Skip to main content

orion v0.4.0

August 26, 2026

The release where Orion runs a ticket end to end without being asked.

Before this, orion guarded an agent you started yourself. Now a label on a Jira ticket is enough: it claims the ticket, works it in a sandbox, answers its own design questions, opens a described pull request, waits for CI, asks Slack for approval, merges, closes the ticket, fast-forwards your checkout and cleans up after itself.

The loop this exists for​

Architect and PM advisors. When the implementer stops to ask something, Haiku classifies the question and either the architect (spec.md, plan.md) or the PM (intent.md) answers it — grounded in committed artifacts, citing the clause, or refusing. The decision is committed to docs/decisions/ and the implementer is resumed in the same session rather than restarted.

A refusal is a success. It says the design is incomplete, which is a human's decision and then an amendment to the document, so the next ticket does not re-ask at full price.

New commands​

  • orion watch — run the queue by itself: reconcile, start one job, sleep
  • orion collect — finish tickets awaiting CI: merge, close, refresh, prune
  • orion sandbox — see, enter or prune the worktrees agents worked in
  • orion logs <KEY> -f — live event log, per project or per ticket
  • orion slack test — prove Slack delivery, or name exactly what is broken

Pull requests describe themselves​

A PR opened by Orion used to just restate the ticket summary and a commit count. It now reads the branch's whole diff and commit history and drafts a real summary / what-changed / why / test-plan via the nj-agents pr-describe skill — read-only, it cannot push or commit anything itself. Any failure in that pass falls back to the old two-line description rather than blocking the PR.

Gates and safety​

  • CI is scaffolded at adoption. orion init writes scripts/test.sh and a workflow that runs it — one entry point for CI, for you, and for the agent. Without checks, Orion has no honest verdict to gate a merge on.
  • Merges need a person. With slack.require_approval, a passing branch asks the project channel; an allowlisted ✅ merges it. An empty allowlist means nobody. A rejection beats an approval however late it arrives.
  • A red build can be sent back to the agent, bounded by an attempt ceiling and by an identical-failure check that stops sooner. Off by default.
  • Merges rebase by default, so each commit keeps its author. Squash collapses a branch into one commit authored by whoever opened the pull request, which erases the agent from the trunk's history.
  • Never --admin. Orion does not bypass branch protection.

Budget​

budget.weekly_tokens no longer has to be invented. Orion reads the plan's own rate_limit_info from each run and a watcher sleeps until the real reset time. There is no percentage available to any process outside the Claude Code TUI, so Orion gets a yes/no and a reset time rather than a warning at 80%.

Observability​

Runs are narrated live. The previous --output-format json emitted a single object at exit, so a forty-minute run was indistinguishable from a hung one. Every event now carries the model that produced it — a single ticket is worked by three.

Fixed​

  • guardrails silently stopped firing after brew cleanup, because hooks were pinned to a versioned Cellar path
  • orion doctor could not say which credential source Jira was failing from, so a stale shell export shadowing config.env looked like a broken token
  • Slack messages were never sent: the channel was created but never recorded, so notify was handed an empty channel and skipped delivery without error
  • every Slack call sent a JSON body, which cursor-paginated reads ignore — private channels were invisible and pagination never advanced
  • the sandbox served stale policy, so a committed config change appeared to do nothing
  • a merged ticket now clears every label Orion owns, not just orion-ci-wait — previously it could keep a stale orion-failed label from an earlier failed attempt, so orion queue reported the same ticket as failed and Done on the same line
  • Slack notifications about the worktree/checkout outcome (removed vs kept, fast-forwarded vs refused) are now told directly rather than assumed, so a Slack message can no longer contradict what the terminal reports one line above it
  • orion watch fcia --max-jobs 1 read the 1 as a second project — a flag value after a value-taking flag could be misread as a positional project key
  • a mistyped project name looked identical to an empty queue; an unknown key is now refused, naming what IS registered
  • orion watch --dry-run without --once looped forever, re-printing an identical rehearsal every two minutes since a dry run changes nothing between ticks — it now rehearses once
  • a permanent (non-transient) error was retried forever; only transient errors keep the retry loop going now

Upgrading​

Nothing in orion.json is required to change; every new field defaults off or to the previous behaviour. To turn on the new gates:

"slack": { "require_approval": true, "merge_approvers": ["U01234567"] },
"ci":    { "auto_fix": true, "max_fix_attempts": 3 }

Slack approvals need reactions:read, reactions:write, channels:history, groups:history and users:read, then an app reinstall — new scopes never apply to an already-issued token.

Download v0.4.0 · Staying current