fix(anthropic): omit 1m context beta for native OAuth by JayGwod · Pull Request #17680 · NousResearch/hermes-agent

JayGwod · 2026-04-30T00:44:33Z

Summary

Omit the context-1m-2025-08-07 beta header for native Anthropic OAuth / Claude Code-style auth.
Keep the existing common beta behavior for API-key auth and non-native Anthropic-compatible routes.
Also omit the 1M context beta from per-request fast-mode headers when the request is OAuth-authenticated.

Why

Some Anthropic OAuth subscriptions reject requests before model/tool execution when the request includes:

anthropic-beta: context-1m-2025-08-07

The observed provider error is:

The long context beta is not yet available for this subscription.

That failure is independent of the model/tool payload. Native Anthropic 1M context availability can be account/subscription gated differently from Bedrock/Azure beta-header requirements, so OAuth requests should not unconditionally inherit this beta header.

Behavior preserved

API-key auth still includes context-1m-2025-08-07 in the common beta list.
Bedrock/Azure and other non-OAuth paths continue using _common_betas_for_base_url(...) unchanged.
OAuth-only betas such as claude-code-20250219 and oauth-2025-04-20 remain present for OAuth requests.

Tests

./venv/bin/python -m py_compile agent/anthropic_adapter.py
./venv/bin/python -m pytest -o 'addopts=' tests/agent/test_anthropic_adapter.py -q

Result:

139 passed

alt-glitch · 2026-04-30T01:00:10Z

Likely duplicate of #17442 — same fix: strip context-1m beta header on OAuth path to avoid 400s on consumer subscriptions.

alt-glitch · 2026-04-30T01:01:01Z

Likely duplicate of #17442 — same fix: strip context-1m beta header on OAuth path to avoid 400s on consumer subscriptions.

JayGwod · 2026-04-30T01:50:28Z

Thanks — agreed the main fix overlaps with #17442.

One small difference in this PR is that it also covers the fast-mode per-request extra_headers path in build_anthropic_kwargs(). That path rebuilds anthropic-beta and overrides the client-level header, so if only the client construction path strips context-1m-2025-08-07, an OAuth + fast_mode=True request can still reintroduce the same beta header.

If maintainers prefer #17442 as the primary PR, I'm happy to close this one and either:

add the fast-mode extra_headers coverage as a small follow-up on top of fix(anthropic): strip context-1m beta on OAuth path #17442, or
keep this open only if you want the broader coverage/test here.

@JayGwod

…#17752) Keep context-1m-2025-08-07 in OAuth requests by default so 1M-capable subscriptions retain full context. When Anthropic rejects a request with 400 'long context beta is not yet available for this subscription', disable the beta for the rest of the session, rebuild the client, and retry once. Addresses #17680 (thanks @JayGwod for the clean reproduction) without forcing every OAuth user off the 1M context window. Changes: - agent/error_classifier.py: new FailoverReason.oauth_long_context_beta_forbidden; pattern matches 400 + 'long context beta' + 'not yet available'. Narrow enough that the existing 429 tier-gate pattern keeps its own reason. - agent/anthropic_adapter.py: _common_betas_for_base_url, build_anthropic_client, build_anthropic_kwargs gain drop_context_1m_beta kwarg. Default=False (1M stays). OAuth OAUTH_ONLY_BETAS unchanged. - agent/transports/anthropic.py: build_kwargs forwards the flag. - run_agent.py: self._oauth_1m_beta_disabled flag, retry-once guard, recovery branch next to the image-shrink path. _rebuild_anthropic_client honors the flag. The main build_kwargs call site threads it through for fast-mode extra_headers. - hermes_cli/doctor.py, hermes_cli/models.py: sibling OAuth /v1/models probes get the same reactive retry — previously they'd falsely report the Anthropic API as unreachable for affected subscriptions. Tests: 2190 tests/agent/ + 94 adjacent integration tests pass. New unit tests cover the classifier pattern (including the collision guard against the 429 tier-gate) and the drop_context_1m_beta adapter behavior (default keeps 1M, flag strips only 1M while preserving every other beta).

teknium1 · 2026-04-30T04:57:16Z

Addressed via PR #17752 (merged as 828d3a3) — thanks @JayGwod for the clean reproduction and the exact error phrasing, both were crucial.

