response_store.db (api server) holds conversation history including tool
payloads, prompts, and results. webhook_subscriptions.json holds per-route
HMAC secrets. Under a permissive umask (e.g. 0o022, default on most
distros) both files were created mode 0o644 — readable by other local
users on shared boxes.
- gateway/platforms/api_server.py: ResponseStore tightens itself + WAL/SHM
sidecars to 0o600 after __init__, then trusts the inode. (Original
contributor patch chmod'd after every _commit() — wasteful on a hot
api_server path; chmod-on-create is sufficient since SQLite preserves
mode bits across writes.)
- hermes_cli/webhook.py: _save_subscriptions writes via tempfile.mkstemp
(which itself creates the file with 0o600), chmods the temp before the
atomic rename, and re-asserts 0o600 on the destination so an existing
permissive file from before this fix gets narrowed.
Tests cover (a) creation under permissive umask leaves 0o600 and (b) an
existing 0o644 webhook_subscriptions.json gets narrowed on next save.
Tests guarded with skipif os.name=='nt' since POSIX mode bits don't apply
on Windows.
Salvaged from PR #30917 by @Hinotoi-agent. Reworked the api_server.py
side from chmod-on-every-commit to chmod-on-create.
Co-authored-by: teknium1 <127238744+teknium1@users.noreply.github.com>
When FEISHU_VERIFICATION_TOKEN is configured, an unauthenticated remote
could previously prove endpoint control by sending a url_verification
payload with any attacker-controlled challenge string — the handler
reflected the challenge BEFORE running the token check.
Move the verification_token check ahead of the url_verification echo so
the challenge response is gated on a valid token. Add a regression test
covering the wrong-token case. Also fix the stale
test_connect_webhook_mode_starts_local_server fixture to set
FEISHU_VERIFICATION_TOKEN (post #30746 webhook mode requires a secret).
Salvaged from PR #29663 by @m0n3r0 — kept the url_verification reorder
and its regression test; dropped the host-conditional weakening of the
#30746 secret guard (we want webhook secrets required regardless of
bind host, not only on 0.0.0.0/::).
Docs updated to call out the gating.
Co-authored-by: teknium1 <127238744+teknium1@users.noreply.github.com>
Operator misconfiguration is a client/setup error, not an internal server
exception. 403 "forbidden" more accurately reflects "this route refuses
to authenticate" than 500 "internal server error" — the latter triggers
incident alerting on operator monitoring and conflates real bugs with
config drift.
Follow-up tweak to PR #29629 by @m0n3r0.
Reject unsigned webhook requests when a route has no effective HMAC secret, even if the request handler is reached without the normal connect-time validation. Add regression coverage for the direct-handler path.
When the 'mcp' Python SDK isn't installed, _run_stdio leaked a bare
'NameError: name StdioServerParameters is not defined' because the
top-level 'from mcp import ...' fails inside try/except ImportError,
leaving the names unbound at module scope.
Mirror the _MCP_HTTP_AVAILABLE gate that _run_http already had: raise
a clear ImportError with install instructions instead.
Fixes#30904
Three test classes lock in the #30963 fix:
1. TestPartialStreamStubFinishReason — drives _interruptible_streaming_api_call
through the two recovery branches and asserts:
- text-only partial → finish_reason="length" (the new behaviour),
- mid-tool-call partial → finish_reason="stop" (unchanged on purpose).
2. TestLengthContinuationPromptBranching — pure-Python check on the branch
that picks the continuation prompt by response.id. Locks the network
error wording for partial-stream-stub vs. the output-length wording
for everything else.
3. TestConversationLoopPartialStreamContinuation — feeds a stub +
continuation pair into run_conversation, verifies the loop makes a
second API call (instead of exiting with text_response(stop)),
confirms the network-error continuation prompt actually reaches the
model on call #2, and that final_response stitches both halves.
Refs: NousResearch/hermes-agent#30963
The length-continue path's user-facing vprint and continuation prompt
both told the model "your response was truncated by the output length
limit." That's a lie when the stub came from a partial-stream network
error (issue #30963) — and a lie the model can detect, leading to "I
wasn't truncated, I'm done" no-op responses that defeat the
continuation entirely.
Detect the partial-stream-stub via response.id and swap in:
- vprint: "Stream interrupted by network error
(finish_reason='length' on partial-stream-stub)"
- prompt: "[System: The previous response was cut off by a network
error mid-stream. Continue exactly where you left off.
