fix(task): preserve command-selected models for command-triggered subtasks

[gmnstr]

Issue for this PR

Closes #18615
Related to #17870

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

This fixes the command/subtask delegation path so a command-selected model is preserved when the task tool spawns a child subagent session.

The bug was that a command-triggered subtask could lose its already-resolved model and fall back to the target subagent model or the parent model instead.

This change keeps the fix narrowly scoped:

• it only forwards the internal delegated model for command-triggered subtasks
• ordinary/manual subtasks keep the existing fallback behavior
• child sessions are not pre-seeded with agent/model, so switch-event and history behavior stay aligned with existing prompt-time resolution

Why this works:

• command-triggered subtasks now carry the resolved model through the subtask path into the child prompt call
• when no delegated model is present, the task tool still falls back through subagent configured model and then parent assistant model
• regression tests cover the command path, fallback behavior, stale state bleed, and session metadata not becoming authoritative

How did you verify your code works?

From packages/opencode:

• bun test --timeout 30000 test/tool/task.test.ts
• bun test --timeout 30000 test/session/prompt.test.ts

Also ran:

• git diff --check

Screenshots / recordings

N/A — no UI changes.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

Thanks for your contribution!

This PR doesn't have a linked issue. All PRs must reference an existing issue.

Please:

1. Open an issue describing the bug/feature (if one doesn't exist)
2. Add Fixes #<number> or Closes #<number> to this PR description

See CONTRIBUTING.md for details.

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Potential Duplicate Found

PR #18752: "fix: preserve explicit model selection in command subtasks"
#18752

This appears to be addressing the exact same issue as the current PR #26117. Both PRs aim to preserve command-selected models for command-triggered subtasks and prevent them from losing their model selection when spawning child subagent sessions.

Why they're related:

• Same core problem: command-triggered subtasks losing their explicit model selection
• Same solution domain: preserving model routing for command-triggered subtasks vs ordinary subtasks
• Both involve task tool spawning behavior and model precedence/fallback logic

[gmnstr]

While reviewing upstream history after opening this, I found that #18752 is already open and overlaps heavily with this PR. Both changes preserve command/subtask model selection through the delegation path and touch the same routing/test files.\n\nThe reason I’m leaving this open for now is that this version tries to keep the fix narrowly scoped to command-triggered subtasks and preserve existing behavior for ordinary/manual subtasks:\n- only forwards the internal delegated model for command-triggered subtasks\n- keeps ordinary/manual task-tool fallback behavior unchanged\n- avoids pre-seeding child session agent/model so switch/history behavior stays aligned with existing prompt-time resolution\n- includes regression coverage for fallback behavior and non-authoritative session metadata\n\nIf maintainers would prefer to converge on #18752 instead of reviewing two overlapping fixes, I’m happy to close this and upstream any useful guardrails/tests there.

[github-actions]

Thanks for updating your PR! It now meets our contributing guidelines. 👍