run_id).await?;
290            if snapshot.status.is_terminal() {
291                return Err(FlowError::RunTerminal(run_id.to_string()));
292            }
293            match snapshot.hooks.get(hook_id) {
294                Some(hook) if hook.status == HookStatus::Active => {
295                    match self
296                        .record_event_at(
297                            run_id,
298                            snapshot.last_sequence,
299                            FlowEvent::HookDisposed {
300                                hook_id: hook_id.to_string(),
301                            },
302                        )
303                        .await
304                    {
305                        Ok(_) => {}
306                        Err(err) if is_event_conflict(&err) => continue,
307                        Err(err) => return Err(err),
308                    }
309                    match self.drive(run_id).await {
310                        Ok(_) => return Ok(()),
311                        Err(err) if is_event_conflict(&err) => continue,
312                        Err(err) => return Err(err),
313                    }
314                }
315                Some(_) => match self.drive(run_id).await {
316                    Ok(_) => return Ok(()),
317                    Err(err) if is_event_conflict(&err) => continue,
318                    Err(err) => return Err(err),
319                },
320                None => {
321                    return Err(FlowError::InvalidTransition(format!(
322                        "hook {hook_id} does not exist for run {run_id}"
323                    )))
324                }
325            }
326        }
327
328        Err(FlowError::ReplayLimitExceeded(self.max_replay_iterations))
329    }
330
331    /// Resume an active hook by its external token.
332    ///
333    /// This is the API webhook handlers normally want: the callback receives a
334    /// token, while `run_id` and `hook_id` remain engine internals.
335    pub async fn resume_hook_by_token(
336        &self,
337        token: &str,
338        payload: serde_json::Value,
339    ) -> Result<(String, String)> {
340        let mut matches = self
341            .store
342            .find_active_hooks_by_token(token)
343            .await?
344            .into_iter()
345            .map(|active| (active.run_id, active.hook.hook_id))
346            .collect::<Vec<_>>();
347
348        match matches.len() {
349            0 => Err(FlowError::HookTokenNotFound(token.to_string())),
350            1 => {
351                let (run_id, hook_id) = matches.remove(0);
352                self.resume_hook(&run_id, &hook_id, payload).await?;
353                Ok((run_id, hook_id))
354            }
355            _ => Err(FlowError::InvalidTransition(
356                "hook token is active in multiple runs (value redacted)".to_string(),
357            )),
358        }
359    }
360
361    /// Dispose an active hook by its external token.
362    ///
363    /// This mirrors [`resume_hook_by_token`](Self::resume_hook_by_token) for
364    /// callback routers that only know the public token.
365    pub async fn dispose_hook_by_token(&self, token: &str) -> Result<(String, String)> {
366        let mut matches = self
367            .store
368            .find_active_hooks_by_token(token)
369            .await?
370            .into_iter()
371            .map(|active| (active.run_id, active.hook.hook_id))
372            .collect::<Vec<_>>();
373
374        match matches.len() {
375            0 => Err(FlowError::HookTokenNotFound(token.to_string())),
376            1 => {
377                let (run_id, hook_id) = matches.remove(0);
378                self.dispose_hook(&run_id, &hook_id).await?;
379                Ok((run_id, hook_id))
380            }
381            _ => Err(FlowError::InvalidTransition(
382                "hook token is active in multiple runs (value redacted)".to_string(),
383            )),
384        }
385    }
386
387    /// List active waits whose `resume_at` is at or before `now`.
388    ///
389    /// Scheduler integrations can use this to inspect due timers before
390    /// deciding how aggressively to drive them.
391    pub async fn list_due_waits(&self, now: DateTime<Utc>) -> Result<Vec<(String, String)>> {
392        let mut due = self
393            .list_due_wakeups(now)
394            .await?
395            .into_iter()
396            .filter(|wakeup| wakeup.kind == ScheduledWakeupKind::Wait)
397            .map(|wakeup| (wakeup.run_id, wakeup.subject_id))
398            .collect::<Vec<_>>();
399        due.sort();
400        Ok(due)
401    }
402
403    /// Complete every due wait and drive the affected workflows.
