Replay-Safe Workflow Logging

Workflow execution may replay the same code many times. Calling Logger directly would duplicate breadcrumbs, so Continuum.AstCheck rejects Logger calls in workflow and Continuum.Pure code. Use the journal-identified logging effect instead:

def run(input) do
  Continuum.log(:info, "starting order #{input.order_id}")
  # ...
end

Continuum.log/2 appends a workflow_log event carrying the command identity, level, and message. Only after that append succeeds does the live execution emit both Logger output and [:continuum, :workflow, :log] telemetry. Replay validates and consumes the event without emitting either output again.

Structured metadata

Continuum.log/3 adds a keyword list:

Continuum.log(:info, "payment accepted", order_id: order.id, cents: total)

The metadata is journaled with the event, merged into the Logger metadata alongside Continuum's own continuum_run_id, continuum_workflow, and continuum_replay keys, and included in the telemetry metadata under :metadata.

Because it is journaled, metadata has to survive the journal: values go through Continuum.DurableTerm, so a PID, reference, port, or function is rejected at the call rather than written and rendered meaningless on replay. Metadata is also part of what replay matches, exactly as the message already is — changing it for a run already in flight is drift, not a silent no-op.

A log/2 call is unchanged on disk: the metadata key is omitted entirely when the list is empty, so existing histories, snapshots, and golden fixtures still match.

Levels follow Logger (:debug through :emergency), so every Logger.<level> call has a direct replacement and the compile error names it. Logger.log/3 and Logger.bare_log/3 map to Continuum.log(level, message), or to Continuum.log(level, message, metadata) when they carried metadata. Logger.metadata/1 and the process-level functions are rejected outright: Logger process state is not durable, so pass context as workflow input instead.

Messages must be durable binaries and are capped at 16 KiB. Logs are workflow breadcrumbs, so avoid putting secrets in them; use an activity when the operation itself is an external side effect rather than observability.