feat(agent-core): enable per-subagent model selection by default and allow [models] aliases - #2439
feat(agent-core): enable per-subagent model selection by default and allow [models] aliases#2439imadcat wants to merge 1 commit into
Conversation
…allow [models] aliases Subagents can now run on a different model than the main agent out of the box: - The secondary-model experiment is promoted to default-on in both agent-core and agent-core-v2 (opt out via KIMI_CODE_EXPERIMENTAL_SECONDARY_MODEL=0 or [experimental] secondary-model = false). - The Agent/AgentSwarm 'model' parameter accepts any configured [models] alias in addition to the symbolic 'primary'/'secondary'. - Profile modelPreference (agent-file 'model_preference') likewise accepts a concrete alias, validated at binding-resolution time. - Unknown aliases fail the spawn up front with a tool-visible error listing the valid choices instead of silently inheriting the parent's model. - Tool descriptions now advertise the configured aliases so the parent model can pick per spawn.
🦋 Changeset detectedLatest commit: 8ce2d93 The changes in this PR will be included in the next version bump. This PR includes changesets to release 6 packages
Not sure what this means? Click here to learn what changesets are. Click here if you're a maintainer who wants to add another changeset to this PR |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 8ce2d93f32
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| throw new Error(unknownSubagentModelMessage(requested, modelIds)); | ||
| } | ||
| const record = config.get<ModelsSection>(MODELS_SECTION)?.[requested]; | ||
| return { model: requested, thinking: record?.defaultEffort ?? own.thinkingLevel }; |
There was a problem hiding this comment.
When a configured alias stores its user-pinned default_effort under [models."<alias>".overrides], passing that alias through the new model/model_preference path reaches this branch but reads only the raw top-level record.defaultEffort, so the subagent inherits the caller's thinking level instead of using the alias's effective default. The model resolver already merges overrides via effectiveModelConfig/effectiveModelAlias, so this branch should use the same effective model view; the v1 helper has the same direct raw read and should be kept in parity.
Useful? React with 👍 / 👎.
1 participant
Related Issue
Resolve #2438
Problem
Subagents spawned via the
Agent/AgentSwarmtools always run on the same model as the main agent unless the operator discovers and enables the experimentalsecondary-modelflag, which is off by default. Even then, themodelparameter only accepts the symbolicprimary/secondarychoices — no way to point one subagent at an arbitrary configured[models]alias — and agent profiles cannot name a concrete model. See linked issue.What changed
Covers both engines (
agent-coreandagent-core-v2, parity-tested):secondary-modelexperiment to default-on (flags/registry.ts,session/subagent/flag.ts). Users who never configure[secondary_model]see zero behavior change (the existing inherit-parent fallback is untouched); the legacy behavior stays available viaKIMI_CODE_EXPERIMENTAL_SECONDARY_MODEL=0or[experimental] secondary-model = false.modelparameter accepts any configured[models]alias inAgentandAgentSwarm(v1collaboration/agent.ts/agent-swarm.ts, v2tools/agent/agent-swarm), in addition to the symbolicprimary/secondary. A concrete alias is honored flag-independently — an explicit choice is never silently ignored.modelPreference/ agent-filemodel_preference(v1profile/types.ts, v2agentProfileCatalog.ts), validated at binding-resolution time rather than profile-load time.resolveSubagentBindingthrows before the child agent is created, so the tool call rejects with e.g.Unknown subagent model "typo". Pass one of: "primary", "secondary", "main", "cheap".instead of silently inheriting the parent's model or failing mid-turn.buildSubagentModelDescriptionsin both engines) so the parent model can pick per spawn.default_effortwhen configured, otherwise inherits the caller's level (symmetric in both engines).This builds on the existing secondary-model plumbing rather than adding a parallel mechanism, which keeps the diff small and both engines symmetric.
Checklist
agent-core:test/session/subagent-host.test.ts,test/config/secondary-model.test.ts,test/tools/agent.test.ts,test/profile/agentfile.test.ts;agent-core-v2:test/app/config/config.test.ts,test/tool/tool.test.ts,test/workspace/workspaceAgentProfileLoader/agentFile.test.ts; full engine suites pass: 510 files / 8452 tests, plus the v1↔v2 parity suite.)gen-changesetsskill, or this PR needs no changeset. (minor changesets for@moonshot-ai/agent-coreand@moonshot-ai/agent-core-v2)gen-docsskill, or this PR needs no doc update. (docs/en+docs/zh:reference/tools.md,customization/agents.md,configuration/config-files.md,configuration/env-vars.md)