404    ///
405    /// Returns the `(run_id, wait_id)` pairs that were resumed. A wait already
406    /// completed by another caller is skipped by [`Self::resume_wait`].
407    pub async fn resume_due_waits(&self, now: DateTime<Utc>) -> Result<Vec<(String, String)>> {
408        let due = self.list_due_waits(now).await?;
409        let mut resumed = Vec::with_capacity(due.len());
410        for (run_id, wait_id) in due {
411            self.resume_wait(&run_id, &wait_id).await?;
412            resumed.push((run_id, wait_id));
413        }
414        Ok(resumed)
415    }
416
417    /// List pending step retries whose `retry_after` is at or before `now`.
418    pub async fn list_due_retries(&self, now: DateTime<Utc>) -> Result<Vec<(String, String)>> {
419        let mut due = self
420            .list_due_wakeups(now)
421            .await?
422            .into_iter()
423            .filter(|wakeup| wakeup.kind == ScheduledWakeupKind::Retry)
424            .map(|wakeup| (wakeup.run_id, wakeup.subject_id))
425            .collect::<Vec<_>>();
426        due.sort();
427        Ok(due)
428    }
429
430    /// List all due wait timers and delayed retries through the store boundary.
431    pub async fn list_due_wakeups(&self, now: DateTime<Utc>) -> Result<Vec<ScheduledWakeup>> {
432        let mut wakeups = self.store.list_due_wakeups(now).await?;
433        wakeups.sort_by(|left, right| {
434            (left.kind, left.run_id.as_str(), left.subject_id.as_str()).cmp(&(
435                right.kind,
436                right.run_id.as_str(),
437                right.subject_id.as_str(),
438            ))
439        });
440        Ok(wakeups)
441    }
442
443    /// Drive every run with a due step retry.
444    pub async fn resume_due_retries(&self, now: DateTime<Utc>) -> Result<Vec<(String, String)>> {
445        let due = self.list_due_retries(now).await?;
446        let mut run_ids = Vec::new();
447        for (run_id, _) in &due {
448            if !run_ids.contains(run_id) {
449                run_ids.push(run_id.clone());
450            }
451        }
452        for run_id in run_ids {
453            self.drive_at(&run_id, now).await?;
454        }
455        Ok(due)
456    }
457
458    pub async fn snapshot(&self, run_id: &str) -> Result<WorkflowRunSnapshot> {
459        let history = self.store.list(run_id).await?;
460        project_run(run_id, &history)
461    }
462
463    pub async fn history(&self, run_id: &str) -> Result<Vec<FlowEventEnvelope>> {
464        self.store.list(run_id).await
465    }
466
467    pub async fn list_run_ids(&self) -> Result<Vec<String>> {
468        self.store.list_run_ids().await
469    }
470
471    pub async fn list_snapshots(&self) -> Result<Vec<WorkflowRunSnapshot>> {
472        let mut snapshots = Vec::new();
473        for run_id in self.store.list_run_ids().await? {
474            snapshots.push(self.snapshot(&run_id).await?);
475        }
476        Ok(snapshots)
477    }
478
479    /// Summarize run state across the active store.
480    ///
481    /// Suspension counters include only non-terminal runs, so a cancelled run
482    /// that still has an old wait or hook in history is not reported as
483    /// actionable work.
484    pub async fn run_summary(&self) -> Result<WorkflowRunSummary> {
485        let snapshots = self.list_snapshots().await?;
486        Ok(WorkflowRunSummary::from_snapshots(&snapshots))
487    }
488
489    /// List open waits, active hooks, and pending delayed retries.
490    ///
491    /// The `due` flag on wait and retry suspensions is computed against `now`.
492    /// Terminal runs are skipped so cancelled histories do not produce
493    /// actionable operator work.
