Note
Access to this page requires authorization. You can try signing in or changing directories.
Access to this page requires authorization. You can try changing directories.
Use the Frontier Tuning plugin for GitHub Copilot to turn a workflow into reviewed skills, samples, and evaluation criteria in a tuning environment. Then evaluate a baseline and inspect the answers and scores to decide whether anything needs to change. You can stop at a satisfactory baseline; model training and publication are optional.
GitHub Copilot hosts the conversation, the plugin supplies tuning guidance, and the Frontier Tuning CLI performs service operations. This is a maker workflow in GitHub Copilot, not the separate expert-contribution journey in Microsoft 365 Copilot Cowork. For context on tuning environments and artifacts, see the tuning process.
Before you begin, confirm access to GitHub Copilot and Frontier Tuning. Review the software and authorization requirements in Prerequisites.
Prerequisites
- GitHub Copilot CLI access, including the entitlement, organizational policy access, installation, and host sign-in described in Installing GitHub Copilot CLI. The GitHub Copilot app is an alternate host.
- Frontier Tuning service access with an account authorized for the actions you intend to take.
- Frontier Tuning CLI version 1.0.17 or later, as specified by the plugin draft. Follow the manual setup procedure before service operations.
- A workflow to describe and representative examples of requests, expected answers, edge cases, and requests the agent should decline.
Install and verify the plugin
Use terminal marketplace installation as the primary route for this draft. GitHub documents generic marketplace and repository installation, listing, updates, and removal in Finding and installing plugins for GitHub Copilot CLI. Use that guidance to set up the host and marketplaces.
Use the terminal marketplace route (primary)
After the named listing is confirmed for your access:
Install the plugin.
copilot plugin install frontier-tuning@awesome-copilotList installed plugins, and check that
frontier-tuningis installed and enabled.copilot plugin listStart a new GitHub Copilot session.
Use an alternate installation route
These alternatives also need confirmation. After installation, check the installed plugin and start a new session before continuing with the same CLI setup and baseline task.
GitHub Copilot app: If this build supports the source's app route, select Customize in the left navigation, go to Plugins, search for
frontier-tuning, and install the Frontier Tuning plugin.Direct repository: If the repository is the confirmed supported route for your access, use the source command:
copilot plugin install microsoft/frontier-tuning-ghcp-plugin
Set up and check the Frontier Tuning CLI
Set up the Frontier Tuning CLI manually for the primary path. GitHub Copilot host sign-in and Frontier Tuning service sign-in are separate.
Follow the official installer instructions in the Frontier Tuning CLI documentation. Review the installer effects before proceeding: the documented installers can acquire Python and dependencies and update
PATHor shell configuration.Check the installed version against the plugin's stated minimum of 1.0.17.
frontier-tuning --versionSign in yourself in your own terminal. Choose the intended account, and accept or decline the preview terms.
frontier-tuning auth login --interactiveCheck authentication and service access.
frontier-tuning auth check
If access is denied, follow the check command's guidance. Confirm the specific consent or access grant needed rather than treating an authentication check as authorization for later changes.
For plugin-assisted setup, review the proposed installation and its effects before approving it. Complete interactive sign-in yourself; don't delegate account selection or acceptance of terms to Copilot.
Create or select a tuning environment
Use the support-ticket example below, or describe your own workflow. The proposed artifacts depend on the conversation, not on a fixed template.
In the new session, enter:
Create a tuning environment for triaging support tickets.Review whether an existing tuning environment fits the task before creating another one. Confirm the target environment and skill using the actual identity returned by the service.
For a new environment, describe the request and what a good answer includes. Review the proposed name, description, skill instructions, first evaluation criterion, and example prompts before approving creation.
Review representative samples, including edge cases and requests the agent should decline. Check which samples are for evaluation and which are for training.
Correct the chat-authored drafts, and accept only the content you want saved. Ask Copilot to save the accepted artifacts and read them back from the target environment.
Chat proposals need review before saving. Service-generated quality checks are a different operation: they save immediately, as described in the next section.
To resume existing work, you can enter "Continue tuning my support ticket skill". Confirm the target and current saved content again; an earlier chat isn't evidence of the environment's current state.
Review quality checks before evaluation
Evaluation criteria define what makes an answer good. For ticket triage, the source examples include assigning the right priority and routing to the correct team.
- Review the proposed checks against the workflow. Correct unclear wording or checks that reward the wrong behavior.
- If you choose service-generated quality checks, approve the cost and immediate save before generation. Unlike chat-authored proposals, this operation saves checks right away.
- Inspect the saved checks and their enabled state before including them in the baseline. Accept their wording before use; don't assume that generation left them disabled.
- Decide whether to request the optional readiness assessment or go directly to the baseline. Review any proposed assessment and possible charges separately.
If expert review would help, arrange it before evaluation. Any proposal to add an expert as a Contributor needs confirmation of the person, the access grant, and your authority to grant that role; content approval alone doesn't grant access.
Evaluate and inspect a baseline
Use a small first evaluation to understand current performance before changing the agent.
- Ask Copilot to propose a baseline evaluation for the saved environment.
- Review the target, samples, enabled quality checks, model, number of repeats, and jobs. The source suggests three repeats when you want to examine variation; three isn't a required minimum.
- Approve the evaluation's scope and possible Frontier Tuning charges before starting it. Command-tool approval isn't cost approval.
- Inspect the returned job and result identities for the intended environment. Ask about progress while the service jobs run, and confirm their final outcome before interpreting results.
- Review each sample's actual answer, score, and grading. Examine runs that weren't scored separately from low-scoring answers. You can ask "Explain my latest evaluation results".
- Decide whether the baseline meets your needs. If it does, stop without making a change.
You can also ask Copilot to write an evaluation report file. Confirm its destination and the data it includes before saving or sharing it; the source doesn't establish a report path or format.
Make one change and compare
Continue only if the baseline reveals a problem worth addressing.
- Ask Copilot to investigate failures before proposing a fix. Check whether each failure comes from the answer, the workflow instructions, or an unclear quality check.
- Review one focused proposed change in full. Start with an instruction edit when it addresses the problem, and identify what should improve and what should stay unchanged.
- Approve the change, then inspect the saved content after Copilot applies and reads it back.
- Review and separately approve a matched reevaluation, including its possible charges. Keep the samples, quality checks, and repeats unchanged. Keep the model and inference strategy unchanged unless one of them is the variable you're testing.
- Compare answers and scores sample by sample against the baseline. Decide whether to keep the change, restore the previous version, try another change, or stop.
If you change the measurement, such as the samples or quality checks, establish a new baseline before claiming improvement. Confirm the saved state after keeping or restoring a version; exact comparison and restore behavior still needs validation.
Other improvements are optional and have different effects:
| Option | Decision before proceeding |
|---|---|
| Instruction refinement. | Review proposed instruction edits before saving them. The source describes refinement as a paid service job, not model-weight training. |
| Inference strategy. | Consider BestOfN or TreeSearch when answers vary. Review model eligibility and recurring usage charges; inference search isn't weight training. |
| Model-weight training. | Consider it only when instruction changes stop helping. The source and CLI reference specify at least 11 Training samples for model tuning, not for the initial baseline. Confirm the applicable counting scope and keep evaluation samples distinct. |
Publish only after a separate decision
Publication isn't required to finish this task. If you choose it, review the current saved version, exact destination, and audience or scope, then give separate publication approval. Inspect the actual service result and distinguish publication success from verified availability to the intended users.
Use the Frontier Tuning CLI documentation for applicable service publication operations. Use plugins with Copilot Cowork covers the generic Cowork lifecycle, not proof of a Frontier-generated package. Changing an inference model is a separate action, not publication.
Understand approvals, usage, and data
Keep these decisions separate throughout the conversation:
- Approve command execution in the host.
- Accept proposed content and saved-state changes.
- Approve the scope and possible cost of each paid operation.
- Authorize access grants or connected-service access.
- Approve publication to a specific destination and scope.
GitHub Copilot conversation and model usage consume GitHub AI Credits according to your plan and policy. See Usage-based billing for individuals or Usage-based billing for organizations and enterprises. Don't assume that every plan has the same included allowance.
Review Frontier Tuning service usage separately from GitHub host usage. The plugin source describes separate service charges for operations such as evaluation, generation, refinement, and training, and describes service usage in tokens rather than currency. Don't use GitHub credit pricing to estimate Frontier Tuning charges.
Saved environment artifacts let you resume work from service state. This doesn't establish that chats, prompts, tool output, reports, or diagnostics are absent from host or service storage.
Update or remove the plugin
Follow GitHub's plugin maintenance guidance, subject to your organization's plugin policy.
Confirm the installed plugin name with
copilot plugin list.To update the source's named plugin, run:
copilot plugin update frontier-tuningStart a new Copilot session after updating.
For removal, GitHub documents copilot plugin uninstall NAME. If the confirmed installed name is frontier-tuning and policy permits removal, the corresponding command is:
copilot plugin uninstall frontier-tuning
In the app alternative, the source places update and removal under Customize > Plugins. Confirm that navigation for the supported app build.
Check setup problems
Start with the symptom rather than a new installation or paid job.
- Ask Copilot to "diagnose my Frontier Tuning setup." Review its proposed checks and the smallest fix before approving changes.
- If Copilot doesn't recognize the plugin, run
copilot plugin list, check whetherfrontier-tuningis enabled, and start a new session. - If sign-in fails or expires, run
frontier-tuning auth login --interactiveyourself in your terminal, then check access again.
The source recommends version, authentication, and job checks first. Confirm their effects for the installed CLI build; a diagnostic operation isn't a guarantee of no local changes, because the CLI reference describes automatic updates during normal commands.