Update the Learning Agent
First update from an older release
Section titled “First update from an older release”Version 0.4.1 does not contain the update command. Close Claude Code, Claude Desktop Code sessions, and Codex. Use this one-time bootstrap:
npx --yes --registry=https://registry.npmjs.org --@iterahq:registry=https://registry.npmjs.org @iterahq/learning-agent@latest updateRequire UPDATED version=0.4.4.
The registry options bypass an existing GitHub Packages scope.
Later updates
Section titled “Later updates”itera-learning updateThe command installs the current public npm release. It detects existing Claude Code and Codex integrations and refreshes only those hooks. It does not add an integration that was absent.
The update preserves these local files:
- The capture spool.
- The Ship journal.
- The shipped ledger and cursors.
- Credentials and consent rules.
- Unrelated Claude Code and Codex settings.
Check the update
Section titled “Check the update”itera-learning helpitera-learning pathsitera-learning consent verifyThe help output must start with itera-learning 0.4.4.
The consent check must print CONSENT_CLI_OK.
Then verify one fresh session.
Upgrade a forced Windows 0.4.2 installation
Section titled “Upgrade a forced Windows 0.4.2 installation”The 0.4.2 update command cannot start npm on Windows.
If you installed 0.4.2 with --force, close Claude Code and Codex.
Run this command once:
npm install -g @iterahq/learning-agent@0.4.4 --registry=https://registry.npmjs.org --@iterahq:registry=https://registry.npmjs.orgInstall each integration that you use, then verify consent:
itera-learning install claude-codeitera-learning install codexitera-learning consent verifyInstall only the integrations that you use.
Require CONSENT_CLI_OK rules=3 protection=windows-acl scope=capture.
The Windows hooks use Node.js directly and do not require a POSIX shell.
After this bootstrap, use only itera-learning update for later releases.
Repair data from 0.4.1
Section titled “Repair data from 0.4.1”If the Ship journal reports invalid lines, close the capture clients and run:
itera-learning repair-spoolThe command records exact originals and invalid lines in a private quarantine. It removes only invalid physical records from the active spool. Valid records can replay safely because the engine deduplicates stable episode IDs.