Plugin improver

For: tenant administrators and plugin users.

Use Plugin improver (ai.oz0/plugin-improver) to propose changes to plugins your tenant owns and keeps in its repositories. It works on a copy of one plugin repository and produces a changeset with evidence for a person to review.

Only after approval does GitHub's push_changes open a draft pull request. A person must merge it. The improver cannot change Orchestrator Zero, another publisher's plugin or itself.

"Verified" means the project tested, pinned and secured this plugin. It does not mean that GitHub, Anthropic or any other vendor endorses it.

Install

oz0 plugin install hub:ai.oz0/github
oz0 plugin install hub:ai.oz0/harness-claude          # or harness-acp, for OpenCode or goose
oz0 plugin install hub:ai.oz0/plugin-improver --accept-reads results --reads-target notes --reads-target triage
                                                      # refused once with what it asks of ai.oz0/github
oz0 plugin install hub:ai.oz0/plugin-improver --accept-reads results --reads-target notes --reads-target triage \
  --accept-uses sha256:...
oz0 setting set PLUGIN_IMPROVER_TARGETS "notes=acme/notes, triage=acme/plugins:plugins/triage@main"
oz0 setting set PLUGIN_IMPROVER_PUBLISHER acme        # only when your plugins carry a hub listing

What it asks for:

Its settings:

Changing them is a plugin configuration change, reviewed as your tenant reviews them. Set the repositories your GitHub connection may write in GITHUB_ALLOWED_REPOS too, so that push_changes refuses any other.

Its agents ask for your smart alias (the session) and fast (the judge), never a vendor's model: map them with oz0 model alias set. The Claude Agent SDK speaks Anthropic's Messages API only, so improver needs an alias that resolves to a provider that takes it; OpenCode and goose speak others.

Start it

A person runs it from the console or the CLI:

oz0 run improve-plugin '{"target": "notes", "task": "The date format eval case fails since the last model change"}'
oz0 run improve-plugin '{"target": "notes", "evidence": ["finding:f-2c91"], "harness": "goose"}'

evidence cites findings, eval runs and comparisons by ID (finding:<id>, eval:<run>, comparison:<id>); the platform keeps the ones that are your tenant's in the change's record, with what it knows of each. An app starts it the same way through the runtime API with its key, such as a CI job of the plugin's repository whose tests failed:

      - name: Ask the improver for a change
        if: failure()
        run: |
          curl -fsS "$OZ0_URL/v1/agents/improve-plugin/jobs?tenant=acme" -H "Authorization: Bearer $OZ0_APP_KEY" \
            -H "Content-Type: application/json" \
            -d '{"target": "notes", "task": "CI test failed on ${{ github.sha }}: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}"}'
        env:
          OZ0_URL: ${{ vars.OZ0_URL }}
          OZ0_APP_KEY: ${{ secrets.OZ0_APP_KEY }}

Routines

No manifest key installs a routine, so the improver ships its routines as files in its folder, plugins/plugin-improver/routines/, which an admin saves (oz0 routine save -f <file>) with an owner_team of the tenant's:

A routine's fire is a job an outside event starts, and such a job never runs a shell and a write in one run (routines). The improver's session writes files, and its changeset ends in a write, so a fire does not run it: improver-notice reads only, works out the target and the cause, and ends with the command that starts improve-plugin, for the person who reads the fire's job.

A CI system that should start the improver itself does it through the runtime API, as above. improver-after-merge runs no session: its routine lists the two platform built-ins its flow calls, oz.report_outcome and oz.propose_update.

Related tasks

• How an improvement runs: Follow evidence collection, candidate checks and the work after a merge.
• Improver limits and records: Look up its boundaries, retained evidence and agent names.