Do not restart or repeat prior text. Finish the answer
directly.]"
Real length truncations still see the original "truncated by output
length limit" prompt — the model needs to know which class of failure
it's recovering from. Same length_continue_retries=3 budget,
truncated_response_parts merging, and final-response stitching
infrastructure on both branches.
Refs: NousResearch/hermes-agent#30963
When the API connection drops mid-stream after text deltas have already
been delivered, chat_completion_helpers returned a stub response with
finish_reason=stop. The conversation loop then classified the stub as a
clean text completion (text_response(finish_reason=stop)) and exited
with iteration budget remaining — even when the goal-judge verdict
came back as "continue" milliseconds later (issue #30963).
Switch the text-only partial-stream stub to finish_reason=length. The
existing length-continuation path (length_continue_retries up to 3,
"continue exactly where you left off" prompt, partial parts merged
into final_response) then fires automatically: the partial assistant
content is persisted, the model is asked to continue from the cut
point, and the loop keeps making progress against the goal.
The mid-tool-call branch keeps finish_reason=stop on purpose — its
user-facing warning ("Ask me to retry if you want to continue") asks
the user to drive the retry rather than auto-replaying a tool call
with possible side effects.
#5544's "no duplicate message" contract is preserved verbatim: the
partial content is reused, never re-emitted as a fresh API call, so
the user never sees two copies of the same delta.
Refs: NousResearch/hermes-agent#30963
PR #29119 dropped the 'not streamed_message' guard unconditionally so
that plugin-transformed responses (transform_llm_output hook) would
reach ACP clients. That regressed test_prompt_does_not_duplicate_streamed_final_message:
when no transform happened, the streamed text was re-sent as a duplicate
final delivery.
Tighten the condition to mirror the gateway side: deliver after streaming
only when response_transformed=True. Otherwise keep the old guard.
Adds test_prompt_delivers_transformed_response_after_streaming so the
transformed path stays covered.
Adds a test that fails without the gateway fix, exercising the
response_transformed=True branch in _finalize_response: a streamed
response whose final text was modified by a transform_llm_output
plugin hook must be edit_message'd in place (not duplicate-sent),
with already_sent=True so the normal final-send is skipped.
Also drops two minor leftovers from the salvaged PR #29119:
* accumulated_text property on GatewayStreamConsumer (unused)
* duplicate _response_transformed=False inside the hook try block
When a transform_llm_output hook appends content after streaming, the previous
fix skipped the final-send suppression which caused the full response to be
sent as a NEW message (duplicate). Instead, edit the existing streamed message
in-place to append the transformed content, then set already_sent=True.
Added stream_consumer.message_id and .accumulated_text public properties.
run_sync() cherry-picks fields from the run_conversation result dict into
a new response dict for the gateway. response_transformed was missing from
the cherry-pick list, so the gateway always saw it as False and suppressed
the final send even though a transform_llm_output hook had modified the content.
When a transform_llm_output hook modifies final_response after streaming,
the gateway was silently discarding the transformed content because
streamed=True / content_delivered=True triggered the final-send
suppression. Three changes:
1. conversation_loop: set `_response_transformed=True` when a
transform_llm_output hook returns a non-empty string, and expose it
as `response_transformed` in the result dict.
2. gateway/run: skip the final-send suppression when
`response_transformed` is True — the transformed response must
reach the client even if streaming already sent the original text.
3. acp_adapter/server: remove `not streamed_message` guard so
final_response is always delivered (ACP path fixed separately).
When streaming is active, streamed_message=True skipped the final_response
update, causing plugin hooks like transform_llm_output to be silently
invisible. Remove the `not streamed_message` guard so the final response
(possibly transformed by plugins) is always delivered to the ACP client.
Closes#31370.
bws defaults to the US identity endpoint, so EU Cloud and self-hosted
machine-account tokens fail with [400 Bad Request] {"error":"invalid_client"}
during 'hermes secrets bitwarden setup'. The token is valid — it's just
being checked against the wrong region.
Add a Bitwarden region step to the wizard between the access-token and
project-list steps:
Step 1 Install bws
Step 2 Provide access token
Step 3 Pick region <-- new (US / EU / self-hosted-custom-URL)
Step 4 Pick project (now talks to the right endpoint)
Step 5 Test fetch
Region is stored in config.yaml as secrets.bitwarden.server_url and
plumbed into every bws subprocess as BWS_SERVER_URL (project list,
secret list, test fetch, and the env_loader startup pull).