494    pub async fn list_open_suspensions(
495        &self,
496        now: DateTime<Utc>,
497    ) -> Result<Vec<WorkflowRunSuspension>> {
498        let mut suspensions = Vec::new();
499        for run_id in self.store.list_run_ids().await? {
500            let snapshot = self.snapshot(&run_id).await?;
501            if snapshot.status.is_terminal() {
502                continue;
503            }
504            for wait in snapshot.waits.values() {
505                if wait.status == WaitStatus::Waiting {
506                    suspensions.push(WorkflowRunSuspension::Wait {
507                        run_id: run_id.clone(),
508                        wait: wait.clone(),
509                        due: wait.resume_at <= now,
510                    });
511                }
512            }
513            for hook in snapshot.hooks.values() {
514                if hook.status == HookStatus::Active {
515                    suspensions.push(WorkflowRunSuspension::Hook {
516                        run_id: run_id.clone(),
517                        hook: hook.clone(),
518                    });
519                }
520            }
521            for step in snapshot.steps.values() {
522                if step.status == StepStatus::Pending {
523                    if let Some(retry_after) = step.retry_after {
524                        suspensions.push(WorkflowRunSuspension::Retry {
525                            run_id: run_id.clone(),
526                            step: step.clone(),
527                            due: retry_after <= now,
528                        });
529                    }
530                }
531            }
532        }
533        suspensions.sort_by(|left, right| {
534            (left.run_id(), left.kind_order(), left.subject_id()).cmp(&(
535                right.run_id(),
536                right.kind_order(),
537                right.subject_id(),
538            ))
539        });
540        Ok(suspensions)
541    }
542
543    /// Return the earliest open wait or delayed retry across non-terminal runs.
544    ///
545    /// This is useful for hosts that want to sleep until the next scheduler tick
546    /// instead of polling at a fixed interval. Active hooks are intentionally
547    /// ignored because they do not have a scheduled wake-up time.
548    pub async fn next_wakeup(&self, now: DateTime<Utc>) -> Result<Option<WorkflowRunSuspension>> {
549        for _ in 0..2 {
550            let Some(wakeup) = self.store.next_scheduled_wakeup().await? else {
551                return Ok(None);
552            };
553            match self.snapshot(&wakeup.run_id).await {
554                Ok(snapshot) => {
555                    if let Some(suspension) = resolve_scheduled_wakeup(&snapshot, &wakeup, now) {
556                        return Ok(Some(suspension));
557                    }
558                }
559                Err(FlowError::RunNotFound(_)) => {}
560                Err(error) => return Err(error),
561            }
562        }
563
564        self.next_wakeup_by_replay(now).await
565    }
566
567    async fn next_wakeup_by_replay(
568        &self,
569        now: DateTime<Utc>,
570    ) -> Result<Option<WorkflowRunSuspension>> {
571        let mut wakeups = self.list_open_suspensions(now).await?;
572        wakeups.retain(|suspension| suspension.scheduled_at().is_some());
573        wakeups.sort_by(|left, right| {
574            (
575                left.scheduled_at(),
576                left.run_id(),
577                left.kind_order(),
578                left.subject_id(),
579            )
580                .cmp(&(
581                    right.scheduled_at(),
582                    right.run_id(),
583                    right.kind_order(),
584                    right.subject_id(),
585                ))
586        });
587        Ok(wakeups.into_iter().next())
588    }
589
590    /// List active external callback hooks across non-terminal runs.
591    ///
592    /// Callback routers and dashboards can use this to discover public hook
593    /// tokens and their audit metadata without projecting every run manually.
594    /// The result is sorted by run ID and hook ID for stable polling output.
595    pub async fn list_active_hooks(&self) -> Result<Vec<ActiveHookSnapshot>> {
596        self.store.list_active_hooks().await
597    }
598
599    /// Replay and dispatch until the run reaches a terminal state or an open
600    /// wait/hook suspension.
601    pub async fn drive(&self, run_id: &str) -> Result<WorkflowRunSnapshot> {
602        self.drive_at(run_id, Utc::now()).await
603    }
604
605    async fn drive_at(&self, run_id: &str, now: DateTime<Utc>) -> Result<WorkflowRunSnapshot> {
606        'replay: for _ in 0..self.max_replay_iterations {
607            let history = self.store.list(run_id).await?;
608            let snapshot = project_run(run_id, &history)?;
609            if snapshot.status.is_terminal()
610                || snapshot
611                    .waits
612                    .values()
613                    .any(|wait| wait.status == WaitStatus::Waiting)
614                || snapshot
615                    .hooks
616                    .values()
617                    .any(|hook| hook.status == HookStatus::Active)
618                || (snapshot.has_future_retry(now) && snapshot.due_retries(now).is_empty())
619            {
620                return Ok(snapshot);
621            }
622
623            let command = self
624                .runtime
625                .run_workflow(WorkflowInvocation {
626                    run_id: run_id.to_string(),
627                    spec: snapshot.spec.clone(),
628                    input: snapshot.input.clone(),
629                    history,
630                })
631                .await?;
632
633            match command {
634                RuntimeCommand::Complete { output } => {
635                    if snapshot.status == WorkflowRunStatus::Cancelling {
636                        return Err(FlowError::InvalidTransition(format!(
637                            "workflow run {run_id} completed after cancellation was requested; cleanup-aware cancellation must return cancel or fail"
638                        )));
639                    }
640                    match self
641                        .record_event_at(
642                            run_id,
643                            snapshot.last_sequence,
644                            FlowEvent::RunCompleted { output },
645                        )
646                        .await
647                    {
648                        Ok(_) => {}
649                        Err(err) if is_event_conflict(&err) => continue,
650                        Err(err) => return Err(err),
651                    }
652                    return self.snapshot(run_id).await;
653                }
654                RuntimeCommand::Fail { error } => {
655                    match self
656                        .record_event_at(
657                            run_id,
658                            snapshot.last_sequence,
659                            FlowEvent::RunFailed { error },
660                        )
661                        .await
662                    {
663                        Ok(_) => {}
664                        Err(err) if is_event_conflict(&err) => continue,
665                        Err(err) => return Err(err),
666                    }
667                    return self.snapshot(run_id).await;
668                }
669                RuntimeCommand::Cancel => {
670                    let cancellation = snapshot.cancellation.as_ref().ok_or_else(|| {
671                        FlowError::InvalidTransition(format!(
672                            "workflow run {run_id} returned cancel without a durable cancellation request"
673                        ))
674                    })?;
675                    match self
676                        .record_event_at(
677                            run_id,
678                            snapshot.last_sequence,
679                            FlowEvent::RunCancelled {
680                                reason: cancellation.request.reason.clone(),
681                            },
682                        )
683                        .await
684                    {
685                        Ok(_) => {}
686                        Err(err) if is_event_conflict(&err) => continue,
687                        Err(err) => return Err(err),
688                    }
689                    return self.snapshot(run_id).await;
690                }
691                RuntimeCommand::Timeout { deadline, reason } => {
692                    match self
693                        .record_event_at(
694                            run_id,
695                            snapshot.last_sequence,
696                            FlowEvent::RunTimedOut { deadline, reason },
697                        )
698                        .await
699                    {
700                        Ok(_) => {}
701                        Err(err) if is_event_conflict(&err) => continue,
702                        Err(err) => return Err(err),
703                    }
704                    return self.snapshot(run_id).await;
705                }
706                RuntimeCommand::RecordProgress { progress } => {
707                    progress.validate()?;
708                    if let Some(existing) = snapshot.progress(&progress.progress_id) {
709                        ensure_progress_matches(run_id, existing, &progress)?;
710                        return Err(FlowError::InvalidTransition(format!(
711                            "workflow rescheduled progress {} without progress",
712                            progress.progress_id
713                        )));
714                    }
715                    match self
716                        .record_event_at(
717                            run_id,
718                            snapshot.last_sequence,
719                            FlowEvent::RunProgressRecorded { progress },
720                        )
721                        .await
722                    {
723                        Ok(_) => {}
724                        Err(err) if is_event_conflict(&err) => continue,
725                        Err(err) => return Err(err),
726                    }
727                }
728                RuntimeCommand::LinkChildOperation { child } => {
729                    child.validate()?;
730                    if let Some(existing) = snapshot.child_operation(&child.reference_id) {
731                        ensure_child_operation_matches(run_id, existing, &child)?;
732                        return Err(FlowError::InvalidTransition(format!(
733                            "workflow rescheduled child operation {} without progress",
734                            child.reference_id
735                        )));
736                    }
737                    match self
738                        .record_event_at(
739                            run_id,
740                            snapshot.last_sequence,
741                            FlowEvent::ChildOperationLinked { child },
742                        )
743                        .await
744                    {
745                        Ok(_) => {}
746                        Err(err) if is_event_conflict(&err) => continue,
747                        Err(err) => return Err(err),
748                    }
749                }
750                RuntimeCommand::ScheduleStep {
751                    step_id,
752                    step_name,
753                    input,
754                    retry,
755                } => {
756                    if let Some(step) = snapshot.steps.get(&step_id) {
757                        ensure_step_command_matches(run_id, step, &step_name, &input, retry)?;
758                        if matches!(
759                            step.status,
760                            StepStatus::Completed | StepStatus::Failed | StepStatus::Cancelled
761                        ) {
762                            return Err(FlowError::InvalidTransition(format!(
763                                "workflow rescheduled terminal step {step_id} without progress"
764                            )));
765                        }
766                    }
767                    match self
768                        .execute_step(
769                            run_id,
770                            &snapshot,
771                            StepExecutionContext {
772                                step_id,
773                                step_name,
774                                input,
775                                retry,
776                                now,
777                            },
778                        )
779                        .await
780                    {
781                        Ok(()) => {}
782                        Err(err) if is_event_conflict(&err) => continue,
783                        Err(err) => return Err(err),
784                    }
785                }
786                RuntimeCommand::ScheduleSteps { steps } => {
787                    ensure_step_batch_valid(&steps)?;
788                    for step in &steps {
789                        if let Some(existing) = snapshot.steps.get(&step.step_id) {
790                            ensure_step_command_matches(
791                                run_id,
792                                existing,
793                                &step.step_name,
794                                &step.input,
795                                step.retry,
796                            )?;
797                        }
798                    }
799                    if steps.iter().all(|step| {
800                        snapshot.steps.get(&step.step_id).is_some_and(|existing| {
801                            matches!(
802                                existing.status,
803                                StepStatus::Completed | StepStatus::Failed | StepStatus::Cancelled
804                            )
805                        })
806                    }) {
807                        let step_ids = steps
808                            .iter()
809                            .map(|step| step.step_id.as_str())
810                            .collect::<Vec<_>>()
811                            .join(", ");
812                        return Err(FlowError::InvalidTransition(format!(
813                            "workflow rescheduled only terminal steps without progress: {step_ids}"
814                        )));
815                    }
816                    match self.execute_step_batch(run_id, &snapshot, steps, now).await {
817                        Ok(()) => {}
818                        Err(err) if is_event_conflict(&err) => continue 'replay,
819                        Err(err) => return Err(err),
820                    }
821                }
822                RuntimeCommand::WaitUntil { wait_id, resume_at } => {
823                    match snapshot.waits.get(&wait_id) {
824                        Some(wait) => {
825                            ensure_wait_command_matches(run_id, wait, resume_at)?;
826                            match wait.status {
827                                WaitStatus::Completed => continue,
828                                WaitStatus::Waiting => return self.snapshot(run_id).await,
829                                WaitStatus::Cancelled => {
830                                    return Err(FlowError::InvalidTransition(format!(
831                                        "workflow rescheduled cancelled wait {wait_id}; cancellation cleanup must use a distinct stable identity"
832                                    )))
833                                }
834                            }
835                        }
836                        None => {
837                            match self
838                                .record_event_at(
839                                    run_id,
840                                    snapshot.last_sequence,
841                                    FlowEvent::WaitCreated { wait_id, resume_at },
842                                )
843                                .await
844                            {
845                                Ok(_) => {}
846                                Err(err) if is_event_conflict(&err) => continue,
847                                Err(err) => return Err(err),
848                            }
849                            return self.snapshot(run_id).await;
850                        }
851                    }
852                }
853                RuntimeCommand::CreateHook {
854                    hook_id,
855                    token,
856                    metadata,
857                } => match snapshot.hooks.get(&hook_id) {
858                    Some(hook) => {
859                        ensure_hook_command_matches(run_id, hook, &token, &metadata)?;
860                        match hook.status {
861                            HookStatus::Received | HookStatus::Disposed => continue,
862                            HookStatus::Active => return self.snapshot(run_id).await,
863                            HookStatus::Cancelled => {
864                                return Err(FlowError::InvalidTransition(format!(
865                                    "workflow rescheduled cancelled hook {hook_id}; cancellation cleanup must use a distinct stable identity"
866                                )))
867                            }
868                        }
869                    }
870                    None => {
871                        self.ensure_hook_token_available(run_id, &hook_id, &token)
872                            .await?;
873                        match self
874                            .record_event_at(
875                                run_id,
876                                snapshot.last_sequence,
877                                FlowEvent::HookCreated {
878                                    hook_id,
879                                    token,
880                                    metadata,
881                                },
882                            )
883                            .await
884                        {
885                            Ok(_) => {}
886                            Err(err) if is_event_conflict(&err) => continue,
887                            Err(err) => return Err(err),
888                        }
889                        return self.snapshot(run_id).await;
890                    }
891                },
892            }
893        }
894
895        Err(FlowError::ReplayLimitExceeded(self.max_replay_iterations))
896    }
897
898    async fn terminate_run(&self, run_id: &str, event: FlowEvent) -> Result<()> {
899        for _ in 0..self.max_replay_iterations {
900            let snapshot = self.snapshot(run_id).await?;
901            if snapshot.status.is_terminal() {
902                return Ok(());
903            }
904            match self
905                .record_event_at(run_id, snapshot.last_sequence, event.clone())
906                .await
907            {
908                Ok(_) => return Ok(()),
909                Err(err) if is_event_conflict(&err) => continue,
910                Err(err) => return Err(err),
911            }
912        }
913        Err(FlowError::ReplayLimitExceeded(self.max_replay_iterations))
914    }
915
916    async fn record_event_at(
917        &self,
918        run_id: &str,
919        expected_sequence: u64,
920        event: FlowEvent,
921    ) -> Result<FlowEventEnvelope> {
922        let envelope = self
923            .store
924            .append_if_sequence(run_id, expected_sequence, event)
925            .await?;
926        self.observer.observe(envelope.clone()).await;
927        Ok(envelope)
928    }
929
930    async fn ensure_hook_token_available(
931        &self,
932        run_id: &str,
933        hook_id: &str,
934        token: &str,
935    ) -> Result<()> {
936        for active in self.store.find_active_hooks_by_token(token).await? {
937            if active.run_id == run_id && active.hook.hook_id == hook_id {
938                continue;
939            }
940            return Err(FlowError::HookTokenConflict {
941                token: token.to_string(),
942                existing_run_id: active.run_id,
943                existing_hook_id: active.hook.hook_id,
944            });
945        }
946        Ok(())
947    }
948}
949
950fn resolve_scheduled_wakeup(
951    snapshot: &WorkflowRunSnapshot,
952    wakeup: &ScheduledWakeup,
953    now: DateTime<Utc>,
954) -> Option<WorkflowRunSuspension> {
955    if snapshot.run_id != wakeup.run_id || snapshot.status.is_terminal() {
956        return None;
957    }
958    match wakeup.kind {
959        ScheduledWakeupKind::Wait => {
960            let wait = snapshot.waits.get(&wakeup.subject_id)?;
961            if wait.status != WaitStatus::Waiting || wait.resume_at != wakeup.scheduled_at {
962                return None;
963            }
964            Some(WorkflowRunSuspension::Wait {
965                run_id: wakeup.run_id.clone(),
966                wait: wait.clone(),
967                due: wakeup.scheduled_at <= now,
968            })
969        }
970        ScheduledWakeupKind::Retry => {
971            let step = snapshot.steps.get(&wakeup.subject_id)?;
972            if step.status != StepStatus::Pending || step.retry_after != Some(wakeup.scheduled_at) {
973                return None;
974            }
975            Some(WorkflowRunSuspension::Retry {
976                run_id: wakeup.run_id.clone(),
977                step: step.clone(),
978                due: wakeup.scheduled_at <= now,
979            })
980        }
981    }
982}
a3s_flow/engine/mod.rs

a3s_flow/engine/
mod.rs
