feat(delegation): opt-in per-task heterogeneous model routing

[TimberFell]

Summary

Adds an opt-in config flag delegation.per_task_routing (default false)
that lets a single delegate_task batch fan out across heterogeneous
models/providers per task. Each task in tasks=[...] may carry its own
provider / model / base_url / api_mode and is resolved independently
through the existing _resolve_delegation_credentials() path, so one batch can
mix e.g. a cheap local model for scanning with a strong model for synthesis,
chosen per task by model strength, hardware, or token budget.

When the flag is off, behaviour is identical to today: per-task routing keys
are ignored and the batch-wide delegation.* credentials apply to every child.

Motivation

_build_child_agent() already accepts a full per-child credential bundle
(override_provider / override_base_url / override_api_key /
override_api_mode / model) and re-derives the API transport when a child's
provider differs from the parent's. However, the batch dispatcher resolved
credentials once before the loop and passed the same bundle to every child —
so the engine supported heterogeneous routing but the dispatcher never exercised
it. This PR closes that gap additively and behind a flag.

What changed

tools/delegate_tool.py

• New helpers: _per_task_routing_enabled() (config > env > default false),
  _task_routing_overrides() (extracts only {provider, model, base_url, api_mode}, string + non-blank), _resolve_task_credentials() (resolves a
  task's routing keys in isolation when enabled; otherwise returns batch-wide
  creds unchanged).
• Dispatcher loop resolves per-task creds inside the loop. Unresolvable per-task
  provider → tool_error naming the task index (fail-loud).
• Schema: tasks[].items gains provider / model / base_url / api_mode, each
  documented as honoured only when the flag is on.
• Tasks description advertises heterogeneous routing only when enabled.

tests/tools/test_delegate.py

• TestPerTaskRouting (12 cases): flag gate, override extraction, fail-closed
  identity passthrough, isolated resolution, no-key-injection invariant,
  fail-loud resolution.

Security / design invariants

• Fail-closed: default off; with it off, byte-for-byte identical to prior
  behaviour.
• API keys never accepted from the model-supplied task dict. Keys resolve
  from the runtime provider system (env / hermes auth) or are inherited from
  the parent. The model selects which backend, never injects a credential.
• Isolated resolution: per-task keys resolved on their own, so a batch-wide
  endpoint/key cannot bleed into a differently-routed task.
• Fail-loud: unresolvable per-task provider aborts with a task-identifying
  error rather than silently using the wrong model.

Testing

• tests/tools/test_delegate.py: 157 passed (145 existing + 12 new).
• Adjacent suites (test_delegate_toolset_scope, test_delegate_composite_toolsets,
  test_async_delegation): 29 passed.
• Runtime checks confirm flag-off = identity passthrough + no schema leakage,
  and flag-on exposes the four per-task fields and the routing hint.

Example

# config.yaml
delegation:
  per_task_routing: true

{"tasks": [
  {"goal": "scan logs for the error signature",
   "model": "gemma-4-12B-it-Q4_K_M.gguf", "base_url": "http://10.90.20.17:8761/v1"},
  {"goal": "synthesize root cause", "provider": "anthropic", "model": "claude-opus-4-8"},
  {"goal": "draft remediation notes", "provider": "openrouter", "model": "anthropic/claude-sonnet-4"}
]}

Backwards compatibility

Fully backwards compatible. No behaviour change unless delegation.per_task_routing
is explicitly enabled.

[alt-glitch]

This was generated by AI during triage.

Related: this is part of a saturated per-task delegation model-routing cluster — #16163 (earliest open per-task model/provider overrides) and #34752 (Phase 7 per-call overrides) implement the same capability via different mechanisms. This PR adds an opt-in delegation.per_task_routing flag with independent per-task credential resolution. Not a duplicate of either (distinct flag-gated approach); flagging the cluster so a maintainer can pick the canonical implementation.

[teknium1]

Thank you for the careful flag-gating, credential-isolation work, and focused tests.

This is an automated hermes-sweeper review. The requested behavior matches the standing delegation-model-routing policy: Hermes does not accept per-call or per-task provider/model routing through delegate_task. The supported configuration is one provider:model choice for all delegated children via delegation.provider and delegation.model in config.yaml.

• Current delegate_task intentionally resolves one delegation credential bundle before building batch children (tools/delegate_tool.py:2453, tools/delegate_tool.py:2511-2534).
• Its live tool instructions explicitly say the subagent model is not selectable per call and point to delegation.provider / delegation.model as the sanctioned configuration (tools/delegate_tool.py:3296).
• The existing docs describe that same installation-wide configuration path (website/docs/user-guide/configuration.md:1936-1944).

A focused contribution improving the existing config-level delegation routing path would remain welcome, but per-task heterogeneous routing cannot be merged.

---

Closed as not-planned per standing maintainer policy (delegation-model-routing). This is a design-direction decision, not a code-quality judgment — see the Contribution Rubric in AGENTS.md for what the project is looking for. If you believe this policy was misapplied to your change, comment here and a maintainer will take a look.