Overview

The Orchestrator loop is synchronous by design — no external queue daemon is required. Each Orchestrator::start() returns once the task is QUEUED (or, in legacy dev mode, RUNNING); a worker (server daemon or browser SharedWorker) calls tick() to drive the task to a terminal state. The deployment choice — server daemon, browser worker, with or without Mercure — lives in Deployment modes. The mechanics below (tick phases, task lifecycle, Mercure publishing) are identical across all three configurations.

The WorkerMode enum at app/Agents/ValueObjects/WorkerMode.php is single-case (Worker) — the legacy WorkerMode::Sync was removed when SPORA_SYNC_MODE was replaced. The new WorkerRuntimeMode enum (server / client) gates which CLI commands and HTTP routes are active; see Environment variables → Worker runtime mode and Client-worker mode.

Tick Structure

Orchestrator::tick() runs in three phases to avoid holding a DB lock during the LLM round-trip:

Phase 1 — Claim (short transaction with lockForUpdate)

• Lock the task row
• Validate status === 'RUNNING'
• Abort early if step_count >= max_steps (flips to ABORTED via TaskStatusWriter::autoAbortTransition($task, 'max_steps_reached', true) so the existing follow-up flow covers continuation; the row carries data.max_steps_reached: true so the UI can label the banner — added in spora-core PR #266)
• Commit → lock released

Phase 2 — Load + LLM call (outside the transaction)

• Load agent, enabled tools, system prompt, and LLMRequest (no DB lock held)
• Task::where('id', $taskId)->increment('step_count')
• Blocking HTTP call to the configured LLM provider
• No DB connection held during the I/O round-trip

Phase 3 — Write results

• If tool calls: appendHistory, execute tools, then either call tick() again or pause for approval (PENDING_APPROVAL)
• If text response: appendHistory, set COMPLETED

safeExecute() reads the calling agent’s row inside the Orchestrator, resolves ownership from it, and threads that into execute() as a PrincipalContext (read $context->ownerUserId) — tools never see a session-derived user id, so async contexts (server-mode daemon, browser-driven SharedWorker, scheduled runs, sub-agent hops) all inherit the same trust boundary. See Architecture → Orchestrator Loop for the invariant.

resume() (PR #173) takes a per-call decisions list and splits it via AgentDecisionProcessor::splitDecisions() into approved and rejected subsets inside a single lockForUpdate() transaction. Approved rows are handed to ApprovedBatchExecutor (PR #171 path — partial-approval semantics preserved: undecided rows stay PENDING_APPROVAL). Rejected rows are stamped REJECTED with rejected_at / rejected_by / reject_reason and a role:'tool' history row is appended carrying toolCallId + toolName so the LLM sees the rejection in its next round-trip. The task transitions back to QUEUED only when the batch leaves no PENDING_APPROVAL rows; partial-approval batches keep the task paused. reject() (task-level bulk) is unchanged in shape but does not write the per-call rejection columns. ApprovedBatchExecutor records worker-mode approvals with executed_at IS NULL as the “approved, awaiting execution” sentinel for the daemon’s next drain (server mode) or the browser’s next /tick (client mode).

Partial approval. When the LLM produced parallel tool calls and the operator approves only some of them, resume() does not transition the task out of PENDING_APPROVAL — it rewrites pending_state with the un-approved tool calls and returns, so the operator can keep deciding on later rounds. Un-approved tool_calls rows stay at status='PENDING_APPROVAL'; they are never silently executed with the LLM’s original arguments, and never auto-rejected.

Worker-mode tool pickup. In Worker mode, resume() persists each approved tool as status='APPROVED' with executed_at=NULL (the worker-pickup sentinel) and returns without running the tool. When the daemon’s task:run worker claims the resulting QUEUED task and runs tick(), TickPhaseRunner::runTick() calls executeApprovedPendingToolsForTask() first — it picks up those APPROVED + executed_at IS NULL rows, validates, executes, and appends the tool result to history — before the LLM round-trip. The next assistant message therefore sees the tool results on the same round-trip. The HTTP POST /api/v1/tasks/{id}/approve response returns within ~100 ms regardless of how long the approved tool takes.

Task Status Lifecycle

stateDiagram-v2
    [*] --> QUEUED : start() / continue()

    QUEUED --> RUNNING : Worker claims task (server daemon or browser)

    RUNNING --> COMPLETED : LLM returns text (no tools)

    RUNNING --> PENDING_APPROVAL : Tool calls need approval

    PENDING_APPROVAL --> RUNNING : resume() / reject() picks up on next tick
    PENDING_APPROVAL --> PENDING_APPROVAL : resume() (partial approval — pending_state rewritten)

    RUNNING --> ABORTED : max_steps (data.max_steps_reached)
    RUNNING --> FAILED : exception

    RUNNING --> AWAITING_SUB_AGENTS : sub_agent tool (SubAgentTool) spawns child
    AWAITING_SUB_AGENTS --> RUNNING : every child TERMINAL (resume on next tick)
    RUNNING --> ABORTED : POST /tasks/{id}/abort (quiescent — resumable)
    AWAITING_SUB_AGENTS --> ABORTED : POST /tasks/{id}/abort-sub-agent (child abort cascades to ancestors)
    ABORTED --> QUEUED : POST /tasks/{id}/continue (clear aborted_at, re-prompt)
    RUNNING --> ABORTED : POST /tasks/{id}/continue (auto-abort + marker row)

Quiescent task states

ABORTED, PENDING_APPROVAL, and AWAITING_SUB_AGENTS together form the quiescent set: in every case the worker is not driving the task — the conversation is waiting on the operator (PENDING_APPROVAL tool call), on sub-agent children (AWAITING_SUB_AGENTS), or on the operator’s next instruction after an explicit halt (ABORTED). The chat detail poller skips the entire set so the browser does not waste cycles fetching a task that is not making progress on its own. The poller re-arms the moment an action moves the task out — approve/reject from the approval bar, sub-agent child reaching a terminal state, or POST /api/v1/tasks/{taskId}/continue resuming an ABORTED task.

AWAITING_SUB_AGENTS is the suspended-while-sub-agent-children-run state set by the SubAgentTool sub_agent op. Each spawn creates a regular Task with parent_task_id and bumps data.sub_agent_expected_count after the child tick. The first spawn also sets data.sub_agent_batch_open to true — a per-batch cross-process lock that prevents the per-child resume hook from racing the batch-boundary hook in worker mode. The resume gate compares the live child count from data.spawned_sub_task_ids against sub_agent_expected_count and only re-enters the loop when every sibling has reached a terminal state; the boundary hook then clears sub_agent_batch_open along with the other batch book-keeping. The next worker pickup (server daemon’s task:run or browser’s /tick) drains it via TickPhaseRunner::maybeResumeParentFromBatchBoundary. See Tool system → Handover for the LLM-facing contract.

Worker CLI

Server mode only. The CLI commands below are the operator’s tools for worker_runtime_mode: server. In client mode the same commands exit with a docs-link error — see Client-worker mode for the browser-driven equivalent.

Entry point: bin/spora (via WorkerRunCommand)

# Default: persistent daemon — continuously polls for scheduled runs and QUEUED tasks
php bin/spora worker:run

# One-shot scheduled runs only, then exit
php bin/spora worker:run --once

# One-shot: scheduled runs + QUEUED tasks, then exit
php bin/spora worker:run --once --include-queue

# Orphan reaper: mark stale RUNNING tasks as FAILED, then exit
php bin/spora worker:run --reap-only

# Options
--limit=N         Max QUEUED tasks per poll cycle (0 = unlimited, default: 0)
--sleep=N         Microseconds to sleep when queue is empty (default: 500000)
--stale-minutes=N Minutes before a RUNNING or AWAITING_SUB_AGENTS task is considered orphaned (0 = disabled; omit to use config default of 60)
--workers=N       Max concurrent child processes (0 = unlimited)
--once            Process due scheduled runs then exit (one-shot)
--include-queue   With --once: also drain the QUEUED task queue
--reap-only       Reap orphaned RUNNING tasks once, then exit
--daemon          Explicit daemon mode (default when no flag is given)

Deployment modes

Command	Scheduled runs	QUEUED tasks	Exit	Typical use
worker:run	✓	✓	Never (until SIGTERM)	VPS/Docker always-on (default daemon)
worker:run --reap-only	—	—	After one iteration	Maintenance: orphan reaping only
worker:run --once	✓	—	After processing	Cron for scheduled runs
worker:run --once --include-queue	✓	✓	After processing	Full cron replacement

For the client-mode equivalent (browser-driven, no daemon), see Client-worker mode and the Deployment modes overview.

Cron setup:

# Full queue drain every minute
* * * * * /usr/bin/php /path/to/spora/bin/spora worker:run --once --include-queue >> /path/to/spora/storage/worker.log 2>&1

The daemon (--daemon) uses the same storage/spora-worker.lock as the one-shot modes, preventing concurrent workers. After each scheduled run, next_run_at is computed using wall-clock now (in the schedule’s timezone) as the cron reference — not the actual last run time — so arbitrarily-delayed cron invocations are handled correctly without drift.

For the full deployment reference (Docker, systemd, supervisord, reaping, single-instance enforcement, monitoring), see the Worker deployment page.

Mercure SSE (Optional — Docker / FrankenPHP)

When SPORA_MERCURE_URL and SPORA_MERCURE_JWT_KEY are set, the Orchestrator publishes task state changes to a Mercure hub after each tick() step (intermediate tool results and PENDING_APPROVAL pauses) and on worker claim / scheduled run dispatch. The frontend subscribes to principal-scoped task topics — one principal/{principalId}/tasks topic per principal the caller can act as (their user-principal plus the group-principals of every group they belong to) — plus the user-scoped user/{userId}/notifications topic for user notifications, for real-time updates instead of polling. The principal-keyed task topic is what lets every group peer with read rights receive the same live task events as the trigger user; the notification topic stays user-keyed because the notifications table is per-user.

When the env vars are not set, MercurePublisher::publish() early-returns false (logged at debug level) — polling remains the default for all deployments.

Env vars (FrankenPHP native Mercure — no separate service needed):

SPORA_MERCURE_URL=https://spora.example.com/.well-known/mercure
SPORA_MERCURE_JWT_KEY=your-shared-secret

Don’t use http://localhost/... (no port) in Docker when SERVER_NAME is a public domain. Caddy’s auto-HTTPS layer redirects every plain-HTTP request — including the in-cluster publish POST — to https://..., and the publisher fails with tlsv1 alert internal error. Use SPORA_MERCURE_PUBLISH_URL=http://localhost:80/.well-known/mercure (loopback with explicit port — the bundled docker/frankenphp.conf adds http://localhost:80 as a second site address so the host matches). Avoid http://spora:80/... (bakes the docker-compose service name into the image) and http://localhost/... (no port).

A separate SPORA_MERCURE_PUBLISH_URL may be set if the publisher posts to a different endpoint than the public hub URL the browser subscribes to (e.g. behind a reverse proxy that disallows loopback).

FrankenPHP bundles a Mercure hub natively — no separate service needed in that configuration.

Subscribing from the browser — the __Secure-mercure_access_token flow

The Mercure hub is not anonymous — every update carries a principal-scoped task topic (principal/{principalId}/tasks) or a user-scoped notification topic (user/{userId}/notifications) and Mercure rejects subscribers whose JWT does not scope to those topics. The browser obtains its JWT through a short exchange with the app rather than embedding a long-lived secret in JS:

1. The Vue app calls GET /api/v1/sse/authorize (session-cookie authenticated).
2. The endpoint returns a short-lived subscriber JWT scoped to the calling user’s topics, and the framework sets it as a __Secure-mercure_access_token HttpOnly cookie.
3. The frontend opens an EventSource against SPORA_MERCURE_URL with credentials: 'include'. The browser attaches the cookie to the hub handshake and Mercure accepts the subscription.
4. The frontend re-runs step 1 periodically to refresh the cookie before expiry; the EventSource reconnect path picks up the new value transparently.

Because the cookie name is __Secure-, browsers only send it over HTTPS — SPORA_APP_URL and SPORA_MERCURE_URL must use https:// in production.

Environment Variables

See Environment variables for the consolidated reference (worker modes, Mercure, logging, database, etc.).

Edit this page on GitHub

Contributors: Fabian Graßl

Previous pageAgent templates

Next pageApp extensions