feat(delegate): per-call model/provider override for subagents by ReqX · Pull Request #3172 · NousResearch/hermes-agent

ReqX · 2026-03-26T09:44:54Z

Summary

Allow delegate_task() to accept model and provider parameters at call time, overriding config.yaml delegation settings for each invocation.

Resolution priority: task-level > top-level > config.yaml > parent agent inherit

Motivation

Previously, subagent model/provider could only be set globally via delegation.model / delegation.provider in config.yaml. This PR enables the LLM (or tool callers) to dynamically route specific subagents to different models/providers per invocation — e.g., delegate cheap fast tasks to Gemini Flash while keeping the parent on Claude.

Changes

`tools/delegate_tool.py` (+74, -15)

Add model and provider optional params to delegate_task() signature
Add model/provider to top-level and batch task item schema properties
Move credential resolution into per-task loop for per-task overrides in batch mode
Extend _resolve_delegation_credentials() with override_model / override_provider params
Per-call values take precedence over config; config is used as fallback

`run_agent.py` (+4)

Fix: Pass model and provider through both hardcoded delegate_task dispatch paths
- _execute_tool_call() — the single tool call dispatch (~line 4957)
- _execute_tool_calls_concurrent() — the concurrent execution path (~line 5309)
These two call sites bypass the tool registry and call _delegate_task() directly.
Without this fix, per-call model/provider were silently dropped, causing all
subagents to fall back to the parent model regardless of override values.

`tests/tools/test_delegate.py` (+249)

New TestPerCallModelProviderOverride class with 12 tests covering:
- Schema validation (top-level + batch items)
- Per-call model/provider pass-through to credential resolution
- Batch per-task model override
- Task-level overrides top-level
- Config fallback when no override
- None override does not clear config
- Credential error with per-call override
- Backward compatibility (no regressions)

`skills/autonomous-ai-agents/model-routing-template/` (new)

Strategic delegation framework skill — the "when and why" companion to the "how"
Includes providers-example.yaml with a worked multi-provider catalog
Provider-agnostic: users fill in their own models, costs, and roles

Test Results

61/61 tests pass (49 existing + 12 new, zero regressions)

Live verification: Tested against running Hermes instance with:

delegate_task(provider="zai", model="glm-5") — correctly routes to glm-5
delegate_task(provider="copilot", model="claude-sonnet-4.6") — correctly routes through GitHub Copilot
Batch mode with per-task overrides — both tasks run on correct model/provider in parallel

Backward Compatibility

Fully backward compatible. All new parameters are optional with None defaults. Existing behavior (config-based delegation, parent inheritance) is unchanged when no per-call overrides are provided.

Allow delegate_task() to accept model and provider parameters at call time, overriding config.yaml delegation settings for each invocation. Resolution priority: task-level > top-level > config > parent inherit. Changes: - Add model/provider to delegate_task() signature and schema - Add model/provider to batch task items schema - Move credential resolution into per-task loop for per-task overrides - Extend _resolve_delegation_credentials() with override params - Add 12 tests for per-call override behavior (61 total, all passing)

…patch paths run_agent.py has two direct call sites for delegate_task that bypass the tool registry. These were not updated in the original commit and silently dropped per-call model/provider overrides, causing all subagents to fall back to the parent model. Fix: pass model/function_args.get('model') and provider/function_args.get('provider') at both call sites: - _execute_tool_call() (single dispatch, ~line 4957) - _execute_tool_calls_concurrent() (concurrent dispatch, ~line 5309) Verified: glm-5, copilot/claude-sonnet-4.6, and batch mode all correctly route to the specified model/provider.

…work Generic template for routing subagent tasks to the right model based on cost, capability, and complexity. Includes: - Decision framework: when to delegate vs handle yourself - 4-tier escalation model: cheap → standard → expensive → frontier - Cost-effective patterns: 'many for one', 'free stack', 'scout party' - Engaged vs autonomous mode: present tradeoffs when user is present, follow ladder strictly when not - providers-example.yaml: worked example catalog showing how to define models, costs, roles, user override phrases, and delegation patterns Designed to be provider-agnostic — users fill in their own catalog. Complements the per-call model/provider override feature in this PR (NousResearch#3172) by providing the strategic 'when and why' to the technical 'how'.

1shanra1 · 2026-03-27T04:30:41Z

I was going to make a PR for this exact feature! Nice to see someone already did it :) This should get merged.

ReqX · 2026-03-27T05:22:23Z

Happy someone else would find it useful. It works quite well for the current limited scope.

Was thinking of adding pre defined personas to it (something like OMO-slim), but kept it low for easier merge at this point, and you can prompt it anyway or add some more instructions to the skill.

SHL0MS · 2026-03-31T19:24:22Z

Another user request for this: "I'd like to change models of subagent by prompting alone — 'Launch subagent to do XYZ with model ABC'." The delegate_task schema (line 713-774) currently has no model parameter. The user can configure a static override in config.yaml under delegation but cannot specify the model per-call in the prompt. See also #3719, #3794, #3984.

HenkDz · 2026-04-01T12:15:36Z

Another user request for this: "I'd like to change models of subagent by prompting alone — 'Launch subagent to do XYZ with model ABC'." The delegate_task schema (line 713-774) currently has no model parameter. The user can configure a static override in config.yaml under delegation but cannot specify the model per-call in the prompt. See also #3719, #3794, #3984.

The #3719 si closed in favor of #3794
The solution in the latter is better and way simpler:

ReqX · 2026-04-01T20:51:10Z

I honestly like #3794, but is it exclusive? The strength path is powerfull but giving us the opp to invoke a specific model is sure valuable for power users?

I can rebase/check how it could fit in if there is interest.

HenkDz · 2026-04-01T22:17:13Z

I honestly like #3794, but is it exclusive?

What do you mean by "Exclusive"?

giving us the opp to invoke a specific model is sure valuable for power users?

Most certainly, at least in my case, I would like to the coding agent to be super powrful, and different from the design one. The default chat one would be different from both and so on.

ReqX · 2026-04-02T12:29:35Z

Overlooked the "Per-call model/provider override" in your PR ;)

That combined with the "strength-routing" is great and makes my PR most likely obsolete (didnt look at the code, on the road).

iRonin · 2026-04-03T19:18:00Z

Great work on this — I opened #4833 which builds on the same model/provider approach and adds two more layers:

Skill-level model routing: delegate_task(skill='code-review') reads model: from the skill's SKILL.md frontmatter automatically
supervisor_model / execution_model config aliases for clear multi-agent intent in config.yaml

Happy to consolidate if maintainers prefer a single PR.

Three interconnected features for multi-agent model routing: 1. delegate_task() — model/provider params (like PR NousResearch#3172) - model, provider: per-call overrides for all subagents in the call - Per-task model/provider in batch tasks array - Priority: per-task > call-level > delegation config > parent inherit 2. delegate_task() — skill/skills params (new) - skill: load a single named skill into the subagent's context - skills: load multiple skills - Skill SKILL.md content is prepended to the subagent system prompt - If SKILL.md frontmatter contains model: or provider: fields, those are used as the model/provider for that subagent call (overridable by the explicit model param) 3. config.yaml — supervisor_model / execution_model aliases - delegation.supervisor_model → model.default (main agent) - delegation.execution_model → delegation.model (all subagents) - Resolved during config load; explicit model.default / delegation.model always take precedence if both are set - Allows clear intent in config: delegation: supervisor_model: anthropic/claude-opus-4-6 execution_model: google/gemini-flash-1.5 Skill frontmatter example: --- name: code-review model: anthropic/claude-opus-4-6 provider: anthropic --- Usage example (agent tool call): delegate_task(goal='review this PR', skill='code-review') delegate_task(goal='summarise docs', model='google/gemini-flash-1.5') delegate_task(tasks=[ {goal: 'fast research', model: 'google/gemini-flash-1.5'}, {goal: 'careful code', model: 'anthropic/claude-opus-4-6'}, ]) Builds on / references PR NousResearch#3172 (ReqX) for the model/provider param idea.

alt-glitch · 2026-05-02T18:34:01Z

Related/competing: #12794 (same feature, rebased onto dispatch refactor). Also see closed #3750 and open issue #5012.

alt-glitch · 2026-05-02T18:34:43Z

Competing PR #12794.

ReqX · 2026-05-24T08:36:00Z

PR #3172 vs #12794

1. Core Feature Parity

Functionally identical resolution chain: per-task > top-level > config > parent inherit. The mechanism is the same, extend _resolve_delegation_credentials() with optional override params, thread them through the schema/signature. No meaningful behavioral difference in the core override itself.

2. Merge-Readiness

Very obvious, right ;)

My feat(delegate): per-call model/provider override for subagents #3172 is stale. It patches two call sites (_execute_tool_call ~L4957, _execute_tool_calls_concurrent ~L5309) that no longer exist in their original form. The _dispatch_delegate_task() refactor (48ecb98) consolidated these. Rebasing feat(delegate): per-call model/provider override for subagents #3172 isn't trivial, it requires reworking the core commits against a different dispatch architecture.
feat(delegate_task): per-subagent model/provider overrides + model observability plugin #12794 is current. Rebased onto 48ecb98, actively maintained (force-pushed 2 days ago), patches the single consolidated method. It merges cleanly today.
Verdict: feat(delegate_task): per-subagent model/provider overrides + model observability plugin #12794 wins on merge-readiness by a wide margin.

4. Test Coverage

Dimension	#3172	#12794
Core override tests	12	~12 equivalent
Plugin lifecycle	0	41
Edge cases (auto-router, mismatch)	0	Yes
E2E verification	No	Yes (live results cited)

#12794 is substantially better tested, especially for the failure modes that matter in production (router resolution mismatches, TTL stash bleed).

Summary

	#3172	#12794
Core feature	✓	✓
Rebased on current main	✗	✓
Maintainer triage labels	No	P2 + comp tags
Observability/verification	No	Yes
Mismatch guards	No	Yes
Test depth	Adequate	Comprehensive
Active maintenance	Stale (March)	Active (May 22)

Bottom line: #12794 is the better PR on every axis except scope minimalism, this is up to maintainers.

Closing here.

ReqX added 3 commits March 26, 2026 09:28

iRonin mentioned this pull request Apr 3, 2026

feat: per-skill model routing + supervisor/execution model config #4833

Open

alt-glitch mentioned this pull request Apr 22, 2026

feat(delegate_task): per-subagent model/provider overrides + model observability plugin #12794

Open

alt-glitch mentioned this pull request Apr 23, 2026

feat: allow delegate_task model overrides #12715

Open

alt-glitch mentioned this pull request May 5, 2026

feat(delegate): per-task model routing for delegate_task #20000

Open

alt-glitch mentioned this pull request May 6, 2026

feat(delegation): per-task model/provider overrides in batch dispatch #20779

Closed

5 tasks

konsisumer mentioned this pull request May 8, 2026

[Feature]: Allow delegate_task to accept provider/model parameters #5012

Open

1 task

alt-glitch mentioned this pull request May 10, 2026

feat(delegate): per-task model and provider overrides #23266

Open

alt-glitch mentioned this pull request May 13, 2026

feat(delegate): per-task model and provider override for delegate_task subagents #25026

Closed

This was referenced May 14, 2026

feat(delegate_task): per-call model/provider override (revives #3719) #25530

Open

feat(delegate): per-task and top-level model override for cost-aware routing #25813

Open

This was referenced May 24, 2026

fix(delegate): wire model/provider params through to _build_child_agent #31284

Closed

feat: allow per-call delegation model overrides #31345

Open

ReqX closed this May 24, 2026

alt-glitch mentioned this pull request May 29, 2026

[TASK] Implement Per-Call Provider and Model Overrides for delegate_task #34462

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat(delegate): per-call model/provider override for subagents#3172

feat(delegate): per-call model/provider override for subagents#3172
ReqX wants to merge 3 commits into
NousResearch:mainfrom
ReqX:feature/delegate-per-call-model

ReqX commented Mar 26, 2026 •

edited

Loading

Uh oh!

1shanra1 commented Mar 27, 2026

Uh oh!

ReqX commented Mar 27, 2026

Uh oh!

SHL0MS commented Mar 31, 2026

Uh oh!

HenkDz commented Apr 1, 2026

Uh oh!

ReqX commented Apr 1, 2026

Uh oh!

HenkDz commented Apr 1, 2026 •

edited

Loading

Uh oh!

ReqX commented Apr 2, 2026

Uh oh!

iRonin commented Apr 3, 2026

Uh oh!

alt-glitch commented May 2, 2026

Uh oh!

alt-glitch commented May 2, 2026

Uh oh!

ReqX commented May 24, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

Conversation

ReqX commented Mar 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Motivation

Changes

tools/delegate_tool.py (+74, -15)

run_agent.py (+4)

tests/tools/test_delegate.py (+249)

skills/autonomous-ai-agents/model-routing-template/ (new)

Test Results

Backward Compatibility

Uh oh!

1shanra1 commented Mar 27, 2026

Uh oh!

ReqX commented Mar 27, 2026

Uh oh!

SHL0MS commented Mar 31, 2026

Uh oh!

HenkDz commented Apr 1, 2026

Uh oh!

ReqX commented Apr 1, 2026

Uh oh!

HenkDz commented Apr 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ReqX commented Apr 2, 2026

Uh oh!

iRonin commented Apr 3, 2026

Uh oh!

alt-glitch commented May 2, 2026

Uh oh!

alt-glitch commented May 2, 2026

Uh oh!

ReqX commented May 24, 2026

PR #3172 vs #12794

1. Core Feature Parity

2. Merge-Readiness

4. Test Coverage

Summary

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

ReqX commented Mar 26, 2026 •

edited

Loading

`tools/delegate_tool.py` (+74, -15)

`run_agent.py` (+4)

`tests/tools/test_delegate.py` (+249)

`skills/autonomous-ai-agents/model-routing-template/` (new)

HenkDz commented Apr 1, 2026 •

edited

Loading