hermes-agent/tools/AGENTS.md

tools/ + toolsets.py + model_tools.py — model tools

Applies on top of the root AGENTS.md: settle the Footprint Ladder before adding anything here. Most capabilities should NOT be core tools. Long-form: website/docs/developer-guide/adding-tools.md, tools-runtime.md.

Registry and discovery

tools/registry.py has no deps and is imported by every tool file; each tools/*.py calls registry.register() at import time; model_tools.py imports the registry and triggers discovery (discover_builtin_tools()), then run_agent.py, cli.py, batch_runner.py, environments/ consume it. Any tools/*.py with a top-level registry.register() is imported automatically — no manual import list. A tool that is a whole package (tools/connectors/) registers from tools/<pkg>/tool.py, the only file discovery scans inside a package; every sibling in the package is a library by construction, and the package needs an __init__.py or discovery skips it with a warning (setuptools would drop it from the wheel). The registry handles schema collection, dispatch (handle_function_call()), availability (check_fn, TTL-cached process-wide), and error wrapping. All handlers return a JSON string.

Adding a core tool (2 files) — only when the user is explicitly contributing a core tool

For custom/local-only tools do NOT edit core: create ~/.hermes/plugins/<name>/plugin.yaml + __init__.py and call ctx.register_tool(...); plugin toolsets are discovered automatically and toggled without touching tools/ or toolsets.py (plugins/AGENTS.md).

1. tools/your_tool.py:

   from agent.secret_scope import get_secret
   from tools.registry import registry
   def check_requirements() -> bool: return bool(get_secret("EXAMPLE_API_KEY"))
   def example_tool(param: str, task_id: str = None) -> str: return json.dumps({"success": True, ...})
   registry.register(name="example_tool", toolset="example",
       schema={"name": "example_tool", "description": "...", "parameters": {...}},
       handler=lambda args, **kw: example_tool(param=args.get("param", ""), task_id=kw.get("task_id")),
       check_fn=check_requirements, requires_env=["EXAMPLE_API_KEY"])
   

2. toolsets.py: add the name to _HERMES_CORE_TOOLS (all platforms) or a new toolset. Required — discovery registers the schema, but a tool is only exposed if a toolset names it. _HERMES_CORE_TOOLS is the default bundle every platform's base toolset inherits, not dead code.

Rules for tool code:

• Schema descriptions must not name tools from other toolsets (browser_navigate saying "prefer web_search"). Those tools may be unavailable (missing key, disabled toolset) and the model hallucinates calls to them. Cross-references are added dynamically in get_tool_definitions() in model_tools.py — see the browser_navigate / execute_code post-processing blocks.
• Paths in schema descriptions use display_hermes_home() (schema is built at import; under multiplex it shows the launch home, which is display-only). State files use get_hermes_home() at call time, never Path.home()/.hermes and never a module constant, so each served profile gets its own state.
• No offset/limit on instructional tools (skills, prompts, playbooks) — models read page 1 and skip the rest (root rubric).
• check_fn answers reachability/opt-in for the profile it runs under, never surface. Results are TTL-cached in registry.py::_check_fn_cache keyed by hermes_home_key(), and one process serves many sessions AND many profiles: a probe reads credentials through agent.secret_scope.get_secret, never bare os.getenv (that answers with the launch profile's .env for everyone). The registry classifies an UnscopedSecretError from current_secret_scope() at the catch site — a boot-time probe with no scope is DEBUG, not a traceback. GUI-only tools go in a named toolset (desktop_ui, project) folded in by _load_enabled_toolsets(platform) (root: capability is a property of the SESSION).
• Agent-level tools (todo, memory) are intercepted before handle_function_call() via the INLINE_TOOL_EXECUTORS table (agent/inline_tool_executors.py; agent/AGENTS.md).
• _last_resolved_tool_names is a process-global in model_tools.py; _run_single_child() in delegate_tool.py saves/restores it around child runs — readers may see it stale mid-delegation.
• New tools integrate with existing setup UX (hermes tools, hermes setup, auto-install) rather than a raw env var; secrets go in OPTIONAL_ENV_VARS (hermes_cli/AGENTS.md).