I took a reactive-recovery approach instead of unconditionally omitting the beta. The concern with the unconditional-omit path is that OAuth users whose subscriptions do include 1M context would silently lose the capability (their next long session gets truncated at 200K with no user-visible signal). Reactive recovery lets Anthropic's own error be the ground truth:

Request goes out with context-1m-2025-08-07 as today.
If Anthropic returns 400 with "long context beta is not yet available for this subscription", the classifier tags it oauth_long_context_beta_forbidden.
The retry loop sets a session-wide _oauth_1m_beta_disabled flag, rebuilds the Anthropic client without the beta, logs 🔕 OAuth subscription doesn't support the 1M-context beta — disabled for this session and retrying..., retries once.
Remaining turns in the session reuse the reduced-beta client — no repeated probing, one failed attempt total per session.

Result: 1M-capable subscriptions keep full 1M; 1M-ineligible subscriptions pay one silent failed call per session and then work transparently.

Also widened to the two sibling OAuth /v1/models probes you didn't touch (hermes_cli/doctor.py and hermes_cli/models.py) — they would have falsely reported Anthropic as unreachable for affected subscriptions.

Full diff: #17752

@JayGwod

…NousResearch#17752) Keep context-1m-2025-08-07 in OAuth requests by default so 1M-capable subscriptions retain full context. When Anthropic rejects a request with 400 'long context beta is not yet available for this subscription', disable the beta for the rest of the session, rebuild the client, and retry once. Addresses NousResearch#17680 (thanks @JayGwod for the clean reproduction) without forcing every OAuth user off the 1M context window. Changes: - agent/error_classifier.py: new FailoverReason.oauth_long_context_beta_forbidden; pattern matches 400 + 'long context beta' + 'not yet available'. Narrow enough that the existing 429 tier-gate pattern keeps its own reason. - agent/anthropic_adapter.py: _common_betas_for_base_url, build_anthropic_client, build_anthropic_kwargs gain drop_context_1m_beta kwarg. Default=False (1M stays). OAuth OAUTH_ONLY_BETAS unchanged. - agent/transports/anthropic.py: build_kwargs forwards the flag. - run_agent.py: self._oauth_1m_beta_disabled flag, retry-once guard, recovery branch next to the image-shrink path. _rebuild_anthropic_client honors the flag. The main build_kwargs call site threads it through for fast-mode extra_headers. - hermes_cli/doctor.py, hermes_cli/models.py: sibling OAuth /v1/models probes get the same reactive retry — previously they'd falsely report the Anthropic API as unreachable for affected subscriptions. Tests: 2190 tests/agent/ + 94 adjacent integration tests pass. New unit tests cover the classifier pattern (including the collision guard against the 429 tier-gate) and the drop_context_1m_beta adapter behavior (default keeps 1M, flag strips only 1M while preserving every other beta).

@JayGwod

…NousResearch#17752) Keep context-1m-2025-08-07 in OAuth requests by default so 1M-capable subscriptions retain full context. When Anthropic rejects a request with 400 'long context beta is not yet available for this subscription', disable the beta for the rest of the session, rebuild the client, and retry once. Addresses NousResearch#17680 (thanks @JayGwod for the clean reproduction) without forcing every OAuth user off the 1M context window. Changes: - agent/error_classifier.py: new FailoverReason.oauth_long_context_beta_forbidden; pattern matches 400 + 'long context beta' + 'not yet available'. Narrow enough that the existing 429 tier-gate pattern keeps its own reason. - agent/anthropic_adapter.py: _common_betas_for_base_url, build_anthropic_client, build_anthropic_kwargs gain drop_context_1m_beta kwarg. Default=False (1M stays). OAuth OAUTH_ONLY_BETAS unchanged. - agent/transports/anthropic.py: build_kwargs forwards the flag. - run_agent.py: self._oauth_1m_beta_disabled flag, retry-once guard, recovery branch next to the image-shrink path. _rebuild_anthropic_client honors the flag. The main build_kwargs call site threads it through for fast-mode extra_headers. - hermes_cli/doctor.py, hermes_cli/models.py: sibling OAuth /v1/models probes get the same reactive retry — previously they'd falsely report the Anthropic API as unreachable for affected subscriptions. Tests: 2190 tests/agent/ + 94 adjacent integration tests pass. New unit tests cover the classifier pattern (including the collision guard against the 429 tier-gate) and the drop_context_1m_beta adapter behavior (default keeps 1M, flag strips only 1M while preserving every other beta).

@JayGwod

…NousResearch#17752) Keep context-1m-2025-08-07 in OAuth requests by default so 1M-capable subscriptions retain full context. When Anthropic rejects a request with 400 'long context beta is not yet available for this subscription', disable the beta for the rest of the session, rebuild the client, and retry once. Addresses NousResearch#17680 (thanks @JayGwod for the clean reproduction) without forcing every OAuth user off the 1M context window. Changes: - agent/error_classifier.py: new FailoverReason.oauth_long_context_beta_forbidden; pattern matches 400 + 'long context beta' + 'not yet available'. Narrow enough that the existing 429 tier-gate pattern keeps its own reason. - agent/anthropic_adapter.py: _common_betas_for_base_url, build_anthropic_client, build_anthropic_kwargs gain drop_context_1m_beta kwarg. Default=False (1M stays). OAuth OAUTH_ONLY_BETAS unchanged. - agent/transports/anthropic.py: build_kwargs forwards the flag. - run_agent.py: self._oauth_1m_beta_disabled flag, retry-once guard, recovery branch next to the image-shrink path. _rebuild_anthropic_client honors the flag. The main build_kwargs call site threads it through for fast-mode extra_headers. - hermes_cli/doctor.py, hermes_cli/models.py: sibling OAuth /v1/models probes get the same reactive retry — previously they'd falsely report the Anthropic API as unreachable for affected subscriptions. Tests: 2190 tests/agent/ + 94 adjacent integration tests pass. New unit tests cover the classifier pattern (including the collision guard against the 429 tier-gate) and the drop_context_1m_beta adapter behavior (default keeps 1M, flag strips only 1M while preserving every other beta).

@JayGwod

…NousResearch#17752) Keep context-1m-2025-08-07 in OAuth requests by default so 1M-capable subscriptions retain full context. When Anthropic rejects a request with 400 'long context beta is not yet available for this subscription', disable the beta for the rest of the session, rebuild the client, and retry once. Addresses NousResearch#17680 (thanks @JayGwod for the clean reproduction) without forcing every OAuth user off the 1M context window. Changes: - agent/error_classifier.py: new FailoverReason.oauth_long_context_beta_forbidden; pattern matches 400 + 'long context beta' + 'not yet available'. Narrow enough that the existing 429 tier-gate pattern keeps its own reason. - agent/anthropic_adapter.py: _common_betas_for_base_url, build_anthropic_client, build_anthropic_kwargs gain drop_context_1m_beta kwarg. Default=False (1M stays). OAuth OAUTH_ONLY_BETAS unchanged. - agent/transports/anthropic.py: build_kwargs forwards the flag. - run_agent.py: self._oauth_1m_beta_disabled flag, retry-once guard, recovery branch next to the image-shrink path. _rebuild_anthropic_client honors the flag. The main build_kwargs call site threads it through for fast-mode extra_headers. - hermes_cli/doctor.py, hermes_cli/models.py: sibling OAuth /v1/models probes get the same reactive retry — previously they'd falsely report the Anthropic API as unreachable for affected subscriptions. Tests: 2190 tests/agent/ + 94 adjacent integration tests pass. New unit tests cover the classifier pattern (including the collision guard against the 429 tier-gate) and the drop_context_1m_beta adapter behavior (default keeps 1M, flag strips only 1M while preserving every other beta).

@JayGwod

…NousResearch#17752) Keep context-1m-2025-08-07 in OAuth requests by default so 1M-capable subscriptions retain full context. When Anthropic rejects a request with 400 'long context beta is not yet available for this subscription', disable the beta for the rest of the session, rebuild the client, and retry once. Addresses NousResearch#17680 (thanks @JayGwod for the clean reproduction) without forcing every OAuth user off the 1M context window. Changes: - agent/error_classifier.py: new FailoverReason.oauth_long_context_beta_forbidden; pattern matches 400 + 'long context beta' + 'not yet available'. Narrow enough that the existing 429 tier-gate pattern keeps its own reason. - agent/anthropic_adapter.py: _common_betas_for_base_url, build_anthropic_client, build_anthropic_kwargs gain drop_context_1m_beta kwarg. Default=False (1M stays). OAuth OAUTH_ONLY_BETAS unchanged. - agent/transports/anthropic.py: build_kwargs forwards the flag. - run_agent.py: self._oauth_1m_beta_disabled flag, retry-once guard, recovery branch next to the image-shrink path. _rebuild_anthropic_client honors the flag. The main build_kwargs call site threads it through for fast-mode extra_headers. - hermes_cli/doctor.py, hermes_cli/models.py: sibling OAuth /v1/models probes get the same reactive retry — previously they'd falsely report the Anthropic API as unreachable for affected subscriptions. Tests: 2190 tests/agent/ + 94 adjacent integration tests pass. New unit tests cover the classifier pattern (including the collision guard against the 429 tier-gate) and the drop_context_1m_beta adapter behavior (default keeps 1M, flag strips only 1M while preserving every other beta).

@JayGwod

…NousResearch#17752) Keep context-1m-2025-08-07 in OAuth requests by default so 1M-capable subscriptions retain full context. When Anthropic rejects a request with 400 'long context beta is not yet available for this subscription', disable the beta for the rest of the session, rebuild the client, and retry once. Addresses NousResearch#17680 (thanks @JayGwod for the clean reproduction) without forcing every OAuth user off the 1M context window. Changes: - agent/error_classifier.py: new FailoverReason.oauth_long_context_beta_forbidden; pattern matches 400 + 'long context beta' + 'not yet available'. Narrow enough that the existing 429 tier-gate pattern keeps its own reason. - agent/anthropic_adapter.py: _common_betas_for_base_url, build_anthropic_client, build_anthropic_kwargs gain drop_context_1m_beta kwarg. Default=False (1M stays). OAuth OAUTH_ONLY_BETAS unchanged. - agent/transports/anthropic.py: build_kwargs forwards the flag. - run_agent.py: self._oauth_1m_beta_disabled flag, retry-once guard, recovery branch next to the image-shrink path. _rebuild_anthropic_client honors the flag. The main build_kwargs call site threads it through for fast-mode extra_headers. - hermes_cli/doctor.py, hermes_cli/models.py: sibling OAuth /v1/models probes get the same reactive retry — previously they'd falsely report the Anthropic API as unreachable for affected subscriptions. Tests: 2190 tests/agent/ + 94 adjacent integration tests pass. New unit tests cover the classifier pattern (including the collision guard against the 429 tier-gate) and the drop_context_1m_beta adapter behavior (default keeps 1M, flag strips only 1M while preserving every other beta).

@JayGwod

…NousResearch#17752) Keep context-1m-2025-08-07 in OAuth requests by default so 1M-capable subscriptions retain full context. When Anthropic rejects a request with 400 'long context beta is not yet available for this subscription', disable the beta for the rest of the session, rebuild the client, and retry once. Addresses NousResearch#17680 (thanks @JayGwod for the clean reproduction) without forcing every OAuth user off the 1M context window. Changes: - agent/error_classifier.py: new FailoverReason.oauth_long_context_beta_forbidden; pattern matches 400 + 'long context beta' + 'not yet available'. Narrow enough that the existing 429 tier-gate pattern keeps its own reason. - agent/anthropic_adapter.py: _common_betas_for_base_url, build_anthropic_client, build_anthropic_kwargs gain drop_context_1m_beta kwarg. Default=False (1M stays). OAuth OAUTH_ONLY_BETAS unchanged. - agent/transports/anthropic.py: build_kwargs forwards the flag. - run_agent.py: self._oauth_1m_beta_disabled flag, retry-once guard, recovery branch next to the image-shrink path. _rebuild_anthropic_client honors the flag. The main build_kwargs call site threads it through for fast-mode extra_headers. - hermes_cli/doctor.py, hermes_cli/models.py: sibling OAuth /v1/models probes get the same reactive retry — previously they'd falsely report the Anthropic API as unreachable for affected subscriptions. Tests: 2190 tests/agent/ + 94 adjacent integration tests pass. New unit tests cover the classifier pattern (including the collision guard against the 429 tier-gate) and the drop_context_1m_beta adapter behavior (default keeps 1M, flag strips only 1M while preserving every other beta).

fix(anthropic): omit 1m context beta for native OAuth

34d30f4

alt-glitch added type/bug Something isn't working P1 High — major feature broken, no workaround comp/agent Core agent loop, run_agent.py, prompt builder provider/anthropic Anthropic native Messages API area/auth Authentication, OAuth, credential pools labels Apr 30, 2026

teknium1 mentioned this pull request Apr 30, 2026

fix(anthropic): reactive recovery for OAuth 1M-context beta rejection #17752

Merged

teknium1 closed this in #17752 Apr 30, 2026

JayGwod mentioned this pull request May 2, 2026

fix(anthropic): use Claude Code-compatible MCP tool prefix for OAuth #17681

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fix(anthropic): omit 1m context beta for native OAuth#17680

fix(anthropic): omit 1m context beta for native OAuth#17680
JayGwod wants to merge 1 commit into
NousResearch:mainfrom
JayGwod:fix/anthropic-oauth-context-1m-beta

JayGwod commented Apr 30, 2026

Uh oh!

alt-glitch commented Apr 30, 2026

Uh oh!

alt-glitch commented Apr 30, 2026

Uh oh!

JayGwod commented Apr 30, 2026

Uh oh!

teknium1 commented Apr 30, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

JayGwod commented Apr 30, 2026

Summary

Why

Behavior preserved

Tests

Uh oh!

alt-glitch commented Apr 30, 2026

Uh oh!

alt-glitch commented Apr 30, 2026

Uh oh!

JayGwod commented Apr 30, 2026

Uh oh!

teknium1 commented Apr 30, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants