Add model.reasoning_echo (default false) and per-fallback-entry
reasoning_echo to preserve assistant reasoning_content when
replaying history to custom providers and OpenAI-compatible gateways
that proxy thinking-mode models (Kimi K3, GLM-5.2, DeepSeek, etc.)
but are not matched by the built-in host-based _REASONING_ECHO_RULES.
The flag is per-active-provider, not a global toggle:
- Primary: read from model.reasoning_echo at init and switch_model
- Fallback: set by try_activate_fallback from the fallback entry
- Restore: restore_primary_runtime copies the switch_model snapshot
Unlike PR #76019 global agent.reasoning_echo toggle, the
per-provider flag travels with the active provider — falling back to
a strict provider (Mistral, Groq, Cerebras) correctly strips
reasoning_content even when the primary had the flag enabled,
because the flag is False for the strict fallback.
Complements PR #27361 (dynamic detection) which fires after the first
API response; this PR covers turn-1 and history-replay-on-fresh-session
where dynamic detection has not fired yet.
Closes#76018
Refs: #27297, #27361, #76019
Signed-off-by: Yingliang Zhang <zhangyingliang@outlook.com>
/simplify-code finding: turn_context evaluated resolve_prompt_cache_scope()
inside set_runtime_main's argument list under the umbrella try/except — a
resolution failure would silently skip the ENTIRE runtime binding
(provider/model/base_url/api_key/session_id for all aux calls that turn),
not just the cache scope.
- prompt_cache_scope: add resolve_prompt_cache_scope_safe() (never raises,
returns None on failure/empty).
- turn_context: resolve the scope into a local via the safe variant BEFORE
the set_runtime_main call, so a failure can only lose the scope.
- chat_completion_helpers: _prompt_cache_scope_for_agent delegates to the
shared safe variant (guarded import retained).
- tests: +1 (hostile-property agent -> None; normal/empty passthrough).
- prompt_cache_scope: memo key now includes DB presence (a lazily attached
_session_db re-resolves instead of staying pinned to the physical id);
_persist_disabled agents (background-review forks that never get a DB row)
memoize the fallback instead of re-querying the lineage per API call;
module docstring cross-references get_conversation_root and why the two
lineage resolvers must not be deduplicated.
- chat_completion_helpers: hoist the triplicated
_prompt_cache_scope_for_agent(agent) call to a single local above the
OpenAI-wire dispatch (after the anthropic/bedrock early returns, which
don't use prompt_cache_key).
- codex transport docstring: x-client-request-id mirrors the derived body
key, not the raw scope id.
- turn_context comment: acknowledge the first-turn pre-persist fallback.
- tests: +2 (persist-disabled memoization; lazy DB attach re-resolution).
Legacy compaction mode (compression.in_place: false) rotates the physical
session_id mid-conversation. The prompt-cache scope introduced in #79161 was
derived from that physical id, so every rotation moved the same conversation
into a fresh cache bucket - the prompt cache went cold at every rotation
boundary (#79017).
Fix: resolve a rotation-stable logical scope - the compression-lineage ROOT
of the current session (SessionDB.get_compression_lineage, fork-aware
post-#79193) - once per turn, memoized per transcript segment, and prefer it
over the physical session_id at every prompt_cache_key derivation site:
- agent/prompt_cache_scope.py (new): resolve_prompt_cache_scope(agent) -
lineage-root walk with per-segment memo; falls back to the physical id
when no DB is attached or the walk fails, degrading to pre-fix behavior.
- transports/codex.py: build_kwargs accepts cache_scope_id and prefers it
for the body prompt_cache_key, the xAI x-grok-conv-id header, and the
Codex x-client-request-id routing header. The Codex session_id header
keeps the raw physical id (transcript identity, #57012 contract).
- transports/chat_completions.py: _add_prompt_cache_key accepts
cache_scope_id with the same precedence.
- chat_completion_helpers.py: build_api_kwargs threads the resolved scope
into all three build_kwargs call sites (codex, profile, legacy).
- auxiliary_client.py: set_runtime_main carries cache_scope; the aux
Responses cache-key site prefers it over the physical session_id.
- turn_context.py: resolves the scope once per turn and threads it through
set_runtime_main (no DB walk on the per-API-call hot path).
Scope semantics preserved from #79161: /new starts a fresh scope (new
lineage), /branch children, delegate subagents, and tool children stay
isolated (explicit-fork exclusion in get_compression_lineage), unrelated
sessions keep distinct buckets, and cron per-fire timestamps still
normalize via _cache_scope_from_session_id.
Default installs compact in place (session_id never rotates), so they hit
the memo and produce byte-identical keys to before.
Fixes#79017
The unconditional 2s join before InterruptedError delayed interrupt
detection when Relay managed execution was not active (CI:
tests/run_agent/test_interrupt_propagation.py — detection took 2.34s
against a <1.0s budget, because the mocked worker sleeps 5s and there
is no Relay scope to unwind).
Extract the join into _join_worker_for_relay_teardown(), which no-ops
unless a Relay runtime exists AND managed execution consumers are
registered — the only case where an orphaned physical scope can corrupt
the LIFO stack (#81521). Applied at all three interrupt sites
(streaming, non-streaming, Bedrock streaming). The regression test now
simulates a live runtime so the join path stays covered.
Follow-up to HexLab98's salvaged commits:
- Apply the same bounded worker join before raising InterruptedError at
the two sibling interrupt sites that share the raise-without-join
shape: the non-streaming API poll loop and the Bedrock streaming poll
loop. Both workers run Relay-managed physical attempts, so raising
immediately allowed turn teardown to race a still-open physical scope
exactly as in the streaming path.
- Address the #81601 review finding (egilewski): the pinned nemo-relay
binding's get_scope_stack() returns a native ScopeStack object which
scope.pop rejects with TypeError, so the orphan drain never drained
under the real binding. current_top() now prefers the version-correct
scope.get_handle() accessor and falls back to the old list-unwrap for
fake/legacy shapes. Handle comparisons go through same_handle(),
comparing by uuid, because native ScopeHandle instances do not
implement value equality.
- Add a real-binding regression test that reproduces the orphaned-scope
session close against the pinned native wheel (skips where the native
binding is unavailable), alongside the existing fake-based coverage.
Empty-stream stalls that trip interrupt were raising InterruptedError
before the stream worker closed its physical LLM scope, corrupting the
Relay LIFO stack and cascading into a CLI EIO redraw storm (#81521).
The unconditional pop from the previous commit also stripped the key when
agent.client was still the real MoA facade, forcing the facade to re-prepare
from scratch — a duplicate reference fan-out per turn (caught by
test_moa_virtual_provider_aggregator_is_actor). Gate the defensive strip on
the same prepare()-capability probe used at the injection site so the
handshake survives on the facade while a swapped-in native client is still
protected.
After a client replacement (credential rotation, dead-connection cleanup,
or fallback+restore), agent.client may become a native OpenAI client
while agent.provider stays "moa". The _moa_prepared_request key was
passed through to the native SDK, causing TypeError on every turn.
Pop the key at the dispatch point (chat_completion_helpers.py:509).
The MoAClient facade already handles a missing key by falling through
to its normal resolution path.
Closes#78382
The keepalive httpx client uses read=None, and stranger-thread abort cannot close FDs, so a DeepSeek stall on the cron inline path waited until TCP died — hours past the 600s watchdog. Inject a per-call read timeout matching the stale budget, walk in-flight pool requests, and clear the socket timeout before shutdown without releasing the FD.
Some OpenAI-compatible providers (DeepInfra, etc.) return validation
errors as in-stream SSE chunks: HTTP 200 with choices=None and
error_type/error_message in model_extra. The streaming loop silently
dropped these chunks, causing EmptyStreamError ("empty stream") and
pointless retries on the same bad request.
Fix: check for error_type/error_message on chunks with no choices
before skipping. When found, raise RuntimeError with the provider's
error message so the error classifier can properly handle it.
Fixes#65631
Maintainer fixup on the #79787 salvage:
- An explicit fb.api_mode of "chat_completions" was silently overridden
by the codex_responses / bedrock re-detection pass (which only skipped
re-detection when the pre-computed mode was non-default). Track
explicitness in fb_api_mode_explicit and gate the whole re-detection
block on it.
- Replace the locals().get('fb_api_mode') dead-code hack with clean code
(fb_api_mode is always bound at that point).
- Restore the post-resolve /anthropic + api.anthropic.com host check for
named custom providers whose base_url comes from config rather than
the fallback entry (#32243, #49247), which the PR's restructure dropped.
- Add regression tests: explicit api_mode honored (incl. explicit
chat_completions not overridden), /anthropic-hint fallback detected
pre-rewrite, api_mode forwarded to resolve_provider_client, plain
fallback unchanged.
Fallback activation determined api_mode from the POST-rewrite client
base_url, losing the Anthropic wire signal for /anthropic endpoints
routed through provider 'custom', and never honored an explicit
fb.api_mode config field. Pre-compute fb_api_mode from the ORIGINAL
fallback base_url hint (before _to_openai_base_url rewriting), honor
the explicit api_mode config field, check provider name before the
base_url gate, and pass api_mode into resolve_provider_client at the
fallback call site.
Salvaged from PR #79787 (chat_completion_helpers.py hunks; the
auxiliary_client.py hunk is redundant with #85466's wrap_base fix).
Salvage of PR #64317 (@deaneeth) onto current main, implementing #64161:
observer-only on_stream_start / on_stream_delta / on_stream_end /
on_interim_message plugin hooks dispatched through a host-owned bounded
queue (one worker per callback) so plugin callbacks never run inline on
the token path. Reasoning deltas are opt-in via
plugins.stream_reasoning_deltas.
Trim verbose comments in conversation_loop.py and run_agent.py to 2 lines
each. Fix the same bug class in the compression summary path at
chat_completion_helpers.py: remove _thinking_prefill from the explicit
pop tuple and move the generic underscore-key sweep to after
_drop_thinking_only_and_merge_users, so the drop pass can recognize
prefill stubs there too.
The Anthropic SDK's streaming accumulator builds ParsedMessage snapshots
whose ParsedTextBlock content doesn't match the generic union pydantic
expects, so model_dump() on stream events (message_stop) emits
PydanticSerializationUnexpectedValue UserWarnings straight into the
user's CLI output mid-response.
Pass warnings=False at every helper that dumps arbitrary SDK models
(relay_llm/_jsonable, relay_tools/_jsonable, anthropic_adapter
_to_plain_data, run_agent _hook_jsonable, chat_completion_helpers
extra_content/reasoning_details sites, chat_completions transport),
with a TypeError fallback for duck-typed model_dump implementations.
Adds regression tests including a precondition test that proves the
fixture still trips the warning without suppression.
Opt-in via compression.codex_responses_native (default: false). When enabled,
gpt-5.6-family models on the direct OpenAI API (api.openai.com) or a ChatGPT
Codex subscription send context_management=[{type: compaction,
compact_threshold: N}] on Responses requests. OpenAI compacts server-side and
returns an encrypted compaction output item; Hermes captures it into the
existing codex_reasoning_items sidecar and replays it on later turns in place
of the pruned history — inheriting persistence, session replay, the
cross-issuer guard, and the encrypted-replay kill switch with zero new state.
Scope is deliberately hard-gated (agent/native_compaction.py, re-checked per
request): gpt-5.6 family only — gpt-5.1/5.2 fail server-side on the field
(HTTP 500 / stream stall, no structured rejection; live-verified) — and
direct OpenAI/Codex routes only; xAI, GitHub/Copilot, OpenRouter, relays,
and local servers never see the field.
Hermes' local compression stays armed as the fallback owner: the native
threshold is clamped ~8K tokens below the local trigger so the server
compacts first, and a structured provider rejection of context_management
disables native compaction for the session and retries without it
(one-shot guard in TurnRetryState).
Live-verified E2E on api.openai.com/gpt-5.6: server compaction fired at a
4K threshold, checkpoints captured and replayed, recall preserved across
3 turns; gpt-5.1 with the flag enabled stays clean (field never sent).
Direction credit: PR #76950 by @laryhorb explored native Responses
compaction; this is a minimal reimplementation on current main.
handle_max_iterations() appends its runtime summary request as a plain
role="user" row, which SessionDB persists verbatim. On later compaction the
synthetic-turn filters only recognized compaction summaries, continuation
rows, and todo snapshots, so the nudge could be selected as the latest
actionable user turn — becoming the task snapshot / auto-focus input and
getting summarized as "User asked: ...", demoting the real human task.
Metadata flags do not survive SessionDB projection (the reason the existing
markers are content-based), so recognition must key off stable content.
Extract the nudge into a shared MAX_ITERATIONS_SUMMARY_REQUEST constant and
teach _is_synthetic_compression_user_turn() to recognize it, mirroring the
continuation/todo markers. Every _is_actionable_user_turn call site already
pairs the synthetic guard, so the single recognizer change covers anchor
selection, auto-focus, and real-user-turn detection.
Fixes#78580
Sibling of the chat_completions zero-byte-args fix (previous commits):
a clean SSE close after content_block_start(tool_use) but before any
input_json_delta / message_delta yields an SDK final-message snapshot
whose content is NON-empty (the tool_use block is present, input={})
and whose stop_reason is None. That shape sailed past both
empty-stream guards (they only fire on empty content) and executed the
tool with empty input — no retry, no error: the same silent-data-loss
class as #80498, one provider transport over.
A legitimate completion always carries a stop_reason, so a
tool_use-bearing message without one is a mid-tool-call stream drop.
Raise EmptyStreamError for it, riding the same bounded stream-retry
(HERMES_STREAM_RETRIES) the eventless-stream case already uses.
Gate checked on both return paths (raw SDK snapshot and
accumulator-modified message). Regression tests cover the dropped
shape (mutation-verified: disabling the gate fails exactly that test),
the legitimate tool_use completion, and the text-only no-stop_reason
shape (pre-existing behavior preserved).
When the stream closes right after a tool call's name arrives but
before any argument bytes are delivered, has_truncated_tool_args
was never set (the existing check required a non-empty, whitespace-
stripped arguments buffer). The call fell through to a normal "stop"
finish_reason, later coerced to "{}" at dispatch and executed
silently with no arguments and no retry.
Route this case through the same dropped-mid-tool-call stub/retry
path already used for partially-truncated JSON.
Follow-up hardening on top of the salvaged #80809 watchdog, porting the
locked state-machine design from #75301 (credit: @Zeraphim):
- All lifecycle transitions (stale/cancelled/done) now happen under
request_client_lock. A user or monitor interrupt marks the request
'cancelled' so a racing stale timer can no longer misclassify the kill
as provider staleness and feed a false +1 into the #58962 cross-turn
circuit breaker.
- 'done' is set under the lock on completion, so a late timer callback
that lost the race to a successful response is inert instead of
leaving a spurious streak=1 behind the reset.
- Registration race closed: if the budget expires while the client is
still being constructed, _make_client aborts the freshly-registered
client and fails the call with a retryable TimeoutError instead of
opening a brand-new socket after the only watchdog already fired.
- _resolve_direct_stale_timeout now fails closed: a raising resolver
propagates (same as the worker path) instead of being swallowed into
an infinite budget that would silently disarm the watchdog and
reinstate the very hang #80759 is about.
4 new regression tests, each verified to fail against the pre-fix
watchdog implementation.
Co-authored-by: Zeraphim <diamantejc87@gmail.com>
Cron turns and delegated children are routed onto direct_api_call, which
ran the request inline with no stale detector. The abort plumbing was
registered but nothing ever invoked it, so a provider that accepted the
request and then went silent — connection held open, zero bytes, no
error — hung the run until an external actor killed it, which also
orphaned the execution row. The httpx read timeout is not a usable bound
(1800s default, and this failure mode never trips it), and the job-level
inactivity monitor was observed not to fire.
Arm a watchdog timer on the same budget the interrupt worker's poll loop
uses, so these turns get exactly the patience every other non-streaming
request already gets. On expiry it only aborts the in-flight sockets
through the already-registered hook — it never issues a request, so the
inline / no-worker property that fixes the nested-pool deadlock is
preserved — bumps the cross-turn stale circuit breaker, and surfaces a
retryable TimeoutError so the outer loop reconnects on a fresh pool.
Fixes#80759
Reasoning-summary models emit one reasoning_content delta per completed
summary part, each a self-contained bold heading. The Responses API delimits
those parts with summary_index; the OpenAI chat wire carries no such field —
verified live against Nous Portal, whose reasoning chunks contain nothing but
delta.reasoning_content — so concatenating them glued every part into one
unspaced, half-bold paragraph.
Re-derive the boundary from the signal the wire does carry: a delta opening a
closed bold heading against a mid-line tail. This matches Hermes own Responses
adapter, which already joins its summary parts with a blank line.
Add activity_hb.join(timeout=2.0) after activity_hb_stop.set() in
direct_api_call's finally block so the heartbeat thread is deterministically
stopped before client teardown. Add test verifying no stray _touch_activity
fires after direct_api_call raises an exception.
Follow-up to PR #78548 by @xxxigm.
Top-level delegate_task runs in the background, and the 450s progress-stall
monitor only sees api_call_count / tool / last_activity_ts. Subagents use
non-streaming direct_api_call, which previously touched activity once and then
went silent — so a healthy local GGUF / long-prefill wait looked frozen and
was interrupted around ~450s as "Operation interrupted: waiting for model
response", even when child_timeout_seconds was raised. Refresh activity while
the inline request is open, and treat last_activity_ts advances as sync
heartbeat progress too.
Review follow-up on the #30223 salvage: the original changed the base
cooldown from 60s to 1800s, benching the primary for 30 minutes on the
FIRST 429 (30x regression in primary-restore latency) and breaking the
existing test_rate_limit_exhaustion_keeps_60s_cooldown contract.
Keep upstream's 60s base and escalate per consecutive rate-limit:
60s -> 2m -> 4m -> 8m -> ... capped at 4h. Counter still resets on
successful primary restore (cicae's mechanism, unchanged).
New tests: escalation doubling, 14400s cap, reset-on-restore.
Existing 60s contract test passes UNCHANGED. Mutation-checked:
escalation disabled -> 2 fail; reset disabled -> 1 fails.
Replace the fixed 60-second cooldown with exponential backoff:
30min → 1h → 2h → 4h cap.
The counter is reset by restore_primary_runtime on successful
primary-provider recovery, so the backoff is strictly for
consecutive failures within a single degradation window.
Closes#29702
agent_init.py's init-time fallback and chat_completion_helpers.py's
try_activate_fallback() still read key_env via raw os.getenv(), missing the
per-profile secret scope installed by the multiplexed gateway (same bug
fixed for fallback_config.py/auxiliary_client.py in this PR). Both now
delegate to hermes_cli.fallback_config.resolve_entry_api_key(), and the
Ollama Cloud OLLAMA_API_KEY read now goes through
agent.secret_scope.get_secret() too.
agent_init.py's fallback loop had no try/except around key resolution
(unlike the other three call sites), so a fail-closed UnscopedSecretError
under multiplexing would have crashed init instead of skipping to the next
fallback entry — added the same skip-and-continue handling.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Convert a completed MoA aggregator response into one valid Chat
Completions delta chunk at the MoA facade boundary, normalize completed
message.tool_calls into indexed stream deltas, and classify these local
MoA adapter-shape errors as non-fallback format errors so a local
compatibility bug cannot silently drift the user's MoA route to a
single model (#55933 follow-up).
Salvaged from PR #74903 by @liusencomic-cyber.
Four hot-path consumers paid a full config deepcopy per read:
- telemetry gate relay_shared_metrics.enabled() — runs 2-3x per agent
turn (2x per API call from lifecycle hooks + 1x per tool call) and
called read_raw_config(), which deepcopies the whole raw config every
call. New read_raw_config_readonly() serves the cached dict directly:
248 us -> 4.6 us per call (54x) on Teknium's real 77-key config.
- interruptible_streaming_api_call local-endpoint stale-timeout branch
called load_config() once per API call for every local-model user.
- gateway get_inbound_media_max_bytes() + _get_ephemeral_system_ttl_default()
called load_config() on per-message paths. All three switched to
load_config_readonly() (345 us -> 12 us; PR #28866 lineage).
Together these account for ~90% of the ~1,900 deepcopy primitives per
turn measured in the 26-call stubbed-LLM profile.
read_raw_config_readonly() keeps the (mtime_ns, size) freshness key so
config edits are picked up next call, and preserves the identity
invariant (cache-miss returns the same object later hits serve) —
regression-tested with 'is', per the PR #28866 identity-bug lesson.
The mutable read_raw_config() is unchanged for save-path callers.
581 targeted tests green (config, relay metrics x2, ephemeral reply,
platform base, new readonly suite).
The streaming hot loop computed len(repr(chunk)) on EVERY chunk to feed
the retry-diagnostic byte counter — a full recursive pydantic repr at
5.5-8.8 us per chunk (measured), ~20-30 ms of pure CPU per 3,000-chunk
response, paid on every streaming response on every platform.
New _estimate_chunk_bytes() sizes the chunk from its delta payload
strings (content / reasoning / tool-call arguments) plus a 40-byte
framing floor: 2.1-2.4 us per chunk (~3x cheaper), independent of
pydantic field count, never raises on unknown shapes (Anthropic events,
stub providers fall back to the floor). Both call sites switched
(chat-completions loop + anthropic event loop).
The counter feeds only the stream-retry diagnostic log line
(agent/stream_diag.py) — an estimate proportional to traffic preserves
its purpose (distinguishing 'died at 0 bytes' from 'died mid-stream').
6 new contract tests; 64 targeted stream tests green.
Yingliang Zhang