Introduction
dev-orchestra is a plugin for Claude Code / Codex / Antigravity that takes a one-sentence request through design, implementation, testing, review by several AIs, and fixes.
This post rounds up the changes in v1.4.0 (2026-11-19). The changes in v1.3.2 are in the previous release post. This release turns the questions before the design into an interview, and brings a set of changes that make review rounds lighter. What follows, numbers included, comes from the CHANGELOG; what I ran myself is grill and a check that the new commands and options exist.
dev-orchestra is open source on GitHub (MIT license). If it looks useful, a star would mean a lot.
Main changes
An interview before the design (grill)
design.ask_questions gains a new value, grill. It asks the questions before the design one per message. When an answer is vague, contradicts an earlier one, or adds a constraint, it asks up to two follow-ups per question. Its own questions, at most five and the most important first, come from perspectives such as the intent, how success is judged, and what is out of scope. After ten questions, follow-ups included, it asks whether to go on.
dev-orchestra config set design.ask_questions grill
When I tried it, I got eight questions, one at a time, and an answer that added a condition got a follow-up. On the other hand, a vague answer was interpreted rather than probed, and I was never asked about the intent, one of the perspectives the docs list. I tried it only once, so this is what happened that time. The steps and output are in part 25 of the series, published the same day: “I Had dev-orchestra Interview Me Before the Design”.
Reviewers are told what was already decided
Every code review prompt now ends with a “Decisions already made” section. It holds three things: the user’s answers to the design questions, the plan’s Risks and Decisions sections, and the findings rejected in earlier rounds of the same review.
Reviewers are told to treat the first two as premises, and to cite one and report it when the code contradicts it or it looks wrong. The rejected findings are passed only as “earlier decisions” and never excuse a critical or high issue. The point is to stop reviewers from reopening, round after round, what the user decided or what was already rejected. The section is at most 6,000 characters per reviewer; when it does not fit, rejected findings are cut first, with a note saying what was cut.
Minor pre-existing problems stay out of triage
A reviewer can now mark a problem with Category: pre-existing when it was there before the change and the change does not make it worse. Medium and low ones are numbered P1, P2, … and listed under “Pre-existing, not triaged” in consolidated.md. They are not in the findings list, the counts, the fix brief or the loop’s decision, so a round with only those ends as clean.
Critical and high ones, and any at a severity in review.blocking_severities, are still triaged as before, pre-existing or not. A problem the change makes worse, or newly exposes, stays a normal finding.
Re-reviews run only the reviewers that matter
A fix-only re-review round now runs a smaller panel. What stays is the general seats, the relevance: always seats, a security seat without relevance: security, and each seat that reported an accepted finding still open. A security seat with relevance: security joins only when the paths the fix touched match the security rule. The other seats sit that round out. A change that touches a high-risk path, --high-risk or --only, and the closing whole-change round keep the usual panel. To turn this off, set optimization.narrow_re_reviews: false.
The CLI picks the light tier itself
When a role has a light tier, run implementer and run review_fixer now pick light themselves without --tier when the conditions are met. For the implementer, the plan touches three files or fewer and none is on a high-risk path. For the review fixer, every accepted finding is medium or lower and none is on a high-risk path. When anything is in doubt, or the previous light run of the same stage failed, the role’s own model runs. The tier chosen and why are printed on a note: line and kept in status and optimization report. Use --no-auto-tier to run the role’s own model once, or optimization.auto_tier: false to turn the choice off.
Continuing in a new session
summary now ends with a Resume: line.
Resume: in a new session, continue workflow 91433992a718: pass --workflow 91433992a718 to every command (or set DEV_ORCHESTRA_WORKFLOW=91433992a718)
The skill recommends a new session at each workflow boundary. If you keep going in one session, the orchestrator’s context keeps growing across workflows.
A short view for triage
review show --brief now shows the findings one line each (the first line of the problem, up to 140 characters), without evidence or fixes. With ids, as in review show F3 F5, only those findings are shown in full. The skill’s triage also starts from this short view.
Smaller changes and bug fixes
- The default wait of
jobs waitcan be set withprogress.wait_seconds(30 to 570 seconds) - If a job turns “may be stuck” during
jobs wait, the wait returns at once even with time left (exit code 4) - Bytecode is now cached even where the host turns bytecode caching off, as Claude Code does. Commands such as
statusstart about 0.3 seconds faster - The implementer and the review fixer run only the tests the change touches, not the whole suite. The whole suite runs once at the Test stage. A project that wants the whole suite in those runs says so in
instructions.implementerand the like - On Antigravity (agy) 1.5.7 and later, shell commands ran even with the setting that skips permission prompts turned off. Now they are refused unless that setting is on
Stabilization
I added tests for the new features and fixed gaps in the triage and re-review checks, among other work to keep things stable.
How to update
How to update depends on the CLI you use. Follow the steps in How to Update with the new version’s tag (v1.4.0). After updating, running dev-orchestra doctor once is a good idea.
Summary
- An interview that asks one question at a time before the design (
design.ask_questions: grill) - In reviews, reviewers are told what was decided, minor pre-existing problems stay out of triage, and re-reviews run only the reviewers that matter
- The CLI picks the light tier itself, a workflow can continue in a new session, and there is a short view for triage
GitHub: istb16/dev-orchestra (a star would be a big encouragement)
Related Posts
[Released] dev-orchestra: A Resumed Design Run Now Records Only Its Own Cost v1.3.2
A roundup of dev-orchestra v1.3.2. When a design was resumed and revised, that run's cost was recorded with the earlier session's cost added in. This is now fixed; tokens show and similar reports could show the architect costing more than it did.
[Released] dev-orchestra: Start the Review While the Tests Run v1.3.1
A roundup of dev-orchestra v1.3.1. The code review can now start without waiting for the tests to finish (off by default). Plan approval now defaults to auto, which skips the approval when nobody is watching, such as in CI, and the conditions are met. Claude write runs now start with only the tools they use and no MCP servers.
[Released] dev-orchestra: A Closing Review of the Whole Change v1.3.0
A roundup of dev-orchestra v1.3.0. When a review round that looked only at the fix finds nothing above Low, the review now ends with one more round over the whole change. It is on by default, and a Medium finding now calls for another round, so a review can take more rounds than before.