Stored tool call ids are minted per turn (terminal:0, terminal:1…), so the
same id recurs on later turns of one session. The Responses adapter replayed
each pair verbatim, putting the same call_id on the wire N times; strict
validators (opencode Console, OpenAI) reject the request with 400
"Duplicate function_call_output for call_id" and every retry fails the same
way (#102629, #111231 bug 1).
Every occurrence past the first now gets a `_dup<n>` wire id and the matching
tool output pops the id its function_call was given, in call order, so pairs
stay intact and stored history is untouched.
Re-applied onto the facade split (helpers moved into _replay_tool_call_items /
_tool_output_items) from PR #102634.
A model that declines mid-stream delivers the explanation on the
structured refusal channel (chat_completions delta.refusal; Responses
response.refusal.delta / refusal content parts) and leaves content
empty. The streaming accumulators dropped that channel entirely, so a
streamed refusal assembled into an empty message and fell into the
empty/invalid-response retry loops - burning paid retries reproducing a
deterministic refusal - while the non-streaming path had already fixed
this class in #46013.
- chat_completions streaming: accumulate delta.refusal (incl.
model_extra), expose message.refusal on the assembled mock response so
ChatCompletionsTransport.normalize_response applies the existing
sole-payload -> content_filter promotion; count refusal deltas in the
zero-chunk guard; carry refusal in the Relay final-response dict.
- Codex Responses stream consumer: collect response.refusal.delta as
answer text so a refusal-only stream no longer raises 'did not emit a
terminal response' with zero usable content.
- Responses normalizer: read type=refusal content parts in
_extract_responses_message_text (attr and dict shapes).
Sabotage-verified: each new test fails with its wiring line disabled.
E2E: refusal-only stream -> terminal content_filter with explanation;
refusal-alongside-content stays a normal usable turn; plain-text
streams unchanged.
`_classify_responses_issuer` reimplemented endpoint canonicalisation with
its own urlsplit/urlunsplit pass. The repo already owns that logic in
`hermes_cli/route_identity.py::normalize_route_base_url` (stdlib-only,
used by agent/backend_identity.py), so delegate to it.
Reasoning items persisted before canonicalisation were stamped with the
raw `other:<agent.base_url>` (trailing slash, host case). Comparing them
verbatim against the now-canonical `current_issuer_kind` marked them
foreign and dropped them on the very endpoint that minted them. Run the
persisted stamp through the same canonicaliser (`_canonical_issuer_kind`,
non-`other:` kinds untouched) before comparing.
Also fixes the `_chat_messages_to_responses_input` docstring, which still
stated the pre-stack rule that legacy endpoint-stamped items drop when the
current model is known; the stack replays them on a matching issuer.
The openai SDK appends a trailing slash to `client.base_url`, so the aux
adapter stamped `other:https://h/v1/` while the main transport stamped
`other:https://h/v1`. On custom Responses endpoints every aux call
(compression, flush_memories) therefore dropped all main-minted reasoning
items as "foreign".
`_classify_responses_issuer` now strips whitespace and trailing slashes and
lowercases scheme+netloc before stamping. The aux adapter also derives its
route flags from `classify_responses_route` — the single owner of the
codex/xai/github predicates — instead of an inline chatgpt.com host check,
and reuses the same flags for the effort clamp.
Native compaction checkpoints and reasoning items persisted before model
stamping existed carry only `_issuer_kind`. Treating a missing `_issuer_model`
as foreign dropped every such item once the current model was known, which
wiped existing sessions' native-compaction context on upgrade (four consumer
tests in test_native_compaction / test_native_preflight_estimate /
test_413_compression went red on the stack).
Trust the endpoint stamp when no model stamp is present, as main does today.
Items minted after this change carry the model stamp and still drop on a
same-endpoint model switch; a wrong guess on a legacy item is caught by the
invalid_encrypted_content 400 classifier and the replay kill switch.
Encrypted reasoning blobs are sealed to the model that minted them, not
only to the endpoint. Switching models on the same custom Responses
endpoint therefore replayed blobs the new model cannot decrypt and the
turn failed with HTTP 400.
Stamp captured reasoning items with `_issuer_model` (the canonical wire
model) alongside `_issuer_kind`, and replay an item only when both the
issuer kind and the model match the current request. Endpoint-stamped
legacy items without model provenance are dropped once the current
model is known (fail closed); ordinary assistant text stays replayable.
The transport threads the effective wire model (request_overrides win)
into conversion and normalization; the auxiliary Codex adapter stamps
and filters against its own model rather than the main agent's. The
400 classifier also recognises the custom-endpoint wording
"encrypted content could not be decrypted or parsed" so recovery strips
the replay state instead of aborting.
Hand-grafted from #95849 (final head d9cf6bcc08) onto current main; the
middleware-model-rewrite half is intentionally left out.
Closes#95834
Every direct-API (api.openai.com) Astra request raised
``TypeError: Responses.create() got an unexpected keyword argument 'prompt_cache_options'``
before reaching the network: openai 2.24.0's Responses.create has no such parameter and no
**kwargs, and neither send path relocates it into extra_body. The PR's tests stopped at
build_kwargs/preflight so the SDK boundary was never crossed.
OpenAI's prompt-caching guide states ``prompt_cache_options.ttl`` accepts only ``30m`` and that
``30m`` is the default, so the field carried no information: sending nothing yields the same
cache lifetime. The sanitizer now only removes what the API rejects (none/minimal effort,
sampling/logprob knobs, the pre-5.6 ``prompt_cache_retention``) and never adds a field, which
also keeps the request body byte-stable for the cache prefix.
Also: none/minimal→low no longer needs a bespoke {"", "none", "disabled", "off"} set —
``clamp_effort`` against CODEX_ASTRA_EFFORTS already resolves to the floor (``low``); and the
auxiliary adapter derives ``is_codex_backend`` from ``classify_responses_route`` (the declared
single owner of that predicate) instead of re-implementing the host test inline.
Tests reshaped to contracts: the two proxy/subdomain cases collapse into one parametrised
"exact host only" test asserting effort and temperature pass through untouched.
A native Responses compaction checkpoint is opaque ciphertext; the rough
preflight estimator counts it as text (5.17M chars -> ~1.29M tokens against
a 204K trigger) and fires local compression on a request whose real prompt is
~116K. Arm the existing one-response real-usage latch when a replayable
checkpoint is captured (build_assistant_message) or restored into a fresh
agent (_hydrate_from_history), honor it in the post-tool gate and idle
compaction, and require non-empty encrypted_content for a checkpoint.
Squash of the author's source commits from #100642 (0e3c234ea0, 771e1b3365,
bb1505a119) plus the fdf140c81d test refresh, re-based onto current main by
patch application. Source delta is byte-identical to the PR head d6ce3e236d.
Fixes#100611
For each issue anchor present in BASE 63279301bc non-test .py and absent on HEAD, the BASE comment/docstring block was re-attached at the HEAD location of the code it explained (matched by the distinctive code line / enclosing def). Sentences already covered by an existing HEAD comment were deduped; the issue number always survives. Insert-only: no code lines changed.
Split _preflight_codex_input_items into per-item-type helpers over a
_PreflightCtx; streaming assembly moves into _CodexResponseAssembler with a
per-event dispatch table; run_codex_app_server_turn loses its session/usage/
interrupt regions to _ensure_codex_session/_finish_codex_turn/
_consume_user_interrupt/_queue_token_counts (counts built lazily so stub agents
without a session DB are never touched). Responses input/normalization and
stream callback order verified byte-identical against merge-base.
Follow-up to the #96217 salvage: the codex/xai/github route checks were
re-implemented inline at four sites (codex_responses_adapter helpers,
chat_completion_helpers kwargs build, _is_openai_codex_backend, the
run_agent silent-reject hint). Consolidate them into
classify_responses_route() / ResponsesRouteFlags in
codex_responses_adapter and migrate every site — backend-identity
predicate class (#22548/#70893/#59561/#72468).
Host checks use exact-host-or-subdomain semantics, never substring
matching.
Automatic preflight used the full durable transcript even when the
Codex Responses request would prune around a native compaction
checkpoint. That false-triggered a 600s local summary against history
the main request never sent. Estimate the converted, checkpoint-pruned
payload when native compaction is eligible, and keep the generic
estimate as the conservative fallback.
/simplify-code reuse+quality reviewers both flagged the byte-identical
fc_->call_ synthesis blocks in the assistant and tool-result branches as
a correctness coupling — the two sites MUST stay in lockstep or pairing
breaks. Extract _canonical_call_id_from_fc() and route both through it.
Mutation check: pairing regression test fails when the tool-branch call
is stubbed out, green after restore.
The sweeper review on #49224 flagged that the assistant branch synthesizes
call_<suffix> from an fc_-only id while the tool-result branch kept the raw
fc_... string — so an oversized pair hashed to two DIFFERENT clamped
surrogates and the function_call_output arrived unmatched (HTTP 400).
Canonicalize the tool-result side to the same call_<suffix> before
clamping. Also fixes the pre-existing short-fc_ pairing mismatch
(call_short123 vs fc_short123). Regression test covers both lengths.
A degenerate tool name stored in conversation history (dots, spaces,
unicode from an earlier model degeneration) bricks every subsequent
Codex Responses turn with a non-retryable HTTP 400:
Invalid input[N].name: string does not match pattern '^[a-zA-Z0-9_-]+'
The 400 replays forever until the user manually starts a new session.
Add _sanitize_replayed_fn_name() — replaces invalid chars with '_'
(runs collapsed), degrades all-invalid names to 'fn' instead of empty
(an empty name would trade one 400 for a preflight ValueError). Applied
at both replay sites: the chat-message converter and the preflight
choke-point. Live tool-definition names are left untouched — they must
match the dispatch registry exactly. Pairing is by call_id, so
renaming a replayed function_call is safe.
call_id overflow (the sibling half of #49224) was already fixed on main
by #73492 (_clamp_responses_call_id); this commit covers the remaining
invalid-name defect.
Credit: @Morad37 (#31678 — identified the bug, the replay sites, and
the regex contract), @lubosxyz (#49224 — replace-not-strip semantics
and 'fn' fallback to avoid the empty-name trap).
Fixes#31666
prune_pre_checkpoint_items() had a hardcoded role=='user' filter that
discarded all non-user messages before a checkpoint — including Hermes'
own compression summaries (role='assistant'), causing total context amnesia
about past conversation summaries.
The fix:
- _is_summary_item delegates to the canonical
agent.context_compressor.is_compaction_summary_message provenance check
(not an ad-hoc heuristic)
- Summaries are retained whole (never byte-sliced) within a 32k token budget
- Idempotent across repeated checkpoints (dedup by identical text)
- _chat_messages_to_responses_input threads item_sources (raw chat messages)
through to the pruner, so it can read summary content directly from the
source when the Responses conversion shape is lossy (tool-result carrier
becomes function_call_output, or stale codex_message_items replay shadows
merged content)
Fixes#90975.
Salvage of #90976 by @JoaoMarcos44.
A captured native-compaction checkpoint lives in the persisted
codex_reasoning_items sidecar, but the wire restructure that follows it
(prune_pre_checkpoint_items) ran unconditionally: the native gate only
decided whether context_management went into the request, and no signal
from it ever reached _chat_messages_to_responses_input.
So a single checkpoint kept deleting every pre-checkpoint item from all
later requests — after a mid-session swap out of the gpt-5.6 family,
after compression.enabled: false, after the rejection kill switch, and
after a session resume that reloads the sidecar from state.db. The model
receiving the opaque blob was no longer the one able to decode it, and
nothing was logged.
Thread a single native_compaction_eligible boolean, derived from the same
value that gates the context_management field, into the converter. When
ineligible: do not replay type: "compaction" items and do not prune. Safe
because native compaction never truncates Hermes' local history, so the
fallback still carries the full conversation.
All Responses call sites are covered: build_kwargs and convert_messages
derive the flag via _native_compaction_active, the auxiliary/compression
client is explicitly ineligible, and the converter defaults to False
(pre-feature wire) so future call sites are safe by construction.
Fixes#85914
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Live verification (gpt-5.6 @ api.openai.com) proved the Responses server
renders NOTHING placed before a replayed compaction checkpoint: a fact
stated in a pre-checkpoint input item is invisible to the model, while the
same item after the checkpoint recalls perfectly. Hermes was replaying the
full pre-checkpoint transcript anyway — dead upload weight, and worse, every
plaintext user ask from before the boundary silently vanished from the
model's view, surviving only inside the opaque server summary. That is the
goal-drift failure mode reported against native compaction sessions.
Codex CLI never hits this because it rebuilds history client-side after
compaction, retaining user messages verbatim under a token budget. This
change is the wire-level equivalent: when a replayed checkpoint is present,
_chat_messages_to_responses_input restructures the input as
[newest checkpoint run] + [retained pre-checkpoint user messages,
newest-first within a 64K-token budget] + [post-checkpoint tail]
Histories without a checkpoint are returned unchanged, so non-native
sessions see a byte-identical wire.
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.
The codex app-server namespaces MCP tool call ids as
codex_mcp__<server>__<tool>_<codex_call_id>. With an exec-<uuid> component the
built-in hermes-tools server alone overflows the Responses API's 64-char
call_id limit, so the request 400s with a non-retryable "string too long".
The offending item sits near the head of the transcript and replays every
turn, permanently bricking the session — the only recovery is /reset.
Sibling defect to #10788, which clamped input[*].id via
_MAX_RESPONSES_ITEM_ID_LENGTH. Apply the same treatment to call_id at both
Responses emit sites in _chat_messages_to_responses_input: a deterministic
surrogate (call_ + sha256[:32]) for ids over the limit, short ids unchanged.
Because the surrogate is a pure function of the original id, a function_call
and its matching function_call_output — which carry the same original id — map
to the same surrogate and stay paired without correlating the two items.
OpenAI documents GPT-5.5 / GPT-5.5 Pro as extended-cache-only: in-memory
prompt cache retention is not available for them, and only
prompt_cache_retention: "24h" is supported. Responses requests that omit
the field see near-zero cached_tokens even with a stable prompt_cache_key
and identical prefixes (observed on an OpenAI-compatible Responses relay:
0 cached across repeated identical calls before; 97% cache reads after).
Send the field for the gpt-5.5 model family (bare and namespaced ids like
openai.gpt-5.5) on OpenAI-compatible Responses routes, mirrored in the
auxiliary Codex adapter, and pass it through preflight normalization.
Skipped for xAI, GitHub/Copilot, and the chatgpt.com Codex backend, which
reject or ignore body-level cache fields.
Map Codex Responses status=incomplete with incomplete_details.reason=content_filter to finish_reason=content_filter so the existing refusal/fallback path runs instead of burning incomplete continuation attempts.
grok-4.x on the xAI /v1/responses surface sometimes ends a turn with only
reasoning items — no message output item, no tool calls — and those
reasoning items carry no encrypted_content. Two compounding problems:
1. The model occasionally emits its final answer INSIDE the reasoning
channel, delimited by grok's internal "<response>" tag. The answer
exists but is classified reasoning-only → finish_reason=incomplete.
2. An interim assistant message holding only plain-text reasoning replays
as nothing in _chat_messages_to_responses_input, so every continuation
request is byte-identical to the one that just failed. The model
deterministically repeats the reasoning-only response until the retry
budget is exhausted and the turn dies with "Codex response remained
incomplete after 3 continuation attempts".
Fixes:
- _normalize_codex_response (xai_responses only): salvage the
<response>-delimited tail from the reasoning text and promote it to
assistant content; the untagged prefix stays as thinking text.
- Codex-incomplete continuation path: when the interim message has
nothing the input converter will replay (no content, no encrypted
reasoning items, no message items), append a user-role nudge so the
retry actually differs and explicitly asks for the final answer /
pending tool call. Mirrors the existing _get_continuation_prompt
pattern used for length truncation.
Observed live with grok-4.20 on xai-oauth (2026-07-13); sibling of the
grok-composer web_search incomplete-loop fix in transports/codex.py.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Follow-up to the salvaged #64449: the status-trusting branch flipped
github_responses to 'stop' alongside unknown relays. Copilot fronts the
same OpenAI model family as codex_backend and shows the same
reasoning-only 'still thinking' degeneration, so it stays on the
continuation path. Only unrecognized (other:*) backends trust
response.status='completed' as terminal.
chelsealong