Summary

Add inter-profile orchestration tools enabling one Hermes profile to delegate tasks to another via HTTP, with full conversation history passing between calls.

New Tools (4 total)

Install

Enable delegation toolset in profile config:

toolsets:
- hermes-cli
- delegation

Enable api_server on worker profile:

platforms:
  api_server:
    enabled: true
    port: 8642
    host: 127.0.0.1

Usage

# In orchestrator profile (p1):
p1 chat

# Delegate a task to another profile
> Ask p2 to calculate 5 times 5
# Returns: {"status": "completed", "result": "25", "task_id": "abc123"}

# Delegate without waiting for result
> Have p2 calculate 6 times 6 in the background
# Returns: {"status": "dispatched", "task_id": "xyz789"}

# Check on a background task later
> What's the status of task xyz789?
# Returns: {"status": "completed", "result": "36"}

# Search across all your profiles for past tasks
> Find my tasks about math calculations
# Returns: [{"task_id": "abc123", "source_profile": "p1", "goal": "calculate 5*5", ...}]

# Search by task ID prefix (auto-detected)
> Look up task abc123
# Returns: [{"task_id": "abc123def", "source_profile": "p1", ...}]

# Multi-turn conversation with memory
> Ask p2 to remember Word A is ALPHA
> Now ask p2 what Word B is BETA and what was Word A?
# Remembers "ALPHA" from conversation history

# Control whether worker can spawn one-off subagents
> Have p2 refactor the auth module without spawning any one-off subagents
# Worker must do all work itself - no delegate_task spawning

Key Features

Delegated agent awareness - Delegated agents now know they're delegated agents and not talking to human users, with visual <DELEGATED from profile X | delegate_task ALLOWED> header.

Cross-Profile Search - profile_task_search searches across ALL profiles' task histories, not just the current one. Auto-detects task ID prefixes (e.g., "abc123") vs natural language queries.

Configurable Depth - Prevent infinite delegation chains with max_depth parameter (default 2).

Delegated agent control - Use allow_subagents=False to force workers to do all work themselves without spawning one-off subagents.

Awareness format

<DELEGATED from profile p1 | delegate_task ALLOWED>

YOUR TASK:
Analyze codebase for security issues

Summary: what you did, found, created, issues.
End with <TASK abc123 COMPLETED>.

Agent: "p1 (profile, not human) delegated this. I am allowed to use delegate_task."

Control Mechanisms

• allow_subagents=True/False — choose if worker can spawn one-off subagents
• max_depth=N — configurable delegation depth (default 2)
• <TASK {id} COMPLETED> footer for completion correlation

Depth Tracking

• X-Hermes-Delegate-Depth header tracks recursion level
• X-Hermes-Max-Depth header sets limit (configurable)
• API server blocks when depth >= max_depth

Cross-Profile Natural Language Search

• profile_task_search("security issues auth.py") — search across all profiles
• Each result includes source_profile field identifying which profile handled it
• Auto-detects task IDs: Short alphanumeric queries (e.g., "abc123", "42eb") search by task_id prefix
• FTS5-powered: Natural language search on goal + result fields with phrases, boolean, and prefix matching
• Hybrid approach: ID-like queries use LIKE task_id% + LIKE %query% on goal/result; natural language uses full-text search

The agent uses ThreadPoolExecutor for reliability instead of aiohttp (which caused asyncio race conditions). Async tasks persist to SQLite so they survive process restarts. If a task isn't found in memory, it falls back to SQLite lookup.

Files

Tests

Handler loads correctly, passes 72 consecutive calls with 100% success rate. Manual testing on Linux (Ubuntu 22.04). Uses stdlib only, no Unix-specific code.

• 15 parallel async dispatches: ✅
• 25 rapid sync calls: ✅
• 15 status checks: ✅
• 6 mixed profile tests: ✅
• Edge cases (unicode, special chars, multiline, fibonacci n=42): ✅
• Delegated agent awareness detection: ✅
• <TASK {id} COMPLETED> detection: ✅
• Depth limit enforcement at MAX_DEPTH=2: ✅
• allow_subagents=False shows NOT ALLOWED: ✅
• Cross-profile natural language search: ✅
• Task ID prefix auto-detection: ✅
• Hybrid search (ID + FTS5): ✅

Related

• Requires: PR feat: add profiles — run multiple isolated Hermes instances #3681 (profile infrastructure) - already merged

Notes

No configuration needed beyond enabling the toolset and api_server — just create profiles and delegate. The orchestrator automatically discovers worker profiles via list_profiles(). If a worker isn't running, it starts automatically via subprocess with HERMES_HOME override. No worker running = delegation fails gracefully with error message.

Architecture

• Uses upstream's existing hermes_cli/profiles.py (from PR feat: add profiles — run multiple isolated Hermes instances #3681)
• profile_orchestrator.py imports from hermes_cli.profiles: get_profile_dir(), ProfileInfo, _check_gateway_running()