fix(delegation): pass target_model to resolve_runtime_provider in _resolve_delegation_credentials by Emidomenge · Pull Request #15320 · NousResearch/hermes-agent

Emidomenge · 2026-04-24T20:06:08Z

Problem

When delegation.model differs from model.default on opencode-go or opencode-zen, delegate_task fails with a 404. _resolve_delegation_credentials calls resolve_runtime_provider(requested=configured_provider) without a target_model, so api_mode is computed from model_cfg.get("default") — the main model — instead of the delegation model.

Example: model.default=minimax-m2.7 (→ anthropic_messages) + delegation.model=glm-5.1 (→ chat_completions). The subagent inherits anthropic_messages, /v1 is stripped from the base URL, and the request hits https://opencode.ai/zen/go/messages instead of https://opencode.ai/zen/go/v1/chat/completions → 404.

Fix

resolve_runtime_provider already accepts target_model for this exact purpose (see its docstring: "Callers performing an explicit mid-session model switch should pass the new model here"). The delegation path just wasn't passing it.

# Before
runtime = resolve_runtime_provider(requested=configured_provider)
# After
runtime = resolve_runtime_provider(requested=configured_provider, target_model=configured_model)

Testing

Verified locally with opencode-go, model.default=minimax-m2.7, delegation.model=glm-5.1:

Before: api_mode=anthropic_messages, base_url=https://opencode.ai/zen/go → 404
After: api_mode=chat_completions, base_url=https://opencode.ai/zen/go/v1 → success

Closes #15319
Related: #13678

…solve_delegation_credentials When delegation.model differs from model.default and the provider is opencode-go or opencode-zen, the wrong api_mode is computed because resolve_runtime_provider falls back to model_cfg.get('default') — the main model — instead of the configured delegation model. For example, with model.default=minimax-m2.7 (anthropic_messages) and delegation.model=glm-5.1 (chat_completions), subagents get anthropic_messages, which strips /v1 from the base URL and causes a 404. resolve_runtime_provider already accepts target_model for exactly this purpose; _resolve_delegation_credentials just wasn't passing it. Fixes NousResearch#15319 Related: NousResearch#13678

ether-btc · 2026-04-27T16:39:28Z

Code Review: fix(delegation): pass target_model to resolve_runtime_provider

Verdict: Approve with nits

This is a correct, minimal fix for a real bug. The change is one argument addition that wires an already-available parameter through to where it's needed. The fix is obviously safe — resolve_runtime_provider already accepts target_model with a None default that preserves existing behavior.

Correctness Analysis

The bug is real and well-diagnosed. _resolve_delegation_credentials (line 2241) calls resolve_runtime_provider(requested=configured_provider) without passing the delegation model. resolve_runtime_provider then falls back to model_cfg.get("default") — the main session model — to compute api_mode. When the delegation model requires a different transport (e.g., glm-5.1 -> chat_completions vs minimax-m2.7 -> anthropic_messages), the subagent gets the wrong api_mode and a mangled base URL.

The fix is correct. configured_model is already resolved at line 2190 from cfg.get("model") and is in scope at line 2241. Passing it as target_model means resolve_runtime_provider uses the delegation model for api_mode computation instead of the stale config default. This matches the documented intent of target_model (line 856-862 in runtime_provider.py):

Callers performing an explicit mid-session model switch should pass the new model here so api_mode is derived from the model they are switching TO, not the stale persisted default.

Call path verified: target_model flows through to:

_resolve_runtime_from_pool_entry (line 980) -> uses it for effective_model (line 195)
_resolve_azure_foundry_runtime (line 897) -> uses it for Azure model-family inference (line 650)
Generic provider fallback (line 1222) -> uses it for opencode_model_api_mode() routing

All three paths fall back to model_cfg.get("default") when target_model is None, so no regression for callers that don't pass it.

Edge case: configured_model is None. Line 2190: str(cfg.get("model") or "").strip() or None. When delegation.model is empty, configured_model is None, which passes target_model=None -> identical to current behavior. Safe.

Issues Found

#	Severity	File	Issue
1	High	`tests/tools/test_delegate.py:630`	Existing test will break. `test_provider_resolves_full_credentials` asserts `mock_resolve.assert_called_once_with(requested="openrouter")` but after this PR the call includes `target_model="google/gemini-3-flash-preview"`. Same issue at line 691 (`test_nous_provider_resolves_nous_credentials`). Both tests will fail with `AssertionError: Expected call: resolve_runtime_provider(requested='openrouter') Actual call: resolve_runtime_provider(requested='openrouter', target_model='google/gemini-3-flash-preview')`.
2	Medium	PR branch	431 commits behind main. The merge-base is `6f1eed39` (Apr 26); main is now at `98d75dea`. Needs rebase before merge. There have been significant changes to `delegate_tool.py` since — at minimum, verify the line numbers haven't shifted.
3	Low	PR diff	No test for the actual bug. The PR adds the fix but no test that verifies `target_model` is passed. A test like `test_delegation_model_determines_api_mode` that sets `cfg={"model": "glm-5.1", "provider": "opencode-go"}` and asserts `mock_resolve` is called with `target_model="glm-5.1"` would prevent regression.

Nit

The PR description focuses on opencode-go / opencode-zen but the bug affects any provider where api_mode varies by model (Minimax, xAI codex_responses, Azure Foundry). The title could be broader: fix(delegation): pass target_model so api_mode matches delegation model, not config default. Not blocking — just improves discoverability.

Scope Check

No credentials, private keys, or wallet addresses in the diff
Single-file change, single-function modification
No new dependencies
No config changes required

Recommendation

Approve after:

Fix the two failing tests (lines 630, 691) — update assert_called_once_with to include target_model
Rebase onto current main
(Nice-to-have) Add a positive test for the bug scenario

The fix itself is correct and minimal. The missing test updates should be straightforward.

Emidomenge · 2026-04-27T17:47:42Z

If this has already been fixed in latest version. This can be closed

teknium1 · 2026-05-04T09:30:50Z

Salvaged via #19623 onto current main - your commit authorship was preserved. Thanks!

…on (NousResearch#15319) When delegation.model differs from the parent model (e.g., delegating to claude-opus-4-6 on an opencode-zen provider while the parent uses a chat_completions model), resolve_runtime_provider needs the target model to select the correct api_mode. Without it, the delegate inherits the parent's api_mode and gets 404 errors. Supersedes NousResearch#15320.

Emidomenge mentioned this pull request Apr 24, 2026

delegate_task 404 when delegation model differs from main model on opencode-go/opencode-zen #15319

Closed

alt-glitch added type/bug Something isn't working P1 High — major feature broken, no workaround tool/delegate Subagent delegation labels Apr 24, 2026

Tranquil-Flow mentioned this pull request Apr 25, 2026

fix(tools): pass target_model to resolve_runtime_provider in delegation (#15319) #15447

Open

19 tasks

ether-btc mentioned this pull request Apr 27, 2026

delegate_task(acp_command=...) silently ignored under Anthropic parent #9712

Closed

teknium1 mentioned this pull request May 4, 2026

fix(delegation): pass target_model to resolve_runtime_provider in _resolve_delegation_credentials #19623

Merged

teknium1 closed this May 4, 2026

This was referenced Jun 12, 2026

fix(cron): pass target_model to resolve_runtime_provider #45244

Open

Cron scheduler omits target_model when resolving runtime provider — same root cause as #18586 #45245

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fix(delegation): pass target_model to resolve_runtime_provider in _resolve_delegation_credentials#15320

fix(delegation): pass target_model to resolve_runtime_provider in _resolve_delegation_credentials#15320
Emidomenge wants to merge 1 commit into
NousResearch:mainfrom
Emidomenge:fix/delegation-target-model-opencode

Emidomenge commented Apr 24, 2026

Uh oh!

ether-btc commented Apr 27, 2026

Uh oh!

Emidomenge commented Apr 27, 2026

Uh oh!

teknium1 commented May 4, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

Emidomenge commented Apr 24, 2026

Problem

Fix

Testing

Uh oh!

ether-btc commented Apr 27, 2026

Code Review: fix(delegation): pass target_model to resolve_runtime_provider

Verdict: Approve with nits

Correctness Analysis

Issues Found

Nit

Scope Check

Recommendation

Uh oh!

Emidomenge commented Apr 27, 2026

Uh oh!

teknium1 commented May 4, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants