Skill versions and rollback
Use standalone Evolver to fetch a skill file and maintain a recoverable installed version. Identify the destination directory and the tasks that use the skill before changing it.
A controlled Skill version is defined by its source asset ID, the actual SKILL.md content, and its installation path. Preview before updating; overwriting a same-name file is not an automatic upgrade.
Version identity includes content and installation location
A skill name describes its purpose, not a reproducible version. Record the source asset ID, installed content, and destination directory. Same-named SKILL.md files in different directories or from different assets require separate checks. Updating one does not establish that every consumer has switched.
Review behavioral differences: new tool calls, broader paths, changed output locations, external dependencies, or altered validation and stop conditions. Successful download establishes retrieval; adoption also requires deciding whether those changes fit the task's permissions and expected behavior.
Scenario: return to the previous version
Before updating, preserve the old file, source asset ID, validation commands, and consuming-task inventory. After installation, use a bounded task to validate the candidate and establish that the environment loads the new file rather than a same-named copy elsewhere.
If validation fails, stop new tasks from using the candidate, restore the old file, and revalidate. Separately reconcile file changes and external outcomes from tasks that already used the new version. File rollback changes future skill use; past actions require task-level handling. Record both the active file version and the disposition of affected tasks.
Inspect a candidate version
evolver skill fetch --asset <asset_id>
Record the asset ID, source, summary, and content diff. Evolver revalidates content addressing before writing, and existing files are not overwritten by default.
Install or update
For a first installation into the confirmed destination:
evolver skill fetch --asset <asset_id> --write --out <dir>
If a file already exists, pause new tasks that use the skill, compare the candidate content, and save the current version. Only when replacement is intended, use:
evolver skill fetch --asset <asset_id> --write --out <dir> --force
Choose the command for the installation case; do not run both in sequence. Track the skill directory in Git or a versioned backup and record the asset ID. After writing, inspect the actual file, directory, and environment that loads the skill, then run the planned validation. Stop if integrity validation fails; overwrite flags do not bypass that check.
Version record
skill name
installed asset_id
source and fetched time
target directory
reviewer and validation result
previous asset_id or backup revision
Roll back
Stop using the new version, restore the last validated SKILL.md, and rerun the original task and validation. The skill fetch flow does not provide a general recycle-bin recovery feature, so rollback depends on a file or repository revision preserved in advance.
Remove
Check projects, automations, and documentation for references before deleting a Skill file. Deleting the local file does not revoke the network asset; revoke trust separately when the asset itself is no longer acceptable.
Completion criteria
The installed file matches the reviewed candidate, validation identifies the current asset ID, and the previous version is recoverable. Re-test after rollback and record the version actually in use. Restoring a file does not undo external actions already performed by the new version.
Related pages
EvoX Docs · Developers · Experience asset management