Every LLM call built a fresh openai.OpenAI wire client (new httpx pool,
TCP+TLS handshake, measured 19.2ms p50 / 35.5ms p95 per call at ~5 calls
per tool-loop turn) and closed it when the request finished. Cache ONE
reusable wire client on the agent, keyed by the effective client kwargs:
- _create_request_openai_client hands back the cached client when the
effective kwargs are identical; any change (credential rotation,
provider failover, vision default_headers) evicts and rebuilds.
- Only a request that produced a response reports a reuse close reason
(request_complete / stream_request_complete); error unwinds report
*_error_cleanup and really close, so a retry after a request error
always builds a fresh pool.
- Cross-thread aborts poison the slot: a pool whose sockets were
shutdown(SHUT_RDWR) from a stranger thread is never reused (#29507) —
the owner-thread close discards it and the next create rebuilds. The
holder read and the abort are atomic (under the holder lock) at all
three abort sites, so a late abort can never poison the NEXT request's
checked-out client.
- Worker-side interrupt breaks close the half-read SSE stream on the
owning thread before building the partial response; a failed close
poisons the slot (otherwise each interrupt leaked one checked-out
connection until the pool hit PoolTimeout). run_codex_stream gets the
same poison-on-close-failure handling.
- Single checked-out slot (in_use): a concurrent call gets an untracked
client with the old per-request lifecycle.
- release_clients() / close() really close the cached client when idle;
if a worker has it checked out they abort the sockets and detach the
slot, deferring the FD release to the worker's own close.
- MoA facade and Mock passthroughs never enter the cache; max_retries=0
is preserved on all request clients.
The concept 'never send a turn that strict wire validation rejects as
empty' was forked across four sites, each with its own predicate and its
own blind spots:
1. build_assistant_message write-time ' ' pad — broke codex commentary
turns (content:'' is a designed state), and a DB-side pad can't
survive _rows_to_conversation's whitespace strip anyway. REMOVED.
2. conversation_loop send-time ' ' pad — main-loop only (summary path
uncovered), ordering-fragile (had to run after whitespace
normalization), assistant-only. REMOVED.
3. stream-stub '[response interrupted]' substitution — defeated the
loop's empty-stub guard (the stub no longer looked empty, entered
history, and the placeholder leaked into the stitched final
response via truncated_response_parts). REMOVED.
4. repair_empty_non_final_messages in sanitize_api_messages — the
unconditional pre-send chokepoint shared by the main loop AND the
summary path, covers user and assistant turns, non-final only,
copy-on-write. This is now the SINGLE OWNER.
The owner's payload predicate (_msg_has_payload) is extended to treat
codex_message_items / codex_reasoning_items as payload, so
designed-empty codex commentary turns are never rewritten on any
api_mode — the failure shape that broke site 1 in CI is encoded in the
owner, not special-cased at a call site.
Tests updated to pin the new contracts: builder stores textless turns
as-is; the empty stream stub stays recognizably empty for the loop
guard; poisoned resumed histories are repaired to the placeholder at
the send boundary; codex item carriers are never rewritten.
Sabotage-verified: unwiring the owner fails 3 regression tests.
Add distinct, greppable log lines at both fix sites so the condition is
observable in the field instead of only inferable from the absence of the old
'Cannot compress further' spiral:
- chat_completion_helpers: warn when an empty partial-stream stub is replaced
with the placeholder (0 chars recovered, no tool call).
- error_classifier: warn when a malformed-body 400 is classified as
format_error rather than context overflow, with num_messages/approx_tokens.
Extend the litellm-shape classifier test to assert the warning is emitted.
A stream that dies after delivering deltas but with 0 recovered chars (and
no captured tool call) produced a content-less assistant stub. That empty
non-final message is invalid for the Anthropic schema, so every later request
400s with 'all messages must have non-empty content' (INVALID_REQUEST_BODY).
On a large session the 400 was misclassified as context overflow, dropping the
loop into a compression spiral that ends in 'Cannot compress further' — a
misleading context-size error on a session nowhere near its limit.
Root cause: substitute a minimal '[response interrupted]' placeholder so the
stub is never empty.
Misclassification: add _INVALID_MESSAGE_BODY_PATTERNS (checked before the
context-overflow heuristic) to classify these as non-retryable format_error,
and teach the body extractors the litellm/Bedrock errorMessage/errorCode/
errorArgs shape so descriptive proxy errors are not mistaken for generic ones.
Adds RED-verified regression tests for the stub path and three classifier tests
(including a guard that real overflows still compress).
Commentary-phase Codex turns persist with content:'' by design (their
text is delivered via the interim assistant callback), and the Responses
wire has no 'assistant must not be empty' validation — padding them
broke test_run_conversation_codex_continues_after_commentary_phase_message
in CI. Both the builder pad and the send-time pad now skip
api_mode=codex_responses. Keying on the ACTIVE api_mode preserves the
repair for codex-written sessions replayed through a strict
chat-completions provider.
A mid-tool-call stream drop with no delivered text produces a
partial-stream stub carrying content:'' and tool_calls=None. The
conversation loop's truncation path appended it to history as
{"role":"assistant","content":""} before the continuation nudge, and
strict providers (Moonshot/Kimi via OpenRouter) reject empty assistant
content with HTTP 400 ("the message at position N with role 'assistant'
must not be empty") on the next replay. Because the message is
persisted, every subsequent turn re-failed — the session was
unrecoverable.
Three layers, smallest blast radius first:
1. conversation_loop (length path): an EMPTY partial-stream stub is no
longer appended as an interim assistant message; only the
continuation user-message is. Stubs that delivered partial text are
still persisted so continuation stitching is unchanged.
2. chat_completion_helpers.build_assistant_message: never serialize a
textless assistant turn with content:'' — pad to a single space, the
same trick as the reasoning_content pad (#15250, #17400). Tool-call
turns are exempt (content:'' alongside tool_calls is accepted
everywhere).
3. conversation_loop send boundary: pad a textless assistant turn's
empty content to a single space AFTER all content-mutating passes
(surrogate sanitize, whitespace normalization, thinking-only drops),
before token estimation. This is the durable repair for sessions
ALREADY poisoned by older builds: the persisted stub rows are rebuilt
to '' on every reload (_rows_to_conversation strips whitespace, so a
DB-side pad can't survive) and only a send-time pad repairs them.
Verified: 485 tests pass across the four affected files; live replay of
a real poisoned session's resumed history against Moonshot via
OpenRouter returns HTTP 200 (was HTTP 400).
Every fallback/dedup/skip decision asks one question — 'is this candidate
the same backend as the one that failed, along the axis that failure
invalidated?' — but it was re-implemented inline at six sites across four
subsystems, each comparing whatever string was locally convenient. Each
incident fixed one site while the others kept the bug: #22548, #70893,
#59561, #72468, #62984/#54250/#57584.
agent/backend_identity.py now owns the concept: BackendIdentity (provider /
model / base_url axes), FailureScope (MODEL / CREDENTIAL / ENDPOINT — each
failure class invalidates a different axis), and should_skip_candidate().
Unknown axes never manufacture a skip (over-skipping strands failover; a
wrong try costs one RTT).
Migrated sites:
- chat_completion_helpers.try_activate_fallback: replaces the provider+model
early-exit (the #62984 bug: ignored base_url, stranding multi-endpoint
pools) AND _fallback_entry_is_same_backend_by_base_url (deleted)
- auxiliary_client._try_configured_fallback_chain +
_try_main_agent_model_fallback: replace label/model comparisons; auth and
payment map to CREDENTIAL scope, keeping the #59561 carve-out
- hermes_cli/fallback_cmd add: primary-match + duplicate checks now identity-
aware (#54250/#57584): same provider+model on a different explicit
base_url is a pool entry, not a duplicate
_mark_provider_unhealthy stays label-keyed deliberately: its only triggers
are confirmed 402s, which ARE credential-scoped.
Owner-level tests pin each incident's semantics by number; sabotage-verified
(removing the base_url axis fails the #62984 test).
Root cause: delegate_task children run through three nested daemon-thread
layers (async-delegation executor -> per-child timeout executor -> the
interrupt worker interruptible_api_call spawns). After multi-day gateway
uptime the deepest layer wedges BEFORE the socket opens — the same
fingerprint as the gateway-cron hang (#62151): zero stale-detector output
(the worker never reaches dispatch), all providers, foreground/restart
works. The cron fix (should_use_direct_api_call) explicitly excluded
delegation 'for lack of evidence' — #60203 is that evidence.
- should_use_direct_api_call: extend the inline gate to delegated
children, detected via the delegation ContextVar set by
_run_single_child (platform='subagent' stamp as fallback). Scope
unchanged otherwise: chat_completions wire only; Codex/Anthropic/
Bedrock/MoA keep their established workers. Interrupts still work —
the inline path registers _active_request_abort, which interrupt()
invokes cross-thread (same mechanism the #72227 stall monitor uses).
- _dump_subagent_timeout_diagnostic: dump ALL thread stacks (bounded,
40), not just the conversation worker — a pre-HTTP wedge is
indistinguishable from a slow provider without seeing where the
nested helper threads sit.
The streaming stale watchdog was calling
_replace_primary_openai_client() from its polling thread, which closes
the shared client's connection pool. Worker threads from previous
stale-killed attempts may still be unwinding their SSL BIOs, causing
TLS application-data to overwrite SQLite file headers via FD reuse.
This is the same corruption vector documented in #67142 for Anthropic,
where the fix was to never close the shared client from a non-owner
thread. Apply the same pattern to the OpenAI-wire path:
- Stale stream watchdog: skip shared client replacement
- Mid-tool-retry cleanup: skip shared client replacement
- Stream retry cleanup: skip shared client replacement
The request-local client is already closed via _close_request_client_once.
The shared client is replaced lazily by _ensure_primary_openai_client
on the next request, which runs on the owning thread.
Closes#70773.
Base-url+model dedup was meant for custom shim aliases, but it also
skipped first-class providers that share an inference host while using
different credentials. That stranded xai-oauth spending-limit failover
to the xai API-key provider when both used the same model slug.
Replace 3 duplicated entry_id resolution blocks (try/except +
entry_id_for_api_key + fallback to None) in agent_init.py,
chat_completion_helpers.py, and switch_model with a single
sync_credential_pool_entry_id(agent) function in agent_runtime_helpers.
Follow-up to #70323.
Track the selected credential by stable pool entry ID so token refreshes and shared cursor movement cannot detach failures from the entry that issued them. Stop unmatched single-entry pools from reporting a no-op rotation as successful recovery.
Co-authored-by: Maxim Esipov <maksesipov@gmail.com>
When MoA mode is active with a Gemini model as the aggregator,
agent.model holds the virtual preset name (e.g. "closed"), not the
actual aggregator model name. The _sanitize_tool_calls_for_strict_api
call uses agent.model to decide whether to keep extra_content
(thought_signature) on tool_calls — since "closed" doesn't contain
"gemini", the thought_signature is stripped and the Gemini aggregator
rejects the next request with HTTP 400 (INVALID_ARGUMENT):
"Function call is missing a thought_signature in functionCall parts."
Fix: resolve the actual aggregator model name from moa_config
(conversation_loop) or last_aggregator_slot (chat_completion_helpers)
and pass it to _sanitize_tool_calls_for_strict_api so the
_model_consumes_thought_signature check sees the real Gemini model.
Closes#65092
The au./apac. additions from #46297 and #65973 covered
is_anthropic_bedrock_model and _normalize_bedrock_model_name; the same
prefix lists exist at two more sibling sites that would still miss
au./ca./sa./me./af. profiles:
- anthropic_adapter._looks_like_bedrock_model_id
- chat_completion_helpers (reasoning stale-timeout floor resolution)
All four sites now share the same 11-prefix set (global/us/eu/apac/ap/
au/jp/ca/sa/me/af, longest-first so apac. wins over ap.). The Bedrock
picker's BEDROCK_GEO_PREFIXES is deliberately untouched: au. absent
there fails open (profile shown), and adding it requires a region-to-geo
remap to avoid hiding Sydney profiles.
Deduplicates the SimpleNamespace stub-response construction that was
copy-pasted between the tool-call-drop guard and the text-only-drop
guard (both in interruptible_streaming_api_call). The third site
(error-handler at ~L3775) has a structurally different shape (different
role/reasoning/model/usage sources + _content_filter_terminated tag)
and is left inline with an existing comment.
When a streaming response ends cleanly (HTTP 200) with no finish_reason
after delivering text but no tool calls, the chunk collector silently
stamps finish_reason='stop' and the conversation loop presents truncated
text as a complete response.
Three stream-drop paths now exist after chunk collection:
1. Zero-chunk → EmptyStreamError, retried (existing)
2. Tool-call in progress, no finish_reason → partial-stream-stub (existing)
3. Text-only, no finish_reason → partial-stream-stub (NEW — this fix)
Path 3 routes through the same PARTIAL_STREAM_STUB_ID + FINISH_REASON_LENGTH
machinery as path 2. The conversation loop shows 'Stream interrupted —
requesting continuation' and injects a continue prompt, giving the model
a chance to resume where the stream dropped.
Observed with DeepSeek provider where CloudFront drops SSE streams
mid-response after delivering partial text.
Deduplicates the sidecar handling across 9 sites:
- substitute_api_content(): 2 API-bound pop+substitute sites
(chat_completion_helpers, transport — conversation_loop keeps its
inline pop because the current-turn compose fallback needs the value)
- drop_stale_api_content(): 4 rewrite-drop sites
(context_compressor x2, replay_cleanup, agent_runtime_helpers)
- extract_api_content_sidecar(): 3 gateway forwarding sites
(gateway/session, gateway/slash_commands, cli_commands_mixin)
Also: restore the eager _pending_cli_user_message clear on the early
row-creation try (was lost when the crash-persist moved after prefetch —
a crash in compression between the two tries would leave a stale staged
input), and fix a comment indentation nit in run_agent.py.
Co-authored-by: Soju06 <qlskssk@gmail.com>
The first LLM call of every gateway turn gets ~0% provider prompt-cache
hit rate (in-turn calls: 97-99%) because the bytes sent for a turn's
user message are not the bytes replayed next turn: memory-prefetch and
pre_llm_call context are injected into the API copy only, the #48677
persist override writes cleaned content to the DB row, and
get_messages_as_conversation sanitize/strips user and assistant content
on load. Any of these diverges the request prefix at that message and
re-prefills everything after it — measured 27.9s for the first call vs
2.4-5.8s cached at a median ~156k-token context.
Persist what you send: a nullable messages.api_content column stores the
exact content string sent to the API when it differs from the clean
stored content, and replay substitutes it verbatim (no sanitize, no
strip). The injection composition lives in one helper
(turn_context.compose_user_api_content); the turn prologue stamps its
output onto the live user message, the api_messages build sends the
stamped bytes, and every outgoing copy pops the field so it never
reaches a provider. The crash-resilience user-turn persist moves after
prefetch/pre_llm_call so the user row is written once with its final
sidecar; _ensure_db_session stays before preflight compression (session
rotation needs the parent row under PRAGMA foreign_keys=ON). The
current-turn index trackers are re-anchored after compaction rebuilds
the message list, in-place preflight compaction backfills the stamp onto
the already-inserted row, and gateway replay forwards the sidecar only
when the replay pipeline did not rewrite the content. Rewrite paths that
would leave stale bytes (historical image strip, merge-summary-into-
tail, consecutive-user repair merge, stale-confirmation redaction) drop
the sidecar; the chat-completions transport and the max-iterations
summary path strip it defensively. codex_app_server and MoA turns are
excluded from stamping because their wire bytes differ from the
composition. A missing or dropped sidecar degrades to today's behavior
(one cache-boundary miss), never to wrong content.
Direct-Anthropic requests used a single shared _anthropic_client, and the
stale/interrupt watchdog closed + rebuilt it from the poll (stranger) thread
at four sites (non-streaming stale/interrupt, streaming stale/interrupt).
Closing a client whose TLS socket a worker thread was still reading released
the FD from a stranger thread; the kernel recycled it under a live SSL BIO,
which then wrote a 24-byte TLS record into an unrelated SQLite header
(cron/executions.db), bricking every cron on the profile. Same shape as the
OpenAI-only #29507 fix, but the Anthropic path never got the owner-thread
contract.
Extend the #29507 ownership contract to Anthropic: build a per-request client
(_create_request_anthropic_client), register it with the request-client
holder tagged by kind, and route _close_request_client_once by kind — a
stranger thread only shuts the request client's sockets down
(_abort_request_anthropic_client), while the owning worker performs the SDK
close (_close_request_anthropic_client). The shared _anthropic_client is now
never closed from inside a request (streaming or non-streaming), including the
worker retry-cleanup sites, since each attempt builds a fresh request client.
The #28161 no-hang guarantee is preserved: the poll-thread socket abort
unblocks the worker immediately.
Salvages the approach from #51688 (@raymondyan-zhijie), reimplemented onto
current main (non-streaming dispatch was refactored into
_dispatch_nonstreaming_api_request; streaming grew _cancel_current_stream_attempt
and worker retry-cleanup sites). Tests updated to the request-local mechanism
(incl. replacing a banned source-reading test with a behavior test) plus new
regression coverage proving the watchdog aborts the request client and never
touches the shared client.
Co-authored-by: raymondyan-zhijie <32435458+raymondyan-zhijie@users.noreply.github.com>
The per-turn `_cleanup_task_resources` unconditionally calls
`cleanup_browser`, closing the browser window immediately after each
bot reply. This makes headed mode (`AGENT_BROWSER_HEADED=1`) unusable
— the window flashes up and disappears on every response.
This mirrors the existing VM persistence pattern: skip per-turn
cleanup when headed mode is active and let the inactivity reaper
handle idle sessions instead. Full-session teardown on gateway
shutdown remains unconditional.
Also adds `browser.headed` config.yaml support and passes `--headed`
to agent-browser in local mode when configured, so users don't need
to rely solely on the `AGENT_BROWSER_HEADED` env var.
Closes#11020 (lead bug)
The _discard_stale_stream_chunk helper emitted both logger.warning and
logger.debug for the first discarded chunk. Gate the debug on
discarded_chunks != 1 so only the warning fires for the first discard.
Second call site (non-streaming / gateway path) now flattens list-type
content with flatten_message_text before the inline <think> regex and the
surrogate sanitizer, matching the interim-text fix from the prior commit.
Adds regression tests (tests/run_agent/test_66267_multimodal_interim.py)
covering:
- build_assistant_message with list content does not raise TypeError
- inline <think> inside list content is extracted + stripped correctly
- _interim_assistant_visible_text is safe for tool messages (list content)
- duplicate_previous_interim dedup guards against tool messages
Verified the tests fail without the fix (TypeError: expected string... got
'list') and pass with it.
Follow-up to PR #66479 salvage. The three new env vars
(HERMES_LOCAL_STREAM_STALE_TIMEOUT, HERMES_GATEWAY_MAX_STARTS,
HERMES_GATEWAY_START_WINDOW_S) were introduced as bare env-var reads with no
config.yaml surface or documentation — violating the .env-is-for-secrets-only
policy (behavioral settings must live in config.yaml, bridged to env internally).
- config.yaml: add gateway.respawn_storm {max_starts, window_seconds} to
DEFAULT_CONFIG, mirroring the existing restart_loop_guard pattern.
- gateway.py: read config.yaml first, env vars override as escape-hatch.
- environment-variables.md: document all three new env vars, noting the
config.yaml alternative for the gateway ones.
- chat_completion_helpers.py: cross-reference the env-var docs from the
local stale timeout comment.
_bedrock_reasoning_stale_floor only matched dashed floor-table keys (opus
'claude-opus-4-6'), so sonnet reasoning models keyed with a dotted version in the
shared table ('claude-sonnet-4.5'/'4.6') got no reasoning floor from the dashed
Bedrock inference-profile id — premature stale-abort if the base timeout is set
below 180s. Generate both version-separator forms (digit-dash-digit <-> digit-dot-
digit, via lookbehind/lookahead so only version numbers flip) and try all
candidates; matches the table however each model is keyed, no edit to the shared
table. Verified: opus->240 (unchanged), sonnet-4.5/4.6->180, haiku->None. New
TestBedrockReasoningStaleFloor (8 cases).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- _spawn_supervised restart cap was lifetime-cumulative (never reset), so a
watcher crashing _MAX_SUPERVISED_RESTARTS times over a days-long process was
permanently abandoned despite the 'consecutive' wording. Make it time-based:
reset the attempt counter when the task ran healthily (>= _SUPERVISED_HEALTHY_SECS
= 300s) before crashing, so only rapid repeated crashes accumulate toward the
ceiling. Reword docstring/log. New regression: healthy-run-then-crash is not
abandoned.
- _derive_stream_stale_timeout read only 'model', so the reasoning stale-floor
never applied to Bedrock (payloads key the model as 'modelId', dotted
us.anthropic.claude-opus-4-6-v1:0 form). Resolve model||modelId and normalize
the Bedrock dotted/region-prefixed form to match the floor. (Sonnet's dotted
vs dashed floor-table key is a pre-existing table mismatch, left out of scope.)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Re-applied against v0.18.2. Upstream now covers the OpenAI-path stall watchdog
(cross-turn give-up breaker #58962) and the tool-batch deadline, so those are
dropped. Two gaps remain:
- Bedrock streaming had NO liveness watchdog and was excluded from the #58962
breaker (it returns before _check_stale_giveup). Add an on_event hook to
stream_converse_with_callbacks (fires per yielded event = true wire-level
liveness), drive a stale timer from it, and wire Bedrock INTO the existing
breaker: entry _check_stale_giveup, _bump_stale_streak on stall, raise to end
the call (invalidate_runtime_client can't abort the in-flight botocore stream,
so the streak escalates across turns like the OpenAI path), and reset the
streak on success.
- local providers still got float('inf') stale timeout (watchdog disabled) — give
a finite ceiling (HERMES_LOCAL_STREAM_STALE_TIMEOUT, default 900s).
- tests: on_event per-event + swallow; Bedrock stall bumps streak + aborts;
pre-elevated streak aborts at entry; success resets streak.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A cron job ("Daily Buzz Report") died with 'AIAgent' object has no
attribute '_claim_stream_writer'. The #65991 single-writer fence lives on
AIAgent (run_agent.py), but the streaming paths that use it live in other
modules — chat_completion_helpers (chat / anthropic / bedrock) and
codex_runtime (codex responses) — and called it directly as
agent._claim_stream_writer() / agent._stream_writer_is_current(). That makes
those modules hard-depend on the method being present on whatever object is
passed as agent.
The fence is an *additive* safety net that may only ever drop a provably
superseded stream, never the sole legitimate writer. But the direct calls
turned any agent that doesn't expose it — a version-skewed checkout (the
streaming helper module newer than run_agent), a hot-reloaded gateway mid
git-pull, a duck-typed agent, or a test double — into a fatal AttributeError
that aborts the whole turn (and, on cron, fails the job).
Route every cross-module claim/check through agent/stream_single_writer.py.
claim_stream_writer(agent) returns 0 when the fence is unavailable (or
raises), and stream_writer_is_current(agent, token) treats a 0 token or an
absent guard as "current" — so a guard-less agent degrades to "no fence"
instead of crashing, while a real AIAgent keeps the full single-writer
protection. Internal self.* uses inside run_agent are unchanged (self is
always a full AIAgent there).
When the stale-stream detector reconnects past a stream whose socket abort
raced (the close never actually stopped the old worker), the superseded stream
and the retry's stream both write deltas into the same turn. The persisted
transcript is then two coherent responses interleaved token-by-token —
de-interleaving the stored text by alternation yields two complete, independent
answers to the same prompt, which is a dual-writer race in the harness, not a
model/context failure (#65991).
The interrupt path already positively cancels before force-closing (#6600), but
the stale-kill path relies only on the socket abort, and nothing fenced late
chunks from a superseded stream out of the shared delta sink.
Enforce a single-writer invariant on the sink itself, guarded by attempt id
rather than only socket state: every streaming attempt (chat_completions,
anthropic_messages, and bedrock paths) claims a monotonic writer token before
it begins consuming its stream. A newer claim supersedes any older one, so the
consume loop bails the instant it is superseded and _fire_stream_delta /
_fire_reasoning_delta / _record_streamed_assistant_text drop chunks from a
stale writer. The token is stored per-thread, so a thread that never claimed
(a non-streaming delta caller) is never fenced — the guard can only ever drop a
superseded stream, never the single legitimate writer. Discards are counted and
logged sparsely so a real provider problem stays visible instead of being
silently swallowed.
Only advertise finite watchdog deadlines that are still in the future, exercise the full MoA heartbeat path, and register the salvaged contributor attribution.
