feat(autofix): Push PR iteration changes and sync completion webhook

src/sentry/analytics/events/autofix_events.py

@@ -7,7 +7,9 @@ class AiAutofixPhaseEvent(analytics.Event):
     project_id: int
     group_id: int
     referrer: str | None
+    run_id: int | None = None
     iteration_index: int | None = None
+    pr_iteration_enabled: bool | None = None
 
 
 @analytics.eventclass("ai.autofix.root_cause.started")

src/sentry/seer/autofix/autofix_agent.py

@@ -87,6 +87,10 @@ class NoSeerQuotaException(Exception):
     pass
 
 
+class PrIterationNoPullRequestException(Exception):
+    pass
+
+
 class AutofixStep(StrEnum):
     """Available autofix steps."""
 
@@ -242,6 +246,10 @@ def get_latest_iteration_index(state: SeerRunState) -> int:
     return 0
 
 
+def is_pr_iteration_enabled(state: SeerRunState) -> bool:
+    return bool(state.metadata.get("pr_iteration_enabled") if state.metadata else False)
+
+
 def get_iteration_for_insert_index(state: SeerRunState, insert_index: int) -> int:
     block = state.blocks[insert_index]
     metadata = block.message.metadata or {}
@@ -274,6 +282,11 @@ def get_autofix_agent_client(
     )
 
 
+def get_autofix_run_state(group: Group, run_id: int) -> SeerRunState:
+    client = get_autofix_agent_client(group)
+    return _get_group_run_state(client, group, run_id)
+
+
 def _validate_run_belongs_to_group(state: SeerRunState, group: Group) -> None:
     group_id = state.metadata.get("group_id") if state.metadata else None
     if group_id != group.id:
@@ -374,23 +387,14 @@ def trigger_autofix_agent(
         pr_iteration_enabled = run_state.metadata.get("pr_iteration_enabled", pr_iteration_enabled)
 
     iteration_index: int | None = None
-    if step == AutofixStep.PR_ITERATION and run_state is not None:
+    if step == AutofixStep.PR_ITERATION:
+        if run_state is None or not run_state.repo_pr_states:
+            raise PrIterationNoPullRequestException()
         if insert_index is not None:
             iteration_index = get_iteration_for_insert_index(run_state, insert_index)
         else:
             iteration_index = get_latest_iteration_index(run_state) + 1
 
-    if config.started_event is not None:
-        analytics.record(
-            config.started_event(
-                organization_id=group.organization.id,
-                project_id=group.project_id,
-                group_id=group.id,
-                referrer=referrer.value,
-                iteration_index=iteration_index,
-            )
-        )
-
     prompt = build_step_prompt(step, group, user_context, run_state=run_state)
     prompt_metadata = {
         "step": step.value,
@@ -443,6 +447,21 @@ def trigger_autofix_agent(
             insert_index=insert_index,
         )
 
+    # Emit the started event after run_id is resolved so it can be joined to
+    # downstream completed/PR events.
+    if config.started_event is not None:
+        analytics.record(
+            config.started_event(
+                organization_id=group.organization.id,
+                project_id=group.project_id,
+                group_id=group.id,
+                referrer=referrer.value,
+                run_id=run_id,
+                iteration_index=iteration_index,
+                pr_iteration_enabled=pr_iteration_enabled,
+            )
+        )
+
     payload: dict[str, Any] = {
         "run_id": run_id,
         "group_id": group.id,
@@ -487,6 +506,8 @@ def trigger_autofix_agent(
                 "step": step.value,
                 "run_id": run_id,
                 "group_id": group.id,
+                "iteration_index": iteration_index,
+                "pr_iteration_enabled": pr_iteration_enabled,
             },
         )
 
@@ -496,6 +517,7 @@ def trigger_autofix_agent(
             "step": step.value,
             "referrer": referrer.value,
             "iteration_index": iteration_index,
+            "pr_iteration_enabled": pr_iteration_enabled,
         },
     )
 
@@ -706,6 +728,7 @@ def trigger_coding_agent_handoff(
             project_id=group.project_id,
             group_id=group.id,
             referrer=referrer.value,
+            run_id=run_id,
             coding_agent=coding_agent_name,
         )
     )
@@ -751,6 +774,7 @@ def trigger_push_changes(
             project_id=group.project_id,
             group_id=group.id,
             referrer=referrer.value,
+            run_id=run_id,
         )
     )
 

src/sentry/seer/autofix/on_completion_hook.py

@@ -21,6 +21,7 @@
     STEP_CONFIGS,
     AutofixStep,
     get_latest_iteration_index,
+    is_pr_iteration_enabled,
     trigger_autofix_agent,
     trigger_coding_agent_handoff,
     trigger_push_changes,
@@ -177,18 +178,24 @@ def _send_step_webhook(
                         project_id=group.project_id,
                         group_id=group.id,
                         referrer=None if current_referrer is None else current_referrer.value,
+                        run_id=run_id,
                     )
                 )
             else:
                 webhook_action_type = SeerActionType.CODING_COMPLETED
                 webhook_payload["code_changes"] = cls._format_code_changes_payload(state)
         elif current_step == AutofixStep.PR_ITERATION:
-            # PR iteration only runs against an existing PR, so there should be pr states.
-            if not state.repo_pr_states:
-                logger.error(
-                    "autofix.on_completion_hook.pr_iteration_missing_repo_pr_states",
-                    extra={"run_id": run_id, "organization_id": organization.id},
-                )
+            # PR iteration only runs against an existing PR, so there must be pr states.
+            assert state.repo_pr_states, "PR iteration completed without any repo PR states"
+            # The iteration agent produces new file patches that still need to be
+            # pushed to the existing PR. The first hook fire (right after the agent
+            # finishes) has unsynced changes; _maybe_continue_pipeline triggers the
+            # push, which re-fires this hook once the PR is updated. Only emit the
+            # completion webhook once changes are synced, or when push failed
+            # terminally for errored repos so we don't wait forever.
+            _, is_synced = state.has_code_changes()
+            if not is_synced and not cls._iteration_terminal_errored_repos(state):
+                return
             webhook_action_type = SeerActionType.ITERATION_COMPLETED
             iteration_index = get_latest_iteration_index(state)
             webhook_payload["pull_requests"] = cls._format_pull_requests_payload(state)
@@ -240,8 +247,16 @@ def _send_step_webhook(
 
