Files
hermes-agent/tools/delegate_tool_config.py
Austin Pickett e7bff4b6d8 fix(cron): a pinned job never falls back to the global fallback chain (#120312)
* refactor(fallback): share the pinned-owner chain rule

delegate_task's _resolve_child_fallback_chain decides which fallback chain
a child may walk: a pinned child never borrows the parent chain, an explicit
[] disables fallback, a declared list is the child's own. Cron needs the
same rule for pinned jobs (#100437), so the body moves to
hermes_cli.fallback_config.scoped_fallback_chain and the delegation helper
becomes a thin caller. Behaviour is unchanged; the delegation matrix test
still pins every cell.

* fix(cron): a pinned job never falls back to the global chain

A job with its own provider, model or base_url is an explicit operator pin
(since 0469740ab3 unpinned jobs store none of these). It still walked the
global fallback_providers chain in two places, so a pinned job could run
on a different provider and model than the one chosen:

- _resolve_job_runtime walked the chain on an AuthError or transient
  network failure while resolving the pinned primary;
- _resolve_cron_agent_setup handed the global chain to every cron agent as
  fallback_model, so the conversation loop's provider ladder could swap a
  pinned job mid-run.

Both now read _job_fallback_chain(job, cfg), which returns no chain for a
pinned job through the same scoped_fallback_chain rule delegate_task uses
for pinned children. The pre-dispatch key check reads it too: the global
chain used to skip that check for every job, so a pinned job with a
missing key now blocks before the agent is built instead of failing in the
resolver. The transient-failure notice for a pinned job says it does not
fall back and names --unpin, instead of "No backup provider succeeded".

Unpinned jobs (including legacy *_snapshot records) and same-provider
credential-pool rotation are unchanged. The two scheduler tests that
asserted atomic provider+model fallback swaps used pinned jobs; they now
use unpinned jobs and keep the same assertions.

No per-job fallback_providers list: jobs have no generic override field
(create_job/update_job, the cronjob tool schema and the CLI enumerate each
field), so an opt-in chain would be a new surface on all of them. The
escape hatch is to leave the job unpinned and pick its model with
cron.model / cron.model_provider.

Co-authored-by: 686f6c61 <6115107+686f6c61@users.noreply.github.com>

* docs(cron): pinned jobs do not use fallback_providers

cron.md "Provider recovery" and the pre-dispatch key check, the cron rows
and section in fallback-providers.md, and the developer notes in
cron-internals.md / provider-runtime.md said every cron job inherits the
global chain. State the new rule, the compatibility note for users who
relied on a pinned job landing on the chain, and the unpinned + cron.model
alternative.

---------

Co-authored-by: 686f6c61 <6115107+686f6c61@users.noreply.github.com>
2026-09-23 10:42:18 -04:00

598 lines
34 KiB
Python

"""Delegation config knobs (delegation.* keys) and child credential/provider resolution."""
from __future__ import annotations
import logging
import os
from typing import Any, Dict, List, Optional
from utils import base_url_hostname, is_truthy_value
from hermes_cli.fallback_config import scoped_fallback_chain
logger = logging.getLogger("tools.delegate_tool") # log-record parity with the origin module
# Runtime-provider sentinel for providers that are not natively known; must
# match hermes_cli.runtime_provider.RUNTIME_PROVIDER_TYPE_CUSTOM.
_RUNTIME_PROVIDER_CUSTOM = "custom"
_DEFAULT_MAX_CONCURRENT_CHILDREN = 10
# One-shot guard: _get_max_concurrent_children() runs on every get_definitions()
# schema rebuild, so the >10 cost advisory would otherwise log on every turn.
_HIGH_CONCURRENCY_WARNED = False
MAX_DEPTH = 1 # flat by default: parent (0) -> child (1); deeper needs max_spawn_depth
_MIN_SPAWN_DEPTH = 1 # floor for the configurable cap; MAX_DEPTH stays the default
_LEGACY_MAX_ASYNC_WARNED = False
# No default wall-clock cap on children: legitimate heavy work (deep reviews, research fan-outs, slow reasoning
# models) was being killed mid-task. Stuck-child detection is the heartbeat staleness monitor;
# delegation.child_timeout_seconds opts back in.
DEFAULT_CHILD_TIMEOUT: Optional[float] = None
def _cfg() -> dict:
"""The ``delegation`` section, read through the origin so tests can patch it."""
from tools.delegate_tool import _load_config
return _load_config()
# ── Subagent approval callbacks ─────────────────────────────────────────────
# Subagent worker threads don't inherit the CLI's threading.local approval
# callback, so prompt_dangerous_approval() would fall back to input() and
# deadlock the parent's prompt_toolkit TUI. Every worker gets a non-interactive
# callback via ThreadPoolExecutor(initializer=...): deny by default, approve when
# delegation.subagent_auto_approve is true. Both warn for audit. Gateway sessions
# are unaffected (they resolve approvals via tools/approval.py's per-session queue).
def _subagent_auto_deny(command: str, description: str, **kwargs) -> str:
"""Auto-deny (safe default): returns 'deny' so the child sees a recoverable refusal."""
logger.warning(
"Subagent auto-denied dangerous command: %s (%s). Set delegation.subagent_auto_approve: true to allow.",
command, description,
)
return "deny"
def _subagent_auto_approve(command: str, description: str, **kwargs) -> str:
"""Auto-approve (opt-in YOLO via delegation.subagent_auto_approve): returns 'once'."""
logger.warning("Subagent auto-approved dangerous command: %s (%s)", command, description)
return "once"
def _get_subagent_approval_callback():
"""Callback for subagent worker threads per delegation.subagent_auto_approve (default False)."""
if is_truthy_value(_cfg().get("subagent_auto_approve", False)):
return _subagent_auto_approve
return _subagent_auto_deny
def _knob(key: str, env_var: Optional[str], parse, default, invalid_msg: str):
"""delegation.<key> > <env_var> > default. A config value that fails ``parse`` logs ``invalid_msg`` (``%r`` = the
value) and yields the default; an env value that fails is silently ignored."""
val = _cfg().get(key)
if val is not None:
try:
return parse(val)
except (TypeError, ValueError):
logger.warning(invalid_msg, val)
return default
env_val = os.getenv(env_var) if env_var else None
if env_val:
try:
return parse(env_val)
except (TypeError, ValueError):
pass
return default
def _warn_once(flag_name: str, message: str, *args: Any) -> None:
"""Module-global one-shot warning (``flag_name`` is the ``_*_WARNED`` global)."""
if not globals()[flag_name]:
globals()[flag_name] = True
logger.warning(message, *args)
def _get_oneshot_max_children() -> int:
"""delegation.oneshot_max_children (total children per finite one-shot session; 0 = unlimited)."""
return _knob(
"oneshot_max_children", None, lambda v: max(0, int(v)), 2,
"delegation.oneshot_max_children=%r is not a valid integer; using default 2",
)
def _get_max_concurrent_children() -> int:
"""delegation.max_concurrent_children > DELEGATION_MAX_CONCURRENT_CHILDREN env > 10.
Floor of 1 is the only bound enforced; there is no ceiling.
"""
result = _knob(
"max_concurrent_children", "DELEGATION_MAX_CONCURRENT_CHILDREN", lambda v: max(1, int(v)),
_DEFAULT_MAX_CONCURRENT_CHILDREN,
f"delegation.max_concurrent_children=%r is not a valid integer; using default {_DEFAULT_MAX_CONCURRENT_CHILDREN}",
)
if result > 10 and _cfg().get("max_concurrent_children") is not None:
_warn_once(
"_HIGH_CONCURRENCY_WARNED", "delegation.max_concurrent_children=%d: each child consumes API tokens "
"independently. High values multiply cost linearly.", result,
)
return result
def _get_independent_completions() -> bool:
"""delegation.independent_completions (bool, default False): split a background call into per-task / per-group
completion messages that land as each finishes. Off = one consolidated message when the whole call is done."""
return is_truthy_value(_cfg().get("independent_completions", False))
def _get_worktree_isolation() -> bool:
"""delegation.worktree_isolation (bool, default False): each child gets its own
git worktree off the parent's HEAD so parallel children never contend for one
working copy. Git-only and local-backend-only; otherwise silently ignored."""
return bool(_cfg().get("worktree_isolation", False))
def _get_max_async_children() -> int:
"""Concurrency cap for background delegations == delegation.max_concurrent_children. At capacity a new async
dispatch is REJECTED (not queued) so a runaway model can't pile up unbounded background work; the caller then
runs synchronously. A leftover ``delegation.max_async_children`` key is ignored with a one-time warning."""
from tools.delegate_tool import _get_max_concurrent_children
if _cfg().get("max_async_children") is not None:
_warn_once(
"_LEGACY_MAX_ASYNC_WARNED", "delegation.max_async_children is deprecated and ignored; "
"delegation.max_concurrent_children now caps background "
"delegations too. Remove the stale key from config.yaml.",
)
return _get_max_concurrent_children()
def _parse_timeout(raw: Any) -> Optional[float]:
"""Seconds → None (<= 0 disables) or max(30, value). Raises on non-numeric."""
parsed = float(raw)
return None if parsed <= 0 else max(30.0, parsed)
def _get_child_timeout() -> Optional[float]:
"""Inactivity cap for one child (seconds of NO progress), or None (default: no cap). Failures should come from
what the child does (API/tool errors, iteration budget), not a stopwatch: the cap restarts on every sign of
progress — a completed call, a tool change, an activity-clock tick — so a slow provider serving multi-minute
completions never loses a live child, and a child frozen for the whole window is still caught. A configured
value pre-empts nothing the heartbeat staleness monitor would not also catch. delegation.child_timeout_seconds
> 0 opts in (floor 30 s); 0 or negative disables. Env fallback: DELEGATION_CHILD_TIMEOUT_SECONDS."""
return _knob(
"child_timeout_seconds", "DELEGATION_CHILD_TIMEOUT_SECONDS", _parse_timeout, DEFAULT_CHILD_TIMEOUT,
"delegation.child_timeout_seconds=%r is not a valid number; using default (no timeout)",
)
def _get_max_spawn_depth() -> int:
"""delegation.max_spawn_depth floored at 1 (no ceiling). Depth 0 is the parent; agents at depths 0..N-1 may spawn,
depth N is the leaf floor. Default 1 is flat. Each extra level multiplies API cost."""
def _floored(v):
ival = int(v)
if ival < _MIN_SPAWN_DEPTH:
logger.warning("delegation.max_spawn_depth=%d below floor %d; using %d", ival, _MIN_SPAWN_DEPTH, _MIN_SPAWN_DEPTH)
return max(_MIN_SPAWN_DEPTH, ival)
return _knob(
"max_spawn_depth", None, _floored, MAX_DEPTH,
f"delegation.max_spawn_depth=%r is not a valid integer; using default {MAX_DEPTH}",
)
def _get_orchestrator_enabled() -> bool:
"""delegation.orchestrator_enabled kill switch (default True): False forces every child to leaf."""
val = _cfg().get("orchestrator_enabled", True)
if isinstance(val, bool):
return val
# Accept "true"/"false" strings from YAML that doesn't auto-coerce.
if isinstance(val, str):
return val.strip().lower() in {"true", "1", "yes", "on"}
return True
def _get_inherit_mcp_toolsets() -> bool:
"""Whether narrowed child toolsets should keep the parent's MCP toolsets."""
return is_truthy_value(_cfg().get("inherit_mcp_toolsets"), default=True)
def _normalized_runtime_url(value: Any) -> str:
return str(value or "").strip().rstrip("/")
def _inherit_parent_capabilities(parent_agent, override_provider, override_base_url) -> Optional[dict]:
"""Parent's endpoint-trust capability map for a child, or None. ``agent.capabilities`` is a trust decision scoped
to one provider+endpoint: inherited ONLY when the child runs the parent's exact route; any provider or base_url
override stays DEFAULT-DENY (matches the /model switch posture).
See #94036, #97292.
"""
if override_provider or override_base_url:
return None
parent_caps = getattr(parent_agent, "capabilities", None)
if not isinstance(parent_caps, dict):
return None
return {key: value for key, value in parent_caps.items() if isinstance(key, str) and isinstance(value, bool)}
def _inherit_parent_endpoint(parent_agent, surface_base_url: Optional[str], surface_api_key: Any) -> tuple:
"""``(base_url, api_key)`` the parent is actually calling, taken from ONE source. ``parent_agent.base_url`` /
``api_key`` can lag the live client (old OpenRouter URL vs local Ollama; a fallback runtime the surface attributes
have not caught up with), and pairing the live endpoint with the surface key hands the child a (base_url, key)
pair that was never valid anywhere — an instant, non-retryable 401 (#90009). The live client's key travels with
its URL; the surface attributes are used only when there is no live OpenAI-wire client (native Anthropic/Bedrock
runtimes keep ``client=None``)."""
client_kwargs = getattr(parent_agent, "_client_kwargs", None)
client = getattr(parent_agent, "client", None)
live_candidates = (
(client_kwargs.get("base_url"), client_kwargs.get("api_key")) if isinstance(client_kwargs, dict) else (None, None),
# OpenAI SDK exposes base_url as httpx.URL — coerce before comparing.
(getattr(client, "base_url", ""), getattr(client, "api_key", None)) if client is not None else (None, None),
)
for raw_url, live_key in live_candidates:
url = _normalized_runtime_url(raw_url)
if url and url.startswith(("http://", "https://")):
return url, (live_key or surface_api_key)
return (surface_base_url or None), surface_api_key
def _loaded_pool(key: Any):
"""``load_pool(key)`` when it holds credentials, else None."""
from agent.credential_pool import load_pool
pool = load_pool(key)
return pool if pool is not None and pool.has_credentials() else None
def _pool_serves_endpoint(pool: Any, provider: Optional[str], base_url: Optional[str]) -> bool:
"""Provider identity AND at least one entry for the child's endpoint; pools without entry metadata pass."""
from agent.credential_pool import (
credential_pool_entry_serves_endpoint as _entry_serves_endpoint, credential_pool_matches_provider,
)
if not credential_pool_matches_provider(pool, provider, base_url=base_url):
return False
entries_fn = getattr(pool, "entries", None)
if not callable(entries_fn):
return True
entries = entries_fn()
return not isinstance(entries, list) or any(_entry_serves_endpoint(entry, base_url) for entry in entries)
def _resolve_child_credential_pool(
effective_provider: Optional[str], parent_agent, effective_base_url: Optional[str] = None,
effective_requested_provider: Optional[str] = None,
):
"""Credential pool for the child: parent's pool (same provider), that provider's own pool, or None (child keeps
its fixed credential). Custom endpoints all collapse to ``provider="custom"``, so they are matched by endpoint
identity (the ``custom:<name>`` pool key) — sharing the parent's pool across different custom endpoints would
overwrite the child's delegated base_url on lease; an unregistered custom endpoint (no custom_providers entry)
keeps the child's fixed credential rather than inherit the parent's.
Custom endpoints are a special case: every direct ``delegation.base_url`` runtime collapses to
``provider="custom"``, so bare provider equality would treat two *different* custom endpoints as
interchangeable and let the child inherit the parent's pool. We therefore resolve custom runtimes by
endpoint identity (the ``custom:<name>`` pool key derived from the base_url) and only share the parent's
pool when both resolve to the *same* custom endpoint. See #7833.
Named custom providers may share one gateway URL with different credentials, so the inherited
``requested_provider`` identity takes precedence over URL-only matching (#45763): the child must not
lease the first pool registered for the shared endpoint.
"""
parent_pool = getattr(parent_agent, "_credential_pool", None)
if not effective_provider:
return parent_pool
parent_provider = getattr(parent_agent, "provider", None) or ""
try:
if effective_provider == "custom":
from agent.credential_pool import get_custom_provider_pool_key
child_key = get_custom_provider_pool_key(effective_base_url, provider_name=effective_requested_provider)
if child_key is None:
return None
parent_key = get_custom_provider_pool_key(
getattr(parent_agent, "base_url", None), provider_name=getattr(parent_agent, "requested_provider", None),
)
if parent_pool is not None and parent_provider == "custom" and parent_key is not None and parent_key == child_key:
return parent_pool
return _loaded_pool(child_key)
if parent_pool is not None and effective_provider == parent_provider:
if not effective_base_url or _pool_serves_endpoint(parent_pool, effective_provider, effective_base_url):
return parent_pool
logger.debug("Parent %s pool has no entry for child endpoint %s; not sharing it",
effective_provider, effective_base_url)
pool = _loaded_pool(effective_provider)
if pool is not None and effective_base_url and not _pool_serves_endpoint(pool, effective_provider, effective_base_url):
return None # child keeps its fixed credential
return pool
except Exception as exc:
if effective_provider == "custom":
logger.debug("Could not resolve custom credential pool for child endpoint '%s': %s", effective_base_url, exc)
else:
logger.debug("Could not load credential pool for child provider '%s': %s", effective_provider, exc)
return None
def _merge_request_overrides(runtime_overrides, explicit_overrides):
"""Merge explicit ``delegation.request_overrides`` OVER runtime-derived ones. Explicit top-level keys win;
``extra_body`` is deep-merged ONE level so provider personality (e.g. ``thinking: {type: disabled}``) survives
unless the explicit dict redefines that exact key. Both sides are deep-copied so transport-side mutation can't
leak into the config/runtime cache. None when both are empty."""
import copy as _copy
runtime_overrides = runtime_overrides if isinstance(runtime_overrides, dict) else None
explicit_overrides = explicit_overrides if isinstance(explicit_overrides, dict) else None
if not runtime_overrides and not explicit_overrides:
return None
merged = _copy.deepcopy(runtime_overrides) if runtime_overrides else {}
explicit = _copy.deepcopy(explicit_overrides) if explicit_overrides else {}
runtime_extra = merged.get("extra_body")
explicit_extra = explicit.pop("extra_body", None)
merged.update(explicit)
if isinstance(runtime_extra, dict) and isinstance(explicit_extra, dict):
runtime_extra.update(explicit_extra)
merged["extra_body"] = runtime_extra
elif explicit_extra is not None:
merged["extra_body"] = explicit_extra
return merged or None
# Native-SDK providers speak their own wire protocol and can't be reached via chat_completions against a base_url:
# always take the runtime-provider path (a configured base_url still flows through it, e.g. a Bedrock region).
_NATIVE_SDK_PROVIDERS = frozenset({"bedrock", "vertex", "google", "google-genai"})
_EXPLICIT_API_MODES = frozenset({"chat_completions", "codex_responses", "anthropic_messages"})
def _require_pinned_command(command: Optional[str], message: str) -> None:
"""A pinned ACP transport command must exist on PATH — refuse loudly rather
than let the child silently fall back to another transport."""
import shutil as _shutil
if command and not _shutil.which(command):
raise ValueError(message)
def _credential_bundle(model, provider, base_url, api_key, api_mode, request_overrides, **extra) -> dict:
"""The child credential dict every branch of ``_resolve_delegation_credentials`` returns."""
return {
"model": model, "provider": provider, "base_url": base_url, "api_key": api_key, "api_mode": api_mode,
"request_overrides": request_overrides, **extra,
}
def _direct_endpoint_credentials(v: dict, explicit_request_overrides) -> dict:
"""``delegation.base_url`` branch: provider/api_mode from URL heuristics."""
# Shared URL-based api_mode detector so Anthropic-compatible direct endpoints (/anthropic suffix: Azure AI
# Foundry, MiniMax, Zhipu, LiteLLM) get the Messages transport instead of 404ing on chat_completions.
# Without this, subagents would default to chat_completions and hit 404s on endpoints that only speak
# the Anthropic Messages protocol. Fixes #10213.
from hermes_cli.runtime_provider import _detect_api_mode_for_url
base_lower = v["base_url"].lower()
host = base_url_hostname(v["base_url"])
provider = "custom"
api_mode = _detect_api_mode_for_url(v["base_url"]) or "chat_completions"
if host == "chatgpt.com" and "/backend-api/codex" in base_lower:
provider, api_mode = "openai-codex", "codex_responses"
elif host == "api.anthropic.com":
provider, api_mode = "anthropic", "anthropic_messages"
elif "api.kimi.com/coding" in base_lower:
api_mode = "anthropic_messages"
# Explicit delegation.api_mode always wins over the URL heuristic; a provider plugin's
# registered dialect counts as explicit.
from agent.transports import registered_api_modes
if v["api_mode"] in _EXPLICIT_API_MODES or (v["api_mode"] and v["api_mode"] in registered_api_modes()):
api_mode = v["api_mode"]
# Preserve the configured provider's request personality on an explicit endpoint.
request_overrides = None
if v["provider"]:
try:
from hermes_cli.runtime_provider import resolve_runtime_provider
runtime = resolve_runtime_provider(requested=v["provider"], target_model=v["model"])
request_overrides = dict(runtime.get("request_overrides") or {}) or None
except Exception as exc:
logger.debug(
"delegation.base_url: runtime resolution for provider '%s' failed; proceeding without request_overrides: %s",
v["provider"], exc,
)
# api_key None → inherited from parent in _build_child_agent
return _credential_bundle(
v["model"], provider, v["base_url"], v["api_key"], api_mode,
_merge_request_overrides(request_overrides, explicit_request_overrides),
)
def _runtime_provider_credentials(v: dict, explicit_request_overrides) -> dict:
"""``delegation.provider`` branch: full bundle via the runtime provider system."""
configured_provider = v["provider"]
try:
from hermes_cli.runtime_provider import resolve_runtime_provider
runtime = resolve_runtime_provider(requested=configured_provider, target_model=v["model"])
except Exception as exc:
raise ValueError(
f"Cannot resolve delegation provider '{configured_provider}': {exc}. "
f"Check that the provider is configured (API key set, valid provider name), "
f"or set delegation.base_url/delegation.api_key for a direct endpoint. "
f"Available providers: openrouter, nous, zai, kimi-coding, minimax."
) from exc
api_key = runtime.get("api_key", "")
if not api_key:
raise ValueError(
f"Delegation provider '{configured_provider}' resolved but has no API key. "
f"Set the appropriate environment variable or run 'hermes auth'."
)
# A pinned ACP transport command must exist — refuse the spawn loudly rather than letting the child
# silently fall back to another transport (#80450).
pinned_command = runtime.get("command")
_require_pinned_command(
pinned_command, f"Delegation provider '{configured_provider}' is pinned to the "
f"'{pinned_command}' command, which was not found on PATH. "
f"Install it or choose a different delegation provider.",
)
# A provider override is assembled into the child as one bundle (see _resolve_child_runtime), so it must
# carry its own endpoint. ACP transports are addressed by command and native-SDK providers by their SDK,
# so neither needs a URL; anything else without one would otherwise be built with no endpoint at all.
if (not runtime.get("base_url") and not pinned_command
and configured_provider.strip().lower() not in _NATIVE_SDK_PROVIDERS):
raise ValueError(
f"Delegation provider '{configured_provider}' resolved without a base_url. "
f"Refusing to build a subagent with an incomplete credential bundle — check the provider's "
f"configuration / auth, or set delegation.base_url for a direct endpoint."
)
return _credential_bundle(
v["model"] or runtime.get("model") or None,
configured_provider if runtime.get("provider") == _RUNTIME_PROVIDER_CUSTOM else runtime.get("provider"),
runtime.get("base_url"), api_key, runtime.get("api_mode"),
_merge_request_overrides(runtime.get("request_overrides"), explicit_request_overrides) or {},
command=pinned_command, args=list(runtime.get("args") or []),
)
def _resolve_delegation_credentials(cfg: dict, parent_agent) -> dict:
"""Child credential bundle from the ``delegation`` config section. Three branches: ``base_url`` set → direct
endpoint (``api_key`` None means inherit the parent's key, so providers keyed outside OPENAI_API_KEY work);
``provider`` set → full bundle via the runtime provider system (same path as CLI/gateway startup); neither →
None values, child inherits everything. ``request_overrides`` is honored on every branch. Raises ValueError
with a user-facing message."""
values = {k: str(cfg.get(k) or "").strip() or None for k in ("model", "provider", "base_url", "api_key")}
values["api_mode"] = str(cfg.get("api_mode") or "").strip().lower() or None
explicit_request_overrides = cfg.get("request_overrides") if isinstance(cfg.get("request_overrides"), dict) else None
is_native_sdk_provider = (values["provider"] or "").strip().lower() in _NATIVE_SDK_PROVIDERS
if values["base_url"] and not is_native_sdk_provider:
return _direct_endpoint_credentials(values, explicit_request_overrides)
if not values["provider"]:
# Pure inherit; explicit request_overrides still merge OVER the parent's.
return _credential_bundle(
values["model"], None, None, None, None,
_merge_request_overrides(getattr(parent_agent, "request_overrides", None), explicit_request_overrides),
)
return _runtime_provider_credentials(values, explicit_request_overrides)
def _load_config() -> dict:
"""The ``delegation`` config section (read-only — do NOT mutate). Prefers the shared ``load_config_readonly()``
(follows HERMES_HOME/profile; no deepcopy, since this runs on every get_definitions() rebuild) over the legacy
``cli.CLI_CONFIG``, which can hide user-set keys — except that ``HERMES_IGNORE_USER_CONFIG=1`` is only honored
by the legacy loader, so it stays authoritative when that flag is set."""
if os.environ.get("HERMES_IGNORE_USER_CONFIG") != "1":
try:
from hermes_cli.config import load_config_readonly
cfg = load_config_readonly().get("delegation") or {}
if isinstance(cfg, dict):
return cfg
except Exception:
pass
try:
from cli import CLI_CONFIG
cfg = CLI_CONFIG.get("delegation") or {}
return cfg if isinstance(cfg, dict) else {}
except Exception:
return {}
# OpenRouter routing filters: inherited from the parent, but reset to these defaults under a pinned provider — parent
# filters (e.g. only=["Anthropic"]) would silently force the child back onto the parent's provider.
# openrouter_min_coding_score stays inherited: model-gated, no-op elsewhere.
_ROUTING_FILTER_DEFAULTS = (
("providers_allowed", None), ("providers_ignored", None), ("providers_order", None), ("provider_sort", None),
("provider_require_parameters", False), ("provider_data_collection", ""),
)
_NOUS_PROVIDERS = frozenset({"nous", "nous-portal", "nousresearch"})
def _resolve_child_fallback_chain(parent_agent, routing_cfg: Any, pinned: bool) -> Optional[List[Dict[str, Any]]]:
"""Fallback chain for a child, owned by the same config block as its route.
Pinned children (provider, endpoint or model override) never borrow the parent chain;
unpinned children inherit it when ``fallback_providers`` is absent/null. An explicit ``[]``
disables fallback either way. Malformed entries are dropped by the canonical normalizer.
Same rule as a pinned cron job (``cron/scheduler.py::_job_fallback_chain``).
"""
return scoped_fallback_chain(
getattr(parent_agent, "_fallback_chain", None),
routing_cfg.get("fallback_providers") if isinstance(routing_cfg, dict) else None,
pinned=pinned, owner="delegation")
def _resolve_child_runtime(
parent_agent, delegation_cfg: dict, parent_api_key: Any, *, model: Optional[str], override_provider: Optional[str],
override_base_url: Optional[str], override_api_key: Optional[str], override_api_mode: Optional[str],
override_acp_command: Optional[str], override_acp_args: Optional[List[str]],
routing_cfg: Optional[Dict[str, Any]] = None,
) -> Dict[str, Any]:
"""Child credentials, transport and routing (config override > parent inherit) as ``AIAgent`` kwargs. Rules that
are easy to break: api_mode is re-derived (not inherited) when the child's provider differs from the parent's
or is Nous Portal (dual-wire); a pinned ``delegation.command`` must exist on PATH or the spawn fails loudly;
``override_provider`` clears the parent's ACP transport, fallback chain and OpenRouter routing filters so the
pinned provider is actually honoured."""
effective_model = model or parent_agent.model
# provider/base_url are one bundle: all from the override, or all from the parent. Per-field fallback built
# children like an override provider pointed at the PARENT's endpoint (e.g. copilot credentials on the
# parent's Codex URL), which 404s on every request and can't be rescued by the fallback chain, whose dedup
# matches provider+model and so skips the entry as a self-loop. _inherit_parent_endpoint recovers the
# parent's live endpoint (with its key), which is meaningless for a different provider.
if override_provider:
effective_provider = override_provider
effective_base_url = override_base_url
elif override_base_url:
effective_provider = getattr(parent_agent, "provider", None)
effective_base_url = override_base_url
else:
effective_provider = getattr(parent_agent, "provider", None)
effective_base_url, parent_api_key = _inherit_parent_endpoint(parent_agent, parent_agent.base_url, parent_api_key)
# api_mode: each provider has its own wire, so a different provider re-derives (None) instead of inheriting (404s
# otherwise). Nous Portal is dual-wire within one provider (anthropic/* → Messages, else chat_completions), so
# same-provider inheritance would pin the child on the wrong wire — re-derive.
# Bug #20558 / PR #20563: api_mode must NOT be inherited when the child uses a different provider than
# the parent — each provider has its own API surface (e.g. MiniMax uses anthropic_messages, DeepSeek
# uses chat_completions). Inheriting the parent's mode causes 404 errors when the child routes to the
# wrong endpoint. Derive the mode from the target provider when it differs. Same-provider inheritance
# would pin a child Hermes/Qwen subagent onto the parent's Claude Messages wire (or the reverse).
# agent_init honors an explicit api_mode above its nous branch, so re-derive here before construction.
_parent_provider = getattr(parent_agent, "provider", None) or ""
if override_api_mode is not None:
effective_api_mode = override_api_mode
elif (effective_provider or "").strip().lower() in _NOUS_PROVIDERS:
from hermes_cli.providers import nous_api_mode
effective_api_mode = nous_api_mode(effective_model)
elif effective_provider != _parent_provider:
effective_api_mode = None # force re-derivation from provider's defaults
else:
effective_api_mode = getattr(parent_agent, "api_mode", None)
# A pinned transport that cannot run must fail the spawn loudly, never fall
# back silently (delegate_task pre-validates; this covers direct callers).
_require_pinned_command(
override_acp_command, f"Pinned delegation command '{override_acp_command}' was not "
f"found on PATH. Install it or remove delegation.command from config.yaml.",
)
effective_acp_command = override_acp_command or getattr(parent_agent, "acp_command", None)
effective_acp_args = list(
override_acp_args if override_acp_args is not None else (getattr(parent_agent, "acp_args", []) or [])
)
# A pinned provider must use direct API calls; inheriting the parent's ACP
# transport would bypass the override credentials entirely.
# Inheriting acp_command unconditionally causes run_agent.py to initialize CopilotACPClient, bypassing
# override credentials entirely (issue #16816).
if override_provider and not override_acp_command:
effective_acp_command, effective_acp_args = None, []
# Defensive: validate trusted delegation.command exists on PATH before honoring it. An explicitly pinned
# transport that cannot run must fail the spawn loudly (#80450) — silently falling back to the default
# transport would run the child somewhere the user explicitly routed it away from. Normally unreachable
# via delegate_task, which pre-validates the command in _resolve_delegation_credentials.
if override_acp_command:
from providers import get_provider_profile
profile = get_provider_profile(effective_provider or "")
# A generic process command does not imply the legacy ACP protocol.
if profile is None or profile.auth_type != "external_process":
effective_provider, effective_api_mode = "copilot-acp", "chat_completions"
# A named provider identity is endpoint-scoped. Preserve it only when the
# child inherits the exact parent route; an override owns its final identity.
effective_requested_provider = effective_provider
if not override_provider and not override_base_url and not override_acp_command:
effective_requested_provider = (
getattr(parent_agent, "requested_provider", None) or effective_provider
)
# Reasoning: delegation.reasoning_effort > parent. Keep the raw value — a
# YAML ``false`` must disable thinking, not coerce to "" and inherit.
child_reasoning = getattr(parent_agent, "reasoning_config", None)
try:
delegation_effort = delegation_cfg.get("reasoning_effort")
if delegation_effort or delegation_effort is False:
from hermes_constants import parse_reasoning_effort
parsed = parse_reasoning_effort(delegation_effort)
if parsed is None:
logger.warning("Unknown delegation.reasoning_effort '%s', inheriting parent level", delegation_effort)
else:
child_reasoning = parsed
except Exception as exc:
logger.debug("Could not load delegation reasoning_effort: %s", exc)
kwargs: Dict[str, Any] = {
"base_url": effective_base_url, "api_key": override_api_key or parent_api_key, "model": effective_model,
"provider": effective_provider, "requested_provider": effective_requested_provider,
"capabilities": _inherit_parent_capabilities(parent_agent, override_provider, override_base_url),
"api_mode": effective_api_mode, "acp_command": effective_acp_command, "acp_args": effective_acp_args,
"reasoning_config": child_reasoning,
# Resolve routing and recovery policy from the same configuration owner. A pinned provider, endpoint, or
# model never borrows the parent's chain; an explicitly declared child chain still remains available.
"fallback_model": _resolve_child_fallback_chain(
parent_agent, delegation_cfg if routing_cfg is None else routing_cfg,
pinned=bool(override_provider or override_base_url or model)),
"openrouter_min_coding_score": getattr(parent_agent, "openrouter_min_coding_score", None),
# Routing filters reset to their defaults under a pinned provider (see _ROUTING_FILTER_DEFAULTS).
**{a: d if override_provider else getattr(parent_agent, a, d) for a, d in _ROUTING_FILTER_DEFAULTS},
}
if not override_provider:
kwargs["provider_data_collection"] = kwargs["provider_data_collection"] or ""
child_max_tokens = getattr(parent_agent, "max_tokens", None)
if isinstance(child_max_tokens, int):
kwargs["max_tokens"] = child_max_tokens
return kwargs