feat(core): add reactive compression on context overflow#3879
Merged
Conversation
Contributor
There was a problem hiding this comment.
Pull request overview
Adds a reactive recovery path for provider context-window overflow errors in the Gemini chat flow: when a request fails due to context length, the chat attempts a forced compression of the current conversation and retries the failed turn once with the compressed context (with a single-attempt guard).
Changes:
- Introduces a context-length overflow classifier (
getContextLengthExceededInfo) that extracts nested messages (including embedded JSON) and parses token counts when available. - Extends chat compression to accept an explicit trigger (
manualvsauto) so forced compressions can still be reported as automatic when appropriate. - Adds reactive compression + retry-once behavior in
GeminiChat.sendMessageStream, with tests covering success, NOOP, throw, and single-retry guard behavior.
Reviewed changes
Copilot reviewed 5 out of 5 changed files in this pull request and generated 1 comment.
Show a summary per file
| File | Description |
|---|---|
| packages/core/src/utils/contextLengthError.ts | New utility to classify context-length overflow errors and extract token counts from provider messages. |
| packages/core/src/utils/contextLengthError.test.ts | Unit tests for overflow classification, timeout exclusions, token parsing, and nested JSON/object extraction. |
| packages/core/src/services/chatCompressionService.ts | Adds CompactTrigger and plumbs an explicit trigger into Pre/Post compact hook reporting. |
| packages/core/src/core/geminiChat.ts | Adds reactive compression-on-overflow and one-time retry behavior during streaming sends. |
| packages/core/src/core/geminiChat.test.ts | Integration tests validating reactive compression behavior and send-lock release on compression failure. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Force-compress chat history after a provider rejects a request for exceeding the context window, then retry the request once with refreshed history. Add provider-agnostic context overflow classification and focused retry coverage. Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
834b032 to
031d52a
Compare
Contributor
Code Coverage Summary
CLI Package - Full Text ReportCore Package - Full Text ReportFor detailed HTML reports, please see the 'coverage-reports-22.x-ubuntu-latest' artifact from the main CI run. |
wenshao
requested changes
May 8, 2026
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
wenshao
requested changes
May 9, 2026
wenshao
approved these changes
May 9, 2026
wenshao
left a comment
Collaborator
There was a problem hiding this comment.
No issues found. LGTM! ✅ — DeepSeek/deepseek-v4-pro via Qwen Code /review
This was referenced May 9, 2026
6 tasks
yiliang114
added a commit
that referenced
this pull request
May 19, 2026
… auto-compaction redesign - OOM reproduction report: root cause confirmed as structuredClone() positive feedback loop during auto-compaction (#3735, #3879), with real debug log evidence from crash session. - Runtime diagnostics benchmark: process-tree RSS sampling results comparing installed CLI vs local rebuilt bundle. - Auto-compaction threshold redesign: proposal for replacing the fixed 70% token threshold with RSS-aware graduated strategy.
yiliang114
added a commit
that referenced
this pull request
May 20, 2026
… auto-compaction redesign - OOM reproduction report: root cause confirmed as structuredClone() positive feedback loop during auto-compaction (#3735, #3879), with real debug log evidence from crash session. - Runtime diagnostics benchmark: process-tree RSS sampling results comparing installed CLI vs local rebuilt bundle. - Auto-compaction threshold redesign: proposal for replacing the fixed 70% token threshold with RSS-aware graduated strategy.
mabry1985
added a commit
to protoLabsAI/protoCLI
that referenced
this pull request
May 27, 2026
Adapted from QwenLM#3879 (+QwenLM#3985 hardening). When a provider rejects a request because the prompt exceeds the context window, force-compress the chat history once and retry the request — instead of surfacing a hard failure. Because our geminiChat has diverged substantially from upstream (≈1/3 the size, bespoke overflow handling, compression owned by client.tryCompressChat rather than a geminiChat.tryCompress method), this is an adaptation, not a patch port: - New `utils/contextLengthError.ts`: provider-agnostic `getContextLengthExceededInfo` that classifies an arbitrary error (incl. nested causes / embedded JSON) as a context-overflow vs. a timeout, and extracts actual/limit token counts. Ported near-verbatim (self-contained) with full test coverage. - geminiChat streaming retry loop: on a classified overflow, compress the LIVE chat via `ChatCompressionService` (force=true) — NOT client.tryCompressChat, which calls startChat() and would swap the chat instance out from under the in-flight generator. Install the compressed history with setHistory, mirror the essential side effects (FileReadCache.clear + telemetry token count), refresh the request contents, and retry once. A one-shot flag prevents any compress/retry loop if the request is still too large after compression. Tests: 25 classifier cases; 3 geminiChat reactive cases (compress→retry succeeds; non-overflow errors don't compress; at-most-once guard). 39 geminiChat + 51 compression-service tests pass; core builds; eslint clean. Co-authored-by: Automaker <automaker@localhost> Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
mabry1985
added a commit
to protoLabsAI/protoCLI
that referenced
this pull request
May 27, 2026
…sion, SDK canUseTool timeout) (#331) Documents behavior shipped in 0.51.x–0.52.0 that lacked coverage: - guides/goal.md: the evaluator's three verdicts (met / not met / impossible), the conservative impossible→abandoned terminal state, and how `/goal` status reports an abandoned goal. (QwenLM#4230) - explanation/agent-harness.md: new "Reactive compression" section — proactive vs. reactive compression and the one-shot force-compress-and-retry on a provider context-overflow rejection. (QwenLM#3879) - reference/sdk-api.md + contributing/sdk-typescript.md: document the `timeout.canUseTool` option and that the SDK now forwards it to the CLI so the control-plane timeout matches the callback's. (QwenLM#4491) (The `--max-tool-calls` / `--max-wall-time` budgets were already documented in guides/run-headless.md when those features shipped.) Co-authored-by: Automaker <automaker@localhost> Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
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
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.
Summary
Validation
reactive-compression-verify-211430:context_length_exceeded, prompt-too-long token-count errors, DashScope input-length errors, timeout/deadline negatives, compression success, compression NOOP, compression exceptions, repeated overflow after retry, and avoiding duplicate retry events after reactive compression.Interactive tmux verification
context_length_exceededfor the first streaming chat completion, returns a non-stream compression/utility summary for non-stream requests, then returns a streamed assistant answer for the retried request.人工验证
qwen-code截图
server截图
Scope / Risk
Testing Matrix
Testing matrix notes:
Linked Issues / Bugs