fix(delegate): read fresh config.yaml so delegation.model edits take effect (#11999) by briandevans · Pull Request #12053 · NousResearch/hermes-agent

briandevans · 2026-04-18T07:44:45Z

What does this PR do?

Fixes #11999.

delegate_task silently ignored delegation.model / delegation.provider overrides in ~/.hermes/config.yaml — subagents kept running on the parent's model regardless of what the user configured, defeating the whole point of cost-tiered delegation (Sonnet subagents under an Opus parent, say). The reporter's exact scenario:

delegation:
  model: claude-sonnet-4-20250514
  provider: anthropic

Parent on claude-opus-4-6. Expected: child runs on Sonnet. Actual: child runs on Opus; result metadata confirms "model": "claude-opus-4-6". Reporter also observed "persists across /reset commands and config reloads" — an important symptom that points at the actual root cause.

Root cause

tools/delegate_tool.py::_load_config() short-circuited through cli.CLI_CONFIG, a dict populated once at module-import time:

def _load_config() -> dict:
    try:
        from cli import CLI_CONFIG
        cfg = CLI_CONFIG.get("delegation", {})
        if cfg:
            return cfg       # <-- returns stale CACHED value
    except Exception:
        pass
    try:
        from hermes_cli.config import load_config
        full = load_config()
        return full.get("delegation", {})    # <-- fresh read (never reached on CLI)
    except Exception:
        return {}

Two independent bugs in this short-circuit:

Cache staleness. CLI_CONFIG is built by cli.load_cli_config() at module import and is never re-read. An edit to ~/.hermes/config.yaml after the CLI starts — including adding a new delegation: section — is invisible to delegate_task. /reset and config-reload commands don't re-run load_cli_config, which is exactly the reporter's "persists across /reset" symptom.
Truthy-but-empty defaults always win. cli.py populates CLI_CONFIG["delegation"] with built-in defaults ({"max_iterations": 45, "default_toolsets": [...], "model": "", "provider": "", "base_url": "", "api_key": ""}) even when the user's config.yaml has no delegation section at all. That's a truthy dict (non-empty), so if cfg: return cfg fires and the persistent hermes_cli.config.load_config() path is never reached. Downstream, configured_model = str(cfg.get("model") or "").strip() or None resolves to None, _resolve_delegation_credentials sets provider: None (inherit-from-parent), and _build_child_agent does effective_model = model or parent_agent.model → the parent model wins.

Reproduced in a minimal sandbox:

At CLI startup, CLI_CONFIG['delegation'] = {'model': '', 'provider': '', ...}
After user edits config.yaml, delegation.model is claude-sonnet-4-20250514
_load_config() returned: {'model': '', 'provider': '', ...}
BUG REPRODUCED: _load_config returned stale model=''

Fix

Drop the stale CLI_CONFIG short-circuit. Always read fresh via hermes_cli.config.load_config():

def _load_config() -> dict:
    try:
        from hermes_cli.config import load_config
        full = load_config()
        delegation = full.get("delegation", {})
        return delegation if isinstance(delegation, dict) else {}
    except Exception:
        return {}

Also hardens against a malformed delegation: not-a-dict value — the old code returned the non-dict verbatim, which then crashed _resolve_delegation_credentials on .get("model"). Coerce to {} so callers safely fall back to inherit-from-parent defaults.

Precedence / compat table

Scenario	Before	After
Edit `delegation.model` after CLI startup	silently ignored (stale CLI_CONFIG)	picked up on next `delegate_task`
Gateway / cron (cli.py not imported)	already read fresh (fallback path)	read fresh (unchanged)
Startup with no `delegation:` section	stale CLI_CONFIG defaults returned → inherit-from-parent	fresh `load_config()` returns DEFAULT_CONFIG delegation → inherit-from-parent (same result)
`delegation.model: ""` explicitly	inherit-from-parent	inherit-from-parent (unchanged)
`delegation: not-a-dict` (malformed)	non-dict leaks to `.get()` → AttributeError	coerced to `{}` → inherit-from-parent
Runtime mutation of `CLI_CONFIG["delegation"]`	picked up	n/a (grep-verified: nothing in the codebase mutates this at runtime)

Narrow scope — explicitly not changed

_resolve_delegation_credentials — downstream of _load_config, already correct given a fresh config dict. No change needed.
_build_child_agent / AIAgent.__init__ — same; the bug is purely at the config-read layer.
cli.CLI_CONFIG — left as-is. Grep confirms no runtime writer to CLI_CONFIG["delegation"] anywhere, so dropping its use in _load_config loses no runtime-only data.
Performance — load_config() is a single YAML parse plus copy.deepcopy(DEFAULT_CONFIG), measured <1 ms. delegate_task is called at most once per user turn; the cost is negligible.

Pre-empt reviewer Q&A

Q: Why not just re-populate CLI_CONFIG on file-change or /reset?
That would solve this bug but introduces a much larger invariant to maintain (every mutation point of every config section would have to either poll or be hooked). Reading fresh per-call is simpler, has a localized fix, and matches how gateway/cron already operate — so we're unifying instead of diverging further.

Q: Could this regress any path that relied on CLI_CONFIG's import-time snapshot?
Grepped every reader of CLI_CONFIG["delegation"] — only tools/delegate_tool.py::_load_config(). Nothing else reads this section, nothing writes to it at runtime.

Q: load_config() calls ensure_hermes_home(). Side-effect concerns like #11983?
For delegate_task specifically: ensure_hermes_home() creates ~/.hermes subdirs and seeds SOUL.md if they don't exist. This is idempotent after first call. delegate_task is only invoked from an AIAgent tool-dispatch loop where the parent agent has already initialized ~/.hermes (it couldn't have gotten to tool-dispatch otherwise). So the side-effect is a confirmed no-op in this call context. Additionally, the pre-fix fallback path (which fires in gateway/cron) already used load_config(), so this side effect was already part of the function's behaviour on the only paths that mattered.

Q: Why coerce non-dict delegation to {} instead of raising?
_load_config is best-effort at all other exit points (catches any exception → {}). Keeping that contract means callers don't have to learn a new failure mode — they still fall through to inherit-from-parent. A warning-log could be nice follow-up but the behavioural contract is maintained.

Regression coverage

tests/tools/test_delegate.py gets TestLoadConfigFreshReads (5 tests):

test_ignores_stale_cli_config_cache — plants a stale cli module with empty-default CLI_CONFIG in sys.modules, writes an override to config.yaml, asserts _load_config() returns the override. Fails on origin/main with the exact reporter symptom: AssertionError: '' != 'claude-sonnet-4-20250514'.
test_picks_up_delegation_override_written_after_startup — end-user scenario: edit config.yaml, call _load_config, get the edit.
test_non_dict_delegation_section_is_coerced_to_empty — hardens .get() callers downstream; also fails on origin/main (old code returned the non-dict string verbatim).
test_returns_dict_when_config_missing_delegation_section — pins graceful handling when delegation: isn't configured at all.
test_end_to_end_resolve_uses_fresh_config — integration through _resolve_delegation_credentials all the way to effective_model = creds["model"] or parent.model.

How to test

Focused suite:

source venv/bin/activate
python -m pytest tests/tools/test_delegate.py tests/tools/test_delegate_toolset_scope.py -q

→ 77 passed (72 pre-existing + 5 new).

Verify the regression is caught on origin/main:
```
git stash push tools/delegate_tool.py
python -m pytest tests/tools/test_delegate.py::TestLoadConfigFreshReads -v
```
→ 2 fail (test_ignores_stale_cli_config_cache, test_non_dict_delegation_section_is_coerced_to_empty). Restore with git stash pop.
Manual repro (the sandbox script above), before & after.
CI-aligned broad suite:
```
python -m pytest tests/ -q --ignore=tests/integration --ignore=tests/e2e --tb=line -n auto
```
→ 12824 passed, 39 skipped, 15 pre-existing baseline failures (dingtalk ×N, matrix, wsl ×2, approve_deny ×2, file_staleness ×2, local_interrupt_cleanup, plus a couple of known flakes in tools/). All reproduce on clean origin/main; none in tools/delegate_tool.py or tests/tools/test_delegate.py.

Tested on: macOS 15 (Darwin 25.5.0), Python 3.11.15.

Related Issue

Fixes #11999

Type of Change

🐛 Bug fix (non-breaking change that fixes an issue)

Changes Made

tools/delegate_tool.py: rewrite _load_config to always read fresh from hermes_cli.config.load_config(); add defensive dict-coercion against malformed config values.
tests/tools/test_delegate.py: add TestLoadConfigFreshReads — 5 regression tests (2 fail cleanly on origin/main).

Adjacent surfaces checked

_resolve_delegation_credentials — downstream; verified correct given fresh config. Unchanged.
_build_child_agent — receives creds["model"]; verified it correctly wins over parent_agent.model when present. Unchanged.
AIAgent.__init__ model resolution (run_agent.py:708, :772) — preserves user-supplied model. Unchanged.
All four preserve_dots=self._anthropic_preserve_dots() call sites — unrelated. Unchanged.
cli.CLI_CONFIG writers — grep-confirmed none for delegation. No data loss from dropping the short-circuit.

Checklist

Code

I've read the Contributing Guide
Commit messages follow Conventional Commits (fix(scope):)
Searched for existing PRs — none on delegate_task ignores subagent model config — always inherits parent model #11999
Only changes related to this fix
Ran the CI-aligned test command and classified failures baseline-vs-change
Added tests for my changes (5 new; 2 fail on origin/main without the fix)
Tested on macOS 15 (Darwin 25.5.0), Python 3.11.15

Documentation & Housekeeping

No docs changes needed — config-read internals, no public API surface
No config-key changes — existing delegation.model / delegation.provider keys behave as their docstrings already promised
No architecture/workflow changes
Cross-platform impact: none — pure Python, no OS-specific paths
No tool descriptions/schemas touched

Notes for reviewers

No standalone lint command is defined; CI runs python -m pytest tests/ -q --ignore=tests/integration --ignore=tests/e2e --tb=short -n auto (matches .github/workflows/tests.yml).
No hermes_cli/** changes — the Nix workflow should not trigger.

…effect (NousResearch#11999) ``tools.delegate_tool._load_config`` short-circuited through ``cli.CLI_CONFIG``, a dict that's populated **once** at module-import time and is never re-read. In long-lived CLI sessions this meant ``delegation.model`` / ``delegation.provider`` edits to ``~/.hermes/config.yaml`` after startup were silently ignored — subagents kept running on the parent's model regardless of the user's override, defeating the entire point of cost-tiered delegation (Sonnet subagents under an Opus parent, say). The reporter's exact scenario (NousResearch#11999): set delegation: model: claude-sonnet-4-20250514 provider: anthropic in config.yaml, run the parent on ``claude-opus-4-6``, call ``delegate_task`` — the child reports ``claude-opus-4-6`` and the result metadata confirms ``"model": "claude-opus-4-6"``. Root cause ---------- The short-circuit at the top of ``_load_config``: try: from cli import CLI_CONFIG cfg = CLI_CONFIG.get("delegation", {}) if cfg: return cfg misfires in two ways: 1. ``CLI_CONFIG`` is a module-level dict built by ``cli.load_cli_config()`` at import time. Subsequent edits to ``config.yaml`` (including adding a new ``delegation:`` section) are never picked up by that process. ``/reset`` and config-reload commands don't re-run ``load_cli_config``, which matches the reporter's observation that the bug "persists across /reset commands and config reloads". 2. The ``if cfg:`` check always fires because cli.py's delegation defaults already populate a truthy dict (``{"max_iterations": 45, "default_toolsets": [...], "model": "", "provider": "", "base_url": "", "api_key": ""}``) even when the user hasn't set any override. That truthy-but-empty dict bypasses the persistent ``hermes_cli.config.load_config()`` path entirely, so ``configured_model = str(cfg.get("model") or "").strip() or None`` resolves to ``None`` and ``_build_child_agent`` falls back to ``parent_agent.model``. Fix --- Drop the stale ``CLI_CONFIG`` short-circuit and always read fresh via ``hermes_cli.config.load_config()``. That's the path gateway and cron code already use, so unifying everyone on it also removes a subtle CLI-only divergence. Performance cost is a single YAML parse + deep- copy of ``DEFAULT_CONFIG``, <1 ms in practice — and ``delegate_task`` is invoked at most once per user turn. Also harden the function so a malformed ``delegation: not-a-dict`` value coerces to ``{}`` instead of leaking a non-dict through to ``_resolve_delegation_credentials`` (which calls ``.get()`` on it). Narrow scope — explicitly not changed ------------------------------------- * ``_resolve_delegation_credentials`` — downstream of ``_load_config``. Verified correct given a fresh config dict; no changes needed. * ``_build_child_agent`` / ``AIAgent.__init__`` — same. * ``cli.CLI_CONFIG`` — leave as-is. Nothing else in the codebase writes to ``CLI_CONFIG["delegation"]`` at runtime (grep-verified), so dropping its use in ``_load_config`` loses no runtime-only data. Regression coverage ------------------- ``tests/tools/test_delegate.py`` gets a new ``TestLoadConfigFreshReads`` class with 5 tests: * ``test_ignores_stale_cli_config_cache`` — plants a stale CLI_CONFIG with empty-default delegation in ``sys.modules["cli"]`` and asserts ``_load_config()`` still reads the fresh user override from ``config.yaml``. On unpatched ``origin/main`` this fails with the exact reporter symptom: ``AssertionError: '' != 'claude-sonnet-4-20250514'``. * ``test_picks_up_delegation_override_written_after_startup`` — end-user scenario: edit config.yaml, call _load_config, get the edit. * ``test_non_dict_delegation_section_is_coerced_to_empty`` — hardens ``.get()`` callers downstream; also fails on ``origin/main`` (the old code returned the non-dict string verbatim). * ``test_returns_dict_when_config_missing_delegation_section`` — pins graceful handling when no delegation section is configured. * ``test_end_to_end_resolve_uses_fresh_config`` — integration through ``_resolve_delegation_credentials`` confirming the config flows all the way to ``effective_model = creds["model"] or parent.model``. Validation ---------- ``source venv/bin/activate && python -m pytest tests/tools/test_delegate.py tests/tools/test_delegate_toolset_scope.py -q`` -> 77 passed. CI-aligned broad suite: ``python -m pytest tests/ -q --ignore=tests/integration --ignore=tests/e2e --tb=line -n auto`` -> 12824 passed, 39 skipped, 15 pre-existing baseline failures (all reproduce on clean ``origin/main``; none in ``tools/delegate_tool.py`` or ``tests/tools/test_delegate.py``). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Copilot

Pull request overview

Fixes delegation config staleness so delegate_task picks up delegation.model / delegation.provider edits from ~/.hermes/config.yaml during long-lived CLI sessions.

Changes:

Remove the cli.CLI_CONFIG short-circuit and always read delegation config via hermes_cli.config.load_config().
Harden _load_config() against malformed non-dict delegation: values by coercing to {}.
Add regression tests covering fresh reads, stale CLI cache behavior, and malformed delegation sections.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
tools/delegate_tool.py	Always loads delegation config fresh from disk and type-checks the `delegation` section.
tests/tools/test_delegate.py	Adds regression tests for #11999 covering config freshness and malformed config handling.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-04-18T07:47:42Z

+    the built-in defaults (empty-string ``model``/``provider`` and a
+    ``max_iterations`` float): those produced a truthy-but-empty dict


The docstring claims the defaults include a "max_iterations" float, but DEFAULT_CONFIG defines delegation.max_iterations as an int (e.g., hermes_cli/config.py:676). This is misleading documentation—please update the wording/type to match the actual config schema.

Suggested change

the built-in defaults (empty-string ``model``/``provider`` and a

``max_iterations`` float): those produced a truthy-but-empty dict

the built-in defaults (empty-string ``model``/``provider`` and an

integer ``max_iterations``): those produced a truthy-but-empty dict

Copilot · 2026-04-18T07:47:42Z

+            prev_home = os.environ.get("HERMES_HOME")
+            prev_key = os.environ.get("ANTHROPIC_API_KEY")
+            os.environ["HERMES_HOME"] = hhome
+            os.environ.setdefault("ANTHROPIC_API_KEY", "sk-test")
+            try:
+                # User adds override
+                with open(os.path.join(hhome, "config.yaml"), "w") as fh:
+                    import yaml as _yaml
+                    _yaml.safe_dump({
+                        "model": "claude-opus-4-6",
+                        "delegation": {
+                            "model": "claude-sonnet-4-20250514",
+                            "provider": "anthropic",
+                        },
+                    }, fh)
+                cfg = _load_config()
+                creds = _resolve_delegation_credentials(cfg, parent)
+                self.assertEqual(creds["model"], "claude-sonnet-4-20250514")
+                # and ``_build_child_agent`` will later do:
+                #    effective_model = creds["model"] or parent.model
+                # which must now land on the delegation model, not parent's.
+                effective = creds["model"] or parent.model
+                self.assertEqual(effective, "claude-sonnet-4-20250514")
+            finally:
+                if prev_home is None:
+                    os.environ.pop("HERMES_HOME", None)
+                else:
+                    os.environ["HERMES_HOME"] = prev_home
+                if prev_key is None:
+                    os.environ.pop("ANTHROPIC_API_KEY", None)
+                else:
+                    os.environ["ANTHROPIC_API_KEY"] = prev_key
+
+


test_end_to_end_resolve_uses_fresh_config currently depends on real provider resolution (_resolve_delegation_credentials -> hermes_cli.runtime_provider.resolve_runtime_provider) and uses os.environ.setdefault("ANTHROPIC_API_KEY", ...), which won’t override an existing-but-empty value. This makes the test environment-dependent/flaky. Consider patching hermes_cli.runtime_provider.resolve_runtime_provider (or agent.anthropic_adapter.resolve_anthropic_token) to return a fixed credential bundle, and/or force-set ANTHROPIC_API_KEY within a patch.dict so the test is hermetic.

Suggested change

prev_home = os.environ.get("HERMES_HOME")

prev_key = os.environ.get("ANTHROPIC_API_KEY")

os.environ["HERMES_HOME"] = hhome

os.environ.setdefault("ANTHROPIC_API_KEY", "sk-test")

try:

# User adds override

with open(os.path.join(hhome, "config.yaml"), "w") as fh:

import yaml as _yaml

_yaml.safe_dump({

"model": "claude-opus-4-6",

"delegation": {

"model": "claude-sonnet-4-20250514",

"provider": "anthropic",

},

}, fh)

cfg = _load_config()

creds = _resolve_delegation_credentials(cfg, parent)

self.assertEqual(creds["model"], "claude-sonnet-4-20250514")

# and ``_build_child_agent`` will later do:

# effective_model = creds["model"] or parent.model

# which must now land on the delegation model, not parent's.

effective = creds["model"] or parent.model

self.assertEqual(effective, "claude-sonnet-4-20250514")

finally:

if prev_home is None:

os.environ.pop("HERMES_HOME", None)

else:

os.environ["HERMES_HOME"] = prev_home

if prev_key is None:

os.environ.pop("ANTHROPIC_API_KEY", None)

else:

os.environ["ANTHROPIC_API_KEY"] = prev_key

with patch.dict(

os.environ,

{"HERMES_HOME": hhome, "ANTHROPIC_API_KEY": "sk-test"},

clear=False,

):

# Ensure credential lookup is deterministic and does not

# depend on any real provider/token resolution in the test

# environment.

with patch(

"agent.anthropic_adapter.resolve_anthropic_token",

return_value="sk-test",

):

# User adds override

with open(os.path.join(hhome, "config.yaml"), "w") as fh:

import yaml as _yaml

_yaml.safe_dump({

"model": "claude-opus-4-6",

"delegation": {

"model": "claude-sonnet-4-20250514",

"provider": "anthropic",

},

}, fh)

cfg = _load_config()

creds = _resolve_delegation_credentials(cfg, parent)

self.assertEqual(creds["model"], "claude-sonnet-4-20250514")

# and ``_build_child_agent`` will later do:

# effective_model = creds["model"] or parent.model

# which must now land on the delegation model, not parent's.

effective = creds["model"] or parent.model

self.assertEqual(effective, "claude-sonnet-4-20250514")

…d-to-end test Addresses both Copilot inline review points on PR NousResearch#12053: 1. Docstring typo fix. The docstring said ``max_iterations`` was a ``float`` in the built-in defaults. It's an ``int`` — ``DEFAULT_CONFIG["delegation"]["max_iterations"] = 50``. Reworded the sentence and dropped the misleading type qualifier. 2. ``test_end_to_end_resolve_uses_fresh_config`` no longer depends on real provider resolution. Replaced the ad-hoc ``os.environ.setdefault("ANTHROPIC_API_KEY", ...)`` (which doesn't override an existing empty value and still calls into the real ``resolve_runtime_provider`` pipeline) with: * ``patch.dict(os.environ, ...)`` to force-set ``ANTHROPIC_API_KEY`` and ``HERMES_HOME`` atomically inside the block, and * ``patch(hermes_cli.runtime_provider.resolve_runtime_provider)`` returning a fixed credential bundle, so the test is hermetic and its outcome only depends on the code path under test — ``_load_config`` → ``_resolve_delegation_credentials`` → ``effective_model`` selection. Validation: ``source venv/bin/activate && python -m pytest tests/tools/test_delegate.py::TestLoadConfigFreshReads -q`` -> 5 passed. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

briandevans · 2026-04-18T07:55:44Z

Addressed both Copilot review points in a7bb79c4:

1. Docstring type fix — reworded the sentence so the description of the built-in-defaults dict no longer labels max_iterations as a float. It's an int (DEFAULT_CONFIG["delegation"]["max_iterations"] = 50 in hermes_cli/config.py:676).