         if current_step is not None and not is_pr_created:
             referrer = current_referrer.value if current_referrer is not None else None
+            iteration_index = get_latest_iteration_index(state)
+            pr_iteration_enabled = is_pr_iteration_enabled(state)
             metrics.incr(
-                "autofix.explorer.complete", tags={"step": current_step.value, "referrer": referrer}
+                "autofix.explorer.complete",
+                tags={
+                    "step": current_step.value,
+                    "referrer": referrer,
+                    "iteration_index": iteration_index,
+                    "pr_iteration_enabled": pr_iteration_enabled,
+                },
             )
             completed_event_cls = STEP_CONFIGS[current_step].completed_event
             if completed_event_cls is not None:
@@ -251,6 +266,9 @@ def _send_step_webhook(
                         project_id=group.project_id,
                         group_id=group.id,
                         referrer=referrer,
+                        run_id=run_id,
+                        iteration_index=iteration_index,
+                        pr_iteration_enabled=pr_iteration_enabled,
                     )
                 )
 
@@ -373,16 +391,19 @@ def _maybe_continue_pipeline(
             )
             if decision is not None:
                 iteration_index = get_latest_iteration_index(state)
+                pr_iteration_enabled = is_pr_iteration_enabled(state)
                 analytics.record(
                     AiAutofixIntrospectionEvent(
                         organization_id=organization.id,
                         project_id=group.project_id,
                         group_id=group.id,
+                        run_id=run_id,
                         referrer=referrer.value,
                         step=current_step.value,
                         action=decision.action.value,
                         reached_stopping_point=reached_stopping_point,
                         iteration_index=iteration_index,
+                        pr_iteration_enabled=pr_iteration_enabled,
                     )
                 )
                 logger.info(
@@ -396,9 +417,19 @@ def _maybe_continue_pipeline(
                         "action": decision.action.value,
                         "reason": decision.reason,
                         "reached_stopping_point": reached_stopping_point,
+                        "iteration_index": iteration_index,
+                        "pr_iteration_enabled": pr_iteration_enabled,
                     },
                 )
 
+        # PR iteration runs against an existing PR rather than the automated
+        # pipeline. Once the agent finishes iterating, push the new changes to
+        # update that PR. _push_changes is a no-op once the repos are synced, so
+        # the hook re-fire after the push doesn't loop.
+        if current_step == AutofixStep.PR_ITERATION:
+            cls._push_changes(group, run_id, state)
+            return
+
         if stopping_point is None or reached_stopping_point:
             # We've reached the stopping point
             return
@@ -450,6 +481,8 @@ def _maybe_continue_pipeline(
                 "current_step": current_step,
                 "next_step": next_step,
                 "stopping_point": stopping_point,
+                "iteration_index": get_latest_iteration_index(state),
+                "pr_iteration_enabled": is_pr_iteration_enabled(state),
             },
         )
         trigger_autofix_agent(
@@ -479,6 +512,28 @@ def run_introspection(
         elif step == AutofixStep.PR_ITERATION:
             return introspect_iteration(organization, run_id, state, group)
 
+    @classmethod
+    def _iteration_terminal_errored_repos(cls, state: SeerRunState) -> list[str]:
+        """
+        Return the errored repos when unsynced repos have terminal push failures.
+
+        Returns an empty list when there are no errored repos, or when some
+        non-errored repo is still unsynced (i.e. a push can still make progress).
+        Used to stop waiting for a synced PR after push errors without retrying.
+        """
+        diffs_by_repo = state.get_diffs_by_repo()
+        errored_repos = [
+            repo
+            for repo in diffs_by_repo
+            if (pr_state := state.repo_pr_states.get(repo)) is not None
+            and pr_state.pr_creation_status == "error"
+        ]
+        if not errored_repos:
+            return []
+        if all(state._is_repo_synced(repo) or repo in errored_repos for repo in diffs_by_repo):
+            return errored_repos
+        return []
+
     @classmethod
     def _push_changes(cls, group: Group, run_id: int, state: SeerRunState) -> None:
         """Push code changes to create PRs."""
@@ -497,16 +552,8 @@ def _push_changes(cls, group: Group, run_id: int, state: SeerRunState) -> None:
             return
 
         # Errored repos are terminal — re-pushing would re-fire this hook in a loop.
-        diffs_by_repo = state.get_diffs_by_repo()
-        errored_repos = [
-            repo
-            for repo in diffs_by_repo
-            if (pr_state := state.repo_pr_states.get(repo)) is not None
-            and pr_state.pr_creation_status == "error"
-        ]
-        if errored_repos and all(
-            state._is_repo_synced(repo) or repo in errored_repos for repo in diffs_by_repo
-        ):
+        errored_repos = cls._iteration_terminal_errored_repos(state)
+        if errored_repos:
             logger.info(
                 "autofix.on_completion_hook.skip_no_pushable_repos",
                 extra={