The /model interactive picker resolved a base_url from user credentials
but never passed it to ProviderProfile.fetch_models(), causing the
picker to always query the provider's hardcoded default endpoint
instead of the user's custom URL (e.g. a company litellm proxy).
- providers/base.py: add optional base_url parameter to fetch_models()
- hermes_cli/models.py: pass resolved base_url to fetch_models()
- Update all subclass overrides for signature compatibility
- Add 6 regression tests covering override, fallback, and integration
Reflect the default-model change in the xAI Grok OAuth guide, the web
search docs (EN + zh-Hans), and the web provider docstring. grok-4.3 is
kept in the model tables as the previous default; the Nous/OpenRouter
aggregator catalog still lists grok-4.3 and is left unchanged.
Switch the default model for the xAI/Grok provider and the xAI web
search backend from grok-4.3 to grok-build-0.1. grok-build-0.1 is
already recognized by the model metadata, so no new model definition
is required; grok-4.3 remains selectable.
Generalizes #32663 (@ehz0ah). The slash-skill scaffolding pollution
affected every auto-syncing memory provider — mem0, hindsight, retaindb,
byterover, honcho, supermemory all store/embed the raw user turn, so a
/skill invocation poisoned their stores with the full skill body, not just
openviking.
- Lift the contributor's parser into agent/skill_commands.py as the canonical
extract_user_instruction_from_skill_message(), co-located with the message
builders so the markers can't drift.
- Strip once in MemoryManager.{prefetch_all,queue_prefetch_all,sync_all} —
fixes the whole provider fan-out, bare /skill turns are skipped entirely.
- OpenViking's _derive_openviking_user_text() now delegates to the shared
helper as defense-in-depth (no duplicated marker literals).
- Marker-drift regression now asserts against the canonical skill_commands
constants; add manager-level coverage proving every provider gets clean text.
'everyone collapses to your peer' read as a promise about all traffic.
pinUserPeer pins the user-side peer and is checked before userPeerAliases
(session.py:335), so a pin overrides every alias — including agent peers.
For a multi-agent operator that silently pools distinct agents onto one
peer, the opposite of intent.
Scopes the wording to 'every non-agent gateway user', notes the pin
overrides aliases, and points agent-mesh operators at pinUserPeer:false +
userPeerAliases instead. Same correction in the wizard menu/echo text,
the plugin README, and the website Honcho page.
* fix(teams): package Microsoft Teams SDK as an installable extra
The Teams adapter imports the microsoft-teams-apps SDK, but it was never
declared as a dependency, so source/local installs hit ImportError and the
adapter silently reported the SDK as unavailable. Add a 'teams' extra
(microsoft-teams-apps==2.0.13.4 + aiohttp) and document 'uv sync --extra teams'.
Per the 2026-05-12 [all] policy, opt-in messaging-platform SDKs are NOT added
to [all] (they would break every fresh install on a quarantined release); the
teams extra is installed on demand like the other platform backends.
Co-authored-by: rio-jeong <rio.jeong@thebytesize.ai>
* chore: map rio-jeong contributor email for attribution (#43945)
* feat(teams): lazy-install the Teams SDK on demand (parity with other channels)
The teams extra alone left Teams as the only messaging platform that wouldn't
auto-install its SDK — every other channel (telegram, discord, slack, matrix,
dingtalk, feishu) lazy-installs via tools.lazy_deps on first connect. Bring
Teams to parity:
- Add 'platform.teams' to LAZY_DEPS (microsoft-teams-apps + aiohttp).
- Replace the passive 'check_teams_requirements = check_requirements' alias with
a real lazy-installer that calls ensure_and_bind('platform.teams', ...),
rebinding all Teams SDK globals on success (mirrors check_slack_requirements).
- Call check_teams_requirements() at the top of TeamsAdapter.connect() so
enabling Teams installs the SDK on demand.
- Keep the passive check_requirements() as the registry check_fn so 'gateway
status' probes never trigger a pip install.
The 'teams' extra remains for packagers / explicit 'uv sync --extra teams'.
Tests: rework the alias test into shortcircuit + lazy-install assertions, and
update test_connect_fails_without_sdk to simulate an uninstallable SDK.
---------
Co-authored-by: rio-jeong <rio.jeong@thebytesize.ai>
Co-authored-by: Teknium <127238744+teknium1@users.noreply.github.com>
Follow up PR #46609's api.minimax.io reasoning report by moving the behavior out of the broad run_agent host gate and into the MiniMax provider profile. Only MiniMax-M3 on the documented OpenAI-compatible /v1 route gets reasoning_split/thinking/reasoning_effort; Anthropic-format MiniMax and non-M3 models keep their existing wire shapes.
Co-authored-by: goku94123 <gooku94123@gmail.com>
Adds an observation_scopes config key (and HINDSIGHT_RETAIN_OBSERVATION_SCOPES
env var) so retained memories can opt into per_tag / all_combinations /
custom scoping instead of Hindsight's default combined pass.
Threaded through _build_retain_kwargs so all three retain paths honor it:
auto-retain and flush-on-switch already use aretain_batch; the tool retain
path is switched from aretain to aretain_batch (functionally equivalent,
aretain just wraps a single-item batch) since aretain doesn't accept the
observation_scopes parameter.
Remove the free Parallel Search MCP path and restore the keyed Parallel backend behavior from before it was introduced.
Also drops the keyless fallback registration/display labeling tests and returns the Parallel SDK pin to the prior version.
GLM-5.2 ships with a 1M (1,048,576) token context window. Without this
entry, Hermes falls through to the generic 'glm' key (202,752 tokens),
under-reporting the context bar and prematurely compressing conversations.
The 1M limit was verified empirically via needle-in-a-haystack retrieval
at 789,240 prompt tokens on api.z.ai/api/coding/paas/v4 — zero errors,
zero truncation, correct retrieval at every tested size (25K through 789K).
Changes:
- agent/model_metadata.py: add 'glm-5.2': 1_048_576 before 'glm' fallback
- hermes_cli/models.py: add glm-5.2 to zai curated models
- hermes_cli/setup.py: add glm-5.2 to setup wizard zai list
- hermes_cli/auth.py: add glm-5.2 to coding plan endpoint probes
- plugins/model-providers/zai/__init__.py: add glm-5.2 to fallback_models
- tests/agent/test_model_metadata.py: context resolution + vendor-prefix tests
Discord enforces a hard cap of 100 global application commands per app.
The adapter registers ~27 native commands plus every gateway-available
entry in COMMAND_REGISTRY plus all plugin commands plus the consolidated
/skill group. On a loaded install (many plugins/quick commands) the
desired set exceeds 100, so tree.sync() / _safe_sync_slash_commands()
hits error 30032 ("Maximum number of application commands reached") and
Discord rejects the ENTIRE batch — silently breaking every slash command,
not just the overflow.
Cap registration at the 100-command limit: native commands (registered
first, highest priority) and the /skill group are always kept; lower-
priority auto-registered COMMAND_REGISTRY and plugin commands are added
only until the cap is reached, with a single concise warning telling the
user how to surface the rest. Since both sync paths read from
tree.get_commands(), bounding the tree fixes the root cause for both.
The Teams adapter only handled image/* attachments — documents (the
application/vnd.microsoft.teams.file.download.info consent-free download
payload and any direct-URL non-image attachment) never reached media_urls
at all, so run.py's document-context injection had nothing to surface.
Completes the class-wide sweep from PR #44695 (Signal/Email/SimpleX).
- download.info attachments: fetch the pre-authed SharePoint downloadUrl
(SSRF-guarded, same guard chain as base.py cache_*_from_url) and route
through cache_media_bytes
- direct-URL non-image attachments: same fetch + classify path
- skip Teams' text/html message-body mirror and adaptive-card attachments
- DOCUMENT > PHOTO > VIDEO > AUDIO precedence for mixed attachments,
matching the Email precedence rationale from #44695
SimpleX tagged unknown files application/octet-stream in media_types
but classification only handled audio/image, leaving msg_type TEXT —
run.py never injected the document context. Same bug class as #12845.