Toolsets (toolsets.py)

Single TOOLSETS dict. Keys today: browser, clarify, code_execution, cronjob, debugging, delegation, discord, discord_admin, feishu_doc, feishu_drive, file, image_gen, kanban, memory, messaging, moa, rl, safe, search, session_search, skills, spotify, terminal, todo, tts, video, vision, web, yuanbao (don't assert the list in tests). Per-platform enable/disable via hermes tools (curses) or tools.<platform>.enabled/disabled in config.yaml. browser_exec replaces the other browser tools when browser.backend is browser-use.

Backends and providers inside tools/

Several tools front pluggable backends: terminal environments in tools/environments/ (local, docker, ssh, modal, daytona, singularity; terminal_tool_backends.py, tool_backend_helpers.py), browser (browser_tool_*.py: cdp, cloud, install, lifecycle, session, real_profile, vision), MCP client (mcp_tool_*.py: config, discovery, transport, registration, content, errors), TTS (tts_tool_providers.py, tts_command_provider.py), skills hub sources (skills_hub_official.py OptionalSkillSource). Adding a backend = a new sibling or provider entry in the existing table, never an elif on a backend name (root shape rules). Remote-backend file visibility problems are fixed at the mount, not by adding a tool.

Native vision embeds are history, not one-shot payloads. vision_tools.py::_vision_analyze_native (and the browser screenshot twins in browser_tool_vision.py / browser_use_cli.py) bake the image into a tool result that is re-sent on every later API call. Size and repeat policy live in vision_tools_history_budget.py (config section vision: embed_target_bytes, max_calls_per_image); the repeat counter is keyed on (session id, resolved source) so region crops share their file's count, and its default cap applies only inside agent.delegation_context.is_delegated_child_process_context(). Put new embed-cost rules there, never a second counter in a tool.

Every spawn goes through one env builder. environments/local.py::build_subprocess_env (+ hermes_constants.apply_subprocess_home_env, env_passthrough.py::resolve_passthrough_value) is how a terminal, execute_code, background process, delegation child, ACP or MCP stdio child gets its environment; a child that acts FOR the served profile (hermes -p X workers, key_cmd helpers, browser drivers, Bot Chat relay turns) uses environments/local.py:: served_profile_child_env(target_home=, inherit_credentials=): launch-profile .env / TERMINAL_* residue dropped (strip_launch_profile_env), the target home pinned, only the target's own secrets overlaid. os.environ.copy() / dict(os.environ) pins the launch profile; contextvars do not cross process boundaries, so resolve before Popen. A child's UnscopedSecretError is a spawn-site bug, never a reason to add environ fallthrough. New threads from scoped code use agent.memory_provider.spawn_context_thread (a bare threading.Thread drops the scope). MCP trust is a per-profile record: mcp_tool_registration.py:: _record_scope_trust keys trust on the home; a secondary never adopts the launch profile's trust for a same-named server, and mcp_tool_handlers.py::_trust_gate_check consults the calling session's profile.

Background-process teardown signals the parent first. process_registry_termination.py:: ProcessTerminationMixin._terminate_host_pid snapshots the descendants, SIGTERMs only the recorded parent, waits terminal.daemon_term_grace_seconds for it to exit and reap its own children, then SIGTERMs the snapshot survivors and SIGKILLs whatever ignored both (so a supervisor that reaps its tree — a Chromium/Electron browser reaping its zygotes, a shell trap — exits cleanly, while a shell whose children ignore SIGHUP still leaves no orphan). Never SIGTERM descendants before the parent: killing a browser's zygote mid-shutdown turns exit 0 into a SIGTRAP core dump. _stop_systemd_unit (scope teardown, kills the worker cgroup) runs only after that PID kill in kill(), or on a parent already proven dead/recycled (session.exited, _signal_kill recycled-PID path, checkpoint recovery) — it is never the first signal a live parent receives.

Delegation (tools/delegate_tool.py)

Spawns a subagent with isolated context + terminal session; the parent waits for the summary unless background=true, which returns a delegation id and re-enters the result via the async-delegation completion queue. Shapes: single (goal + optional context, toolsets) or batch (tasks: [...], concurrency capped by delegation.max_concurrent_children, default 3). A background batch returns as ONE completion by default; with delegation.independent_completions it is split into completion units (delegate_tool_dispatch._units_of): tasks sharing a group join and report together; each ungrouped task reports alone as it finishes. Units of one call share ONE pool slot (slot_key in async_delegation._dispatch) — never count units against capacity; the executor is sized by live UNITS and the stall clock arms when the runner starts, so a queued unit is never judged stalled. Roles: leaf (default; no delegate_task, clarify, memory, send_message, cronjob; keeps execute_code) and orchestrator (keeps delegate_task; gated by delegation.orchestrator_enabled, bounded by delegation.max_spawn_depth, default 2). Config knobs under delegation:: max_concurrent_children, independent_completions, max_spawn_depth, child_timeout_seconds, orchestrator_enabled, subagent_auto_approve, inherit_mcp_toolsets, max_iterations. Child processes: a child's background processes are killed at its teardown and their notices are suppressed in the parent; process_manage(action="handoff") (children only) flips ProcessSession.owner_task_id to the parent under the registry lock (process_registry.transfer_ownership) so the completion routes and reaps by the new owner; un-handed leftovers land on the result as orphaned_processes, exited-but-never-read notify processes as unread_completions (_ChildRun.account_background_processes, before cleanup kills them). Child kernels: a child's execute_code kernels are keyed <parent-owner>::child::<child-session-id>, pinned against the max_session_kernels LRU cap while the child runs and disposed by cleanup (code_kernel.shutdown_kernels_for_delegated_child) — never let a finished child's kernel squat the cap. output_schema: a miss after the one retry keeps status: completed with the raw text in summary plus schema_valid: false / schema_errors / schema_note — never discard a child's result. Durability: background delegation is process-local; work that must survive restart uses cronjob or terminal(background=True, notify_on_complete=True). heartbeat=N on the same spawn emits a heartbeat event (type on the completion queue; delta output only, HEARTBEAT_MIN_SECONDS floor, one daemon timer thread for all sessions in process_registry.py::ProcessRegistry._heartbeat_loop) — every surface that renders watch_match must render heartbeat (process_registry_notifications.py, gateway/run.py::_drain_gateway_watch_events, tui_gateway/session_notifications.py::_DEDUP_EXTRA_FIELDS). Stop fan-out: a turn's hard interrupt reaches only _active_children; background units are detached at dispatch, so every stop surface (gateway /stop busy AND idle paths, TUI session.interrupt, ACP cancel, CLI /stop via interrupt_all) calls async_delegation.interrupt_for_session too. Depth>0 delegations are always synchronous (_model_background_value), so the stop recurses through the child's own _active_children fan-out; an interrupted entry's summary is the child's last real assistant text (_build_result_entry). API: website/docs/developer-guide/subagent-lifecycle-api.md.

Tests

tests/tools/. Test the handler through the registry (real dispatch), not the bare function only; assert contracts ("every registered tool has a toolset", "no schema description names a tool from another toolset") rather than tool counts. Approval/security-boundary tools are E2E'd with real imports against a temp HERMES_HOME (see tests/tools/test_approval_config_readonly.py).

Surface capability is a property of the SESSION, never of the process env

A tool that works only because of who is on the other end (desktop panes, in-app browser, message reactions, Projects) must resolve availability from the session's own source, not from an env var on the backend. Client and backend are separate machines: the desktop app may drive a locally spawned backend, one over SSH, one behind URL + token, or Hermes Cloud, and only the first two carry HERMES_DESKTOP=1. An env-keyed gate is a silent no-op on the other topologies — the tool is stripped from the schema while the platform hint tells the model it is "inside the Hermes desktop app". The pattern:

• The toolset is the surface gate. Keep such tools off _HERMES_CORE_TOOLS and in a named toolset (desktop_ui, project); the GUI gateway's _load_enabled_toolsets(platform) folds it in when the session's platform says GUI. One resolver, every topology.
• check_fn answers reachability or opt-in, not surface. "Is the bridge wired?" — fine. "Was I spawned by Electron?" — not. check_fn results are TTL-cached process-wide (tools/registry.py); a per-session answer does not belong there.
• Ask which identity you mean. HERMES_DESKTOP=1 legitimately means "this backend was spawned by the app" (cron ticker, web-dist handling). It does NOT mean "a GUI is watching"; the embedded terminal pane (hermes --tui against that backend) is the counterexample.

Test: if the capability still makes sense with the client on another machine, it is session-scoped. Assert the GUI session gets the tool with the env var absent.