feat(amazon-bedrock-mantle): add known context windows for open-weight Mantle models#68725
feat(amazon-bedrock-mantle): add known context windows for open-weight Mantle models#68725wirjo wants to merge 4 commits intoopenclaw:mainfrom
Conversation
Greptile SummaryAdds Claude Opus 4.7 support on Bedrock Mantle by enriching model discovery with a static context-window map, an image-input allowlist, and reasoning pattern entries for the new Claude generation. The tests are well-structured and cover the new code paths. Two minor P2 observations about the allowlist design are noted inline. Confidence Score: 5/5Safe to merge — all findings are P2 style suggestions with no correctness impact on current behavior. The changes are well-scoped, all current models are covered correctly, and the new test suite directly validates context windows, image input, and reasoning flags. The only issues are maintenance-oriented P2 concerns about how future Claude models will be handled. No files require special attention. Prompt To Fix All With AIThis is a comment left during a code review.
Path: extensions/amazon-bedrock-mantle/discovery.ts
Line: 51-56
Comment:
**Inconsistent matching strategy vs. stated intent**
The PR description states "all Claude models on Mantle support image input," but `resolveInputModalities` uses an explicit allowlist with `modelId.includes(p)`. Meanwhile, `REASONING_PATTERNS` uses short substrings like `"claude-opus-4-7"` that naturally cover future variants. A new Claude model added to Mantle but not to `IMAGE_INPUT_MODELS` will silently get `input: ["text"]` — contradicting the stated intent. A `modelId.includes("anthropic.claude")` pattern check would be consistent with `REASONING_PATTERNS`' approach and eliminate the need to keep two parallel lists in sync.
How can I resolve this? If you propose a fix, please make it concise.
---
This is a comment left during a code review.
Path: extensions/amazon-bedrock-mantle/discovery.ts
Line: 39-45
Comment:
**Two parallel lists to keep in sync**
`IMAGE_INPUT_MODELS` and the keys of `KNOWN_CONTEXT_WINDOWS` are identical, which means every new Claude model on Mantle requires updating both structures. Since `resolveInputModalities` already iterates, consider deriving image support directly from `KNOWN_CONTEXT_WINDOWS` (i.e., `Object.keys(KNOWN_CONTEXT_WINDOWS).some((k) => modelId.includes(k))`) or consolidating into a single model metadata map to reduce the maintenance surface.
How can I resolve this? If you propose a fix, please make it concise.Reviews (1): Last reviewed commit: "feat(amazon-bedrock-mantle): add Claude ..." | Re-trigger Greptile |
| function resolveInputModalities(modelId: string): Array<"text" | "image"> { | ||
| if (IMAGE_INPUT_MODELS.some((p) => modelId.includes(p))) { | ||
| return ["text", "image"]; | ||
| } | ||
| return ["text"]; | ||
| } |
There was a problem hiding this comment.
The PR description states "all Claude models on Mantle support image input," but resolveInputModalities uses an explicit allowlist with modelId.includes(p). Meanwhile, REASONING_PATTERNS uses short substrings like "claude-opus-4-7" that naturally cover future variants. A new Claude model added to Mantle but not to IMAGE_INPUT_MODELS will silently get input: ["text"] — contradicting the stated intent. A modelId.includes("anthropic.claude") pattern check would be consistent with REASONING_PATTERNS' approach and eliminate the need to keep two parallel lists in sync.
Prompt To Fix With AI
This is a comment left during a code review.
Path: extensions/amazon-bedrock-mantle/discovery.ts
Line: 51-56
Comment:
**Inconsistent matching strategy vs. stated intent**
The PR description states "all Claude models on Mantle support image input," but `resolveInputModalities` uses an explicit allowlist with `modelId.includes(p)`. Meanwhile, `REASONING_PATTERNS` uses short substrings like `"claude-opus-4-7"` that naturally cover future variants. A new Claude model added to Mantle but not to `IMAGE_INPUT_MODELS` will silently get `input: ["text"]` — contradicting the stated intent. A `modelId.includes("anthropic.claude")` pattern check would be consistent with `REASONING_PATTERNS`' approach and eliminate the need to keep two parallel lists in sync.
How can I resolve this? If you propose a fix, please make it concise.| const IMAGE_INPUT_MODELS = [ | ||
| "anthropic.claude-opus-4-7", | ||
| "anthropic.claude-opus-4-6-v1", | ||
| "anthropic.claude-sonnet-4-6", | ||
| "anthropic.claude-sonnet-4-5-20250929-v1:0", | ||
| "anthropic.claude-haiku-4-5-20251001-v1:0", | ||
| ]; |
There was a problem hiding this comment.
IMAGE_INPUT_MODELS and the keys of KNOWN_CONTEXT_WINDOWS are identical, which means every new Claude model on Mantle requires updating both structures. Since resolveInputModalities already iterates, consider deriving image support directly from KNOWN_CONTEXT_WINDOWS (i.e., Object.keys(KNOWN_CONTEXT_WINDOWS).some((k) => modelId.includes(k))) or consolidating into a single model metadata map to reduce the maintenance surface.
Prompt To Fix With AI
This is a comment left during a code review.
Path: extensions/amazon-bedrock-mantle/discovery.ts
Line: 39-45
Comment:
**Two parallel lists to keep in sync**
`IMAGE_INPUT_MODELS` and the keys of `KNOWN_CONTEXT_WINDOWS` are identical, which means every new Claude model on Mantle requires updating both structures. Since `resolveInputModalities` already iterates, consider deriving image support directly from `KNOWN_CONTEXT_WINDOWS` (i.e., `Object.keys(KNOWN_CONTEXT_WINDOWS).some((k) => modelId.includes(k))`) or consolidating into a single model metadata map to reduce the maintenance surface.
How can I resolve this? If you propose a fix, please make it concise.Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!
…t models Mantle's /v1/models endpoint returns only model IDs with no token limit metadata. All models get a hardcoded 32K context window, which is wrong for most models (MiniMax M2: 1M, Qwen Coder: 256K, etc.). Add a KNOWN_CONTEXT_WINDOWS lookup table for open-weight models currently available on Mantle and raise the default fallback from 32K to 128K for unknown models. Note: Claude models are NOT included — they use the separate /anthropic/v1/messages endpoint on Mantle, not the OpenAI-compatible /v1/chat/completions endpoint that this extension handles. Claude model metadata is managed by the amazon-bedrock extension.
ecbaa9c to
ab9b0d0
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: ecbaa9c795
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| function resolveKnownContextWindow(modelId: string): number | undefined { | ||
| return KNOWN_CONTEXT_WINDOWS[modelId]; |
There was a problem hiding this comment.
Match known Claude context windows against ID variants
resolveKnownContextWindow does an exact key lookup, so any Mantle ID variant (for example a prefixed or suffixed Claude ID such as ...-v1:0) falls through to the 128K default even when it belongs to a model family you already marked as 1M/200K elsewhere in this change. In those cases discovery returns inconsistent metadata (reasoning/image may match by substring while context does not), which can cause unnecessary prompt truncation for affected Claude IDs.
Useful? React with 👍 / 👎.
|
@vincentkoc Ready for review — no CI failures, single-extension change. Happy to address any feedback. |
Cross-referenced against known Mantle model list. Added: - qwen.qwen3-next-80b-a3b-instruct (128K) - qwen.qwen3-vl-235b-a22b-instruct (128K) - mistral.mistral-large-3-675b-instruct (128K) - mistral.devstral-2-123b (128K) Reordered entries by context window size (1M first) for readability. Models at 128K default (GPT-OSS, Kimi, Mistral small, etc.) are intentionally omitted — the 128K fallback handles them correctly.
Validated against live Mantle /v1/models endpoint (39 models). Added missing entries: - anthropic.claude-opus-4-7 (1M) — appears in /v1/models; uses /anthropic/v1/messages for inference, not Chat Completions - minimax.minimax-m2.5 (1M) - nvidia.nemotron-super-3-120b (128K) - zai.glm-5 (128K)
Consistent with openclaw#65952. 200K matches the Claude model floor — ensures new models get a sensible default without table updates.
|
Raised default from 128K → 200K, consistent with #65952. Claude is the most popular model family and 200K is the floor for all current Claude models — this future-proofs against new releases without requiring table updates. |
|
Codex review: needs changes before merge. Summary Reproducibility: yes. The Next step before merge Security Review findings
Review detailsBest possible solution: Land a rebased Mantle-local metadata update that maps known open-weight model IDs, keeps fallback behavior conservative, preserves the current Claude Anthropic Messages handling, and adds focused tests plus a changelog entry. Do we have a high-confidence way to reproduce the issue? Yes. The Is this the best way to solve the issue? No, not as written. A static known-context-window table is the right narrow shape, but this branch needs conservative fallback semantics, aligned Nemotron metadata, and release-note/test coverage before merge. Full review comments:
Overall correctness: patch is incorrect Acceptance criteria:
What I checked:
Likely related people:
Remaining risk / open question:
Codex review notes: model gpt-5.5, reasoning high; reviewed against 3980eaa1c22e. |
Problem
Mantle's
/v1/modelsendpoint returns only model IDs — no token limit metadata. Discovery hardcodescontextWindow: 32000for every model, which is wrong for most:This causes the same premature context overflow and wrong compaction thresholds as #65952 (the amazon-bedrock equivalent).
Fix
Add a
KNOWN_CONTEXT_WINDOWSlookup table for open-weight models currently available on Mantle. Raise the default fallback from 32K → 128K for unknown models.Claude models are intentionally excluded — they use a separate Anthropic-native endpoint on Mantle (
/anthropic/v1/messages), not the OpenAI-compatible/v1/chat/completionsendpoint that this extension handles. Claude model metadata is managed by theamazon-bedrockextension (#65952).Changes
extensions/amazon-bedrock-mantle/discovery.tsresolveKnownContextWindow(), default bumpSingle file, no type system changes.
References
/anthropic/v1/messageson Mantle, not Chat Completions