Repository navigation
Description
Bug Description
This is likely an edge case rather than a common plain-gateway startup bug. I am filing it because the repro uses only upstream modules and points to a cache invalidation boundary: tool availability can depend on environment variables, but the quiet-mode tool-definition cache does not include those env values.
Specifically, the cronjob tool can remain hidden if tool definitions are inspected before gateway.run sets HERMES_EXEC_ASK=1.
I hit this in a downstream setup that performs tool-visibility/status checks around the gateway. A plain hermes gateway run startup may not hit this ordering if nothing probes tool definitions before the gateway module finishes initializing.
However, the minimal repro below uses only upstream modules. The issue is not specific to any private tool or custom configuration.
Steps to Reproduce
- In a clean detached
v2026.5.28worktree, run this script:
import os
for key in ("HERMES_GATEWAY_SESSION", "HERMES_EXEC_ASK", "HERMES_INTERACTIVE"):
os.environ.pop(key, None)
from model_tools import get_tool_definitions
pre = [
tool["function"]["name"]
for tool in get_tool_definitions(
enabled_toolsets=["cronjob"],
quiet_mode=True,
)
]
import gateway.run
post = [
tool["function"]["name"]
for tool in get_tool_definitions(
enabled_toolsets=["cronjob"],
quiet_mode=True,
)
]
print({
"pre": pre,
"exec_ask": os.getenv("HERMES_EXEC_ASK"),
"post": post,
})- Observe that importing gateway.run sets HERMES_EXEC_ASK=1.
- Observe that cronjob is still absent from the second tool-definition lookup.
- I also reproduced the same ordering on current origin/main (b1a2540) using skip_tool_search_assembly=True.
Expected Behavior
After gateway.run marks the process as a messaging gateway approval context via HERMES_EXEC_ASK=1, cronjob should become visible, or any stale pre-gateway tool availability cache should be invalidated.
Actual Behavior
cronjob remains hidden because the earlier quiet_mode=True tool-definition result is reused.
Observed output from a clean detached v2026.5.28 worktree:
{'pre': [], 'exec_ask': '1', 'post': []}Observed output from a clean detached current origin/main worktree:
{'pre': [], 'exec_ask': '1', 'post': []}Affected Component
Tools (terminal, file ops, web, code execution, etc.)
Messaging Platform (if gateway-related)
No response
Debug Report
Not included. I did not run `hermes debug share` because the repro is a minimal import/cache-order script using upstream modules, and my live deployment contains private/customized configuration.Operating System
macOS 26.3
Python Version
Python 3.11.15
Hermes Version
v2026.5.28; also reproduced on origin/main b1a2540
Additional Logs / Traceback (optional)
No traceback. Minimal repro output:
{'pre': [], 'exec_ask': '1', 'post': []}Root Cause Analysis (optional)
tools.cronjob_tools.check_cronjob_requirements() enables the tool only when one of these env vars is truthy:
HERMES_INTERACTIVEHERMES_GATEWAY_SESSIONHERMES_EXEC_ASK
model_tools.get_tool_definitions(..., quiet_mode=True) memoizes the filtered tool list.
In gateway/run.py, HERMES_EXEC_ASK=1 is set during module initialization, but not before every possible prior import/probe of model_tools. If another path in the same process calls get_tool_definitions() first, cronjob can be cached as unavailable. Importing gateway.run later sets HERMES_EXEC_ASK=1, but the stale cached tool list is still reused.
Proposed Fix (optional)
A few possible approaches:
- Set the regular gateway env markers (
HERMES_QUIET,HERMES_EXEC_ASK) earlier, before any imports/config validation paths can probe tool availability, and invalidate already-loaded tool/check caches after setting them. - Include the relevant tool-availability env fingerprint in the
get_tool_definitionscache key, so env changes that affect check functions cannot reuse stale results. - Expose a small explicit cache invalidation call for gateway startup after it sets process-level tool context.
Are you willing to submit a PR for this?
- I'd like to fix this myself and submit a PR
Activity
alt-glitch commentedon May 30, 2026
Related: #33423 (fix: invalidate tool caches on dotenv reload — same
_tool_defs_cachecode path, different trigger), #22349 (quiet-mode cache not scoped by sender identity — closed).This issue identifies a new invalidation boundary: env-var-gated tool availability (
HERMES_EXEC_ASK) is not reflected in the quiet-mode tool-definition cache after the env changes mid-process.bbopen commentedon Jul 5, 2026
Opened draft PR #58635 for this: #58635
The patch covers two repro paths from this issue class:
cronjobas absent, thenHERMES_EXEC_ASK=1appears andcronjobbecomes visiblecronjobas absent, then a Matrix-style gateway ContextVar session appears andcronjobbecomes visibleIt also keeps the fix narrower than broad dotenv/runtime env reload invalidation by keying availability caches on explicit gateway/session context plus the existing explicit invalidation path.
GottZ commentedon Jul 28, 2026
Summary
Two PRs address #35561’s stale quiet-mode tool-availability cache after gateway/session context changes. #35605 covers the reported process-environment flip but misses current task-local ContextVar state and outer-cache invalidation boundaries, while #58635 coordinates both cache layers across environment, ContextVar, profile, TTL, and explicit-invalidation transitions.
