camunda-process-mgmt

Camunda Process Management

Runtime operations for Camunda 8.8+ clusters via c8ctl: deploy resources, start instances, inspect state, resolve incidents, complete tasks, publish messages, cancel instances. Single skill for everything that happens after the model is written.

Prerequisites

• Camunda 8.8+ cluster (local via c8run, SaaS, or Self-Managed)
• c8ctl CLI installed and a profile configured for the target cluster — see camunda-c8ctl for setup

Cross-References

• camunda-c8ctl: Use for c8ctl setup, cluster safety, and shared conventions (flags, profiles, output modes)
• camunda-bpmn: Use for fixing BPMN process issues found during debugging
• camunda-dmn: Use for fixing DMN decision issues — EXTRACT_VALUE_ERROR / DECISION_EVALUATION_FAILED incidents typically trace back to the decision file
• camunda-connectors: Use for fixing connector configuration issues found during debugging
• camunda-job-workers: Use when an incident traces back to handler code rather than the process model — the worker side of zeebe:taskDefinition type
• camunda-connectors-development: Use when an incident traces back to a custom connector's runtime / registration / hosting rather than its template configuration
• camunda-feel: Use for diagnosing FEEL evaluation errors (EXTRACT_VALUE_ERROR, CONDITION_ERROR) in incidents
• camunda-process-test: Use for testing processes against an embedded Zeebe engine

Instructions

Deploying Resources

Cluster safety — ask before deploying to anything that isn't local c8run.

The globally-active c8ctl profile might still point at a shared cluster (prod, staging, customer name, …) from a previous session. A bare c8ctl deploy process.bpmn then silently writes there.

• Run c8ctl which profile first. If the name suggests a shared environment, confirm with the user before deploying.
• Always pass --profile=<name> explicitly on deploy and other mutating commands (run, cancel, resolve, complete, publish, watch).
• For validation deploys (just running the change once to check it works), prefer --profile=local against a running c8run cluster — c8ctl cluster start if not running.

See camunda-c8ctl for the authoritative cluster-safety rules.

Example — deploy a single BPMN file to the local profile:

c8ctl deploy process.bpmn --profile=local

Deploy multiple resources at once (BPMN, DMN, forms):

c8ctl deploy process.bpmn decision.dmn approval-form.form

Deploy an entire directory (recursive):

c8ctl deploy ./my-project

After deploying, verify:

c8ctl list pd
c8ctl search pd --iname="MyProcess"

Each deployment creates a new version of any resource it contains. Use c8ctl list pd --fields Key,Name,Version to see versions.

Starting Process Instances

Create an instance for a deployed process:

c8ctl create pi --id MyProcess

With input variables:

c8ctl create pi --id MyProcess --variables '{"orderId": "ORD-123", "amount": 1500}'

Deploy and start in one step:

c8ctl run process.bpmn --variables '{"orderId": "ORD-123"}'

Start and block until the instance completes (useful for tests / smoke checks):

c8ctl await pi --id MyProcess --variables '{"orderId": "ORD-123"}'

With a custom request timeout:

c8ctl await pi --id MyProcess --requestTimeout 60000

Not for long-running activities. await pi uses the cluster's start-and-wait REST endpoint, which times out before any single activity that runs longer than the request timeout (LLM agents, slow HTTP calls, user tasks). The timed-out response carries no instance key, so you can't follow up on the partial run. For those processes, use c8ctl create pi and poll c8ctl get pi <key> until terminal state, or fall back to c8ctl search pi --state=ACTIVE to recover the orphaned instance after a failed await.

Watch Mode (Development)

Auto-redeploy on file changes during local development:

c8ctl watch ./my-project

Watches .bpmn, .dmn, and .form files and redeploys on save.

Inspecting Process Instances

List active instances:

c8ctl list pi

Search by process name (i prefix means case-insensitive substring match):

c8ctl search pi --iprocessDefinitionName="MyProcess"

Get instance details with variables:

c8ctl get pi <instanceKey> --variables

Dump variable values for a specific instance (use --fullValue to avoid truncation):

c8ctl search variables --processInstanceKey=<key> --fullValue

Retrieve the deployed BPMN XML for a process definition:

c8ctl get pd <key> --xml

Resolving Incidents

Incidents are the cluster's way of pausing an instance when something fails non-recoverably (FEEL error, missing variable, connector failure, job timeout exceeded retries). Always inspect before resolving.

First-response triage — when a user reports "instance X is stuck", run these four in sequence to assess before deciding to resolve:

PI=<instanceKey>
c8ctl get pi $PI                                              # state (ACTIVE = stuck on a step), processDefinitionId
c8ctl search inc --processInstanceKey=$PI                     # any incidents on this PI, with errorType
c8ctl search variables --processInstanceKey=$PI --fullValue   # variables at the failure point
c8ctl get inc <incidentKey> --json                            # full error message

The four together usually localise the cause without needing to open Operate.

List active incidents:

c8ctl list inc
c8ctl search inc --state=ACTIVE

Filter by error message:

c8ctl search inc --ierrorMessage="Connection refused"

Debug workflow when a process has an incident:

1. Find the incident and read the error type + message:

   c8ctl search inc --state=ACTIVE
   

2. Inspect the instance variables at the failure point:

   c8ctl search variables --processInstanceKey=<key> --fullValue
   

3. Identify root cause. Common categories:

   • FEEL error (EXTRACT_VALUE_ERROR / CONDITION_ERROR) → cross-ref camunda-feel for expression fixes; check for dotted-variable shadowing or null inputs
   • Job worker missing or timing out → check the <zeebe:taskDefinition type="..."> matches a running worker
   • Connector misconfiguration → cross-ref camunda-connectors to fix the element template binding
   • BPMN logic bug → cross-ref camunda-bpmn to fix and redeploy

4. Fix the underlying issue (BPMN edit, connector reconfig, worker fix, or variable correction).

5. Resolve the incident to retry the failed step:

   c8ctl resolve inc <incidentKey>
   

Resolving without fixing the root cause just re-triggers the same incident. Always fix first.

Completing User Tasks

List pending user tasks:

c8ctl list ut

Complete with output variables:

c8ctl complete ut <taskKey> --variables '{"approved": true, "comments": "Looks good"}'

Variable keys must match the form's component key values (see camunda-forms) — extra keys are ignored; missing required keys block completion downstream.

Publishing Correlation Messages

Publish a message to wake up a waiting message event in an instance:

c8ctl publish msg <messageName> --correlationKey=<key> --variables '{"status": "approved"}'

The correlation key must match the value resolved from the process's message subscription expression (typically a FEEL expression over a variable, e.g. =orderId).

Cancelling Instances

c8ctl cancel pi <instanceKey>

JSON Output for Scripting

Pass --json per call, optionally narrowed with --fields. Don't toggle session-wide output mode — see camunda-c8ctl for the why.

c8ctl list pi --json --fields=key,state,bpmnProcessId
c8ctl list pd --json --fields=key,bpmnProcessId,version
c8ctl get inc <key> --json

Troubleshooting

Deployment fails — Run c8ctl bpmn lint process.bpmn first (see camunda-bpmn). Most deploy errors are caught by the linter pre-deploy.

Process won't start — Verify the process ID matches a deployed definition (c8ctl list pd). Check required input variables.

Instance hangs at a service task — No worker is registered for that task type, the worker crashed, or the job timed out beyond max retries. Check c8ctl list jobs --processInstanceKey=<key> and confirm a worker process is running for the matching task type.

Repeated identical incidents — You're resolving without fixing. Read the error message every time; if it doesn't change, the root cause hasn't been addressed.