backnotprop/plannotator v0.27.18

on GitHub

Follow @plannotator on X for updates

What's New in v0.27.18

A short release with four pull requests. Model pickers now show the models your installed Claude and Codex actually offer, including Opus 5.5, Fable 5.1, and the GPT-6 family, instead of a list we typed by hand. Unsent comments on a pull request review now survive new pushes. The Claude Agent SDK behind Ask AI also moves to a newer version.

Model pickers come from your installed tools

Every model picker in Plannotator used to read from lists written into the code. Ask AI, the review agents, Guided Review, and Code Tour each had their own copy for Claude and Codex, and the copies fell behind whenever a new model shipped. Opus 5.5 and Fable 5.1 could not be picked at all, code review still defaulted to Opus 5, and the Codex pickers were missing the GPT-6 models.

The pickers now ask the tools you already have. Claude's list comes from your installed claude, and Codex's list comes from your installed codex. When you update either one, new models appear in Plannotator without a Plannotator release. The lookup runs once, the first time a picker opens, never at startup or on page load. It gives up after a few seconds and falls back to a short built-in list if the tool is missing, signed out, or broken. Ask AI and all the launchers share one list, so they can no longer disagree.

Claude entries now carry their version: Opus 5.5 (latest), Fable 5.1 (latest), Sonnet 5 (latest), Haiku 4.5 (latest). The number comes from the model each name currently runs, so it updates on its own when Claude does. Code review defaults to the latest Opus, and Codex defaults to the model Codex marks as its default. Each picker only offers the effort levels the chosen model supports, and fast mode only appears where it works.

Choices you saved before this release keep working. A saved pick that your tool no longer lists moves to the latest model in the same family, so an old Opus pick becomes the latest Opus and an old Sonnet pick stays Sonnet. A regular pick never turns into a 1M-context one. A retired Codex model moves to Codex's current default with fast mode turned off.

(#1593)

Unsent PR review comments survive new pushes

Draft review comments were saved under a fingerprint of the exact diff. When a teammate pushed to the pull request, the diff changed, the fingerprint changed, and the next session could not find the comments. They stayed on disk but nothing pointed to them anymore.

In PR reviews the draft is now also saved under the pull request itself (host, repository, number, and layer or full-stack view). Reopen the PR after new commits and your unsent comments come back. Each line comment remembers the code it was written on and a few lines around it. If that code is unchanged, the comment stays where it was. If it changed, the comment gets an Outdated tag in the sidebar, where you can still edit or delete it. Outdated comments are never moved to a guessed line. When you post the review, they go in the review summary with the code they were written on, not onto whatever line now sits at that number.

Switching PRs, or switching between layer and full-stack view, in the same tab brings back that PR's unsent comments automatically with a short notice. Sending, approving, or closing clears the saved draft for every PR the session touched, so submitted comments do not reappear after a later push. After a push, files you had marked Viewed are unmarked, since new commits may have changed them. Local reviews are unchanged. Drafts saved before this release are still restored when the diff is unchanged, but not after a push.

(#1592, part of #1590, reported by @RobertoArtiles)

Additional Changes

• Claude Agent SDK 0.3.273. Ask AI's Claude provider moves from 0.3.260. One behavior change comes with it: when Claude changes directory during a conversation, follow-up questions now start in that directory instead of resetting. (#1594)
• Docs. The AI features guide no longer lists models by name, since the pickers now show whatever your installed tools offer.
• @plannotator/ui 0.46.0 for hosts. The hand-written model list exports are removed in favor of the shared catalog in @plannotator/core 0.25.6. packages/ui/HANDOFF.md lists every removed export.

Install / Update

macOS / Linux:

curl -fsSL https://plannotator.ai/install.sh | bash

Windows:

irm https://plannotator.ai/install.ps1 | iex

Claude Code Plugin: Run /plugin in Claude Code, find plannotator, and click "Update now".

Pi: Update @plannotator/pi-extension to 0.27.18 and restart Pi.

OpenCode: Clear cache and restart:

rm -rf ~/.bun/install/cache/@plannotator

What's Changed

• chore(deps): bump @anthropic-ai/claude-agent-sdk to 0.3.273 by @backnotprop in #1594
• feat(ai): discover models from installed tools instead of hand-maintained lists by @backnotprop in #1593
• fix(review): keep unsent PR review comments across new commits by @backnotprop in #1592
• fix: v0.27.18 QA regressions (PR drafts, discovered model catalogs) by @backnotprop in #1595

Community

@RobertoArtiles reported #1590 with an exact reproduction and a proposed design: key review drafts by what is being reviewed rather than the diff's content, and mark comments outdated instead of dropping them, the way GitHub does. This release ships that design for pull request reviews. Local reviews are still open under the same issue.

Full Changelog: v0.27.17...v0.27.18