What does this PR do?

Implements named delegation profiles for delegate_task: users define
specialised subagent roles in cli-config.yaml and route tasks to them by
name — selecting a different model, provider, toolset, iteration budget, and
system prompt per role without touching application code.

Full design details in the sections below.

Related Issue

Closes #9459

Also addresses: #10995, #21827, #31793, #30652

Type of Change

• ✨ New feature (non-breaking change that adds functionality)

Changes Made

• tools/delegate_tool.py — _resolve_profile(), ACP + API routing, toolset priority chain, per-profile max_iterations, system prompt / constraints injection, schema enum injection from configured profiles
• run_agent.py — forward profile arg through _dispatch_delegate_task()
• cli-config.yaml.example — full delegation.profiles usage example (API + ACP modes, all fields)
• tests/tools/test_delegate_profiles.py — 37 new tests (new file)

How to Test

1. Add a delegation.profiles block to ~/.hermes/config.yaml (see example in cli-config.yaml.example)
2. Start a chat session and call delegate_task(goal="...", profile="<name>")
3. Verify the child agent runs with the configured model / toolsets / iteration budget
4. For ACP mode: set acp_command: "claude" in a profile and verify the local runner is spawned via the terminal tool

Unit tests: scripts/run_tests.sh tests/tools/test_delegate_profiles.py

Implementation details

New delegation.profiles config section (nested under the existing
delegation block — see rationale below):

delegation:
  profiles:
    coder:
      nickname: "👷 Coder"          # shown in tool description (informational)
      summary: "Writes code, TDD"   # shown to LLM orchestrator in the schema enum
      model: deepseek-v4-flash
      provider: deepseek
      toolsets: [terminal, file]
      max_iterations: 80            # overrides delegation.max_iterations for this profile
      system_prompt: |
        You are a focused developer. Follow TDD strictly.
      constraints: |
        - No changes outside the assigned task scope

    critic:
      nickname: "🔍 Critic"
      summary: "Code review, bug hunting"
      model: claude-sonnet-4-6
      provider: openrouter
      toolsets: [file]
      max_iterations: 20

Usage:

delegate_task(goal="implement auth module", profile="coder")
delegate_task(tasks=[
    {"goal": "write tests", "profile": "coder"},
    {"goal": "review PR",   "profile": "critic"},
])

Two routing modes per profile (mutually exclusive):

Toolset priority chain:
explicit call arg > profile.toolsets > inherited from parent

Iteration budget:
profile.max_iterations > delegation.max_iterations (global). The
orchestrating LLM cannot override this — the budget is always config-controlled
to keep costs predictable.

System prompt injection:

• system_prompt — replaces the first line of the child's system prompt (persona layer)
• constraints — prepended before the task instruction block

Schema enum injection:
Configured profiles surface in the delegate_task JSON schema as an enum with
summary descriptions, so the orchestrating LLM can choose the right profile
autonomously.

How this differs from other open proposals

There are several open PRs in this space (#18522, #34650, #34681, #34752, and
others). This PR takes a different position on two points that the others don't
cover:

ACP routing. Most proposals only address API-based subagents (model +
provider overrides). This PR also lets a profile launch a local agent runner
— Claude Code, aider, a second Hermes instance — via a terminal command wrapper
(acp_command + acp_args). That covers the cloud-local hybrid pattern from
#31793 and the "right tool for the right job" use case from #21827 without
requiring a second gateway process.

LLM-autonomous profile selection. Configured profiles are injected into the
delegate_task JSON schema as a typed enum with summary labels. The
orchestrator LLM sees the available profiles and their descriptions at call
time, and can select one based on task content — no hard-coded routing logic
required in the skill or system prompt.

Both features are absent from the per-call model/provider override approach
(#34681, #34752) and from the fixed-tier approach (#34650). They're also absent
from the closest structural cousin (#18522), which uses agent_profiles: as a
top-level key and does not implement ACP mode, schema enum injection, or
per-profile iteration budgets.

Why delegation.profiles and not agent_profiles

The issue proposes a top-level agent_profiles key. Nesting under delegation
is a better fit because:

1. Profiles inherit the delegation family settings (max_iterations,
   max_spawn_depth, subagent_auto_approve) without repeating them.
2. The delegation block already owns all subagent configuration — profiles
   are a natural extension, not a parallel concept.
3. Keeps the root config namespace flat.

What's not in v1 (explicit stub)

• proxy field is accepted and logged as a warning but not applied.
  Routing through a local proxy requires additional client wiring (v2).

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits
• I searched for existing PRs to make sure this isn't a duplicate (aware of feat(delegation): agent profiles for delegate_task (#9459) #18522, feat(delegation): delegate_task model/provider routing by fixed tier set #34650, Add model parameter to delegate tool for per-call provider overrides #34681, feat(hermes): Phase 7 Task Delegation + Phase 3 Model Capability Scoring #34752)
• My PR contains only changes related to this feature
• I've run scripts/run_tests.sh and all tests pass (37/37)
• I've added tests for my changes (37 new tests in tests/tools/test_delegate_profiles.py)
• I've tested on my platform: Ubuntu 24.04

Documentation & Housekeeping

• I've updated cli-config.yaml.example with the new delegation.profiles config keys
• I've considered cross-platform impact — scripts/check-windows-footguns.py clean, no POSIX-only syscalls introduced
• I've updated tool descriptions/schemas (schema enum injection adds profiles to delegate_task schema)
• README / docs updates — N/A (config example is self-documenting)
• CONTRIBUTING.md / AGENTS.md — N/A