perf: use load_config_readonly() at read-only call sites in agent/ by Stoltemberg · Pull Request #56085 · NousResearch/hermes-agent

Stoltemberg · 2026-07-01T05:42:21Z

What does this PR do?

load_config() performs a defensive deepcopy (~135μs per call) of the cached config dict. The codebase has load_config_readonly() specifically for read-only callers, but most call sites in agent/ still used load_config().

Replaced load_config() → load_config_readonly() at 31+ read-only call sites across 20 files in agent/. Each saves ~135μs per call. In the agent loop, config is read 20-50x per conversation.

Type of Change

⚡ Performance improvement

Changes Made

20 files in agent/ — all confirmed read-only (only .get() lookups, no mutation):

agent_init.py (4 aliased imports), auxiliary_client.py (11 calls), curator.py (2), prompt_builder.py (2), and 16 other files (1 each)

How to Test

uv run --extra dev ruff check agent/
uv run --extra dev python -m pytest tests/test_project_metadata.py tests/test_packaging_metadata.py -q

Checklist

All call sites confirmed read-only
ruff check passes
19 tests pass
No behavior change — same config values, just no deepcopy

teknium1

Thanks for identifying the cached-config deepcopy cost; the performance premise still holds on current main.

Problems

Blocking: agent/agent_runtime_helpers.py:1722 changes _sm_cfg to load_config_readonly() and passes it to get_compatible_custom_providers(). Current hermes_cli/config.py:4699-4700 and :4714-4722 mutate provider-entry dictionaries in place, while load_config_readonly() explicitly forbids nested mutation (hermes_cli/config.py:6766-6781). This can corrupt the shared config cache for camelCase or api_key_env provider configurations.
The cleanup is now incomplete: current read-only-only paths remain in agent/auxiliary_client.py:2175-2181, :2196-2202, and agent/coding_context.py:369-376.

Suggested changes

First make the provider normalizer operate on a copied entry (and add a regression test that alias normalization does not mutate its input), then re-audit the remaining current-HEAD call sites during conflict resolution.

Automated hermes-sweeper review.

teknium1 · 2026-07-15T11:52:28Z

-            from hermes_cli.config import load_config, get_compatible_custom_providers
-            _sm_cfg = load_config()
+            from hermes_cli.config import load_config_readonly, get_compatible_custom_providers
+            _sm_cfg = load_config_readonly()


load_config_readonly() returns the shared cached dict, but this value is passed to get_compatible_custom_providers(). On current main, its normalizer writes alias keys into provider-entry dicts (hermes_cli/config.py:4699-4700, :4714-4722), so this path can corrupt the config cache. Please make that normalizer/caller non-mutating before using the readonly loader here.

@Stoltemberg

Salvaged from #56085 (@Stoltemberg), rebased onto current main: sites main had already converted (credential_pool, auxiliary_client MoA paths, model_metadata, moa_loop, agent_runtime_helpers) resolve to main's versions; the remaining ~29 read-only sites across 16 agent/ files swap to the no-deepcopy readonly loader (~135us saved per call). Full per-site mutation audit performed (every enclosing function read, escapes traced): 23 SAFE, 5 ESCAPES with read-only consumers, 1 UNSAFE path (init_agent -> get_compatible_custom_providers -> normalizer in-place alias writes) fixed by the preceding no-mutate commits, which make the normalizer copy-safe for ALL callers.

teknium1 · 2026-07-29T22:28:33Z

Thanks @Stoltemberg — this landed on main via #74322 with your commit preserved as author. Salvage notes: rebased onto current main (several of your sites had been independently converted in the meantime and resolve to main's versions; ~29 survived), and a full per-site mutation audit found one path where the swap was unsafe — init_agent → get_compatible_custom_providers → _normalize_custom_provider_entry mutated cached provider sub-dicts in place. That normalizer bug was fixed first (via @golldyck's #57096) so your swaps are safe everywhere. Measured 332µs → 11.9µs per read (28x) on a real config. Second salvaged PR of yours today (#55176 → #74194) — appreciate the consistent, well-aimed perf work!

The terminal-command guard path loaded config 2-3x per invocation via load_config(), which pays a defensive deepcopy of the entire config on every call (~356us of the ~376us warm-cache cost measured on a real config.yaml). All six swapped call sites were audited read-only — every caller takes scalar reads or iterates the returned structures; none mutate (the save path at save_permanent_allowlist keeps load_config) — so they now use load_config_readonly(), the API built for exactly this (precedent: NousResearch#74211, NousResearch#74322; the one unsafe-site lesson from NousResearch#56085's salvage is covered by the mutation audit and a cache-integrity test). Measured (real config.yaml, warm cache): load_config 376.0us -> load_config_readonly 19.9us (18.9x); full guard pass check_all_command_guards('ls -la','local') 930.7us -> 241.8us (3.85x). Tests: new test_approval_config_readonly.py drives the real functions against a temp HERMES_HOME — readonly call counts per function, a no-deepcopy pin for the full guard pass, and cache-identity/integrity checks. Existing test mocks retargeted from load_config to load_config_readonly (same injection intent). Note: 6 test_approval_mode_parity failures are pre-existing ordering flakes — identical with the change stashed on clean main.

