Skip to main content

orion v0.8.3

August 31, 2026

Added​

  • orion doctor checks that the Claude CLI is AUTHENTICATED, not merely present. A logged-out CLI passed the old check and failed every run; it is now a FAIL naming the account and the command that repairs it.

  • delegation.inherit_operator_config in orion.json: stages or actors whose runs get your own Claude Code configuration, plugins and MCP servers included. Empty by default; the choice is recorded in the event log.

  • The run-start event now states what the run was given — the tool count and the MCP servers, read off the CLI's own init frame. Until now this could only be discovered by reading a raw transcript.

  • orion release add <version> <KEY>... attaches tickets to a milestone from the command line. It takes bare keys, comma- or space-separated, and INCLUSIVE ranges — orion release add v0.8.3 OR-100 OR-133 OR-140..OR-145 — which is the point: attaching one work block of thirty-six consecutive tickets previously meant a click each in the Jira UI or a scripted REST loop, and a milestone that is expensive to maintain drifts. release could already create a version and report on one; this is what puts something in it.

  • The command resolves every key before it writes any, and prints what it will add, what it will MOVE (a ticket already on another milestone is leaving that one, which is a different sentence from an add and is reported as such), what is already on the target, and what names no ticket at all. A range can quietly include a ticket you did not picture, so the plan is readable before it is applied. Re-running is a no-op that says so, the same property release create has.

  • Adding to an already-released version is refused unless --force is given: it rewrites a history that has already shipped. A version that does not exist is refused naming the versions that do, and a range whose ends are reversed or that spans two projects is refused with the reason rather than silently expanding to nothing.

  • --project is optional and inferred from the ticket keys, which say it unambiguously; passing one that contradicts the keys is refused rather than silently preferred.

  • orion settle <KEY> unsticks a ticket's worktree without you ever needing its path. It finds the worktree for the ticket, reports what is holding the branch, and commits it as an unverified snapshot so orion collect can rebase again. Nothing is verified and nothing is pushed. Use --dry-run to see what it would do first.

    It refuses rather than guesses when the worktree is mid-merge, mid-rebase, holding unmerged paths, or on a detached HEAD, and prints the command that resolves each. Until now the only recovery from a worktree left dirty by a killed process or a full disk was to cd into a hashed path under ORION_HOME and run git against an agent's branch by hand.

Changed​

  • orion watch and orion work now print the run rather than the agents' tool-call transcript. At concurrency 4 the console was about 60% "read", "ran" and "edited" lines, and the ten lines that decide whether a person has to act were scattered through them. The default now prints stage boundaries, outcomes, escalations, failures and anything awaiting a person — on the order of 15 lines for a ticket instead of roughly 200. Add --verbose for the full stream as before. Nothing is removed from the record: the event log, the per-run log and orion logs are complete at either level, so triage and history are unaffected.

  • A run of lines from the same actor on the same ticket states the name and model once instead of on every line. The columns stay where they are, so the layout is unchanged and the eye tracks the ticket key and the verb, which are what vary. Any change of ticket, actor or model — and every stage boundary — states the identity again.

  • Consecutive identical lines collapse to one line with a count (edited cmd/orion/aiops_test.go (x4)) rather than repeating.

  • A long absolute path in a message prints as its file name (ran cat bxpkzaome.output, not the full sandbox temp path, whose only useful token was the part the line clipped away). The full path is in the event log.

  • A branch is now rebased onto its integration branch immediately before its first push, so it arrives on the remote current and the first CI run is the one that counts. An agent run takes ten to forty minutes; at concurrency 4 something else usually merges inside that window, so the branch used to be pushed at a base that had already moved, CI ran on it, and the landing pass then rebased it and force-pushed — a second full CI run on a second commit, for a reason that was avoidable. This narrows that window rather than closing it: the base can still move between the rebase and the push, and the landing queue still handles what gets through.

  • The pre-push rebase never refuses the push. If the branch does not replay cleanly, the original is pushed and the pull request opens anyway, with the conflict reported and the exact commands to resolve it — finished work hidden behind an unresolved conflict leaves you with nothing to look at. A worktree carrying .orion-manual-lock is never rewritten, and an unreachable remote degrades to pushing as it stands rather than failing the run. A branch that is already current costs one extra fetch and prints nothing.

  • slack.merge_approvers accepts a Slack user ID, a username, a display name or an email address, and each is resolved to the member ID a mention needs. A name lookup needs the users:read scope and an email needs users:read.email; without them the request still sends and still names the person in plain text, and the run reports which approver could not be mentioned. A user ID needs no scope at all, so it is the form that always works. Each approver is resolved at most once per run.

  • Only the approval request mentions anybody. A merged notice, a CI report and a cost report do not, and no message ever uses @channel or @here — a room tagged for everything is a room that gets muted, and in a channel created per project a broadcast reaches people with no standing to approve.

Fixed​

  • orion release status and orion release verify no longer report every ticket in an already-shipped release as undocumented. Collation writes the fragments into CHANGELOG.md and deletes them, so .changelog.d/ is the one place a released version's notes are guaranteed not to be — and that was the only place the check looked. Re-running orion release verify on a version that is tagged, published and marked released reported it unsafe to promote, and every past release stayed blocked from then on. The check now asks whether the change is documented rather than whether a file exists: a ## <version> section in CHANGELOG.md counts, and so does an uncollated fragment. Which state applies is read from CHANGELOG.md itself, so the check needs no tracker call and answers the same offline. A version still awaiting collation behaves exactly as before, and a done ticket the collated section does not name is still reported — as a warning rather than a blocker, because a note folded into another ticket's bullet reads identically and a published release cannot be corrected by refusing it.

  • An expired Claude login is reported as what it is. A run that dies because the CLI is not signed in now says claude is not authenticated: <the CLI's own reason>. Run: claude, sign in, then restart the watcher. instead of claude exited 1, which was indistinguishable from a crash, a bad prompt or a sandbox denial.

  • Tickets are no longer labelled orion-failed for it. Nothing was attempted -- no turn, no token, no branch work -- so the claim is released and the ticket goes back to the queue, the way the quota back-off already behaves.

  • The watcher stops instead of draining the queue. Every subsequent ticket would fail identically until a human signs in, so continuing to claim them turned one fixable problem into a queue of released tickets.

  • The Slack approval request now @-mentions the people on slack.merge_approvers, so Slack actually notifies them. It named them in plain text before — "Only navjyot can approve" — and Slack raises a notification for the member-ID form <@U012ABCDEF> and for nothing else, so a bare username was styled like any other word and reached nobody. The merge then waited on a person who was never told, and the ticket sat in ci-wait repeating "nobody has approved it yet".

  • A run that ends with uncommitted tracked changes in its worktree now commits them as an unverified snapshot, whatever ended the run and whatever the breaker says. The cleanup used to fire only when a breaker trip was still flagged, and that flag is erased by several unrelated things — an unverified-edits trip clears itself when a verify passes, orion reset clears it by hand, and every agent session writes its own state file. On OR-217 the flag had already self-cleared, so 163 lines of staged work were left behind, orion collect refused to rebase the branch on every poll for over fifteen minutes, and two healthy branches starved behind it in the landing queue. A trip, where one is still on record, now decides only the commit message and what is reported — not whether the cleanup happens.

  • The QA stage now commits the tests it wrote before it reports its verdict. The QA prompt already asked for that commit and the agent did not reliably make it — two consecutive tickets left 163 and 110 lines of test code sitting in the worktree, once staged and once not — and everything after the stage reads commits rather than the worktree. So an uncommitted test made the red-before-green check report "QA did not add or change a test file", a false negative on precisely the run where a test was written; it left a dirty worktree, which is what collect's rebase refuses; and the branch was pushed without it, so CI could go green on a pull request whose own evidence was still on disk. The commit happens on every exit from the stage, including when findings are still open at the round ceiling: a red pull request is the correct outcome for a change QA found a defect in, and must not be avoided by leaving the failing test behind.

Security​

  • A supervised run no longer inherits the operator's Claude Code configuration. Every agent Orion launches now gets a config directory Orion builds ($ORION_HOME/agent-config, holding the nj-agents skills and agents and nothing else) and is started with --strict-mcp-config, so it has no MCP servers. Runs previously loaded ~/.claude in full: on a measured run that was 179 tools, 148 of them MCP tools with write access to the operator's own authenticated accounts — createJiraIssue, editJiraIssue, createConfluencePage — none of which Orion's breaker, sandbox or approval path could see.

Download v0.8.3 · Staying current