feat(delegate): add per-call model/provider/base_url overrides

[mddragon18]

Summary

Add model, provider, and base_url parameters to delegate_task, allowing per-call overrides for subagents — both at the top level and per-task in batch mode. api_key is kept in the code path (works via delegation.api_key config) but excluded from the schema for security.

Changes

tools/delegate_tool.py:

• Extended delegate_task() function signature with model, provider, base_url params
• Added fields to DELEGATE_TASK_SCHEMA at top-level and per-task — api_key excluded from schema to prevent LLM secret leakage
• Credential resolution happens per-child: per-task > top-level per-call > delegation config > parent inherit
• All pre-existing code preserved (cost rollup, ACP transport safety, ACP override clearing)

tests/tools/test_delegate.py:

• Updated test_schema_valid to assert new fields
• Added test_per_call_model_override

Security

api_key is NOT exposed to the model via the schema. The code path still supports it internally — set via delegation.api_key in config.yaml — but the LLM cannot inject or read it through tool calls.

Example

# Single task with model override
delegate_task(goal="Do X", model="gemini-flash-latest", provider="google")

# Batch with per-task overrides
delegate_task(tasks=[
    {"goal": "Research", "model": "fast-model"},
    {"goal": "Analyze", "model": "reasoning-model", "provider": "openrouter"},
])

Verification

• All 121 delegate tests pass

[Copilot]

Pull request overview

Adds per-invocation runtime overrides to delegate_task so callers can choose a different model/provider (and related connection settings) for spawned subagents, including per-task overrides in batch mode.

Changes:

• Extended delegate_task() to accept model, provider, base_url, and api_key, with per-task override precedence in batch delegation.
• Updated DELEGATE_TASK_SCHEMA to expose the new override fields at both top-level and per-task item level.
• Added/updated tests covering schema fields, per-call overrides, and subagent cost rollup behavior.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
tools/delegate_tool.py	Adds per-call/per-task runtime override parameters + schema exposure; includes subagent cost rollup and ACP transport handling updates.
tests/tools/test_delegate.py	Updates schema validation assertions and adds coverage for per-call overrides and cost rollup.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[mddragon18]

check the updated PR again for this

[mddragon18]

Rebased onto latest main to resolve merge conflicts. Clean diff now: +117 lines across 2 files.

Changes since initial review:

• model: None no longer clears the configured delegation model (task dict only includes keys when value is not None)
• api_key removed from schema for security — still works via delegation.api_key config, but the LLM can not inject or read it through tool calls
• .get("model") is not None guards in per-child loop as belt-and-suspenders

[gigi206]

The patch in delegate_tool.py is correct, but it also needs forwarding in run_agent.py _dispatch_delegate_task() on line 9645. This function intercepts the tool call (delegate_task is in _AGENT_LOOP_TOOLS) and didn't forward model, provider, base_url, api_key to the actual delegate_task() function.

Without this, the parameters arrive in function_args but are never passed through. The comment at line 9648 even says: "New DELEGATE_TASK_SCHEMA fields only need to be added here to reach all invocation paths."

The fix is adding:

model=function_args.get("model"),
provider=function_args.get("provider"),
base_url=function_args.get("base_url"),
api_key=function_args.get("api_key"),

before parent_agent=self in that function.

[teknium1]

We do not want this