VS Code 1.135: a second model reviews your agent's work
VS Code 1.135 ships Rubber Duck, an experimental second-opinion pass from a complementary model, plus per-turn token accounting. Build the same loop in CI.
The most useful thing in VS Code 1.135 is a slash command with a joke name. /rubber-duck sends your coding agent's work to a different model and asks it what the first one missed. Microsoft filed it under experimental. We'd file it under the pattern every team running coding agents should have already built.
What actually happened
Per the VS Code 1.135 release notes, shipped August 26, Rubber Duck runs inside Copilot agent host sessions and pulls "a second opinion from a complementary model" to surface overlooked details and edge cases. Complementary is the operative word — the point is a model that fails differently from the one that wrote the code.
Three other changes in the same release:
- External agent sessions. You can continue a Copilot or Claude agent session started in another application, gated behind
chat.agentSessions.showExternal. The session is no longer trapped in the tool that opened it. - Per-turn token accounting. Hover a chat response footer and you get input, cached input, and output tokens broken out by model, for that single turn. GitHub gave you this at the billing level on August 11. Now it's at the turn level, next to the thing that spent it.
- Agent Host. Agent harnesses run in a dedicated process over the Agent Host Protocol, with the Copilot agent itself running on the Copilot SDK. Same plumbing that opened the Agents window without a GitHub account in 1.133.
Why a second-model review matters for your business
One model checking its own work is not review. It shares the training data, the tokenizer quirks, and the blind spots that produced the bug. A second model from a different family is the cheapest independent check you can buy — and it costs one extra call, not one extra engineer.
Don't leave the pattern in the editor. Rubber Duck fires when a developer remembers to type it. The same call belongs in your pipeline: agent writes the diff, a second model reviews it against the ticket and the tests, the result lands as a PR comment. That runs on every change, not on the ones someone felt uneasy about.
Per-turn tokens make the second pass a line item you can defend. You can now see exactly what the review call costs against what the generation call cost. If the reviewer catches one production bug a quarter, the math is not close.
Key takeaways
- VS Code 1.135 shipped August 26 with `/rubber-duck`, an experimental second-opinion pass from a complementary model
- External agent sessions let you resume Copilot or Claude sessions started in other apps (`chat.agentSessions.showExternal`)
- Chat response footers now show input, cached input, and output tokens per model, per turn
- Agent harnesses run in a dedicated Agent Host process over the Agent Host Protocol
- Pick a reviewer model from a different family than the generator — shared blind spots defeat the point
- Put the second-model review in CI, not just in the editor, so it runs on every diff
An editor command that fires when someone remembers it is not a review process. We wire cross-model review into the pipeline — generator, independent reviewer, verdict on the PR — so the check runs whether or not anyone types the slash command. See how we build the review loop, or send us your current agent setup and we'll tell you where it's unguarded.
Sources: VS Code 1.135 release notes.
- #vs-code
- #ai-coding
- #code-review
- #dev-tools
- #token-costs
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Omacom Foundation's $10M: who holds your tool's trademark?
Ten tech founders put $1M each behind a Linux distro's foundation. The interesting line isn't the money — it's that the foundation holds the trademarks.
Read itMicroduck: a $399 open-source physical-AI dev kit
Hugging Face's 25cm biped ships lidar, 15 motors, and the full reinforcement learning stack under Apache-2.0. Robotics prototyping is now an expense line, not a project.
Read it