fix(aux): title generation recovers from stacked parameter rejections on primary and fallback routes (#78273, #72351, #109774, salvage #72515) by teknium1 · Pull Request #113958 · NousResearch/hermes-agent

teknium1 · 2026-09-17T07:53:13Z

Auxiliary title generation (and every other aux task) now recovers when a model rejects more than one request field, when a fallback candidate rejects a field, and when a strict-schema gateway rejects the reasoning field — instead of surfacing a misleading first 400 and leaving the session untitled.

agent/auxiliary_client.py::_ladder_parameter_rungs: the parameter rungs (temperature / structured-output format / reasoning field / output cap) are a table walked until no rung matches, each field stripped at most once — N rejected fields recover in N retries in whatever order the provider raises them. _param_rung_accepts now admits a temperature 400 raised by a later retry.
agent/auxiliary_client.py::_is_unsupported_parameter_error recognises the strict-pydantic phrasing Extra inputs are not permitted, field: '…', so the existing reasoning-strip rung fires for gateways that reject the generic extra_body.reasoning fallback.
New sibling agent/auxiliary_fallback_recovery.py: fallback candidates (per-task chain, main chain, discovery; sync + async) run the same parameter rungs around their request; auth / payment / connection errors still reach the caller's existing refresh-and-quarantine handling.
agent/title_generator.py sends the provider-default temperature instead of forcing 0.3 (salvage fix(agent): use provider-default temperature for title generation (#72351) #72515), so default-only models (gpt-5.x) accept the first title request and the daemon-thread retry no longer races CLI cleanup.

Live repro:

before (origin/main, real gpt-5-mini title_generation call_llm): reasoning_effort 400 → reasoning stripped → temperature 400 re-raised, task failed / after: reasoning stripped → temperature stripped → 200 hello.
before (origin/main, real gpt-5-mini via _call_fallback_candidate_sync and _async with temperature=0.3): BadRequestError: Unsupported value: 'temperature' raised, no retry / after: one retry without temperature → 200 on both.
before (origin/main, stand-in replaying Fireworks' documented 400 Extra inputs are not permitted, field: 'reasoning' — no Fireworks credential here, so a local OpenAI-compatible server replays the body logged in Auxiliary title_generation sends reasoning extra_body to providers without a reasoning-aware profile (Fireworks HTTP 400: Extra inputs are not permitted, field: 'reasoning') #109774; docs: https://docs.fireworks.ai/api-reference/post-chatcompletions): 1 request, HTTP 400, title failed / after: 2 requests, second without reasoning, 200 title.
Control: the stand-in's first request still carries reasoning (the generic fallback is unchanged for gateways that accept it); the temperature-accepting path is untouched (tests/agent/test_unsupported_temperature_retry.py green).

Root cause: the parameter rungs ran once in a fixed order, so a field rejected after an earlier rung had run was never stripped; fallback candidates bypassed the rungs entirely; and the strict-schema "Extra inputs are not permitted" phrasing matched no unsupported-parameter predicate.

Tests: tests/agent/test_auxiliary_parameter_rung_chaining.py (2 invariants, red on origin/main via source swap); focused suites test_unsupported_temperature_retry.py, test_auxiliary_auth_rung_fallthrough.py, test_title_generator.py, test_compression_fallback_budget.py, test_fast_compression_lane.py, test_auxiliary_client.py → 307 passed.

Fixes #78273
Fixes #72351
Fixes #109774

Salvages #72515 (@JonthanaHanh, cherry-picked) and #109807 (@huklaa, Fireworks profile override — ported in the review follow-up, extended with the effort mapping). Supersedes #78321 (@686f6c61, +259/-90 peel loop — same intent, landed here as the table walk on the current generator ladder) and #110809 (@sy0u1ti, host allow-list gate): the reactive rung covers every strict gateway without a per-vendor table.

Dropped hunks

fix(agent): peel stacked unsupported auxiliary params (temp + max_tokens) #78321: max_tokens → max_completion_tokens translation on retry — the preflight auxiliary_max_tokens_param already picks max_completion_tokens for gpt-5-family model names; the retry drops the cap as before (re-probed: gpt-5-mini receives max_completion_tokens on the first request).
fix(auxiliary): gate extra_body reasoning to supported providers (#109774) #110809: hard-coded host allow-list for extra_body.reasoning — replaced by the reactive strip so unknown strict gateways recover too.
fix(auxiliary): map Fireworks reasoning disable #109807: initially not taken; now ported (see review follow-up) — Fireworks documents top-level reasoning_effort (none disables thinking), so the profile override removes the extra 400 round-trip the reactive rung alone left in place.

Review follow-up (99ff3b9):

_param_rung_accepts lacks _is_rate_limit_error (MAJOR) → fixed. Reproduced on c187a45 with the reviewer's probe (max_tokens strip, retry raises openai.RateLimitError 429 → raised out of _ladder_parameter_rungs; origin/main's max_tokens rung accepted rate limits). Added _is_rate_limit_error(exc) to the predicate; invariant test test_rate_limit_after_parameter_strip_falls_through_to_later_rungs in tests/agent/test_auxiliary_parameter_rung_chaining.py (red on the old facade via source-file swap, green now).
Fixes #109774 while the generic extra_body.reasoning still went to Fireworks unconditionally (MINOR) → fixed by porting fix(auxiliary): map Fireworks reasoning disable #109807 (@huklaa, Co-authored-by): FireworksProfile.build_api_kwargs_extras emits top-level reasoning_effort (none for thinking-off, the clamped effort otherwise) and marks the profile reasoning-aware, so the transport omits the generic fallback on that route. Parametrized TestFireworksReasoning::test_auxiliary_reasoning_wire_shape covers thinking-off, effort, and a control that a profile-less route still gets extra_body.reasoning. Fixes #109774 stands.
Positional _LadderRoute construction → refuted / already resolved on c187a45 (route built by field name; test builds it via _LadderRoute._fields).

Infographic

…2351) Title generation hardcoded temperature=0.3 in its call_llm() call. Models like GPT-5.6 only accept their server-side default temperature and reject explicit values with "Unsupported value: 'temperature'". While call_llm() has a retry that strips temperature on error, the daemon thread races with session cleanup in short-lived CLI sessions, causing the retry to fail with a connection error. Fix: pass temperature=None so the provider uses its own default. This avoids the unsupported-temperature error entirely and eliminates the need for the retry path. Fixes #72351

…ack requests Reasoning models reject several request fields at once (gpt-5: temperature AND max_tokens, #78273), a reasoning-strip retry can then 400 on temperature (#72351), and strict-schema gateways such as Fireworks reject the generic extra_body.reasoning fallback with "Extra inputs are not permitted, field: 'reasoning'" (#109774) — a phrasing none of the unsupported-parameter predicates recognised, so the reasoning rung never fired and every title call 400'd. The parameter rungs were a fixed single-pass order (temperature → structured output → reasoning → max_tokens): a field rejected AFTER an earlier rung had already run was never stripped, and _param_rung_accepts did not admit a temperature 400 raised by a later retry. They now form a table walked until no rung matches, each field stripped at most once, so N rejected fields recover in N retries in whatever order the provider raises them. Fallback candidates (per-task chain, main chain, discovery) got a single shot with the caller's temperature/max_tokens/reasoning fields and raised on the first parameter 400; agent/auxiliary_fallback_recovery.py now runs the same parameter rungs around a candidate's request (sync + async), leaving auth / payment / connection errors to the caller's existing handling. Live: real gpt-5-mini title_generation call (reasoning_effort 400 → temperature 400 → 200), real gpt-5-mini fallback candidate sync+async (temperature 400 → 200), and a stand-in replaying Fireworks' documented 400 (reasoning stripped → 200) — all failed on origin/main.

github-actions · 2026-09-17T07:53:46Z

૮ >ﻌ< ა ci review

_{ran on 9e6a818 — fix(aux): max_tokens rung retries even when the wire kwargs}

all good!

A positional _LadderRoute(...) in auxiliary_fallback_recovery breaks as soon as the route tuple gains a field (#113968 adds timeout); construct by name so fields the parameter ladder never reads default to None regardless of tuple width.

…lds the route by name base_info is the endpoint identity every rung reads for per-route bookkeeping; the fallback ladder left it empty. The test mirrored the positional construction that broke under a wider tuple.

@huklaa

… reasoning_effort The pre-ladder max_tokens rung accepted payment|connection|rate_limit errors on its retry, so a 429 after a stripped retry reached the credential and provider-fallback rungs. The ladder's shared `_param_rung_accepts` dropped rate_limit: a 429 on the retry raised out of the primary call and skipped the whole fallback chain. Accept it there too, with a rung-level test that the stripped kwargs and the 429 are handed on rather than raised. The body claimed `Fixes #109774` while the aux client still sent the generic `extra_body.reasoning` to Fireworks on every call and only recovered reactively (an extra 400 round-trip per request). Port @huklaa's profile override from #109807: Fireworks documents top-level `reasoning_effort` (`none` disables thinking), and overriding `build_api_kwargs_extras` marks the profile reasoning-aware so the transport omits the generic fallback on this route. Extends the salvage with the effort mapping so enabled-with-effort takes the same wire, with a control that a profile-less route keeps the fallback. Salvages #109807 (@huklaa). Co-authored-by: Hukla <129692708+huklaa@users.noreply.github.com>

… carry the cap The rung table refused to re-send an unchanged request, but the Codex Responses route translates the caller cap away and gateways inject their own: the 400 names max_tokens while the kwargs show none, and the identical retry is what completes (tests/agent/test_injected_param_strip_retry_registry.py, red in CI on 99ff3b9).

…ved by the rung table Since #113958 every fallback candidate runs the shared parameter rung table (send_with_parameter_rungs), whose structured-output rung strips the field and records the rejection through its remember column. The fallback-side copy could never fire any more; delete it rather than keep two seams for one behaviour.

JonthanaHanh and others added 2 commits September 17, 2026 00:47

alt-glitch added type/bug Something isn't working comp/agent Core agent runtime: loop, agent_init, prompt builder, context-compression, responses endpoint P2 Medium — degraded but workaround exists labels Sep 17, 2026

teknium1 and others added 3 commits September 17, 2026 01:15

teknium1 mentioned this pull request Sep 17, 2026

fix(aux): structured-output 400s no longer kill fallback candidates or cost a doomed first request on DeepSeek (#83390, #105191, salvage #92908) #113966

Merged

teknium1 merged commit cbe2413 into main Sep 17, 2026
34 checks passed

teknium1 deleted the fix/baux-B-title-param-rejections branch September 17, 2026 15:45

teknium1 mentioned this pull request Sep 17, 2026

fix(agent): omit reasoning effort for custom title routes #112789

Closed

9 tasks done

Navigation Menu

fix(aux): title generation recovers from stacked parameter rejections on primary and fallback routes (#78273, #72351, #109774, salvage #72515) - #113958

fix(aux): title generation recovers from stacked parameter rejections on primary and fallback routes (#78273, #72351, #109774, salvage #72515)#113958
teknium1 merged 6 commits into
mainfrom
fix/baux-B-title-param-rejections

teknium1 commented Sep 17, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Sep 17, 2026 •

edited

Loading

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

teknium1 commented Sep 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Dropped hunks

Infographic

Uh oh!

github-actions Bot commented Sep 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

૮ >ﻌ< ა ci review

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

teknium1 commented Sep 17, 2026 •

edited

Loading

github-actions Bot commented Sep 17, 2026 •

edited

Loading