Also:
- Non-interactive: 'hermes secrets bitwarden setup --server-url ...'
- Pre-existing BWS_SERVER_URL in the shell is detected and reused
- Cache key includes server_url so EU/US fetches don't collide
- 'hermes secrets bitwarden status' shows the configured region
- 'invalid_client' / '400 Bad Request' from bws now triggers a hint
pointing at the region setting instead of looking like a bad token
PR #6a1aa420e coupled `display.tool_progress: verbose` (a per-tool display
toggle for full args / results / think blocks) to `self.verbose` — which
controls root-logger DEBUG level. Result: setting tool_progress: verbose
in config silently flipped every module in the process to DEBUG and
flooded the terminal with internal logging, far beyond just full tool
calls.
The two concepts are separate:
- `tool_progress_mode == 'verbose'` → display behavior (tool rendering)
- `self.verbose` → logging behavior (root logger → DEBUG, line 9795)
This change keeps PR #6a1aa420e's argparse.SUPPRESS / config-fallback
plumbing but severs the verbose-display → debug-logging link.
Changes:
- cli.py:2868 — `self.verbose` only follows explicit `verbose=` arg; no
longer auto-True when tool_progress_mode == 'verbose'.
- cli.py:_toggle_verbose — slash-cycle through tool progress modes no
longer flips `self.verbose` / `agent.verbose_logging` / `agent.quiet_mode`.
- cli.py:9355 — fix misleading label (drop 'and debug logs').
- tui_gateway/server.py:_make_agent — same decoupling on the TUI side
(verbose_logging no longer derived from tool_progress_mode).
- tests/cli/test_tool_progress_scrollback.py — invert the test that
asserted the broken coupling; add coverage for explicit `--verbose`
still enabling DEBUG independent of tool_progress.
Live verified:
- tool_progress: verbose, no --verbose flag → 0 DEBUG/INFO log lines
- --verbose flag explicit → 32 DEBUG/INFO log lines (as expected)
When asyncio.sleep() fires just before Task.cancel() is called, CPython
sets _must_cancel=True but cannot cancel the already-completed sleep
future, so CancelledError is delivered at the next await (handle_message)
rather than at the sleep. By that point the superseded task has already
popped the merged event from _pending_text_batches, so the superseding
task sees an empty batch and silently drops the message.
Fix: add a synchronous task-registry check between the sleep and the pop.
No await between the check and the pop means no other coroutine can
interleave, so the guard is race-free.
When WeCom returns errcode=40001 (invalid credential) or 42001 (token
expired), send() was returning a failure without evicting the bad token
from _access_tokens. All subsequent sends then kept using the same
invalid cached token until its TTL naturally expired (~7200s).
Fix: on the first token-rejection errcode, evict the cache entry and
retry once with a freshly fetched token. Non-token errcodes fail
immediately as before. If the refreshed token also fails, the error
is returned without looping further.
Adds four regression tests covering: successful retry on 40001,
successful retry on 42001, no retry on unrelated errcode, and clean
failure when the refresh does not help.
* fix(profiles): cross-profile soft guard on file-write tools + system-prompt hint
Adds a soft guard so an agent running under one Hermes profile cannot
silently edit a different profile's skills/plugins/cron/memories.
Three layers:
A. agent/file_safety.classify_cross_profile_target
Classifies a write target against the active HERMES_HOME. Returns
a {active_profile, target_profile, area, target_path} dict when the
path lands in another profile's scoped area. PROFILE_SCOPED_AREAS =
(skills, plugins, cron, memories). get_cross_profile_warning()
wraps it into a model-facing error string that names both profiles,
names the area, and points at the cross_profile=True bypass.
Defense-in-depth, NOT a security boundary — the terminal tool runs
as the same OS user and can write any of these paths directly. The
guard exists to prevent confused-agent corruption, not to stop a
determined attacker. SECURITY.md §3.2 (terminal-bypass posture)
still applies.
Wired into tools/file_tools.write_file_tool and patch_tool with a
cross_profile=False kwarg. WRITE_FILE_SCHEMA and PATCH_SCHEMA both
advertise cross_profile so the model can pass it after explicit
user direction. patch_tool extracts target paths from V4A patch
bodies before checking (same shape as the existing sensitive-path
check).
skill_manage is already scoped to the active profile's SKILLS_DIR
by construction, so no extra guard wiring is needed there. The
D-side error message (below) still names other profiles when the
skill exists elsewhere.
B. agent/system_prompt
One deterministic line near the environment-hints block names the
active profile and tells the model not to modify another profile's
skills/plugins/cron/memories without explicit direction. Profile
name is stable for the lifetime of the AIAgent, so the line is
prompt-cache-safe.
D. tools/skill_manager_tool._skill_not_found_error
Replaces the bare "Skill 'X' not found." with a message that:
- names the active profile,
- searches OTHER profiles' skills dirs for the same name,
- names the profile(s) where the skill exists and the path,
- suggests `hermes -p <name>` to switch profiles, or
cross_profile=True for an explicit edit.
All 5 "not found" sites in skill_manager_tool (edit, patch, delete,
write_file, remove_file) now go through the helper.
Reference incident (May 2026): a hermes-security profile session
edited skills under both ~/.hermes/profiles/hermes-security/skills/
AND ~/.hermes/skills/ (the default profile's skills) without
realizing the second path belonged to a different profile. Three of
the four skill files needed manual restoration afterward.
What this PR does NOT do:
* No hard block. The terminal tool can still touch any of these
paths with no guard — same posture as the dangerous-command
approval flow. SECURITY.md §3.2 applies.
* No regex sweep on terminal commands for cross-profile paths.
That direction is a Skills-Guard-style arms race (cd + relative
paths, base64, etc.) and would false-positive on legitimate
cross-profile reads. Filed as a follow-up.
* No on-disk path migration. ~/.hermes/skills/ remains the
default profile's skills dir; this PR is about telling the
agent about that boundary, not changing the layout.
Tests:
tests/agent/test_file_safety_cross_profile.py (16 tests)
- _resolve_active_profile_name covers default/named/failure paths
- classify_cross_profile_target covers all four scoped areas,
both directions (default → named, named → default, named → named),
non-Hermes paths, and root-level config files
- get_cross_profile_warning covers in-profile no-op, cross-profile
message shape, and the defense-in-depth self-documentation
tests/tools/test_cross_profile_guard.py (12 tests)
- write_file: in-profile allow, cross-profile block, cross_profile=True
bypass, non-Hermes pass-through
- patch: replace-mode block, cross_profile=True bypass, V4A patch
path extraction
- skill_manage: error names the other profile (single + multiple),
missing-everywhere falls back to skills_list hint
- system prompt: contract-level checks (both branches present,
cross_profile=True mentioned, ~/.hermes/profiles/ referenced)
All 207 existing tests in file_safety/file_operations/skill_manager
still pass. 10 system-prompt tests still pass.
E2E verified: the exact incident scenario (security profile editing
default's hermes-agent-dev skill) is now blocked with the warning
message; cross_profile=True unblocks.
* fix(code_execution): add cross_profile to write_file/patch stubs
The cross_profile kwarg added to write_file_tool/patch_tool needs to
flow through the execute_code sandbox stubs in _TOOL_STUBS so the
test_stubs_cover_all_schema_params drift test passes. Without this,
scripts running inside execute_code couldn't pass cross_profile=True
through hermes_tools.write_file().
Caught by CI on PR #31290.
Adds an --ids flag to 'hermes kanban promote' mirroring the existing
block/schedule convention, so the marquee use case from issue #28822
(promote all children of a closed organizational parent in one shot)
doesn't require a shell loop. Single-id JSON output stays a flat
object for back-compat; bulk emits a list. Dedupes positional + --ids
so the same id can't be promoted twice in one call. 5 new CLI-level
tests cover bulk happy path, partial-failure exit code, JSON shapes,
and dedup.
Also adds the thedavidmurray noreply-email -> github-login mapping in
scripts/release.py so the salvage cherry-pick passes the AUTHOR_MAP
contributor-credit check.
Adds `hermes kanban promote <task_id>` for manual lifecycle recovery
when an auto-promote daemon misses the parent-done transition (issue
#28822). Refuses promotion unless every parent dep is done/archived
(override with --force). Emits a `promoted_manual` audit event distinct
from the automatic `promoted` kind, so audit consumers can filter
human-driven from system-driven promotions. Supports --dry-run and
--json for orchestration. Does not mutate assignee/claim state — the
dispatcher picks the card up via its normal ready polling path.
Closes#28822.
The post-turn background reviewer prompt listed pinned skills under
'Protected skills (DO NOT edit these)' alongside bundled and
hub-installed skills, with the instruction to say 'Nothing to save.'
if only protected skills needed updating. This meant the reviewer
would refuse to patch a pinned skill even when the user explicitly
wanted that skill improved.
The underlying tool layer already gets this right: skill_manage's
_pinned_guard only fires on delete; patch/edit/write_file go through
on pinned skills. Curator archive/consolidation still skips pinned
at the data layer (agent/curator.py), which is the correct place for
that protection — pin's job is anti-deletion, not anti-improvement.
Both _SKILL_REVIEW_PROMPT and _COMBINED_REVIEW_PROMPT now explicitly
tell the reviewer that pinned skills can be patched, with rationale,
so it doesn't bail out of an improvement just because the target is
pinned.
Two independent bugs caused the slash-command autocomplete to render
`/goal` as `/goa` (and `/gquota` as `/gquot` for that matter) in the TUI:
1. `tui_gateway/server.py` was forwarding `c.display` from
prompt_toolkit's `Completion` straight into the JSON-RPC payload.
prompt_toolkit normalizes `display=` into `FormattedText` (a `list`
subclass), so the wire format became `[["", "/goal"]]` instead of
the `string` that `CompletionItem.display` in the TUI declares.
`meta` already went through `to_plain_text` — `display` did not.
2. The dropdown row in `appOverlays.tsx` used `flexDirection="row"`
with the display `<Text>` and the (very long) meta `<Text>` as
siblings. When the meta overflows the row width, Ink/Yoga shrinks
the *first* column by one cell, lopping the trailing character off
the command name. `/goal` triggers it reliably because its meta
string is the longest of any built-in command (description +
embedded `[text | pause | resume | clear | status]` usage hint).
Wrapping the display column in `<Box flexShrink={0}>` keeps it at
its natural width and lets the meta wrap or truncate instead.
If Nous Portal is the recommended way to run Hermes Agent, it deserves
more than a sub-section buried under `## Inference Providers`. Add two
new pages and shrink the existing providers.md section to a stub that
points at them.
New pages:
- `website/docs/integrations/nous-portal.md` — landing page. What's in
the subscription (300+ model catalog table, Tool Gateway breakdown,
Nous Chat, cross-platform parity, no-dotfile-credentials). Hermes 4
recommendation note. Setup paths (fresh install, existing install,
headless / SSH, profiles). Day-to-day usage (portal status / portal
tools / portal open, switching models, mixing gateway with own
backends, subscription management). Configuration reference. Token
handling. Troubleshooting. Cross-links. Sidebar-position 1 — first
entry under Integrations.
- `website/docs/guides/run-hermes-with-nous-portal.md` — task script.
Eight numbered steps: subscribe → setup --portal → verify with
portal status → first chat → switch models → customize gateway
routing → voice mode → cron/always-on. Per-step troubleshooting.
'What this gets you in plain numbers' comparison table. Sidebar
position 1 — first entry under Guides & Tutorials.
Existing providers.md:
- Replace the 80-line `### Nous Portal` deep-dive with a 13-line stub
that summarizes the value prop, lists the three CLI commands, and
links to the new pages. Saves ~6KB. Other provider sections and
callouts (Codex Note, Two Commands, Tool Gateway tip) preserved.
Sidebar:
- `integrations/nous-portal` inserted right after `integrations/index`,
before `integrations/providers`.
- `guides/run-hermes-with-nous-portal` inserted first in Guides &
Tutorials.
The original PR #17194 description claimed test_display_tool_preview.py
but only ever shipped test_display_todo_progress.py. Add the missing
coverage for the failure-suffix path:
- _trim_error: whitespace strip, length cap, File-not-found path collapse
- _detect_tool_failure: terminal exit codes, memory full, structured
{error}/{message} extraction, malformed JSON, None result
- get_cute_tool_message E2E: read_file failure, terminal exit-only,
terminal stderr message, memory full, success path, no-result path
Also update test_tool_progress_scrollback.test_error_suffix_on_failed_tool
to reflect the new behavior: the generic '[error]' fallback in cli.py
has been removed; failure suffixes now come from the result-aware
_detect_tool_failure (e.g. '[exit 1]', '[File not found: x]').
Parse the todo_tool result summary to display completion progress in
CLI tool preview lines:
Read: ┊ 📋 plan 3/4 task(s) 0.5s
Update: ┊ 📋 plan update 3/4 ✓ 0.5s
Create: falls back to plain count when no completed tasks
Falls back gracefully to the existing 'N task(s)' format when the
result is missing, malformed, or has no completed items.
Originally proposed in PR #17194 by Albert.Zhou; salvaged onto current
main.
Co-authored-by: Albert.Zhou <albert748@gmail.com>
Improves the failure suffix on tool completion lines. Instead of always
showing '[error]' for non-terminal failures, parse the tool's JSON result
and surface the actual message:
Before: ┊ 📖 read foo.py 0.1s [error]
After: ┊ 📖 read foo.py 0.1s [File not found: foo.py]
Before: ┊ 💻 $ ls bad 0.1s [exit 127]
After: ┊ 💻 $ ls bad 0.1s [ls: cannot access 'bad'...]
Adds a _trim_error helper that strips long absolute paths down to the
filename and caps the suffix at 48 chars so it stays readable on narrow
terminals.
Threads the tool result through the tool.completed progress callback so
agent/display.get_cute_tool_message can inspect it. The cli.py [error]
post-suffix is removed in favor of the richer suffix _detect_tool_failure
now produces directly.
Originally proposed in PR #17194 by Albert.Zhou; salvaged onto current
main with the dead-code preview-length bumps dropped (tool_preview_length
config already strictly caps previews, so the per-tool n= defaults are
unreachable).
Co-authored-by: Albert.Zhou <albert748@gmail.com>
`terminal(background=true)` without `notify_on_complete=true` or
`watch_patterns` runs the process SILENTLY — the agent has no way
to learn it finished short of calling `process(action='poll')`
explicitly. That's correct for genuine long-lived processes (servers,
watchers, daemons) but is a footgun for every bounded task (tests,
builds, deploys, CI pollers, batch jobs), which is the vast majority
of background uses.
Hit on May 23, 2026 (PR #31231 incident): agent launched a CI-watch
loop with `background=true` only. The poller ran fine, exited green
6 minutes later, agent never noticed. User had to surface 'we are
green CI, you can merge.' Memory and skill docs said *what* to do
(poll in background) but not *how* to receive the result. The
`notify_on_complete=true` flag exists and works, but is easy to
forget when bg seems sufficient on its own.
Two changes here, mutually reinforcing:
1. Runtime nudge: tool result for `background=true` w/o notify or
watch_patterns now includes a `hint` field explaining the silent-
process failure mode and pointing at the corrective flag. Agent
sees it on the same turn and self-corrects without needing the
user to surface anything. Cost for legitimate server cases is one
ignored read (~50 tokens); cost for forgot-notify cases is
prevented blindness (potentially many turns, or a user nudge).
False positives << false negatives.
2. Schema/description rewrite: top-level TERMINAL_TOOL_DESCRIPTION
and the `background` field description now lead with 'Almost
always pair with notify_on_complete=true' instead of presenting
it as one of two equally-likely patterns. The two legitimate
non-notify shapes (long-lived servers; watch_patterns mid-process
signals) are still documented, but as the minority case.
Tests cover all four shapes: bg-only emits hint, bg+notify doesn't,
bg+watch_patterns doesn't, foreground doesn't. 4 new tests; full
suite of background/process tests stays green (160/160 across the
relevant 6 test files).
AI Card "tool progress" cards created with finalize=False were left in
streaming state on DingTalk's UI after a gateway restart because
disconnect() called _streaming_cards.clear() without first closing
them via _close_streaming_siblings.
Move the finalization loop before self._http_client.aclose() so the
HTTP client is still available when the finalize requests are sent.
Adds a regression test that asserts the HTTP client is alive during
finalization.
Reorder the per-provider subsections under '## Inference Providers'
so Nous Portal — the recommended setup — leads the list, and Google
Gemini via OAuth (which carries a policy-risk warning) drops to last
position right before the '## Custom & Self-Hosted LLM Providers'
section. All other provider sections keep their relative order. Pure
section move; no content changes.
