You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Let custom agents inherit the selected model by default #5747
Some custom agents set a specific model in their frontmatter. I want to choose the model for my session without having a reusable agent profile select a different one or fail because its preset is unavailable.
For example, SE: Security sets model: GPT-5. Invoking it in Copilot returned:
Model 'GPT-5' is not available.
The available model list did not include that name. This does not mean every preset fails in every client.
Other examples at the same revision:
Agent
Model preset
Context Architect, SE: Security, SE: Tech Writer
GPT-5
Custom Agent Foundry
Claude Sonnet 4.5
Meta Agentic Project Scaffold
GPT-4.1
Agent Governance Reviewer
gpt-4o
Client-specific model overrides and personal profiles can work around this. I would rather not maintain those overrides just to use shared agent instructions.
Suggested change: omit model by default so the agent inherits the selected/default model. If an agent genuinely depends on a particular model, document why and the supported alternatives instead of making that constraint the default for general-purpose profiles.
GitHub's configuration reference documents default-model inheritance when model is unset. VS Code's reference also documents selected-model inheritance and ordered model preferences. Behavior differs by harness, so this is not a claim that presets cannot be overridden.
The value is portability: the same instructions and tools should remain usable as model availability, account access and user preferences change.
I confirmed a workaround for the failed SE: Security launch: explicitly pass the current parent model in the delegation call. My successful test included:
Use the current parent model's supported ID, not this particular ID everywhere. The override was accepted and the agent launched without the previous GPT-5 error. The tool does not separately expose backend model telemetry.
CLI 1.0.102 also documents /subagents for per-agent model preferences, including inherit. Its noninteractive copilot config command rejects the nested subagent model setting, so that is not an equivalent workaround.
I have not verified that changing only the desktop's main model picker overrides an agent preset. If an explicit override fails, stop and surface the error rather than silently switching models.
This makes the preset avoidable in the tested invocation path, but default inheritance would still remove the need for callers to maintain these overrides.
This is most likely an indication of items in the repo that are in need of a review due to staleness. If you're finding situations like this, send through a PR to uplift the agent, but otherwise we're working in the confines of what the client supports.
This is most likely an indication of items in the repo that are in need of a review due to staleness. If you're finding situations like this, send through a PR to uplift the agent, but otherwise we're working in the confines of what the client supports.
Sure thing.
Would a style guide or some sort of linter be able to guard against this?
We have a linter on skills, but that's the only thing that has a "spec" of sort that we could do conforming with.
We have an agentic workflow that runs weekly that surfaces a staleness report, but that requires someone to have the bandwidth the actually action against it
Activity
arielvalentin commentedon Oct 27, 2026
I confirmed a workaround for the failed
SE: Securitylaunch: explicitly pass the current parent model in the delegation call. My successful test included:{ "agent_type": "SE: Security", "model": "gpt-6.1-sol" }Use the current parent model's supported ID, not this particular ID everywhere. The override was accepted and the agent launched without the previous
GPT-5error. The tool does not separately expose backend model telemetry.CLI 1.0.102 also documents
/subagentsfor per-agent model preferences, includinginherit. Its noninteractivecopilot configcommand rejects the nested subagent model setting, so that is not an equivalent workaround.I have not verified that changing only the desktop's main model picker overrides an agent preset. If an explicit override fails, stop and surface the error rather than silently switching models.
This makes the preset avoidable in the tested invocation path, but default inheritance would still remove the need for callers to maintain these overrides.
aaronpowell commentedon Oct 28, 2026
This is most likely an indication of items in the repo that are in need of a review due to staleness. If you're finding situations like this, send through a PR to uplift the agent, but otherwise we're working in the confines of what the client supports.
arielvalentin commentedon Oct 28, 2026
Sure thing.
Would a style guide or some sort of linter be able to guard against this?
aaronpowell commentedon Oct 29, 2026
We have a linter on skills, but that's the only thing that has a "spec" of sort that we could do conforming with.
We have an agentic workflow that runs weekly that surfaces a staleness report, but that requires someone to have the bandwidth the actually action against it