Feature/enhance automation nodes

[pikann]

Closes #366

Summary

Adds synchronous AI-agent execution and the rest of the automation node types requested in #366, then simplifies the automation status model and cleans up a few things found along the way.

New automation capabilities

• Synchronous AI agent execution — a trigger_ai_agent action now pauses the graph walk until the conversation it started actually finishes, instead of continuing to the next node the instant the conversation is created. Resumes via a new durable StreamAgentConversationStatus stream once the conversation reaches a terminal status.
• Wait/Delay action node — pauses a graph walk for a configured number of minutes before continuing, polled by a new WaitScheduler (same leader-lock/ticker pattern as the existing due-date/cron schedulers).
• Sprint automation — four new sprint triggers (created/started/completed/deleted), five new sprint-scoped condition fields, and two new actions (update_sprint, complete_sprint), fed by a new durable StreamSprintActivities stream so sprint events reach the automation engine the same way task events already do.
• {{variable}} interpolation — node config fields (agent messages, Call API URL/body/headers, task titles, sprint name/goal) can reference {{task.title}}, {{sprint.name}}, and similar placeholders from the current run.

Automation status: active/inactive only

Replaced the three-state draft/active/archived lifecycle with a single active/inactive toggle. archived previously locked a graph from being edited at all, which no longer made sense next to the pause/resume nodes above — inactive automations stay fully editable, matching how draft already worked. The three lifecycle endpoints (activate/archive/revert-to-draft) collapse to two idempotent ones (activate/deactivate), and the builder page's badge-plus-three-buttons cluster becomes a single switch.

Other fixes and cleanup

• Bugfix: a sprint-triggered automation that paused at a Wait or trigger_ai_agent node lost its sprint context on resume, failing the downstream update_sprint/complete_sprint action with "no sprint in context". The resume path only ever reconstructed the walk's task, never its sprint.
• Generalized the pause/resume context (previously one column per field — task_id, then sprint_id) into a single context JSONB column, so a future context type doesn't need its own migration.
• Consolidated the pending-wait table migrations into one, since none had shipped anywhere outside this branch yet.
• Builder page header split into two rows (identity + status; view tabs + actions) instead of one crowded strip, and the trigger picker groups its 13 built-in types into Task events / Sprint events / Schedule & webhook sections.
• All new/changed strings translated across all 9 locales, including a couple of pre-existing gaps found while auditing.

Testing

• go build ./... && go vet ./... && gofmt -l . && go test ./...
• Full e2e suite (go test ./test/e2e/...), including new coverage for synchronous agent execution, Wait nodes, Sprint triggers/conditions/actions, variable interpolation, and the active/inactive lifecycle
• Frontend tsc --noEmit and biome check .

[pullfrog]

Reviewed PR #375. This is a large but well-scoped change: sprint-scoped triggers/conditions/actions, wait/trigger_ai_agent pause/resume, variable interpolation, and collapsing the automation lifecycle to active/inactive.

The test coverage is strong and the migrations are safe, but I found a critical durability bug in the pause/resume path and a high-severity fan-out AI-agent semantics bug that should be fixed before merge. I’ve also flagged a handful of medium/low issues inline.

Blocking issues

1. Pause/resume can permanently stall a run — both ClaimPendingAgentWait and ClaimDueDelays delete the pending row before the graph walk resumes. If the process crashes or resumeWalk/resumeAfterDelay errors after the delete, the row is gone and the run can never continue.
2. Fan-out trigger_ai_agent resumes downstream on the first finished conversation — when the action fans out to multiple conversations, each terminal status event walks the same outgoing edges, so downstream actions run before the remaining conversations have resolved.

Other concerns

• Front-end wait node accepts fractional minutes while the backend stores an *int.
• *target = *updated in applyUpdateSprint mutates a repository-returned pointer in place.
• call_api interpolates user values into URLs/headers without encoding/sanitization.
• Sprint status from the activity stream is cast without enum validation.
• WaitScheduler.Start takes a context.Context it never uses.
• Custom-field diff uses fmt.Sprintf("%v", …) comparison.
• Removing /archive and /revert-to-draft is a breaking API change; it should be documented or aliased for one release.

Please address the two blocking issues at minimum; the rest can be handled here or in follow-ups.

  ^{｜ Fix it ➔ ｜ View workflow run ｜ Using Kimi K2 (free via Pullfrog for OSS) ｜ 𝕏}

[pullfrog]

Following up on the review above with line-anchored details on the blocking correctness/durability issues. Please address at least the two blocking items (pause/resume durability and fan-out AI-agent semantics) before merging; the rest are inline suggestions.

  ^{｜ Fix it ➔ ｜ View workflow run ｜ Using Kimi K2 (free via Pullfrog for OSS) ｜ 𝕏}

[pullfrog]

Line-anchored follow-up to the review above. The two blocking issues are #1 and #2; please address them before merging.

---

1. Critical — pause/resume can permanently stall a run (automation_consumer.go:982)

ClaimPendingAgentWait deletes the pending row before resumeWalk runs. If the consumer crashes or resumeWalk errors after the DELETE, the stream message is redelivered, but the row is gone (ClaimPendingAgentWait returns nil), so the handler just acks and returns. The run is left running with no pending waits and downstream actions never execute.

Fix: Do not delete the pending row until resumeWalk succeeds, or use a lease/outbox pattern so resume is retryable.

2. High — fan-out trigger_ai_agent resumes downstream on the first finished conversation (automation_consumer.go:1528)

When trigger_ai_agent fans out to multiple conversations, each terminal status event calls resumeWalkFrom and walks the same node's outgoing edges. The first finished conversation therefore executes downstream actions while the other conversations are still in flight. There is no test covering this case.

Fix: Track the pending conversation count per (run_id, node_id) and only walk outgoing edges after the last pending wait is cleared.

3. Medium — wait node accepts fractional minutes in the UI (automation-node-config-panel.tsx:2190)

const minutes = Number(waitMinutes); accepts 1.5, but the backend stores WaitMinutes as *int, so the API will reject fractional values.

Fix: Enforce integer minutes (e.g., parseInt, Math.floor, or step={1}).

4. Medium — sprint pointer mutated in place (automation_consumer.go:2268)

*target = *updated mutates the sprint pointer returned by resolveSprintFor, which may come from the repository/cache.

Fix: Copy the updated sprint into a fresh local variable.

5. Low — variable interpolation is not URL/header-safe (automation_consumer.go:1767)

vartemplate.Render interpolates user values directly into URLs and headers. Values with spaces, &, or newlines can produce malformed URLs or header injection.

Fix: URL-encode values in the URL path/query and strip newlines from header values.

6. Low — sprint status from stream is not validated (automation_consumer.go:665)

Status: sprintdom.SprintStatus(status) casts the stream value directly. Malformed events can propagate invalid statuses through condition evaluation.

Fix: Validate against the allowed enum and ack/drop bad messages.

7. Low — WaitScheduler.Start ignores its context (wait_scheduler.go:70)

Start(ctx context.Context) accepts a context but run() uses context.Background(), so cancellation cannot shut it down cleanly.

Fix: Use the passed context as the parent.

8. Low — loose custom-field comparison (automation_consumer.go:2108)

fmt.Sprintf("%v", existing) != fmt.Sprintf("%v", v) can miscompare numeric types or structured values.

Fix: Use a normalized comparison such as JSON marshal or reflect.DeepEqual.

9. Note — breaking API change

/archive and /revert-to-draft were removed in favor of /deactivate. Please document this in release notes, or keep the old endpoints as aliases for one release if backward compatibility matters.

  ^{｜ View workflow run ｜ via Pullfrog ｜ Using Kimi K2 (free via Pullfrog for OSS) ｜ 𝕏}

[pullfrog]

ℹ️ One concurrency issue remains. The prior review's blocking issues (pause/resume durability, fan-out gating, wait-minutes mismatch, call_api escaping, sprint cache aliasing, sprint status validation, custom-field diff, WaitScheduler context) are all resolved in the commits since the last review.

The remaining issue is a read-modify-write race in the fan-out gate: resumeWalk counts pending agent waits before deleting the completed row, so two concurrent consumers can both see stillPending > 1 and skip calling resumeWalkFrom, leaving the run stalled. The fix is to delete the resolved row first, then count remaining siblings.

Suggested inline at services/api/internal/worker/automation_consumer.go:1079.

  ^{｜ Fix it ➔ ｜ View workflow run ｜ Using Kimi K2 (free via Pullfrog for OSS) ｜ 𝕏}

[pullfrog]

Blocking feedback converted to a PR comment: the fan-out gate in resumeWalk has a read-modify-write race.

At services/api/internal/worker/automation_consumer.go:1076-1085, the code counts pending agent waits before deleting the completed row. With multiple consumers, two sibling completions can both see stillPending > 1 and skip resumeWalkFrom, leaving the run stuck.

Suggested change:

	// Delete this wait before counting so the count reflects only siblings,
	// and concurrent resumes see a consistent remaining total.
	if err := c.repo.DeleteAgentWaitForAutomationNode(ctx, wait.ID); err != nil {
		return fmt.Errorf("delete completed agent wait: %w", err)
	}
	stillPending, err := c.repo.CountPendingAgentWaitsForNode(ctx, wait.RunID, wait.NodeID)
	if err != nil {
		return fmt.Errorf("count pending agent waits for node: %w", err)
	}
	if stillPending > 0 {
		return nil
	}

This deletes the resolved wait first, then resumes the parent only when no pending siblings remain.

  ^{｜ View workflow run ｜ via Pullfrog ｜ Using Kimi K2 (free via Pullfrog for OSS) ｜ 𝕏}

[pullfrog]

✅ No new issues found.

Reviewed changes

• Closed the remaining read-modify-write race in the trigger_ai_agent fan-out gate by replacing the separate count-then-delete with an atomic DeletePendingAgentWaitAndCountRemaining operation. A transaction-scoped Postgres advisory lock keyed on (runID, nodeID) serializes sibling resolutions across consumer replicas, so the last wait always sees remaining == 0 and proceeds exactly once.

  ^{｜ View workflow run ｜ Using Kimi K2 (free via Pullfrog for OSS) ｜ 𝕏}

[pullfrog]

ℹ️ No critical issues — minor suggestions inline.

Reviewed changes

• Hardened WaitScheduler's Redis leader lock with per-acquisition tokens, compare-and-delete release, and per-delay lease renewal so a long backlog can't outlive the lock.
• Made StopConversation publish a terminal "stopped" status to StreamAgentConversationStatus so a trigger_ai_agent node paused waiting on this conversation resumes when the conversation is explicitly stopped.
• Added E2E regression coverage for the explicit-stop resume path.

  ^{｜ Fix all ➔ ｜ Fix 👍s ➔ ｜ View workflow run ｜ Using Kimi K2 (free via Pullfrog for OSS) ｜ 𝕏}