Skip to content
Apkar
ع

Changelog

Every release, in order.

What shipped in Apkar, when, and what it changed. Twelve releases on a two-week cadence.

  1. 2.4.0
    NewImproved

    Parallel runs

    Independent branches in a run graph now execute at the same time instead of in declaration order.

    • Branches with no shared dependency run concurrently, cutting median wall-clock time on fan-out graphs by roughly three times.
    • Concurrency is capped per workspace and configurable per agent.
    • The run graph draws parallel branches on their own tracks.
  2. 2.3.2
    Fixed

    Retry budgets no longer double-count

    A retried step could consume budget twice when the failure came from a nested tool call.

    • Nested tool failures now attribute their cost to the parent step once.
    • Runs that hit a cost ceiling mid-retry stop cleanly instead of finishing the attempt.
  3. 2.3.1
    FixedImproved

    Faster replay on long runs

    Replaying a run with more than 200 steps took several seconds to open.

    • Step payloads load on demand rather than all at once.
    • The replay timeline virtualises past 100 steps.
  4. 2.3.0
    New

    Human in the loop

    Pause a run at any step, review the state, approve or amend, and resume from the same point.

    • Approval gates can be declared per step or triggered by a predicate.
    • Paused runs hold their checkpoint for 72 hours.
    • Approvals are recorded in the audit trail with the reviewer's identity.
  5. 2.2.4
    Improved

    Sharper cost attribution

    Token spend is now attributed to the step that caused it rather than the run as a whole.

    • Per-step cost appears in the node inspector.
    • Workspace billing breaks down by agent and by step.
  6. 2.2.3
    Fixed

    Webhook delivery ordering

    Run-completed webhooks could arrive before the final step's own webhook.

    • Delivery is now ordered per run.
    • Failed deliveries retry with the same backoff policy as tool calls.
  7. 2.2.2
    ImprovedFixed

    Audit export

    Audit trails export as JSONL for a date range, suitable as SOC 2 evidence.

    • Exports stream rather than buffering, so a year of runs no longer times out.
    • Redaction rules apply to exported payloads.
  8. 2.2.0
    New

    Cost guards

    A hard ceiling per run. The agent stops at the limit instead of quietly overspending.

    • Ceilings can be set per agent, per workspace, or per invocation.
    • A run that stops on budget is marked distinctly from one that failed.
    • Alerts fire at 80 percent of the ceiling.
  9. 2.1.3
    Fixed

    Checkpoint restore on cold start

    A run resumed after a worker restart could replay its last completed step.

    • Checkpoints are committed before the step is marked complete, not after.
  10. 2.1.1
    Improved

    Policy files accept per-tool overrides

    Retry policy can now be narrowed to a single tool without duplicating the whole file.

    • Tool-level overrides inherit from the agent default.
    • Policy validation runs at deploy time and rejects unreachable rules.
  11. 2.1.0
    New

    Run graph

    Every step, branch, and retry, drawn as it happens, with a node inspector on each.

    • Click any node for its exact inputs, outputs, latency, and token count.
    • The graph streams live and is shareable by link.
  12. 2.0.0
    NewImproved

    Durable runs

    Runs are checkpointed after every tool call and survive worker restarts, deploys, and regional failover.

    • The execution model moved from in-memory to durable by default.
    • Breaking: agents defined with the 1.x builder need one migration pass.
    • Retention raised from 14 to 90 days on all paid plans.