Inbound events key the tracker by the DM chat GUID (any;-;+1555...),
but home-channel react calls address the same space by bare E.164 —
normalize both to the phone so add_reaction's last-inbound default
resolves regardless of which form the caller uses (mirrors the
sidecar's phoneTargetFromSpaceId).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Add `action='react'` to `send_message` tool and expose `add_reaction`/
`remove_reaction` on the Photon adapter.
- Track latest inbound message id per chat (`_last_inbound_by_chat`,
bounded to 200 entries) so the agent can react without threading
message ids through tool calls
- New `add_reaction`/`remove_reaction` public methods on PhotonAdapter;
unlike the lifecycle tapbacks, these are not gated by PHOTON_REACTIONS
- `send_message` gains `action='react'` with `emoji` and optional
`message_id` params; resolves target via existing channel-directory
and home-channel logic; requires a live gateway adapter
Prevents "Future attached to a different loop" errors when
_sidecar_call is invoked from a worker thread via _run_async in
send_message_tool. The persistent _http_client remains in use for
the inbound streaming loop, which always runs on the gateway's loop.
A hard gateway exit (crash, SIGKILL, supervisor restart) left the
detached Node sidecar running with a token the next gateway run doesn't
know, so it could never be told to /shutdown. Every replacement spawn
then died on EADDRINUSE, failing each 30→300s reconnect attempt while
the orphan kept consuming the inbound gRPC stream.
Two layers:
- Lifetime binding: the adapter now holds the sidecar's stdin as a
pipe, and the sidecar (PHOTON_SIDECAR_WATCH_STDIN=1) shuts down on
stdin EOF — fired by the OS on any parent death, including SIGKILL.
- Startup reaping: before spawning, the adapter probes the port and
terminates a stale listener, but only after verifying its command
line is a Photon sidecar; a foreign listener raises a clear error
instead of being signalled.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Pin spectrum-ts to exactly 3.0.0 (was ^1.18.0 plus an `npm install
spectrum-ts@latest` on every setup) so breaking SDK majors can't take
down fresh installs silently; `hermes photon setup` now runs `npm ci`.
Upgrade procedure documented in the README.
Migrate resolveSpace to the v3 namespace API: `im.space.create(phone)`
for DMs and `im.space.get(id)` for everything else — group spaces are
now rehydratable from their persisted id after a sidecar restart, which
v1 could not do.
Markdown: replies go out via the v3 `markdown()` builder (iMessage
renders natively; other Spectrum platforms degrade to plain text).
`PHOTON_MARKDOWN=false` reverts to the stripped plain-text path.
Reactions, behind PHOTON_REACTIONS (default off): lifecycle tapbacks
(👀 while processing, 👍/👎 on completion) via new sidecar /react and
/unreact endpoints with per-target reaction-handle tracking, and user
tapbacks on bot-sent messages routed to the agent as synthetic
`reaction:added:<emoji>` events.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(discord): recover from runtime gateway task exits
Salvaged from #39416 (AMEOBIUS) — cherry-picked only the task-exit
recovery; the original PR was 1081 commits behind with 28 unrelated
commits.
A post-ready discord.py WebSocket crash left the gateway split-brained:
producers stayed active while Discord stopped responding. After this fix
the adapter calls _set_fatal_error(retryable=True) + _notify_fatal_error()
so the existing GatewayRunner reconnect watcher replaces the dead adapter.
Also adds _wait_for_ready_or_bot_exit() so startup failures (SOCKS/proxy
errors, invalid tokens) surface fast instead of burning the full ready
timeout. Because connect() no longer waits via asyncio.wait_for on that
path, test_connect_releases_token_lock_on_timeout is updated to trigger
the timeout through the new helper (same lock-release contract).
3 tests pass (2 new runtime-failure tests + the updated timeout test);
test_discord_connect.py and test_discord_slash_commands.py green.
Co-Authored-By: ameobius <ameobius@local.host>
* fix(test): patch _wait_for_ready_or_bot_exit in timeout cancel test
connect() no longer uses asyncio.wait_for for the ready handshake, so
test_connect_timeout_cancels_bot_task was hanging for 30s in CI.
Co-authored-by: Cursor <cursoragent@cursor.com>
---------
Co-authored-by: ameobius <ameobius@local.host>
Co-authored-by: Cursor <cursoragent@cursor.com>
Follow-up to #44389: the generic 'except Exception' branch in connect()
had the same orphaned-task hazard as the timeout branch. Extract the
cancel-and-await logic into _cancel_bot_task() and call it from all
three sites (timeout branch, exception branch, disconnect()).
Also adds deaneeth to AUTHOR_MAP.
When connect() times out waiting for the Discord ready event, the background
asyncio.Task running client.start() was not cancelled. discord.py's internal
reconnect loop can ignore client.close() while a WebSocket handshake is in
flight, so the orphaned task eventually completes and fires on_ready.
A later successful reconnect then leaves two live Discord clients in the same
process — each with its own on_message handler and MessageDeduplicator instance
— so every @mention creates two threads because the per-adapter dedup caches
cannot catch cross-client duplicates.
Fix: explicitly cancel and await _bot_task in two places:
1. The asyncio.TimeoutError handler inside connect() — catches the case where
the adapter's own inner wait_for fires before the gateway's outer timeout.
2. The start of disconnect() — the load-bearing path, always reached via
_dispose_unused_adapter regardless of which timeout fired first.
Root cause confirmed from production logs: a Jun 8 network outage caused three
consecutive connect() timeouts. The first attempt's bot_task completed its
handshake 4 minutes later ("Connected as") with no preceding watcher line,
then the watcher's real reconnect also connected 90 seconds after that. The two
clients ran continuously for 41+ hours, confirmed by the same user message
appearing as two separate inbound events in two different thread IDs 357ms apart.
Regression tests added to tests/gateway/test_discord_connect.py:
- test_connect_timeout_cancels_bot_task: simulates a connect() timeout with a
NeverReadyBot and asserts _bot_task is None afterward
- test_disconnect_cancels_running_bot_task: injects a live zombie task, calls
disconnect(), and asserts the task is cancelled and the attribute cleared
Replace the hermes-identifying clientInfo/User-Agent/session-id prefix on
the keyless Parallel Search MCP path with a neutral 'mcp-web-client'
identity. Project policy forbids third-party usage attribution without an
explicit user opt-in (see telemetry PR policy); MCP requires a clientInfo,
so a generic one satisfies the spec without attributing traffic.
Also adds the contributor AUTHOR_MAP entry and refreshes uv.lock against
current main (parallel-web 0.6.0).
Make Parallel the web search/extract backend with a zero-setup free tier:
- Keyless (no PARALLEL_API_KEY): web_search/web_extract work out of the box via
Parallel's free hosted Search MCP (search.parallel.ai/mcp), and parallel
becomes the default backend when no other web credentials are configured
(ahead of ddgs, which is search-only). A small hand-rolled Streamable-HTTP
JSON-RPC client speaks the MCP's web_search/web_fetch tools; the existing
web_search/web_extract tools are the only tools registered.
- Keyed (PARALLEL_API_KEY set): uses the Parallel v1 REST endpoints
(client.search / client.extract with advanced_settings.full_content) — no beta.
Bumps parallel-web 0.4.2 -> 0.6.0.
- Attribution: on the free path only, results carry provider/attribution and the
CLI tool line reads "Parallel search" / "Parallel fetch"; the paid path is
unbranded.
- Selection/registration: web tools register unconditionally (free MCP backstop)
while check_web_api_key remains a real usability probe; explicit per-capability
backends are honored (so misconfig surfaces) rather than masked by the fallback.
Tested: live web_search/web_extract against search.parallel.ai in keyless and
keyed modes; unit suites for the MCP client, backend selection, and display
labeling; full agent run shows the "Parallel search" label on the free path.
Drop pinPeerName from the key table (now a deprecated-alias note), and replace
the single/multi/hybrid 'deployment shapes' section with the gateway-gated
intent tree the wizard actually presents, including the [e] raw-edit hatch and
the un-pin pooling steer.
The single/multi/hybrid 'deployment shape' was a misnomer: these keys only
affect the gateway (the one entrypoint supplying a runtime user ID), and the
three preset names stamped a lossy taxonomy onto three orthogonal knobs while
hiding which keys got written.
Replace it with an intent-led tree gated on gateway detection:
- _gateway_platforms() lazily inspects the gateway config (best-effort, no
hard dependency); the step auto-skips when no platform is connected.
- 'who talks to this?' → just me / me+others (pooled?) / only others, deriving
pinUserPeer + userPeerAliases + runtimePeerPrefix and echoing the result.
- [e] drops to a raw-knob editor for power users.
- The single→multi orphan guard survives as a pooling steer.
The setup wizard wrote the legacy pinPeerName even though pinUserPeer is
the canonical key that outranks it in the resolver — so it had to scrub
the canonical key afterward to stop it winning. Write pinUserPeer directly
and migrate any legacy pinPeerName onto it on touch (setup load + clone),
which removes the precedence-fighting entirely.
Resolver still reads pinPeerName as a back-compat alias; that's deferred.
* fix(openrouter): route reasoning_effort to verbosity for adaptive Anthropic models
Reasoning-mandatory Anthropic models (Claude 4.6+/fable/mythos-class) over
OpenRouter ignore reasoning.effort and use adaptive thinking. #42991 correctly
stopped Hermes from sending a reasoning field to them (it 400s), but put nothing
in its place — leaving agent.reasoning_effort a silent no-op on the OpenRouter
path: the model always ran at its adaptive default (high) regardless of config.
OpenRouter honors the requested effort on the top-level verbosity field instead
(maps to Anthropic output_config.effort). Route the existing
reasoning_config[effort] there for these models while still never emitting a
reasoning field, preserving the #42991 fix. No new config arg — the value the
user already sets via agent.reasoning_effort now flows to verbosity.
- low/medium/high/xhigh/max pass through verbatim (OpenRouter accepts the
extended scale for Claude; verified live HTTP 200 + monotonic token spend).
- effort unset/none/disabled omits verbosity so the model keeps its default.
- native Anthropic transport already correct; unchanged.
Fixes#43432
* test(openrouter): cover real effort range (add minimal, frame max as passthrough)
Adversarial review noted the verbosity tests looped over 'max' — a value
parse_reasoning_effort can never produce — while omitting 'minimal', which it
can. Align the routing test with the real config range
(VALID_REASONING_EFFORTS = minimal/low/medium/high/xhigh) and keep a separate
value-agnostic passthrough test that documents why xhigh/max must survive
verbatim (TypedDict, no runtime literal validation; OpenRouter accepts the
extended scale for Claude).
* docs: explain reasoning_effort -> verbosity routing for adaptive Anthropic models
Document that reasoning_effort transparently maps to OpenRouter's verbosity
field for adaptive-thinking Anthropic models (Claude 4.6+/Fable/Mythos), where
reasoning.effort is ignored. Note xhigh is the configurable ceiling (max is wire-
only). Add verbosity as a top-level-kwarg example in the provider-plugin guide.
Follow-ups on top of #26016's expensive-model guard:
- gateway/slash_commands.py: typed '/model <name>' now routes through the
expensive-model confirmation gate (slash-confirm buttons / text fallback)
instead of bypassing the guard the pickers enforce. Cancel leaves the
session override and --global config untouched.
- telegram/discord/web_server: run expensive_model_warning() via
asyncio.to_thread — it can hit models.dev or a /models endpoint on a
cache miss, which would otherwise block the event loop.
- telegram: picker callback no longer toasts 'Model switched!' when the
switch callback raised (both mm: and mc: paths).
- tests: new tests/gateway/test_model_command_expensive_confirm.py pins
the typed-path gate (prompt, confirm-once, cancel, cheap-model no-op).
Rebased onto current main and re-ported across the restructured
surfaces: model flows now thread confirm_provider/base_url/api_key
through hermes_cli/model_setup_flows.py, the Discord picker lives in
plugins/platforms/discord/adapter.py, and the web dashboard picker
applies chat-mode switches via config.set so the expensive-model
confirmation can ride the response.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
DISCORD_ALLOWED_ROLES was checked by the Discord adapter (_is_allowed_user)
but gateway._is_user_authorized only read DISCORD_ALLOWED_USERS, so
role-authorized users were rejected with "Unauthorized user" at the
gateway layer despite passing the adapter gate.
- Add role_authorized: bool = False to SessionSource
- Add role_authorized param to build_source (base.py)
- Compute _role_authorized in on_message when user passes via role not user ID
- Thread _role_authorized through _handle_message -> build_source
- Check source.role_authorized early in _is_user_authorized (run.py)
Fixes#33952
The voice inactivity timer (VOICE_TIMEOUT) only counted the bot's OWN audio
playback as activity. Under /voice off (text-only replies, but still in the
channel — leaving is /voice leave) nothing ever reset it, so every 300s the bot
disconnected and spammed "Left voice channel (inactivity timeout)."
The adapter now learns the live voice-reply mode via a getter wired from run.py
and skips the auto-disconnect while mode is off. It also resets the timer when a
user actually speaks to the bot, so an active listener (incl. voice-on
text-only sessions that never play audio) isn't dropped mid-conversation.
The Langfuse SDK treats `data:*;base64,...` strings as media and tries to
decode them. `_truncate_text` was slicing those strings mid-payload, producing
invalid base64 and noisy "Error parsing base64 data URI" logs. Observability
only needs the metadata, not raw image/audio bytes, so redact the whole data
URI (type, media_type, length) before it reaches the SDK.
Salvaged the Langfuse fix from #39682 onto current main as a standalone,
single-concern change (the dashboard `dist/**` and plugin-discovery parts of
that PR already landed separately on main).
Co-authored-by: foras910521-lab <foras910521-lab@users.noreply.github.com>
The previous fix (#42991) only omitted reasoning when it was being disabled.
But reasoning-mandatory Anthropic models (Claude 4.6+, fable) 400 with
thinking.type.disabled on EVERY tool-continuation turn even when reasoning is
enabled: chat_completions never replays signed thinking blocks, so the prior
assistant tool_call has no thinking, and OpenRouter resolves "reasoning
requested but history has none" by emitting thinking.type.disabled — which
these models reject. Result: first turn works, every turn after the first tool
call dies (HTTP 400, non-retryable).
OpenRouter ignores reasoning.effort for adaptive Anthropic models anyway (the
model self-decides), so the reasoning field is pointless for them on every turn
and harmful on tool-replay turns. Omit it entirely → adaptive default.
- openrouter profile: drop the reasoning field for reasoning-mandatory Anthropic
models regardless of enabled/disabled; legacy Anthropic + non-Anthropic models
unchanged.
- tests: assert omission across enabled/disabled/effort variants; parity tests
switched to a non-Anthropic reasoning model (deepseek) since Anthropic 4.6+ no
longer carries a reasoning field.
Verified live end-to-end: a tool-replay turn on anthropic/claude-fable-5 with
reasoning enabled now builds extra_body=None and returns HTTP 200 (was 400).
New Anthropic models without a recognized version substring (claude-fable-5
and future named/numbered releases) were classified as legacy and routed down
the manual-thinking path, which made OpenRouter emit thinking.type.disabled —
a form reasoning-mandatory Claude models reject with a non-retryable HTTP 400.
Invert the brittle version-substring allowlists to default-to-modern (mirroring
_get_anthropic_max_output): unknown Claude models get the adaptive/xhigh/
no-sampling contract, with an explicit legacy list for older families. Non-Claude
Anthropic-Messages models (minimax, qwen3, …) keep the manual path.
- anthropic_adapter: _supports_adaptive_thinking / _supports_xhigh_effort /
_forbids_sampling_params now default unknown Claude models to modern; legacy
families enumerated in _LEGACY_MANUAL_THINKING_CLAUDE_SUBSTRINGS.
- openrouter profile: omit reasoning entirely (→ adaptive default) instead of
forwarding {enabled:false} for reasoning-mandatory Anthropic models; legacy
Anthropic + all non-Anthropic models still pass the disable form through.
- model_metadata + output-limit table: register claude-fable-5 (1M ctx, 128K out).
Tests assert the invariant ("unknown Claude model -> modern contract; legacy
stays manual; non-Claude unaffected"), not specific model names.
* fix(photon): override transitive CVEs in the sidecar deps
`npm audit` flagged 7 high-severity transitive CVEs (protobufjs code injection
GHSA-66ff-xgx4-vchm + outdated @opentelemetry OTLP exporters) pulled in via
spectrum-ts -> @photon-ai/otel. npm's suggested fix downgrades spectrum-ts to a
version that targets the decommissioned spectrum host, so instead pin patched
versions via `overrides` (protobufjs 8.6.1, @opentelemetry/* 0.218.0) without
touching spectrum-ts. `npm audit` -> 0; spectrum-ts + provider still import.
* fix(photon): harden the sidecar bridge + bound the dedup cache
- constant-time sidecar control-token comparison (was `!==`, timing-attackable).
- cap the control-channel request body (2 MiB) so a compromised local peer can't
OOM the sidecar.
- wrap the inbound gRPC stream consumer in a re-subscribe loop with capped
exponential backoff + jitter — if the async iterator throws/ends it would
otherwise stop inbound forever (the adapter dedupes any replay).
- add an unhandledRejection handler so a stray rejection logs instead of killing
the process.
- dedup cache (adapter) was a true bounded LRU only for expired entries; a burst
of unique ids within the window grew it without limit. Evict oldest at the cap.
* chore: add AUTHOR_MAP entry for PhilipAD
---------
Co-authored-by: PhilipAD <philipadsouza@gmail.com>
Based on #42658 by @mnajafian-nv.
Preserves the real downstream provider/tool exception when NeMo Relay's
managed adaptive execution wraps a failing callback as an internal runtime
error. Without this, the original exception (and its retry-classification
signal, e.g. status_code) is lost behind Relay's wrapper.
Salvage changes on top of the original PR:
- Tolerant Relay-wrapper match: _is_relay_wrapped_callback_error now uses
str.startswith on the "internal error: <cls>: <msg>" prefix instead of
exact equality, so a future Relay version appending a traceback/suffix
doesn't silently defeat the unwrap. On a total format change it returns
False and falls back to the pre-fix behavior (surfacing Relay's error)
rather than masking it.
- Deduplicated the LLM and tool execute paths into a shared
_run_managed_with_downstream_preservation helper, removing ~20 lines of
copy-pasted nonlocal/try-except scaffolding that could drift out of sync.
- Added a real-middleware regression guard
(test_nemo_relay_downstream_unwrap_matches_real_middleware_wrapper_shape)
that drives hermes_cli.middleware._run_execution_chain and asserts the
plugin's _original_downstream_error unwraps the actual private
_DownstreamExecutionError wrapper. The original synthetic tests modeled the
wrapper with a local class, so a rename or shape change in core middleware
would not have been caught; this test fails loudly if that contract drifts.
Co-authored-by: mnajafian-nv <mnajafian@nvidia.com>
