fix(gateway): model-aware overflow threshold (unhardcode 200k context)#478
Merged
Conversation
## Bug
``DEFAULT_THRESHOLD.contextWindow`` was hardwired to 200_000 tokens
— right for Claude Opus 4.x but wrong for every other model. See
``bugs/known-issues.md#context-window-hardcoded``.
Operators switching the planner to GPT-4o-mini (128k context)
saw a provider-side ``context_length_exceeded`` 400 before
compaction fired, because the threshold thought there was still
~70k of runway. Operators switching to Gemini Flash (1M) paid
for summarization runs that weren't necessary.
## Fix
``thresholdForModel(model, override?)`` in
``src/session/overflow.ts`` resolves the context window from a
lookup table keyed on ``provider/modelId``. Populated with the
models the gateway actually runs against today:
- Anthropic 4.x (Opus / Sonnet / Haiku) — 200k
- OpenAI GPT-4o / GPT-4o-mini — 128k
- OpenAI GPT-4.1 / 4.1-mini — 1_047_576
- OpenAI o3 / o3-mini — 200k
Unknown models fall back to 128k. The fallback is deliberately
smaller than every common default — triggers compaction slightly
earlier than strictly necessary, never lets the request hit a
provider-side overflow. Strictly-safer default.
Operators with exotic models can:
- Pass ``override.contextWindow`` at the call site (used by
tests, planner config overrides).
- Declare the full threshold in ``gateway.config.json`` and
route it through the existing ``compactIfNeeded({threshold})``
signature — no schema changes needed.
``compactIfNeeded`` now resolves the threshold from ``input.model``
instead of the single ``DEFAULT_THRESHOLD`` constant. One-line
change at the call site; behavior unchanged for Claude Opus.
## Tests
New: ``tests/session/overflow-threshold.test.ts`` (14 tests).
Covers:
- Lookup for every provider/model in the table
- Unknown-model fallback to the 128k default (regression guard)
- Override precedence (operator escape hatch)
- Before/after overflow math at specific token counts — the
key regression test: GPT-4o-mini @ 100k now correctly
overflows where before it would NOT have.
Full suite: 127/127 pass.
## Known-issues entry
Marked fixed in ``bugs/known-issues.md#context-window-hardcoded``
with a pointer to the new resolver + test file.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
Caution Review failedThe pull request is closed. ℹ️ Recent review info⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Pro Run ID: 📒 Files selected for processing (4)
📝 WalkthroughWalkthroughThe pull request implements model-aware context window threshold resolution. It adds a lookup table mapping provider/model identifiers to specific context windows, introduces a Changes
Estimated code review effort🎯 3 (Moderate) | ⏱️ ~20 minutes Poem
✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
3 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What this fixes
known-issues.md#context-window-hardcoded — the overflow threshold was hardwired to 200k tokens (Claude Opus 4.x's window). Operators swapping to a different model got the wrong trigger point:
anthropic/claude-opus-4-7openai/gpt-4o-miniopenai/gpt-4.1How
New
thresholdForModel(model, override?)in gateway/src/session/overflow.ts resolves from a lookup table.compactIfNeededin gateway/src/session/compaction.ts now passes the planner's actual model through.Operators with unusual models still have two escape hatches:
compactIfNeeded({threshold: {contextWindow: 2_000_000}})— one-offgateway.config.json → agent.plannerwith acontextWindowfield — persistentTests
14 new tests in gateway/tests/session/overflow-threshold.test.ts guarding both the lookup table and the overflow math itself — the key regression being GPT-4o-mini @ 100k tokens now correctly triggers compaction where it would not have before.
Full suite: 127/127 pass.
Known-issues
Marked ✅ fixed in bugs/known-issues.md with pointers to the resolver + test file.
🤖 Generated with Claude Code
Summary by CodeRabbit
Bug Fixes
Tests