Keiko 1.3.1

Keiko 1.3.1 Release Notes

High · Fixes

• A long coding run keeps its live activity feed and its browser pairing. A busy feed waits for the browser instead of closing, a dropped feed reconnects on its own, and a quiet model turn no longer signs the window out.
• In Full access on macOS, when the isolated verifier cannot install a project's dependencies, the model is told to run the project's own checks in its shell, and the Activity Log names the reason.
• The task timeline says why a call failed, such as the command's exit code, and model reasoning no longer carries a warning-coloured badge.
• Routine Coding Workbench actions are recorded as interaction events instead of browser diagnostics, so real browser failures stand out in the Activity Log.
• The editor answers go to definition, find references, rename, code actions, signature help and go to line again.
• Chats that use a Git change can connect, send, and review, approve and apply proposals again, and pull request description review finds its proposal.
• Applying a pull request description no longer reports a failure when Keiko runs in GitHub Codespaces or on a hosted runner.
• On Linux, strict sandbox runs keep temporary files outside the worktree, so verified commits are no longer reported as failed. Below Full access, governed research can ask for its network approval again.
• A restored Coding Workbench window picks up the project's repository, a signed-out window shows the pairing guidance, and an open Changes section stays open while it refreshes.
• keiko start in a project installation no longer warns about a stale binary when it runs the installed package through npm's bin link.
• A Coding Workbench run that passed the repository's own combined check in its shell, such as npm run check running the type check, lint and tests, now counts every check that line ran. Before, such a run could be reported as not verified because only its tests counted.
• Operators of trusted model deployments can start Keiko with KEIKO_CONTENT_REDACTION=off so Chat keeps authorized document details such as phone numbers, account numbers, configuration values and code. Credentials, private keys, URL credentials and Authorization values are still removed, memory proposals keep their redaction, and the default is unchanged. The setting applies to every configured model provider, and the Activity Log records it at startup.
• The Coding Workbench hands its agent the model's answers as written, so a quoted change in a closing summary no longer appears as [REDACTED]. Only credentials, private keys, URL credentials and Authorization values are removed, in every installation.
• A command that ran and returned a non-zero exit code, such as a test run that reproduces a defect, shows its exit code in the task timeline instead of a red Failed, as OpenCode reports it. A tool that reported an error still shows Failed with its reason.
• Updating from 1.2.7 or 1.3.0 needs no action. An installation on 1.2.6 or earlier still runs the verified updater bootstrap once before automatic updates.

Normal · Update Notes

• Download Keiko for Windows x64, Linux x64, macOS arm64, and macOS x64 and start from the bundled launcher; npm remains available for developer and compatibility workflows.

Technical release metadata

• Package: @oscharko-dev/keiko@1.3.1
• Release tag: v1.3.1
• npm dist-tag: latest
• Registry: https://registry.npmjs.org/
• Generated by: npm run release:publish

Structured impact entries:

• 2026-10-27-keiko-1.3.1-coding-workbench-reliability: high / fixes; supported from 0.2.0, 1.2.5, 1.2.6, 1.2.7, 1.3.0; affected state stores ui-store-schema, workspace-search-index; remediation no-action-required; default patch notes yes; traceability catalog:2026-10-27-keiko-1.3.1-coding-workbench-reliability; review:github-issue-comment:#3897#6443332275