Related pull requests
partial— (+62/-3) — partial fix: Tracks process-level availability environment variables and adds the focused HERMES_EXEC_ASK regression, but it does not cover the current ContextVar gateway-session path or fully coordinate registry and outer schema-cache invalidation. Consistent with the contributor keep_open review, its regression should be salvaged into fix(tools): scope availability cache by session context #58635 rather than merging this branch as-is; the author also accepted that disposition.best fix— (+486/-35) — fixes: Keys registry availability by the relevant profile/session context, propagates availability identity and expiry into the outer schema cache, advances an epoch on explicit invalidation, and adds env-flip, ContextVar-transition, TTL/grace, profile-isolation, and explicit-invalidation regressions. The contributor keep_open review rates this approach highly salvageable and confirms that the diff directly covers both stale-cache boundaries.Duplicates
#35605 and #58635 target the same HERMES_EXEC_ASK stale-cache cause, but #58635 supersedes #35605 by also covering task-local ContextVar sessions and coordinated registry/outer-cache invalidation.
Suggested consolidation
Merge #58635 as the comprehensive fix for #35561; close #35605 as superseded after retaining its focused HERMES_EXEC_ASK regression coverage. This follows the contributor review on #35605, which explicitly recommends consolidation into #58635, while the keep_open review on #58635 confirms that its diff addresses the relevant cache layers and current ContextVar path.
Complex graph
flowchart LR classDef open fill:#dbeafe,stroke:#1d4ed8,color:#1e3a8a classDef merged fill:#dcfce7,stroke:#15803d,color:#14532d classDef closed fill:#e5e7eb,stroke:#6b7280,color:#1f2937 classDef unverified fill:#f3f4f6,stroke:#9ca3af,color:#374151 classDef best stroke-width:3px,stroke:#b45309 classDef target stroke-width:3px,stroke:#4338ca I35561(["issue #35561 (open)"]) subgraph Dup35605 ["PRs duplicating each other"] P35605["PR #35605 (open)"] P58635["PR #58635 (open)"] end P35605 -.->|partial| I35561 P58635 -->|best fix| I35561 class I35561 open class I35561 target class P35605 open class P58635 open class P58635 best click I35561 "https://github.com/NousResearch/hermes-agent/issues/35561" click P35605 "https://github.com/NousResearch/hermes-agent/pull/35605" click P58635 "https://github.com/NousResearch/hermes-agent/pull/58635"Graph: solid arrow = fixes / best fix, dashed arrow = partial or unverified (see edge label); boxed group = PRs duplicating each other; amber border = best fix; indigo border = target; gray node = closed or no verify verdict yet (state tag in the node label).
Cross-PR triage: Reviewed 2 pull requests and 1 issue in this complex. Each diff was read against this issue; Assessment working set: 37 kB of PR diffs, 11 kB of issue/PR text, 7 kB of discussion (12 comments), 4 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.
GottZ commentedon Jul 30, 2026
Delta since our previous triage comment
Since the previous comment, @kiranmagic7 closed #35605 as stale WIP and confirmed that its process-environment fingerprint is not suitable for the current availability-context and outer-cache-epoch architecture. This reinforces rather than changes the earlier partial-fix assessment; #58635 remains the best existing fix.
Changed pull requests
partial— (+62/-3) — partial fix, now closed as stale WIP: Despite the earlier keep_open review on fix(tools): refresh cache when session env changes #35605, the diff only fingerprints process environment values and does not cover the current ContextVar/session and coordinated outer-cache boundaries; @kiranmagic7 explicitly accepted that limitation and closed the branch rather than landing it as-is. It remains relevant for its focused HERMES_EXEC_ASK regression and as the narrower implementation superseded by fix(tools): scope availability cache by session context #58635.Suggested consolidation
Recommendation updated for the current backlog lane: keep #58635 open, consistent with its keep_open review, as the best existing fix pending maintainer/author activation rather than merge it now; #35605 is already closed as superseded.
Complex graph
flowchart LR classDef open fill:#dbeafe,stroke:#1d4ed8,color:#1e3a8a classDef merged fill:#dcfce7,stroke:#15803d,color:#14532d classDef closed fill:#e5e7eb,stroke:#6b7280,color:#1f2937 classDef unverified fill:#f3f4f6,stroke:#9ca3af,color:#374151 classDef best stroke-width:3px,stroke:#b45309 classDef target stroke-width:3px,stroke:#4338ca I35561(["issue #35561 (open)"]) subgraph Dup35605 ["PRs duplicating each other"] P35605["PR #35605 (closed)"] P58635["PR #58635 (open)"] end P35605 -.->|partial| I35561 P58635 -->|best fix| I35561 class I35561 open class I35561 target class P35605 closed class P58635 open class P58635 best click I35561 "https://github.com/NousResearch/hermes-agent/issues/35561" click P35605 "https://github.com/NousResearch/hermes-agent/pull/35605" click P58635 "https://github.com/NousResearch/hermes-agent/pull/58635" classDef changedSince stroke:#dc2626,stroke-width:3px,stroke-dasharray: 6 3 class P35605 changedSinceGraph: solid arrow = fixes / best fix, dashed arrow = partial or unverified (see edge label); boxed group = PRs duplicating each other; amber border = best fix; indigo border = target; gray node = closed (state tag in the node label).
Dashed red border = changed since our previous triage comment (1 added/updated, 1 removed graph elements).
Cross-PR triage: Reviewed 2 pull requests and 1 issue in this complex. Each diff was read against this issue; Assessment working set: 37 kB of PR diffs, 11 kB of issue/PR text, 6 kB of discussion (11 comments), 3 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.