The terminal-command guard path loaded config 2-3x per invocation via load_config(), which pays a defensive deepcopy of the entire config on every call (~356us of the ~376us warm-cache cost measured on a real config.yaml). All six swapped call sites were audited read-only — every caller takes scalar reads or iterates the returned structures; none mutate (the save path at save_permanent_allowlist keeps load_config) — so they now use load_config_readonly(), the API built for exactly this (precedent: #74211, #74322; the one unsafe-site lesson from #56085's salvage is covered by the mutation audit and a cache-integrity test). Measured (real config.yaml, warm cache): load_config 376.0us -> load_config_readonly 19.9us (18.9x); full guard pass check_all_command_guards('ls -la','local') 930.7us -> 241.8us (3.85x). Tests: new test_approval_config_readonly.py drives the real functions against a temp HERMES_HOME — readonly call counts per function, a no-deepcopy pin for the full guard pass, and cache-identity/integrity checks. Existing test mocks retargeted from load_config to load_config_readonly (same injection intent). Note: 6 test_approval_mode_parity failures are pre-existing ordering flakes — identical with the change stashed on clean main.

@Stoltemberg

Salvaged from NousResearch#56085 (@Stoltemberg), rebased onto current main: sites main had already converted (credential_pool, auxiliary_client MoA paths, model_metadata, moa_loop, agent_runtime_helpers) resolve to main's versions; the remaining ~29 read-only sites across 16 agent/ files swap to the no-deepcopy readonly loader (~135us saved per call). Full per-site mutation audit performed (every enclosing function read, escapes traced): 23 SAFE, 5 ESCAPES with read-only consumers, 1 UNSAFE path (init_agent -> get_compatible_custom_providers -> normalizer in-place alias writes) fixed by the preceding no-mutate commits, which make the normalizer copy-safe for ALL callers.

The terminal-command guard path loaded config 2-3x per invocation via load_config(), which pays a defensive deepcopy of the entire config on every call (~356us of the ~376us warm-cache cost measured on a real config.yaml). All six swapped call sites were audited read-only — every caller takes scalar reads or iterates the returned structures; none mutate (the save path at save_permanent_allowlist keeps load_config) — so they now use load_config_readonly(), the API built for exactly this (precedent: NousResearch#74211, NousResearch#74322; the one unsafe-site lesson from NousResearch#56085's salvage is covered by the mutation audit and a cache-integrity test). Measured (real config.yaml, warm cache): load_config 376.0us -> load_config_readonly 19.9us (18.9x); full guard pass check_all_command_guards('ls -la','local') 930.7us -> 241.8us (3.85x). Tests: new test_approval_config_readonly.py drives the real functions against a temp HERMES_HOME — readonly call counts per function, a no-deepcopy pin for the full guard pass, and cache-identity/integrity checks. Existing test mocks retargeted from load_config to load_config_readonly (same injection intent). Note: 6 test_approval_mode_parity failures are pre-existing ordering flakes — identical with the change stashed on clean main.

@Stoltemberg

Salvaged from NousResearch#56085 (@Stoltemberg), rebased onto current main: sites main had already converted (credential_pool, auxiliary_client MoA paths, model_metadata, moa_loop, agent_runtime_helpers) resolve to main's versions; the remaining ~29 read-only sites across 16 agent/ files swap to the no-deepcopy readonly loader (~135us saved per call). Full per-site mutation audit performed (every enclosing function read, escapes traced): 23 SAFE, 5 ESCAPES with read-only consumers, 1 UNSAFE path (init_agent -> get_compatible_custom_providers -> normalizer in-place alias writes) fixed by the preceding no-mutate commits, which make the normalizer copy-safe for ALL callers.

perf: use load_config_readonly() at read-only call sites in agent/

1964040

alt-glitch added type/perf Performance improvement or optimization comp/agent Core agent runtime: loop, agent_init, prompt builder, context-compression, responses endpoint P3 Low — cosmetic, nice to have labels Jul 1, 2026

hashbender mentioned this pull request Jul 1, 2026

perf: use load_config_readonly() at read-only call sites in agent/ hashbender/hermes-agent#166

Open

5 tasks

golldyck mentioned this pull request Jul 2, 2026

fix(config): stop _normalize_custom_provider_entry mutating the caller's dict #57096

Closed

8 tasks

hashbender mentioned this pull request Jul 2, 2026

fix(config): stop _normalize_custom_provider_entry mutating the caller's dict hashbender/hermes-agent#881

Open

8 tasks

teknium1 reviewed Jul 15, 2026

View reviewed changes

teknium1 added sweeper:risk-compatibility Sweeper risk: may break existing users, config, migrations, defaults, or upgrades sweeper:blast-broad Sweeper blast radius: broad — a core path most sessions hit labels Jul 15, 2026

teknium1 mentioned this pull request Jul 29, 2026

perf(agent): readonly config loader at 29 call sites, 28x per read — mutation-audited (salvage #56085 + #57096) #74322

Merged

teknium1 closed this Jul 29, 2026

github-actions Bot mentioned this pull request Jul 29, 2026

🦞 OpenClaw 生态日报 2026-07-30 junlinzhao327-oss/big_model_radar#195

Open

This was referenced Aug 1, 2026

perf(tools): use load_config_readonly on the approval guard path spfcraze/hermes-agent#22

Open

perf(tools): use load_config_readonly on the approval guard path #76194

Closed

Navigation Menu

Uh oh!

perf: use load_config_readonly() at read-only call sites in agent/ - #56085

perf: use load_config_readonly() at read-only call sites in agent/#56085
Stoltemberg wants to merge 1 commit into
NousResearch:mainfrom
Stoltemberg:perf/use-readonly-config

Stoltemberg commented Jul 1, 2026

Uh oh!

teknium1 left a comment

Uh oh!

teknium1 Jul 15, 2026

Uh oh!

teknium1 commented Jul 29, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Uh oh!

Conversation

Stoltemberg commented Jul 1, 2026

What does this PR do?

Type of Change

Changes Made

How to Test

Checklist

Uh oh!

teknium1 left a comment

Choose a reason for hiding this comment

Problems

Suggested changes

Uh oh!

teknium1 Jul 15, 2026

Choose a reason for hiding this comment

Uh oh!

teknium1 commented Jul 29, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants