delegate_task model override ignored — subagents use wrong model, consume unauthorized credits

[Nandoca95]

Description

The model parameter in delegate_task is completely ignored. Subagents always run on the session's default model, regardless of per-task overrides.

Steps to Reproduce

1. Call delegate_task with explicit model override:

{
  "goal": "Reply exactly: OK_TEST",
  "model": {"provider": "nvidia", "model": "moonshotai/kimi-k2.6"},
  "toolsets": []
}

2. Check the returned model field in the result.

Expected Behavior

The subagent should run on moonshotai/kimi-k2.6 via NVIDIA NIM.

Actual Behavior

The subagent runs on mimo-v2.5-pro (session default). The model parameter is completely ignored.

Verified Behavior Table

Specified	Actually Ran
{"provider": "nvidia", "model": "moonshotai/kimi-k2.6"}	mimo-v2.5-pro (session default)
{"provider": "nvidia", "model": "deepseek-ai/deepseek-v4-pro"}	qwen/qwen3.6-plus (Nous Portal — PAID)
No model specified	mimo-v2.5-pro or qwen/qwen3.6-plus (unpredictable)

Impact

Financial: System routed subagent tasks to qwen/qwen3.6-plus (Nous Portal) without authorization, consuming 676,664 paid tokens.

Functional: Multi-model workflows are impossible because delegate_task always uses the same model.

Root Cause

The delegation config has empty fields causing session default inheritance.

Workaround

Use hermes --yolo chat via terminal:

hermes --yolo chat -Q -q "prompt" -m "moonshotai/kimi-k2.6" --provider nvidia

Environment

• Hermes version: v0.10.0+ (2026-06-19)
• OS: macOS 13.7.8
• Session model: mimo-v2.5-pro (Xiaomi)
• Providers: xiaomi, nvidia (NIM), nous (OAuth), openai-codex (OAuth)

Evidence

• Session 20260619_170959_942278: subagent used Qwen 3.6 Plus without authorization
• Direct test: delegate_task with Kimi K2.6 specified returned MiMo
• Internal docs confirm: references/delegate-task-model-override-broken.md

Requested Fix

1. Make model parameter actually override delegation config.
2. Don't silently route to paid providers when default is rate-limited.
3. Add delegation.fallback_model config option.

Related

• references/delegate-task-model-override-broken.md
• references/debugging-fallback-model-drift.md
• references/nous-qwen-auxiliary-route.md

[alt-glitch]

Duplicate of #23467 — same bug: the model override passed to delegate_task is silently discarded and subagents inherit the parent/session model. Tracking the fix under #23467 (see also the open fix PRs #48867, #34472, #27805, #19719). Noting the financial angle reported here (unauthorized paid-token consumption via fallback routing) on the canonical issue for visibility.

[kshitijk4poor]

Verified against current main. The premise here — that a model parameter is accepted and then silently discarded — does not hold: delegate_task exposes no model parameter (top-level or per-task), and its function signature carries none. The per-call model arg was removed intentionally in fb0f579b1 (refactor: remove model parameter from delegate_task function) to keep delegation behavior consistent. So this is intentional design, not a dropped param.

On the financial angle: subagents inherit the parent's fallback chain by design (merged in #19601), so the Nous Portal token spend is the user's own configured fallback_providers recovering from a rate limit — exactly as the top-level agent would — not an unauthorized route. The supported way to pin subagents to a cheaper/local model is config-level: delegation.provider / delegation.model in config.yaml.

The one genuine gap was documentation: the tool never said how subagent model is controlled, which is what keeps generating these reports. Fixed in #49489 (docs-only, one bullet added to the tool description). Closing as a duplicate of #23467, which remains open to track the per-call-override feature requests for a maintainer decision.