feat(task): allow selecting a model per task invocation (1.18.23 lineage) - #4
Merged
leoncheng57 merged 2 commits intoAug 26, 2026
Conversation
Adds an optional `model` parameter ("provider/model") to the native task
tool, so a delegating agent can route one subtask to a different model
without changing the subagent's configuration.
Precedence is explicit per-invocation model > subagent's configured model
> invoking assistant's model. Omitting the parameter is behaviourally
identical to before.
The model is resolved once, before the foreground/background branch, so a
BackgroundJob child runs the same model a foreground one would and the
`{ providerID, modelID }` tool metadata stays correct on every return path.
Validation is early: an explicit model goes through the same
Provider.getModel the prompt path calls later, but before the child
session is created, so an unavailable model fails without leaving an
orphaned child. Omission skips that call entirely and leaves the existing
fallback chain's late validation untouched.
The parent's variant is only inherited when the child actually runs the
parent's model, so a variant cannot leak onto a different model.
Provider.node is already a tool-registry dependency, so acquiring
Provider.Service adds no new wiring; the test layer is updated to match.
Wire shape: the new parameter is optional, string-typed, and rejects a non-string; the JSON Schema snapshot records what the calling model sees. Behaviour, against a config-declared provider so model lookups resolve deterministically without network access or loading an SDK: - explicit model beats the invoking assistant's model - explicit model beats the subagent's configured model - omission preserves both existing fallbacks, including that the parent's variant is still inherited only when no model overrides it - an explicit model on a resumed task_id applies to that invocation and creates no second child, and does not persist onto the child session - running and result metadata both report the selected pair - a model id containing extra slashes parses as provider + rest - a background launch runs the child on the selected model, asserted on the prompt actually addressed to the child rather than on the hand-back prompt injected into the parent Validation is pinned to what the code really does rather than to the aspiration: an unknown or malformed model fails with Provider's own ProviderModelNotFoundError before any child session exists, before the child is prompted, and before any running metadata is emitted.
leoncheng57
marked this pull request as ready for review
August 26, 2026 17:19
leoncheng57
merged commit Aug 26, 2026
d677b7e
into
fix/subagent-effective-deny-inheritance
4 of 9 checks passed
|
Thanks for updating your PR! It now meets our contributing guidelines. 👍 |
This was referenced Aug 26, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Issue for this PR
Closes leoncheng57/custom-dca-opencode#90. Upstream tracking issue: anomalyco#6651.
Fork-internal PR: this targets
fix/subagent-effective-deny-inheritance, the branch our local patched binary is built from. It is not a contribution to anomalyco/opencode — see the note at the bottom.Type of change
What does this PR do?
Adds an optional
modelparameter to thetasktool so one delegation can run on a different model than the subagent's configured default.Resolution order becomes: explicit
model> subagent's configured model > invoking assistant's model. Omitting the parameter takes the same branch as before, so existing behaviour is untouched.The model is resolved once, before the foreground/background split, so a background
BackgroundJobchild runs on the same model a foreground one would, and both report it in tool metadata. Metadata still carries the normalized{ providerID, modelID }object, so downstream consumers that read task provenance need no change.Validation is early:
provider.getModelis called before the child session is created, so an unknown model fails withProviderModelNotFoundErrorand leaves no orphaned child. This is the same call the prompt path makes later, so it changes when the failure happens, not whether.On a resume via
task_id, an explicit model applies to that invocation only and is not persisted as a session default.How did you verify your code works?
bun test test/tool/task.test.ts test/tool/parameters.test.ts— 91 pass, 0 failbun run typecheck(tsgo) — exit 0packages/opencodesuite — 3329 pass, 1 fail; the failure ischunkTimeout ... SSE body stalls, which reproduces identically on the pristine basemodelappears in the task tool's properties and is absent fromrequired; the old binary has nomodelkey at allBackground coverage needed one correction: the completion handler injects a hand-back prompt into the parent, so a naive last-prompt capture asserts the wrong message. The test now asserts on the prompt addressed to the child.
Screenshots / recordings
Not a UI change.
Checklist
Note on upstream
This is deliberately not proposed to anomalyco/opencode. That repository already has an active implementation of the same feature in anomalyco#34947, which additionally gates the parameter behind a
model_overridepermission defaulting to deny. Earlier raw-parameter attempts (anomalyco#26535, anomalyco#29447) were closed as superseded by it. This fork carries the change only so the locally pinned binary has it.