2. End-to-end test hermeticity — replaced the environment-dependent os.environ.setdefault("ANTHROPIC_API_KEY", "sk-test") + real resolve_runtime_provider() call with:

patch.dict(os.environ, {"HERMES_HOME": …, "ANTHROPIC_API_KEY": "sk-test"}, clear=False) so the env vars are force-set (not just set-if-missing) for the duration of the test, and
patch("hermes_cli.runtime_provider.resolve_runtime_provider", return_value={…fixed bundle…}) so the test outcome depends only on the code path under test (_load_config → _resolve_delegation_credentials → effective_model), not on whatever auth state happens to be resident in the runtime-provider pipeline.

Validation: python -m pytest tests/tools/test_delegate.py::TestLoadConfigFreshReads -q → 5 passed.

The unrelated test CI failure in the previous run (tests/tools/test_accretion_caps.py::TestReadTrackerCaps::test_live_cap_applied_after_read_add - KeyError: 'read_timestamps') is a baseline flake — passes in isolation both on this branch and on clean origin/main:

$ python -m pytest tests/tools/test_accretion_caps.py::TestReadTrackerCaps::test_live_cap_applied_after_read_add -v
1 passed in 10.56s

That test exercises tools/file_tools.py read-tracker bookkeeping under concurrent state mutations and is unrelated to this PR's tools/delegate_tool.py change.

…xplicit provider When a subagent is configured with a custom base_url but no explicit provider, it should use 'custom' as the provider rather than inheriting the parent's provider. This fixes credential resolution for subagents using custom endpoints like NVIDIA NIM. Also fixes _load_config() to always read fresh from config.yaml instead of using the stale CLI_CONFIG cache. This ensures delegation config changes are picked up without restarting the CLI. Fixes subagent credential resolution when using custom base_url without explicit provider (complements PR NousResearch#12053).

briandevans · 2026-05-13T14:08:18Z

Closing to keep the queue clean — happy to reopen if this is still useful.

Copilot AI review requested due to automatic review settings April 18, 2026 07:44

Copilot started reviewing on behalf of briandevans April 18, 2026 07:45 View session

Copilot AI reviewed Apr 18, 2026

View reviewed changes

briandevans mentioned this pull request Apr 19, 2026

fix(models): validate Alibaba coding-endpoint switches via catalog when /models 404s (#12272) #12287

Closed

alt-glitch mentioned this pull request Apr 21, 2026

fix(delegation): _load_config falls through to hermes_cli config when CLI_CONFIG has no delegation overrides (#11999) #12105

Closed

alt-glitch added type/bug Something isn't working P2 Medium — degraded but workaround exists tool/delegate Subagent delegation comp/agent Core agent loop, run_agent.py, prompt builder area/config Config system, migrations, profiles labels Apr 24, 2026

alt-glitch mentioned this pull request Apr 25, 2026

fix(delegate): _load_config reads from disk every call, not cached CLI_CONFIG #15540

Open

briandevans closed this May 13, 2026

This was referenced May 22, 2026

fix: delegation config reads from disk instead of frozen CLI_CONFIG snapshot #30581

Open

_load_config() in delegate_tool.py never reads persistent config — breaks all delegation.* settings in v0.14.0 #32671

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fix(delegate): read fresh config.yaml so delegation.model edits take effect (#11999)#12053

fix(delegate): read fresh config.yaml so delegation.model edits take effect (#11999)#12053
briandevans wants to merge 2 commits into
NousResearch:mainfrom
briandevans:fix/delegate-task-model-override

briandevans commented Apr 18, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Apr 18, 2026

Uh oh!

Copilot AI Apr 18, 2026

Uh oh!

briandevans commented Apr 18, 2026

Uh oh!

briandevans commented May 13, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

		the built-in defaults (empty-string ``model``/``provider`` and a
		``max_iterations`` float): those produced a truthy-but-empty dict

Conversation

briandevans commented Apr 18, 2026

What does this PR do?

Root cause

Fix

Precedence / compat table

Narrow scope — explicitly not changed

Pre-empt reviewer Q&A

Regression coverage

How to test

Related Issue

Type of Change

Changes Made

Adjacent surfaces checked

Checklist

Code

Documentation & Housekeeping

Notes for reviewers

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Apr 18, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 18, 2026

Choose a reason for hiding this comment

Uh oh!

briandevans commented Apr 18, 2026

Uh oh!

briandevans commented May 13, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants