v0.10.82 · 2026-10-26
A reliable runtime, clearer evidence, and one verdict
Keep agents working through setup failures, choose a stable repository version, and see what current evidence actually supports.
What we heard
A team using CodeLedger in cloud coding sessions found that a failed upgrade left their agent unable to work or repair its setup. We also heard that conflicting status messages and optimistic value claims made it difficult to tell whether a session was ready to continue.
These problems have the same practical cost: time spent interpreting or repairing the tool instead of completing the task.
What we did
- Keep agents working through setup failures. Missing, outdated, broken or slow integrations report a short repair hint and let work continue. Managed enforcement remains an explicit choice; only an actual policy decision can block an edit. Repair commands stay available.
- Make the repository version predictable. Repositories can keep a chosen CodeLedger version. Upgrades prepare and validate the replacement before switching, retain a rollback copy, and refuse accidental downgrades. Routine commands leave the committed runtime alone. Repairing a missing launcher preserves session state.
- Use consistent MCP setup. Supported client configurations launch the repository's chosen runtime. Lightweight Linux environments can use the generated launchers without installing Bash. The release also prepares version-specific MCP Registry metadata. Windsurf receives a configuration to import into its client settings.
- Give local checks one verdict. Verification, task review and pre-PR checks distinguish new blocking problems from inherited debt. Problems introduced or worsened by the current work can block; existing debt stays visible without being presented as a new failure. Missing evidence is labeled, and strict compatibility remains available.
- Explain whether context can be trusted. Reports identify moved revisions, a different branch or worktree, an outdated base, mismatched tasks and older evidence, with a suggested repair. Recall reports distinguish missing evidence from measured results, and refreshing context preserves the session's work history.
- Describe value honestly. Descriptive bundle sizes and measured outcomes stay distinct. Estimates remain estimates, and a small bundle is not presented as proof of a percentage saved across the whole repository.
What this release does for you
CodeLedger becomes easier to keep stable across local and cloud sessions, and its reports make the evidence and remaining limitations easier to assess. After installing v0.10.82, run codeledger upgrade in each repository and commit the exact files it lists so cloud sessions receive the same runtime.
Managed enforcement still has a documented limitation for opaque shell commands. A configuration export is not evidence that a client has imported it, and missing activation or measurement evidence is not a successful check. We will keep these distinctions visible as the product develops.
Tell us what would make your next coding session easier on the feedback page.