The transport wiring (build_kwargs model -> _last_issuer_model -> normalize_response
stamp) and the new error_classifier string had no test that fails when either is
removed; these two do.
OpenAI documents credit_balance_exhausted / *_spend_limit_exceeded /
organization_usage_limit_exceeded as HTTP 429 responses, so the invariant
must hold through _status_429 (which short-circuits _by_error_code), not
only on the status-less body path.
Clean-room port of the billing-code coverage from zed-industries/zed#63208: credit_balance_exhausted, organization_spend_limit_exceeded, project_spend_limit_exceeded, organization_usage_limit_exceeded now classify as billing (rotate + fallback) instead of falling through to generic buckets.
Port from code-yeongyu/oh-my-openagent#6677 (credit: @niStee).
LiteLLM proxies stamp a structured `terminal_quota_exhausted` code on
hard-cap 429s. Hermes' `_status_429` handler always returns a verdict, so
`_by_error_code` (which maps _BILLING_ERROR_CODES to billing) never saw
the code: the exhausted key classified as rate_limit, earned the 429
cooldown, and got retried against a wall that cannot clear until someone
pays. Upstream this respawned duplicate subagent sessions.
- `_status_429` now honors a structured billing code first (decisive
signal outranks message heuristics).
- `terminal_quota_exhausted` joins _BILLING_ERROR_CODES so every path
(429, 402, status-less) agrees.
- "hard billing limit" free text joins _BILLING_PATTERNS ("billing hard
limit" was already there; providers use both orders). "terminal billing
limit" text is deliberately NOT matched: substring rules cannot negate
the "non-terminal billing limit" wording — the structured code covers it.
* fix(auth): close the free tier's gaps against the gateway's welcome-tier contract
The inference gateway's welcome tier (NousResearch/api DOCS/anon-tier/plan.md) serves an
anonymous account exactly one model on its own host, refuses everything else with a structured
429, cross-refuses a request on the wrong host with a 400 (403 while the tier is dark), and
tells a signed-in account that still asks for `nous/welcome` what to switch to in an
`x-nous-model-switch` header. Four client-side gaps against that contract:
- Auxiliary calls were refused on every session. The auxiliary client asked the welcome host
for the Portal's recommended compaction/vision model, a guaranteed 429 `model_not_free`
before each fallback. On the welcome host it now uses `nous/welcome` (its backing model
covers auxiliary work) and skips Nous for vision, which the welcome model does not take.
- The structured 429 body was never read. The classifier now parses `reason` /
`retry_after` / `alternates` / `upgrade_url`: `model_not_free` and `feature_not_free` are
non-retryable gates that fall back; `at_capacity`, `admission_closed` and `rate_limited`
are rate limits that honour `retry_after` and never rotate the free tier's only credential.
The wrong-host 400 and the dark-tier 403 are deterministic, so they abort this route and
fall back instead of retrying or re-exchanging. The terminal paths say what happened and
name the sign-in (`/login` in a chat, `hermes auth upgrade` in a terminal).
- The `x-nous-model-switch` header was ignored. The chat-completions transport records it
beside the rate-limit and credits headers; the next call moves the session, and the config
default when it still names `nous/welcome`, to the backing model the gateway named.
- A guest fell back to the paid host. With `inference_base_url` absent from the exchange or
outside the host allowlist, routing defaulted to inference-api, where every request is a
400. A guest now defaults to the welcome literal at the exchange, in the shared store's
shape, and in effective routing.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
(cherry picked from commit fc758aad7efceff6223fc144a9b5c69f13e41bd8)
* feat(auth): the free tier is set up on request; nous.guest_setup decides whether also on first use
A caller that names nous/welcome on a Nous route with no Nous identity in reach — the guided
setup's session (provider=nous, which skips the resolver's nothing-configured rung), the free-tier
picker row, a bare --provider nous pointed at it — is asking for the free tier. The OAuth runtime
rung now sets it up there instead of failing "not logged in", so the guided chat no longer races
the root profile's first-run mint.
nous.guest_setup is the policy seam: "auto" (default) keeps today's first-use setup wherever
nothing else is configured; "on-request" mints only when the free tier is asked for by name
(nous/welcome, /login, hermes auth upgrade, replacing a retired identity). Implicit callers —
the resolver's last rung, the first-run check, free_tier.status, the CLI's background setup, the
connector token path — still adopt what the shared store holds, so every profile follows the one
identity the guided setup created, but never create one on their own.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
(cherry picked from commit ae915ddc65ecdb81b81e29b604671d15cd49233c)
(cherry picked from commit 62ad1ff3ab200ea064975a32c502041b25910165)
* feat(auth): the guided setup provisions the free tier explicitly; nous.guest_setup is auto | explicit
Two questions govern the free tier: may it exist (nous.guest) and who may CREATE the identity
(nous.guest_setup). "auto" (default) keeps today's first-use setup wherever nothing else is
configured. "explicit" means Hermes never creates one on its own: the only creator is the new
provision_free_tier() primitive, exposed as the free_tier.provision RPC, which the guided setup
on Hermes Desktop calls as its first step — on the root gateway, before the setup profile and
before the guided chat exists — so the identity lands in the root store every profile reads
through and is there before any session asks for nous/welcome. That closes the race against the
backend's own setup, and makes "only when the setup-bot flow is used" literally true.
The earlier "on-request" tier is replaced: it minted whenever any caller named nous/welcome
(the hermes model row, --provider nous), which treated a model name as intent and was broader
than the guided setup. Under "explicit" a nous/welcome request with no identity fails "not
logged in" as before the free tier existed, and /login or hermes auth upgrade report nothing to
sign in from. Implicit callers still adopt an identity the shared store holds, and a retired
credential is replaced (a continuation, not a creation).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
(cherry picked from commit c63d2c935c1e59016164fdfb90cf70b4094466a0)
* fix(auth): remove the nous.guest_setup knob; the free tier is created on first use
`nous.guest_setup: auto | explicit` decided who may CREATE the free-tier identity. Under its
default every line it added was inert (`may_mint` always true), nothing in tree set `explicit`,
unknown values read as `auto`, and under `explicit` a CLI-only install could never get an
identity, which contradicts the first-run contract (first command mints, then chats).
The mint race the knob accompanied is already benign: every caller takes the profile lock then
the shared-store lock, and the loser adopts what the winner wrote. What makes the guided setup
win deterministically is `provision_free_tier()` behind the `free_tier.provision` RPC, which
stays. `nous.guest` remains the only free-tier policy.
Removed: `guest_setup_policy()` and its constants, the `explicit=` / `may_mint=` threading through
`ensure_portal_identity` and `_reconcile_and_provision`, the flag at the three replacement call
sites (now no-ops), the config default, the docs section, and the four `guest_setup` test-config
entries. The three policy tests that hold regardless of the knob are kept under
`TestExplicitProvision`; the two that only tested the knob are deleted.
(cherry picked from commit d8a50526d93c374c0067dd935b5a65055e0af261)
* fix(gateway): a server-driven model switch off nous/welcome does not evict the cached agent
When a signed-in account still asks the paid host for `nous/welcome`, the inference gateway
serves the current backing model and names it in `x-nous-model-switch`. `apply_model_switch`
moves the live session to that model and moves `config.yaml`'s default off the alias in the
same step. The messaging gateway's fallback-eviction check compares the agent's model with the
config default and evicts on any mismatch that is not a /model override, so when the config
write did not land (unreadable config, lock) the cached agent was evicted once per turn, and
prompt caching with it.
`apply_model_switch` now stamps the alias it moved the session off on the agent, and
`_is_intentional_model_switch` treats "agent moved off the alias the config still carries" as
deliberate, beside the existing /model override case. The check takes the agent and the config
model instead of a bare model string; its one caller in `_run_agent_evict_on_fallback` passes them.
(cherry picked from commit 696d1ec86b69db28bf002c841e9389b85178a954)
* fix(auth): the free tier outranks implicit host credentials in provider resolution
On a fresh install with a leftover ~/.aws profile, resolve_provider("auto")
reached the Bedrock rung before the free-tier rung, so the first turn ran on
Bedrock and failed 403 while the free tier was still being minted in the
background at agent setup (NS-829). Live on a Mac with ~/.aws present: 28 s,
three retries, no answer; the next process then switched to nous/welcome.
The free-tier rung now sits directly above the Bedrock chain: when nous.guest
is on, an existing free-tier identity answers, else a blocking mint runs, and
only then does the boto chain get a say. Everything above is unchanged and
still wins: CLI creds, config.yaml model.provider, env keys, the OpenRouter
pool, a logged-in active_provider. nous.guest: false skips the rung, and a
failed mint still falls through to Bedrock and the no-provider guidance.
Tests: six precedence cases (identity present, fresh mint, free tier off, env
key still wins, sign-in still wins, failed mint falls through). The opt-out
test now neutralizes the AWS chain like the precedence tests do; on a machine
with ~/.aws it was failing for the same reason as the bug.
Live after the fix, same Mac, AWS credentials visible, isolated shared store:
identity minted 2 s in, turn on model=nous/welcome provider=nous, answer in
11 s.
(cherry picked from commit a04b05260cd334dd7199ad9b6cd5b2538364c75a)
* fix(auth): review follow-ups for the free-tier rung (NS-829)
- tests/agent/test_bedrock_integration.py: the Bedrock auto-detect test switches
the free tier off; its contract is the boto chain, and the free tier now
sits above it.
- gateway/run_notifications.py: the free-tier startup line reads auth.json
before consulting the resolver, so a gateway boot on a machine with AWS
credentials never mints or refreshes over the network.
- hermes_cli/anon_auth.py: module docstring says where the free tier sits in
the ladder instead of "the ladder is untouched".
- tests/hermes_cli/test_provider_precedence.py: two invariant tests instead of
six (parametrized ladder cases; a failed mint that returns None or raises
falls through to Bedrock).
scripts/run_tests.sh on the five affected files: 147 passed, 0 failed.
(cherry picked from commit 10790d148c60ada11b9ecdde2cd2c836c6a82a11)
* feat(auth): HERMES_GUEST_ONBOARDING=1 is the one launch gate for the free tier; HERMES_FORCE_GUEST is gone
The free tier is pre-GA. Until GA it must not exist for anyone who did not
ask for it: no identity minted, no portal traffic, no free-tier copy on any
surface. One environment variable now decides that, and one function reads it.
`guest_enabled()` returns False unless `HERMES_GUEST_ONBOARDING` is exactly
"1"; only then does `nous.guest` (the user's off switch) get consulted. Every
free-tier site already funnels through `guest_enabled()`, so the gate closes
minting, routing, connector entitlement, status lines and the picker row in
one place. With the variable unset, `resolve_provider("auto")` on a fresh
install raises `no_provider_configured` exactly as upstream does.
`HERMES_FORCE_GUEST` and `force_guest_mode()` are removed. They inverted the
gate (forced the tier ON over `nous.guest: false`), their "new" value re-minted
identities as a side effect of provider resolution, and `_has_any_provider_
configured` read them ahead of every other check, making the CLI a second
reader of a flag that must have exactly one. `_forced_new_done` and the
`force` parameter of `_reconcile_and_provision` go with them.
Supersedes the dev lever introduced in fcf9d11679 (rung 1) and hardened in
b5c162c3ec. Ruling: NS-845 Q1.1 (recorded on NS-847).
Not a user preference: the variable is never written to config.yaml or .env
and never shown in setup. It is deleted at GA together with its comment in
anon_auth.py. This is a deliberate, temporary exception to the "no new
HERMES_* env vars for non-secret config" rule.
Tests: fixtures set the gate instead of deleting the old lever; one new
invariant (`test_launch_gate_off_means_no_free_tier_at_all`) proves that "",
"0", "true" and "new" all leave the tier off with zero portal calls, red on the
previous commit. The `HERMES_FORCE_GUEST=new` re-mint test is deleted with the
feature.
* feat(auth): the free-tier identity is created in one place, at boot; every other site is a read
Before this commit eight sites could create a Nous free-tier identity as a
side effect of something else: resolving a provider, the CLI's first-run
check, the CLI's session setup (in the background beside an own key), a
connector bearer read, the desktop polling `free_tier.status`, the sign-in
precondition, the desktop's `free_tier.provision`, and the dead-credential
re-mint. A poll could mint. Provider resolution could hit the network. Two
of them raced each other on a fresh install.
Now `hermes_cli/free_tier_bootstrap.py::run_bootstrap` is the only creator.
`hermes serve` runs it on a daemon thread from `_lifespan` beside the other
background boots; `cmd_chat` runs it synchronously before the first-run
guard. It inventories credentials first (`resolve_provider("auto",
skip_free_tier=True)`: what would carry inference if the free tier did not
exist), creates the identity only when `guest_enabled()`, resolves inference,
records a `SetupRecord` in process memory and broadcasts ONE `setup.ready`
event. It runs on every boot; only the mint is gated.
`ensure_portal_identity` now requires `explicit=True` and raises otherwise.
Its callers are the bootstrap, the desktop's `free_tier.provision` (the
explicit retry when the boot could not create the identity) and the two
dead-credential replacements (`auth_nous.resolve_nous_runtime_credentials`,
`managed_tool_gateway._replace_dead_guest_token`). The background thread
path and `provision_free_tier` are deleted with their last callers.
Reads that used to mint and now only read: `auth.py::resolve_provider`
rung 7 (an existing identity still outranks the Bedrock chain, NS-829
ordering kept), `main.py::_has_any_provider_configured`,
`cli_agent_setup_mixin._ensure_runtime_credentials`,
`managed_tool_gateway.read_nous_access_token` (no identity -> None),
`anon_sign_in.run_sign_in` (no identity -> Unavailable),
`methods_free_tier` `free_tier.status`.
`setup.status` answers from the record for the launch profile, blocking up
to 8 s while the bootstrap is in flight so a client's first poll lands after
the identity exists rather than racing it; a named profile, or a process
that never ran the bootstrap, keeps today's live probe. The record's fields
ride along additively (`ready`, `free_tier`, `other_providers`,
`inference_provider`).
Identity and inference are decoupled (NS-845 Q1.3): the mint sets
`active_provider="nous"` only when the inventory found nothing else usable
(`_mint_locked(carries_inference=)`); an adopted account always does. A token
refresh no longer re-elects the provider it refreshed
(`_save_provider_state_to_source` writes credentials, not the user's
choice) — that write was how an own-key install ended up on the free tier
after the first connector call.
Supersedes the mint sites in fcf9d11679, a42d0748fc (first-run check),
bbbaa8935a (CLI background setup), 0179efc989 (`free_tier.status` mint),
62ad1ff3ab / c63d2c935c / d8a50526d9 (the `nous.guest_setup` knob and
`provision_free_tier`), and a04b05260c (blocking mint in the resolver).
Ruling: NS-845 Q1.2 + Q1.3, recorded on NS-847.
Tests: `TestBootstrapIsTheOneCreator` (one mint per process; own key keeps
inference; reads never reach the portal; a refused mint is memoised),
`free_tier.status` fails loudly if it ever calls the creator, the resolver
stub fails loudly if resolution ever mints, `setup.status` reads the record,
`skip_free_tier` proves the inventory question. The three sign-in tests for
the deleted pre-mint collapse into one (`no identity -> Unavailable, zero
portal calls`). Live: real `_lifespan` boot with a fake portal, gate on and
off (/tmp/ns847-recon/evidence/e2e-rung5-c2-serve-boot.txt), and the CLI
matrix incl. an own-key cell (e2e-rung5-c2-bootstrap.txt), 20/20.
* fix(credits): the welcome host is free-tier evidence, so a free-tier identity never sees "run /topup"
A free-tier identity carries $0 by design, so the portal seed reports
`paid_access=False` for it. `is_free_tier_model` did not know the welcome
host, read that as a depleted account, and every free-tier turn ended with
the credits-depleted notice telling the user to top up an account they do
not have.
Rule (4) in `is_free_tier_model`: a `base_url` on the Nous welcome host
(`anon_auth.route_is_welcome_host`) is the free tier. The host is the
evidence, not the model name: the paid inference host can serve
`nous/welcome` to a named account and that account's depletion is real, so
`("nous/welcome", <inference host>)` stays False. Local data only, like the
three rules above it.
Restores the two contracts dropped by hermes-magic 674e11d1eaa (the
prototype line ran without unit tests): the welcome host is free without
any pricing evidence; the model name alone is not. The first is red without
this fix.
* fix(copy): free-tier text stops promising a connector transfer and never names the config key
Sign-in copy on every surface said "Sign in to keep your connectors" and
ended with "Your connectors are kept." The transfer registry that would
make that true is empty (NS-821): nothing carries over today. The copy now
says what signing in does give ("unlock more models and tools") and the
completion line names the account, not a transfer. The docs page loses the
"connectors carry over" paragraph for the same reason.
The picker's off-state line exposed `nous.guest: false` and the word
"guest"; user copy names the free tier only (R-USR-1).
The docs page gains the pre-rollout note: until GA nothing on it happens
without `HERMES_GUEST_ONBOARDING=1`. Its "first command mints" and
"replaced on next use" sentences now describe the boot bootstrap.
zh is a strict locale: the `freeTier` block was English placeholder text
copied from `en`; it is now Chinese. `connectorsKept` is renamed
`completedBody` since it no longer talks about connectors.
* feat(desktop): the free-tier launch flag is decided once in Electron and stamped onto every backend spawn
The Python backend reads `HERMES_GUEST_ONBOARDING` and treats exactly "1"
as on. Until now nothing in the desktop set it, so a packaged app could
never turn the free tier on, and a backend spawned by the app could
disagree with the app about whether the tier was live.
`electron/guest-onboarding.ts` owns the decision: `guestOnboardingEnabled`
is true when the launch env has `HERMES_GUEST_ONBOARDING=1` or argv has
`--guest-onboarding` (the packaged-app spelling). It is read ONCE at launch
into a module constant. `desktopBackendSpawnEnv` wraps every backend env
as the outermost call and writes the flag LAST, as "1" or an explicit "0",
so no earlier spread (`process.env`, `backend.env`) can resurrect a stray
value from the parent shell.
Stamped onto all three spawn sites: the primary `serve` spawn, the pooled
per-profile spawn, and the remote SSH `exec env ...` command (which gains
` HERMES_GUEST_ONBOARDING=1` only when on). The embedded terminal PTY and
the backend probes are not backend spawns and do not get it: a
`hermes --tui` typed in the pane must not mint.
The renderer learns the same fact read-only through the existing
`hermes:launch-flags` sync IPC (`guestOnboarding`) and preload
(`window.hermesDesktop.guestOnboardingEnabled`).
Ruling: NS-845 Q1.1 / Q2 (env var is the contract, `--guest-onboarding`
maps to it in main). Two invariant tests on the pure helpers: only "1" or
the argv flag enables; the spawn env carries "1"/"0" as the last word and
preserves every other key.
* feat(desktop): the renderer learns free-tier readiness from one `setup.ready` push, not a 60 s poll
The backend's boot bootstrap now announces `setup.ready` once, after it has
created (or refused) the free-tier identity and resolved the inference
route. The renderer used to discover both by polling `setup.status`,
`setup.runtime_check` and `free_tier.status` every 60 s from
`useStatusSnapshot`; a fresh install's chip, notice strip and onboarding
overlay could sit stale for up to a minute after boot, and three RPCs a
minute per window kept asking a question whose answer changes only at
boundaries the backend already announces.
`handleLifecycleEvent` routes `setup.ready` (active source only, like
`skin.changed`) to `notifySetupReady()`, a one-shot tick atom in
`live-sync.ts` beside the other change ticks. `useStatusSnapshot` listens
to it and runs one readiness round at once (`setup.status` +
`setup.runtime_check` + `free_tier.status`). The readiness legs also run
once on open and on return from another app, as today. The 60 s tick keeps
only `getStatus()`.
`SetupStatusSnapshot` types the record's additive fields (`ready`,
`free_tier`, `other_providers`, `inference_provider`); readiness semantics
are unchanged and still key on `provider_configured` + `runtime_check`.
Ruling: NS-845 Q1.2 (renderer half). Tests: the lifecycle branch fires one
refresh from the active source and none from another; the snapshot hook's
contract is three legs on open, one leg on the tick.
* fix(cli): the banner names the free tier's model instead of "no model configured"
The welcome banner prints before credentials resolve, so on a fresh install
`model` is empty and the banner said, in red, "no model configured — run
/model or hermes setup". Under the free tier that is false: the route is
already known from local state (identity on disk, tier on), and the first
message will run on `nous/welcome`.
`_banner_left_lines` now asks the route the same question when `model` is
empty (`guest_carries_inference()`, a local read) and shows `welcome · Nous
Research`. When nothing resolves the red line stays. Ruling: NS-845 ("the
banner's 'no model configured' line reads the resolved route").
Live: fresh HERMES_HOME + fake portal, gate on -> `welcome · Nous Research`;
gate off -> the red line, zero portal calls.
* fix(aux): vision on the free tier uses nous/welcome too
The text-only modality on the gateway's `nous/welcome` row is DeepSeek V4 Flash's, the
backing model until the repoint; `z-ai/glm-5.3-flash` is natively multimodal and the
repoint declares the welcome row `text+image->text`. Skipping Nous for vision on the
welcome host would have sent every image step past the free tier for no reason, so the
auxiliary client pins the route's one model for every lane. A backing model that takes
no images answers with the upstream's own error, which the ladder handles as it always has.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
(cherry picked from commit 7456e028faba55480db43015dc2c8df3e393a415)
* fix(gateway): hermes gateway run is a boot owner of the free tier too
Rung 5 made every demand-time free-tier site a read: resolve_provider,
the connector token, the /login precondition. That is only correct if
every process that can reach those sites ran the bootstrap first. The
CLI (cmd_chat) and hermes serve (_lifespan) did; the standalone
messaging gateway did not. A fresh HERMES_HOME with the gate on and
`hermes gateway run` reached provider resolution with no identity to
consume, and /login returned Unavailable. Reported by @andrexibiza on
#107697 (P1).
GatewayRunner.start now runs `free_tier_bootstrap.run_bootstrap` on an
executor thread right after startup recovery and BEFORE any adapter
connects, so a fast first DM cannot arrive with nothing to resolve. It
is its own step, not part of the turn-machinery warm-up: the warm-up is
an optimisation with an off switch (HERMES_STARTUP_WARMUP_TIMEOUT<=0);
the bootstrap is correctness and must always run. With the gate unset it
is a local inventory and no network.
Live, real GatewayRunner.start against a fake portal in a fresh home:
gate on -> 1 create, identity persisted, resolve_runtime_provider=nous,
/login precondition sees the identity
gate off -> 0 portal calls, no identity, no_provider_configured
Before the fix the gate-on row was identical to the gate-off row.
Test: the bootstrap seam runs before _start_prefilter_platforms and
delegates to the one creator. Red on 5554eb6993 (no seam), green here.
---------
Co-authored-by: Robin Fernandes <robin@soal.org>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Azure Foundry (gpt-6-astra) answers a request that replays encrypted
reasoning from more than one prior response with HTTP 400 "Conflicting
authenticated continuation identities", code invalid_value. The classifier
mapped that to a non-retryable client error, so the turn aborted and every
later turn in the session failed the same way.
Map the message to invalid_encrypted_content. The existing recovery in
turn_recovery then strips the replay state and retries once.
Refs #105369
Novita returns HTTP 429 with message 'server overload, please try again
later' and error type 'server_overload' when its server is genuinely busy.
Neither phrase was in _OVERLOADED_PATTERNS, so the 429 fell through to
_V_RATE_LIMIT and set should_fallback=True + should_rotate_credential=True —
rotating the credential / falling back early instead of retrying the same
key. Add 'server overload' and 'server_overload' to the overload tuple so
this reaches the existing _V_OVERLOADED verdict (retryable, no rotation).
Closes#106205
oMLX/MLX prefill memory-guard rejections name an allocation peak in BYTES but
close with "Reduce context length", so _CONTEXT_OVERFLOW_PATTERNS claims them
and the turn enters the compress-and-shrink loop. Compression cannot lower a
prefill peak — the prompt is usually far below the window — so it burns the
compression budget, re-hits the wedged server on every attempt and ends in
"Cannot compress further" plus a destructive session reset.
Classify them as `overloaded` instead: retry with backoff, no compression, no
session reset (mirrors 503/529 recovery).
The guard runs before the overflow check AND before the usage-limit
disambiguation. The second ordering matters more than it looks: "memory limit
exceeded" contains "limit exceeded", so a status-less rejection — a proxy that
flattened the body — is currently classified as `billing` and rotates a
healthy credential.
Sites covered:
- _OVERFLOW_AS_5XX_RULES, which _400_TAIL_RULES extends → 400, 500, 502,
503, 529
- _MESSAGE_HEAD_RULES for the status-less path (ahead of usage-limit)
- _ERROR_CODE_VERDICTS for the structured oMLX codes
- _classify_400, because _by_status runs before _by_error_code, so a body
whose wording a proxy stripped would otherwise fall through to
format_error
Every pattern names memory/allocation in bytes, never a token or window count,
so the list stays disjoint from _CONTEXT_OVERFLOW_PATTERNS. Both oMLX wordings
are kept: 0.5.6 says "Prefill would require ~13.87 GB peak", 0.5.7 reworded it
to "predicted peak would require/exceed" and both are in the field. A test
pins that a genuine window overflow still compresses.
Refs #52261. Supersedes #52289, which predates the classifier rewrite and can
no longer be merged.
The ChatGPT Codex backend answers a rejected encrypted-reasoning replay with the
same bare 400 {code: invalid_prompt, message: "Request blocked."} it uses for
genuine blocks. classify_api_error bucketed it as format_error, so the one-shot
repair that already exists for invalid_encrypted_content (disable replay, strip
cached codex_reasoning_items, retry once) never ran and the turn aborted.
Classify exactly that envelope from provider openai-codex — SDK 400 body, SSE
``error`` frame, or the ``response.failed`` text — as invalid_encrypted_content
while keeping format_error's abort-and-fallback hints, so the only behavioural
delta is turn_recovery's replay strip, which still requires cached reasoning
items. A block with nothing to strip aborts exactly as before; other providers,
other messages/codes and the #18028 safety refusals are untouched.
evals/codex_masked_replay_review.py drives a real AIAgent (codex_responses)
against a synthetic local Responses server for the before/after check.
Refs #92353, #92357, #94077.
Reasoning-mandatory routes answer reasoning: {enabled: false} with HTTP 400
"Reasoning is mandatory for this endpoint and cannot be disabled". Hermes
sends that disable for /reasoning none, agent.reasoning_effort: none, and the
one-shot thinking-exhaustion continuation override (which GLM-5.3-flash
triggers on its own). The Nous profile's catalog guard swallows the disable
only when its per-process capability cache already says mandatory; a gateway
that warmed the cache before the route flipped kept sending it, and the 400
was classified as a non-retryable format_error that aborted the turn.
- error_classifier: new reasoning_mandatory reason (retryable, no fallback,
no compression), matched before the request-validation branch.
- conversation_loop: one-shot recovery — set agent._reasoning_disable_rejected,
queue a catalog refresh for the provider, retry.
- chat_completion_helpers: _reasoning_config_for_wire drops every disable
(configured or ephemeral) once the route has rejected one.
- hermes_cli/models: refresh_reasoning_caps_async(provider) forces a
background re-fetch of the Nous/OpenRouter catalog.
- openrouter profile: omit a disable when the catalog marks the route
mandatory (parity with the Nous profile).
Live: z-ai/glm-5.3-flash on the Portal with a poisoned mandatory:false cache.
Before: turn aborted with the 400. After: one retry, thinking stays on, turn
completes.
The permanent-brick class in #69078: xAI returns 'Invalid PNG image'
when a re-serialized image part in replayed history becomes
undecodable. The existing image-error patterns cover only Anthropic
'exceeds max dimension' wordings and 'model does not support images'
strings, so the classifier lands on a generic non-retryable 400 and
neither the shrink path nor the strip path fires. Every subsequent
turn (even bare text) fails identically because the poison stays in
history — the session is permanently wedged until deleted.
Two recovery layers, deliberately separate:
- Semantic split: new FailoverReason.image_corrupt with
_IMAGE_CORRUPT_PATTERNS ('invalid png image' / 'invalid jpeg image'),
checked BEFORE _IMAGE_TOO_LARGE_PATTERNS in both _classify_400 and
_classify_by_message. Corrupt bytes route to strip-and-retry, never
to the shrink path (shrinking corrupt bytes cannot help).
- Generic fallback: any non-retryable 400 whose outgoing messages
still contain image parts gets one strip-and-retry via the existing
_strip_images_from_messages helper, guarded by a new
stripped_images_this_turn one-shot flag on TurnRetryState. This
un-bricks the session for any current or future provider wording
without adding another pattern list to maintain.
Item 3 from the report (multimodal-part integrity across FTS
persistence + compaction handoff) is a separate investigation and
remains follow-up work.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: paultaki <paultaki@users.noreply.github.com>
Consolidates the 429-quota-classifier cluster on top of the merged #93419
Anthropic core. Three independent contributor findings salvaged into one
coherent change to the single 429 branch:
- Broaden the 429 usage-limit check from the narrow 'usage limit' string to
the full _USAGE_LIMIT_PATTERNS ('quota', 'limit exceeded', 'key limit
exceeded') and add _BILLING_PATTERNS detection on 429 ('insufficient
credits' wrapped in a 429 instead of 402), guarded by a _RATE_LIMIT_PATTERNS
exclusion so an explicit 'Rate limit exceeded' never promotes to
non-retryable billing. (credit @Pluviobyte, #39441 — earliest submitter)
- Add 'resets in' to the transient signals: Codex's 'Weekly usage limit
reached. Resets in 6hr 29min.' wrongly read as terminal billing because
main only had 'reset in' (no substring match). (credit @LeonSGP43, #63021)
- Add 'reset after' / 'available in' / 'per minute' / 'per second' transient
signals. (credit @jtstothard, #74785)
Supersedes #65633 (defective branch placement, no tests). The aux-client
path already covers these shapes (_is_payment_error catches weekly/quota
walls; _is_rate_limit_error treats 'resets in' as transient), so no change
there.
Tests: 6 new cases (generic quota wall, insufficient-credits 429, rate-limit
guard, Codex resets-in, extra transient phrases). Guard sabotage-verified.
Co-authored-by: Pluviobyte <Pluviobyte@users.noreply.github.com>
Co-authored-by: LeonSGP43 <LeonSGP43@users.noreply.github.com>
Co-authored-by: jtstothard <jtstothard@users.noreply.github.com>
The Codex OAuth backend (chatgpt.com/backend-api/codex) intermittently
injects prompt_cache_retention into its own upstream call and then rejects
it, returning HTTP 400 invalid_parameter. Hermes never sends that field on
this route (see agent/transports/codex.py::_default_prompt_cache_retention_
for_request, which only sets it for api.meta.ai and bedrock-mantle hosts).
Reproduced live: a minimal 1-message request carrying no cache parameters
at all failed 4/20 (20%) with this error, so the rejection is not
deterministic and retrying the identical request is the correct recovery.
Previously the catch-all in _classify_400 returned format_error/
retryable=False, which tripped the is_client_error abort gate in
conversation_loop and killed the turn on the first attempt - burning an
entire large-context request (~550k tokens) per failure.
Classify these as retryable server_error (should_compress=False - the
request shape was never the problem). The same guard is applied to the
sibling 5xx request-validation branch, where a fronting proxy can surface
the identical rejection.
Deliberately narrow: keyed on parameters we only send on specific routes,
and skipped when the current provider is one that legitimately sends them,
so a genuine client-side bad parameter (max_tokens on GPT-5) still fails
fast as a format_error.
Port from anomalyco/opencode#40707: connection-establishment and DNS
failure messages wrapped in generic exceptions (RuntimeError from local
shims, MCP bridges, SDKs re-raising without chaining) fell through to
FailoverReason.unknown, which misses the retry loop's eager transport
fallback — the full retry budget burned against a dead endpoint before
provider fallback.
New _CONNECTION_MESSAGE_PATTERNS (connect refused, no route, network
unreachable, DNS phrasings across Python/glibc/macOS/Node, fetch failed,
Envoy upstream connect error) classify as retryable timeout via
_classify_by_message, mirroring _TIMEOUT_MESSAGE_PATTERNS. Mid-stream
disconnect strings are deliberately excluded — they keep their
_SERVER_DISCONNECT_PATTERNS routing (large-session compression).
Review follow-up (egilewski): the previous commit only hedged the guidance
text; the exact Anthropic 400 was still classified, persisted, and surfaced
as confirmed billing exhaustion. Carry the ambiguity all the way through:
- agent/error_classifier.py: 'out of extra usage' matches on the 400 and
status-less paths now attach error_context {billing_unverified,
possible_content_filter}. Reason stays FailoverReason.billing (rotation +
fallback remain the right recovery either way); ClassifiedError grows a
billing_unverified property.
- agent/credential_pool.py: new FAILURE_REASON_BILLING_UNVERIFIED. An
unverified billing exhaustion gets the short transient cooldown instead of
the one-hour bench, regardless of pool size: a content-filter rejection
leaves the credential healthy and fails identically on every key, and the
hour-long sole-credential latch is what replayed the stored error and made
real fixes look ineffective. A true 402 keeps the full bench. The marker
persists with the entry so a restart cannot upgrade it back to a bench.
- agent/agent_runtime_helpers.py + run_agent.py: recover_with_credential_pool
threads billing_unverified and hands the pool 'billing_unverified' as the
persisted failure_reason.
- agent/conversation_loop.py: the fallback-switch status, max-retries status,
terminal label, and both structured terminal results hedge when the verdict
is unverified. New _billing_terminal_label + _billing_failure_result build
the returned terminal response in one place; the result dict now carries
billing_unverified and the billing_block gains 'unverified': true. The
confirmed-billing path (a real 402 or an API-key credit depletion) keeps
the original assertive wording, so the caveat no longer dilutes it.
Regression tests: classifier marking (400 + status-less + unambiguous-body
negative), pool cooldown TTLs + persistence round-trip, pool failure_reason
plumbing, and the returned terminal response for both unverified and
confirmed verdicts.
Note: tests/agent/test_credential_pool_routing.py::TestFailureAttribution::
test_unmatched_key_does_not_retry_only_pool_entry fails identically on
current main without this change (pre-existing, unrelated).
Run the expensive-model warning for explicit startup `-m` / `--provider`
overrides before the chat loop starts, and fail closed for non-interactive
invocations that select an expensive or known-confusing model.
Also classify Nous paid-model 404s that say credits are required as billing
exhaustion so they fail fast with billing guidance.
Tested:
- scripts/run_tests.sh tests/hermes_cli/test_cli_startup_model_cost_guard.py tests/hermes_cli/test_model_cost_guard.py tests/agent/test_error_classifier.py -- --tb=short -q
Add regressions for the Discord-shaped applyPromptTemplate 400 that embeds
"No user query found", bare no-user-query on a large session, and the genuine
llama.cpp unable-to-generate-parser grammar path that must stay recoverable.
Selecting an NVIDIA NIM model whose id reached config without the nvidia/
prefix produced a bare "HTTP 404: 404 page not found" — retried three times,
never naming the model. It reads exactly like an outage or an auth failure,
which is where the Discord thread spent its time before the id was spotted.
normalize_model_for_provider() had no branch for nvidia, so a bare id passed
straight through to the API. Repair it from the provider's curated catalogue:
a bare name that matches exactly one entry modulo the prefix gets it back.
That's a lookup, not a guess — build.nvidia.com also fronts local NIM
containers and third-party models, and anything absent from the catalogue is
left alone. Because the repair runs on every runtime setup, an already-broken
config self-heals on the next turn and prints what it changed.
If a bare id still reaches the wire, the 404 now explains itself. The
classifier consults the same catalogue: a prefix-less id the provider only
serves as vendor/model is a deterministic failure, so it classifies as
model_not_found instead of burning three retries on a retryable "unknown",
and the error trace names the id to use.
Fixes#78796
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).
Port from anomalyco/opencode#37848 (+ dev-branch twin #37840): expand
context-overflow patterns and guard against rate-limit messages that
mention tokens.
- 'Throttling error: Too many tokens, please wait before trying again.'
(AWS Bedrock / proxy shape) classified as context_overflow and routed a
healthy session into compression on every throttle. Added 'throttling'
to _RATE_LIMIT_PATTERNS, which the message-only path checks BEFORE the
overflow list.
- 'Input length N exceeds the maximum allowed input length of M tokens.'
(Together/Fireworks shape) fell through to unknown — no compression
recovery. Added 'maximum allowed input length' to overflow patterns.
- 'request_too_large' / 'Request exceeds the maximum size' (Anthropic 413
type re-wrapped without a status code by aggregators/proxies) fell
through to unknown. Added to _PAYLOAD_TOO_LARGE_PATTERNS.
All three shapes proven live on main before the fix; 265 classifier +
bedrock tests and 238 sibling rate-guard/compression tests pass.
Port from anomalyco/opencode#35671: Z.AI / Zhipu GLM returns
'tokens in request more than max tokens allowed' (error code 1210) on
context overflow. This matched no pattern in _CONTEXT_OVERFLOW_PATTERNS,
so the error classified as unknown/retryable — the agent would retry the
oversized request instead of triggering context compression.
Proven live on main before the fix: classify_api_error() returned
FailoverReason.unknown for the exact Z.AI error shape; now returns
context_overflow.
Keep each xAI OAuth auth-add login as an independent manual device-code pool entry and recognize xAI personal-team spending-limit 403 responses as billing exhaustion. Preserve the structured top-level error message so the failed credential is quarantined and the next healthy account is selected without attempting a pointless token refresh.
Route direct xAI HTTP consumers through the credential pool as well. Proactive and 401-reactive refreshes update the exact issuing manual entry, preserve validated xAI base URL overrides, and serialize single-use refresh-token rotation across concurrent pool instances.
The cross-turn stale-call circuit breaker (_check_stale_giveup in
agent/chat_completion_helpers.py) raises a RuntimeError when the
provider has been unresponsive for N consecutive stale attempts.
This error was classified as FailoverReason.unknown (retryable=True,
should_fallback=False), causing the retry loop to burn all max_retries
against the same dead provider — each retry hitting the circuit breaker
instantly with zero network overhead — before fallback was attempted.
Add a classification rule (section 7b in classify_api_error) that
recognizes the stale-breaker RuntimeError by its signature phrases and
classifies it as FailoverReason.timeout with retryable=False,
should_fallback=True. This makes the retry loop skip retries and go
straight to fallback provider activation on the first hit.
Test: test_stale_breaker_runtime_error_triggers_fallback_not_retry
_classify_by_status() routes every other transient HTTP status to a retryable
reason (500/502 -> server_error, 503/529 -> overloaded, 429 -> rate_limit,
413 -> payload_too_large), but 408 Request Timeout fell through to the generic
`400 <= status < 500` branch and was classified as a non-retryable
format_error -- the same bucket as a 400 Bad Request.
A 408 is a transient timing failure the server itself flags as safe to retry
(RFC 9110 15.5.9), not a malformed request, so the retry loop aborted the turn
when a simple retry would recover. Common trigger: a reverse proxy in front of
a self-hosted backend (llama.cpp / Ollama / vLLM) returns 408 when a long
generation outruns the proxy's request-read window.
Route 408 to the existing FailoverReason.timeout (rebuild client + retry).
Add a regression test plus a boundary test asserting 400 stays non-retryable.
Inspired by Claude Code v2.1.199 (July 2, 2026): SSL certificate errors
(TLS-inspecting proxies, missing CA bundles, expired certs) no longer
burn retries before showing actionable guidance — they fail immediately
with the fix hint.
- agent/error_classifier.py: new FailoverReason.ssl_cert_verification +
_SSL_CERT_VERIFY_PATTERNS, checked BEFORE the transient-SSL patterns
(cert-verify messages also contain '[SSL:' and previously retried
forever as timeout). Non-retryable, no compression, no fallback churn.
- agent/conversation_loop.py: dedicated status line + per-cause fix
hints (corporate proxy CA bundle, certifi refresh, self-signed local
endpoints) on the non-retryable abort path.
- 7 new tests incl. regression guards (transient alerts still retry,
large-session cert failure doesn't trigger compression).
Review nit (helix4u): the fix covers 500/502/503/529 but the positive tests
only asserted 500 and 503. Parametrize over all four so 502/529 are covered
too; keep the plain-5xx negatives.
Local inference servers (llama.cpp/llama-server, vLLM/Ollama behind a
Cloudflare/Tailscale hop) report context overflow with HTTP 500/502/503/529
instead of 400/413. _classify_by_status returned server_error/overloaded and
retried blindly, then dropped the turn with no compaction. Route explicit
_CONTEXT_OVERFLOW_PATTERNS matches on those 5xx codes to context_overflow
(should_compress=True); plain 500 stays server_error, plain 503 overloaded.
Anthropic returns HTTP 400 with "You're out of extra usage. Add more at
claude.ai/settings/usage and keep going." when the account's extra-usage
allowance is depleted. The existing _BILLING_PATTERNS list did not
include this wording, so classify_api_error fell through to generic
format_error — non-retryable and should_fallback=False — causing the
agent to abort instead of engaging the configured fallback chain.
Add the pattern and a regression test covering the exact Anthropic body.
OpenRouter returns 429 in two shapes: an account-level throttle on the
user's key, and an upstream-provider throttle (DeepSeek/Anthropic/etc.
rate-limiting OpenRouter's aggregate traffic). The classifier treated
both identically and rotated/exhausted OPENROUTER_API_KEY on every 429 —
burning the key for ~24min and silently disabling auxiliary features
(compression, summarization, vision) on an upstream throttle where the
key was healthy.
Add a FailoverReason.upstream_rate_limit classified from OpenRouter's
unambiguous wrapper message "Provider returned error" (the same signal
the metadata-raw parser already trusts). Recovery skips credential
rotation and defers to the fallback chain to switch models instead.
Co-authored-by: Hermes Agent <127238744+teknium1@users.noreply.github.com>
Z.AI / Zhipu reuse HTTP 429 for server-wide overload. The 429 status
path classified these unconditionally as rate_limit with
should_rotate_credential=True, so an overloaded provider exhausted the
credential pool after two errors — fatal for a single-key user, who has
nothing to rotate to.
The credential is valid; the server is just busy. Disambiguate the 429
body against a shared _OVERLOADED_PATTERNS list and route overload
language to FailoverReason.overloaded (retryable, no rotation), matching
the existing 503/529 path and the message-only path (#52890). Genuine
rate limits (no overload language) still rotate.
Extracted the inline overloaded tuple #52890 added into the shared
_OVERLOADED_PATTERNS constant so the status-code and message paths use
one list.
Closes#14038.
Anthropic returns a 400 when the thinking/redacted_thinking blocks in the
latest assistant message are mutated upstream: 'thinking or redacted_thinking
blocks in the latest assistant message cannot be modified. These blocks must
remain as they were in the original response.'
The classifier's thinking_signature branch only matched on the substring
'signature', so this variant fell through to a non-retryable client error
and hard-aborted the turn -- even though the existing strip-reasoning_details
-and-retry recovery would have healed it.
Broaden the 400 match to also catch 'cannot be modified' / 'must remain as
they were' (still gated on 'thinking'), routing it to the same recovery.
Adds a negative-case test so unrelated 'cannot be modified' 400s are not
swept in.
Defense-in-depth, orthogonal to the root-cause work in #35975 / #17861
(which prevent the block mutation in the first place). Only changes a
terminal-failure into a one-shot recovery.
Signed-off-by: Ian Culling <ian@culling.ca>
A GPT-5 model rejecting max_tokens returns a 400 whose message contains the
literal substring 'max_tokens' — one of the _CONTEXT_OVERFLOW_PATTERNS. The 400
path in _classify_400 checked overflow patterns before any request-validation
check (which only existed on the 5xx path), so the parameter error was routed
into the compression loop, re-sent with the same bad param, and ended in
'Cannot compress further' on a tiny context.
Hoist a request-validation guard (unsupported/unknown parameter) above the
context-overflow check in _classify_400. Deliberately excludes the generic
invalid_request_error code, which OpenAI also stamps on real overflow 400s, so
genuine overflows still compress. Pairs with the max_completion_tokens param
fix that stops the bad request at the source.
Also adds AUTHOR_MAP entry for the salvaged PR #13902 commit.
* fix(agent): fallback immediately on provider content-policy blocks
Provider safety-filter refusals (e.g. OpenAI Codex 'flagged for possible
cybersecurity risk', OpenAI moderation 'violates our usage policies',
Anthropic safety-system rejections, Azure content_filter) are
deterministic decisions about a specific prompt. Retrying the same
prompt up to api_max_retries times just reproduces the same refusal and
burns paid attempts before surfacing the generic 'API failed after 3
retries — <provider message>' to Telegram / cron with no indication that
the failure came from the model provider rather than Hermes itself.
Classify these as a new FailoverReason.content_policy_blocked
(non-retryable, should_fallback=True) and route them through the
existing is_client_error path so the loop:
- skips the 3x retry backoff
- activates a configured fallback model immediately
- emits a clear provider-safety message to the user (not the generic
'Non-retryable error (HTTP None)') and surfaces actionable guidance
when no fallback is configured (rephrase, narrow context, or set
fallback_model in hermes config)
- returns a final_response that explicitly tells the user this came
from the model provider, so gateway delivery is unambiguous and
cron last_status reflects the safety block rather than a vague
'agent reported failure'
Patterns are intentionally narrow — verbatim refusal phrasings keyed to
specific provider safety pipelines, not generic words like 'policy' or
'violation' that would collide with billing / format / auth errors.
Regression guards in test_18028_content_policy_blocked.py verify
billing 402s, generic 400s, and OpenRouter account-level
provider_policy_blocked remain distinct classifications.
Salvaged from #18164 onto current main (file restructure: loop logic
moved from run_agent.py to agent/conversation_loop.py, _emit_status →
_buffer_status), broadened patterns beyond the original OpenAI Codex
cybersecurity case to cover OpenAI moderation, Anthropic safety system,
and Azure content_filter; added user-actionable guidance and a clear
final_response so cron/gateway surfaces the policy block instead of a
generic non-retryable error, and added a regression-guard test module
mirroring the is_client_error predicate.
Addresses #18028.
Co-authored-by: Kuan-Chieh Huang <kchuang1015@users.noreply.github.com>
* chore: add kchuang1015 to AUTHOR_MAP
---------
Co-authored-by: Kuan-Chieh Huang <kchuang1015@users.noreply.github.com>
* fix(codex-responses): gracefully recover from invalid_encrypted_content (salvage #10144)
When an OpenAI-compatible Responses API surface accepts an initial
request but later rejects the replayed `codex_reasoning_items`
encrypted blob with HTTP 400 `invalid_encrypted_content`, the
session previously got stuck retrying the same poisoned payload.
Recovery: classify the error as a dedicated FailoverReason, and on the
first hit disable encrypted reasoning replay for the rest of the
session, strip cached items from message history, and retry once.
Changes:
* error_classifier: add FailoverReason.invalid_encrypted_content
branch in _classify_400 (before context_overflow so the messages
that mention 'encrypted content … could not be verified' don't trip
context heuristics), in _classify_by_error_code, and extend
_extract_error_code to peek inside wrapped JSON in error.message and
ignore the bare '400' as a code.
* agent_init: initialize `_codex_reasoning_replay_enabled = True` on
every agent.
* run_agent: add AIAgent._disable_codex_reasoning_replay() helper
that flips the flag and pops cached items.
* codex_responses_adapter: thread a `replay_encrypted_reasoning`
kwarg through _chat_messages_to_responses_input so that when the
flag is False we don't replay codex_reasoning_items.
* transports/codex.py: read `replay_encrypted_reasoning` from params,
thread it into the adapter, and gate the
`include=['reasoning.encrypted_content']` request hint on it.
* chat_completion_helpers: pass the agent's replay flag through to
the transport.
* conversation_loop: in the retry loop, add an
invalid_encrypted_content recovery branch that fires once per
session, only when api_mode == codex_responses, only when replay is
still enabled, and only when at least one assistant message in
history actually carries cached reasoning items (otherwise the 400
has nothing to do with our cache and the normal retry path handles
it).
Tests:
* test_error_classifier: new wrapped-JSON _extract_error_code case;
new TestClassifyApiError cases proving the 400 is retryable with
no fallback, that the broad message match doesn't catch a generic
'parsed' message, and that the error code match is
case-insensitive.
* test_run_agent_codex_responses: end-to-end test of the recovery
branch firing once and disabling replay, plus a sibling test that
proves the branch does *not* fire (and the flag stays True) when
history has no cached reasoning items.
Salvages PR #10144 onto the post-refactor module layout
(error_classifier / codex_responses_adapter / transports/codex /
conversation_loop / agent_init) since the original diff was written
against the pre-refactor monolithic run_agent.py.
* chore(release): map victorGPT in AUTHOR_MAP for #10144 salvage
---------
Co-authored-by: victorGPT <wuxuebin1993@gmail.com>
Standard OpenAI returns request-validation failures (unknown/
unsupported parameter, malformed request) as 4xx. Some
OpenAI-compatible gateways return them as 5xx instead — codex.nekos.me
returns 502 for an unknown parameter.
The generic '5xx -> retryable server_error' rule then misfires: the
error is deterministic (every retry gets the identical rejection), so
the retry loop burns all 3 attempts, the transport-recovery path
resets the counter and burns 3 more, and the result is a request
flood against a request that can never succeed.
Fix: when a 500/502 body carries an unambiguous request-validation
signal — 'unknown parameter' / 'unsupported parameter' /
'invalid_request_error' in the message text, or invalid_request_error
/ unknown_parameter / unsupported_parameter as the structured error
code — classify as a non-retryable format_error so the loop fails
fast and falls back. Genuine 502 Bad Gateway with no such signal
stays retryable as before.
Origin: local-author
Upstream-PR: none
Patch-State: local-only
Some providers (Xiaomi MiMo, some Alibaba endpoints, a long tail of
OpenAI-compatible servers) follow the OpenAI spec strictly and require
tool message `content` to be a string — they reject our list-type
content (text + image_url parts) with HTTP 400 'text is not set' /
'tool message content must be a string'.
Instead of an allowlist of known-good providers (maintenance burden,
guaranteed to miss aggregators like OpenRouter where the underlying
model determines support, not the aggregator name), this lands a
reactive recovery:
1. New `FailoverReason.multimodal_tool_content_unsupported` with a
small pattern list covering the common 400 wordings.
2. `AIAgent._try_strip_image_parts_from_tool_messages` walks the API
message list, downgrades any `role:tool` message whose content is
list-with-image to a plain text summary (preserves text parts) in
place, AND records the active (provider, model) in a session-scoped
`_no_list_tool_content_models` set.
3. `_tool_result_content_for_active_model` short-circuits to a text
summary when (provider, model) is in the cache — so after the first
400 + retry, subsequent screenshots in the same session skip the
round trip entirely.
4. Retry hook in `agent.conversation_loop` mirrors the existing
`image_too_large` recovery: detect the reason, run the helper,
retry once, fall through to the normal error path if no list-type
tool content was actually present.
Cache is transient (per-session) by design — next session retries in
case the provider added support, no persistent state to maintain.
Fixes#27344. Closes#27351 (allowlist approach superseded by reactive
recovery).
kshitijk4poor