feat(models): support Grok 4.5/4.6 on xAI and OpenRouter by PeterDaveHello · Pull Request #2871 · The-PR-Agent/pr-agent

PeterDaveHello · 2026-08-28T15:51:03Z

Summary

register published Grok 4.5 and 4.6 model IDs, aliases, and canonical OpenRouter routes with 500K context windows
clamp always-on Grok reasoning effort to model-specific levels for native xAI and OpenRouter requests
add a targeted compatibility workaround for the LiteLLM grok-build-latest registry gap
update the xAI configuration guide with aliases, defaults, global effort impact, and routing limitations

Behavior changes

PR-Agent sends its configured reasoning effort, medium by default, to Grok 4.5 and 4.6; high restores the native xAI default but also affects other registered reasoning models
none and minimal map to low because reasoning cannot be disabled
max and xhigh map to high on Grok 4.5; Grok 4.6 retains xhigh and maps max to xhigh
OpenRouter-specific precedence is resolved before the same Grok clamp is applied
OpenRouter routing suffixes such as :nitro currently require custom_model_max_tokens

Testing

focused token, reasoning, OpenRouter, and handler suites: 239 passed
Python compilation and git diff whitespace checks passed

References

qodo-code-review · 2026-08-28T15:52:02Z

PR Summary by Qodo

Support Grok 4.5 and 4.6 across xAI and OpenRouter

✨ Enhancement 🧪 Tests 📝 Documentation ⚙️ Configuration changes 🕐 20-40 Minutes

AI Description

• Registers Grok 4.5/4.6 routes with 500K context windows.
• Clamps always-on reasoning effort across native xAI and OpenRouter requests.
• Documents configuration behavior and verifies token, routing, and LiteLLM compatibility.

Diagram

sequenceDiagram
    participant C as Configuration
    participant H as LiteLLM Handler
    participant R as Model Registry
    participant G as Grok Clamp
    participant L as LiteLLM
    participant P as xAI or OpenRouter
    C->>H: Model and effort
    H->>R: Resolve capabilities
    R-->>H: Context and levels
    H->>G: Normalize final effort
    G-->>H: Supported effort
    H->>L: Provider request
    L->>P: Route completion
    P-->>L: Model response
    L-->>H: Completion result

High-Level Assessment

The model-specific capability table plus shared clamp helpers fits the existing registry-based design and keeps native xAI and OpenRouter behavior consistent. Relying solely on LiteLLM was considered but is unsuitable while the pinned registry omits grok-build-latest and does not encode the required always-on, generation-specific effort semantics.

Files changed (6) +260 / -2

Enhancement (2) +67 / -0

__init__.pyRegister Grok models and reasoning capabilities +20/-0

Register Grok models and reasoning capabilities

• Adds native xAI and canonical OpenRouter Grok 4.5/4.6 IDs with 500K context windows. Registers reasoning-capable aliases and defines generation-specific accepted effort levels.

pr_agent/algo/init.py

litellm_ai_handler.pyClamp and route Grok reasoning effort +47/-0

Clamp and route Grok reasoning effort

• Adds Grok model normalization and effort clamping for native xAI and OpenRouter requests after provider-specific precedence is resolved. It also allowlists reasoning_effort for grok-build-latest to bridge the pinned LiteLLM registry gap.

pr_agent/algo/ai_handlers/litellm_ai_handler.py

Tests (2) +184 / -0

test_get_max_tokens.pyVerify Grok 500K context windows +20/-0

Verify Grok 500K context windows
• Adds parameterized coverage confirming all registered xAI aliases and canonical OpenRouter routes resolve to 500,000 tokens.
tests/unittest/test_get_max_tokens.py

test_litellm_reasoning_effort.pyCover Grok reasoning and LiteLLM compatibility +164/-0

Cover Grok reasoning and LiteLLM compatibility
• Tests generation-specific effort levels, clamp mappings, native xAI forwarding, OpenRouter precedence and budgets, routing suffixes, invalid overrides, and the grok-build-latest LiteLLM allowlist workaround.
tests/unittest/test_litellm_reasoning_effort.py

Documentation (1) +8 / -2

changing_a_model.mdDocument Grok 4.5/4.6 setup and reasoning behavior +8/-2

Document Grok 4.5/4.6 setup and reasoning behavior
• Updates the xAI example to Grok 4.6 and documents registered IDs, aliases, 500K context windows, reasoning-effort clamping, OpenRouter precedence, and routing-suffix token limitations.
docs/docs/usage-guide/changing_a_model.md

Other (1) +1 / -0

configuration.tomlClarify always-on reasoning clamping +1/-0

Clarify always-on reasoning clamping
• Documents that explicit none effort is clamped to the lowest supported level for always-on models such as Grok.
pr_agent/settings/configuration.toml

qodo-code-review · 2026-08-28T15:54:02Z

Code Review by Qodo

🐞 Bugs (0) 📘 Rule violations (2) 📜 Skill insights (0)

1. reasoning_effort comment uses narrative phrasing 📘 Rule violation ⚙ Maintainability ⭐ New

Description

The added comment says the setting “keeps” reasoning disabled and describes Grok behavior
narratively rather than using imperative phrasing. Rewrite it as an instruction, such as “Use none
to keep reasoning disabled...” while retaining the Grok exception.

Code

pr_agent/settings/configuration.toml[R423-425]

+# Explicit openrouter.reasoning_effort = "none" keeps reasoning disabled, except on
+# Grok 4.5/4.6: there "none" is clamped to the lowest supported
+# effort first, so a positive budget wins over it.

Relevance

●●● Strong
Recent configuration-comment precedents consistently accepted imperative rewrites, including
matching phrasing-rule findings.
PR-#2823
PR-#2796
PR-#2862

ⓘ Recommendations generated based on similar findings in past PRs

Evidence
Compliance rule 2694688 requires newly added behavior-describing comments to use imperative
phrasing. The added comment at lines 423-425 narrates what the setting does rather than instructing
the reader how to configure it.
Rule 2694688: Docstrings and comments must use imperative phrasing
pr_agent/settings/configuration.toml[423-425]

Agent prompt

The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
The new `openrouter.reasoning_effort` comment uses descriptive third-person phrasing (`keeps`, `is clamped`) instead of imperative phrasing.

## Issue Context
Preserve the documented Grok 4.5/4.6 exception and budget precedence while rewriting the comment as an instruction.

## Fix Focus Areas
- pr_agent/settings/configuration.toml[423-425]

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools

2. Grok docs exceed 120 characters 📘 Rule violation ⚙ Maintainability

Description

The newly added Grok documentation paragraphs on lines 196, 198, and 200 each exceed the
120-character maximum. These paragraphs should be wrapped across physical lines to satisfy the
repository-wide line-length requirement.

Code

docs/docs/usage-guide/changing_a_model.md[196]

+Grok 4.5 and Grok 4.6 are registered with a 500K token context window (`xai/grok-4.5`, `xai/grok-4.5-latest`, `xai/grok-build-latest`, `xai/grok-4.6`, `openrouter/x-ai/grok-4.5`, `openrouter/x-ai/grok-4.6`). xAI publishes `grok-4.5-latest` and `grok-build-latest` aliases for Grok 4.5; Grok 4.6 currently has no published alias.

Relevance

● Weak
Recent repository precedents explicitly rejected identical requests to wrap long Markdown
documentation lines for the 120-character rule.
PR-#2817
PR-#2774

ⓘ Recommendations generated based on similar findings in past PRs

Evidence
PR Compliance ID 2694690 requires every non-generated modified source line to be at most 120
characters. The PR adds three unwrapped Markdown paragraphs whose physical lines substantially
exceed that limit.
Rule 2694690: Enforce maximum line length of 120 characters
docs/docs/usage-guide/changing_a_model.md[196-196]
docs/docs/usage-guide/changing_a_model.md[198-198]
docs/docs/usage-guide/changing_a_model.md[200-200]

Agent prompt

The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
The newly added Grok documentation paragraphs exceed the required 120-character maximum line length.

## Issue Context
Wrap the Markdown source without changing the rendered paragraph content or meaning.

## Fix Focus Areas
- docs/docs/usage-guide/changing_a_model.md[196-196]
- docs/docs/usage-guide/changing_a_model.md[198-198]
- docs/docs/usage-guide/changing_a_model.md[200-200]

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools

Author self-review: I have reviewed the code review findings, and addressed the relevant ones.

Context sources

✅ Compliance rules (platform): 34 rules

Review mode: ⚖️ Balanced: The push changes runtime model registration and reasoning/precedence behavior across native xAI and OpenRouter paths, creating real compatibility and request-shaping risk, but not enough independent logic for extended review.

Tip of the day

💡 Did you know, you can group findings by type and pick your Finding display, from Minimal to Full

More tips ↗ | Customize Qodo ↗ | Qodo docs ↗

Previous reviews

Review updated until commit 7b3168d ⚖️ Balanced

Results up to commit 08e4918 ⚖️ Balanced

🐞 Bugs (0) 📘 Rule violations (1) 📎 Requirement gaps (0) 🎨 UX issues (0) 📜 Skill insights (0)

1. Grok docs exceed 120 characters 📘 Rule violation ⚙ Maintainability

Description

The newly added Grok documentation paragraphs on lines 196, 198, and 200 each exceed the
120-character maximum. These paragraphs should be wrapped across physical lines to satisfy the
repository-wide line-length requirement.

Code

docs/docs/usage-guide/changing_a_model.md[196]

+Grok 4.5 and Grok 4.6 are registered with a 500K token context window (`xai/grok-4.5`, `xai/grok-4.5-latest`, `xai/grok-build-latest`, `xai/grok-4.6`, `openrouter/x-ai/grok-4.5`, `openrouter/x-ai/grok-4.6`). xAI publishes `grok-4.5-latest` and `grok-build-latest` aliases for Grok 4.5; Grok 4.6 currently has no published alias.

Relevance

● Weak
Recent repository precedents explicitly rejected identical requests to wrap long Markdown
documentation lines for the 120-character rule.
PR-#2817
PR-#2774

ⓘ Recommendations generated based on similar findings in past PRs

Evidence
PR Compliance ID 2694690 requires every non-generated modified source line to be at most 120
characters. The PR adds three unwrapped Markdown paragraphs whose physical lines substantially
exceed that limit.
Rule 2694690: Enforce maximum line length of 120 characters
docs/docs/usage-guide/changing_a_model.md[196-196]
docs/docs/usage-guide/changing_a_model.md[198-198]
docs/docs/usage-guide/changing_a_model.md[200-200]

Agent prompt

The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
The newly added Grok documentation paragraphs exceed the required 120-character maximum line length.

## Issue Context
Wrap the Markdown source without changing the rendered paragraph content or meaning.

## Fix Focus Areas
- docs/docs/usage-guide/changing_a_model.md[196-196]
- docs/docs/usage-guide/changing_a_model.md[198-198]
- docs/docs/usage-guide/changing_a_model.md[200-200]

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools

IsmaelMartinez

Thanks, this is carefully done and the tests earn it: 29 of the 35 go red when the source is reverted, and of the six that stay green, four are the litellm registry pins your docstring flags as deliberate.

One suggestion inline on configuration.toml, then this is good to go.

Qodo's line-length finding on changing_a_model.md I would leave alone. The same rule has been raised and let through on #2817 and #2774, and that file already carries lines of 1098 and 785 characters.

qodo-code-review · 2026-08-29T11:23:14Z

Code review by qodo was updated up to the latest commit 7b3168d

naorpeled · 2026-08-29T19:48:59Z

Hey @PeterDaveHello,
Thanks for this!
Could you please resolve the merge conflicts?

Register published xAI model IDs and aliases, plus their canonical OpenRouter routes, with 500K context windows so token budgeting works without custom limits. Clamp always-on Grok reasoning values to model-specific levels for native xAI and OpenRouter calls. Allow reasoning_effort through pinned LiteLLM for the grok-build-latest alias missing from its registry. PR-Agent sends its configured effort, medium by default, to these models; use high to retain xAI's native default. References: - https://docs.x.ai/developers/models/grok-4.5 - https://docs.x.ai/developers/models/grok-4.6 - https://openrouter.ai/docs/guides/best-practices/reasoning-tokens

qodo-code-review · 2026-08-30T04:19:20Z

Code Review by Qodo

Sorry, something went wrong

We weren't able to complete the code review on our side. Please try again manually by commenting /agentic_review on this PR.

naorpeled · 2026-08-30T05:00:19Z

Thanks Peter!

@IsmaelMartinez

…dels Addresses @IsmaelMartinez's review on The-PR-Agent#2530. All three findings verified against the code and reproduced before fixing. 1. OpenRouter double-configuration. He wrote: "grok_model is the basename, so openrouter/x-ai/grok-4 lands here too and gets reasoning_effort from [config] while the OpenRouter block below is already setting extra_body.reasoning from [openrouter]." Reproduced exactly -- with the flag on, openrouter/x-ai/grok-4.5 came out with BOTH top-level reasoning_effort='high' (from [config]) and extra_body.reasoning={'effort':'high'} (from [openrouter]). Adopted his inline `not is_openrouter` suggestion; the OpenRouter block is now the sole owner of that path. 2. allowed_openai_params too broad. He wrote: "grok-2 and grok-3 have reasoning_effort dropped without the allowlist and included with it ... grok-4.5 does not need the allowlist at all; grok-4.20-multi-agent genuinely does." Confirmed against the pinned litellm: get_supported_openai_params reports reasoning_effort for xai/grok-4.5 and xai/grok-4.20-multi-agent but not for xai/grok-3. The allowlist is now added only when litellm does not already report the param, mirroring how main does it for its own Grok path. 3. MAX_TOKENS entries. Added xai/grok-4.20-multi-agent (main's The-PR-Agent#2871, merged today, added the 4.5/4.6 entries but not this one), so the documented xhigh-capable model can actually run without custom_model_max_tokens. Also, and this narrows the PR substantially: main gained native always-on Grok 4.5/4.6 reasoning support in The-PR-Agent#2871 while this sat. The opt-in branch now also skips any model main already covers via GROK_REASONING_EFFORT_LEVELS, so it no longer overrides main's clamping. The residual scope is xAI-direct Grok families main does not cover. Rebased onto main (a64730a); the configuration.toml conflict he flagged is resolved (kept main's add_user_to_requests + "max" effort alongside our flag). Tests: adds the OpenRouter-isolation regression, an allowlist-narrowing test, and a test that natively-supported models are left to main. Adopts the _restore_litellm_globals autouse fixture per his third note. tests/unittest: 2702 passed, 1 pre-existing failure (TestLiteLLMReasoningEffortGrok::...[grok-4.6-xhigh-False]) that reproduces identically on pristine upstream/main and is untouched by this branch. Co-authored-by: IsmaelMartinez <IsmaelMartinez@users.noreply.github.com>

github-actions Bot added the feature 💡 label Aug 28, 2026

qodo-code-review Bot added the 🕐 20-40 Minutes label Aug 28, 2026

IsmaelMartinez reviewed Aug 28, 2026

View reviewed changes

Comment thread pr_agent/settings/configuration.toml

PeterDaveHello force-pushed the feature/grok45And46Support branch from 08e4918 to 7b3168d Compare August 29, 2026 11:21

PeterDaveHello requested a review from IsmaelMartinez August 29, 2026 11:26

naorpeled approved these changes Aug 29, 2026

View reviewed changes

PeterDaveHello force-pushed the feature/grok45And46Support branch from 7b3168d to cb8fa07 Compare August 30, 2026 04:19

naorpeled merged commit 505997f into The-PR-Agent:main Aug 30, 2026
5 checks passed

PeterDaveHello deleted the feature/grok45And46Support branch August 30, 2026 10:44

Kyzcreig mentioned this pull request Aug 30, 2026

feat(litellm): forward reasoning effort to Grok #2530

Closed

IsmaelMartinez mentioned this pull request Sep 4, 2026

qodo-code-review[bot] has failed on every PR for seven days #3047

Closed

Navigation Menu

Uh oh!

Uh oh!

feat(models): support Grok 4.5/4.6 on xAI and OpenRouter - #2871

feat(models): support Grok 4.5/4.6 on xAI and OpenRouter#2871
naorpeled merged 1 commit into
The-PR-Agent:mainfrom
PeterDaveHello:feature/grok45And46Support

PeterDaveHello commented Aug 28, 2026

Uh oh!

qodo-code-review Bot commented Aug 28, 2026

Uh oh!

qodo-code-review Bot commented Aug 28, 2026 •

edited

Loading

Uh oh!

IsmaelMartinez left a comment

Uh oh!

Uh oh!

qodo-code-review Bot commented Aug 29, 2026

Uh oh!

naorpeled commented Aug 29, 2026

Uh oh!

qodo-code-review Bot commented Aug 30, 2026 •

edited

Loading

Uh oh!

Uh oh!

naorpeled commented Aug 30, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Uh oh!

Uh oh!

Conversation

PeterDaveHello commented Aug 28, 2026

Summary

Behavior changes

Testing

References

Uh oh!

qodo-code-review Bot commented Aug 28, 2026

PR Summary by Qodo

Uh oh!

qodo-code-review Bot commented Aug 28, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Code Review by Qodo

Uh oh!

IsmaelMartinez left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

qodo-code-review Bot commented Aug 29, 2026

Uh oh!

naorpeled commented Aug 29, 2026

Uh oh!

qodo-code-review Bot commented Aug 30, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Code Review by Qodo

Sorry, something went wrong

Uh oh!

Uh oh!

naorpeled commented Aug 30, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

qodo-code-review Bot commented Aug 28, 2026 •

edited

Loading

qodo-code-review Bot commented Aug 30, 2026 •

edited

Loading