fix(tokens): accurate input_tokens for context management#107
Open
juslintek wants to merge 7 commits intojwadow:mainfrom
Open
fix(tokens): accurate input_tokens for context management#107juslintek wants to merge 7 commits intojwadow:mainfrom
juslintek wants to merge 7 commits intojwadow:mainfrom
Conversation
The prompt_tokens reported to clients (used by Claude Code's /context command) were wildly inaccurate because they were derived from Kiro's contextUsagePercentage, which returns unreliable values. Instead, count tokens from the complete serialized Kiro request payload using tiktoken. This includes system prompt, messages, tools, and all other payload fields — matching what actually gets sent to the API. - Replace request_messages/request_tools params with pre-counted prompt_tokens across all streaming functions - Count tokens from full kiro_request_body in both OpenAI and Anthropic route handlers - Remove dependency on contextUsagePercentage for token counting - Update tests to match new function signatures
Claude Code calls this endpoint before each request to check conversation size and decide whether to trigger compaction. Without it, the gateway returns 404, Claude Code cannot estimate context usage, and long conversations eventually hit the upstream CONTENT_LENGTH_EXCEEDS_THRESHOLD error (400). The endpoint builds the full Kiro payload and counts tokens on the serialized JSON using tiktoken, consistent with the token counting approach used in the messages endpoint. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace validate_tool_names (400 error) with deterministic truncation for tool names exceeding 64-char Kiro API limit. Names are shortened to 55 chars + '_' + 8-char md5 hash. Mapping is reversed in responses so clients receive original names. Fixes MCP plugins with auto-generated names like mcp__plugin_cloudflare_cloudflare-docs__search_cloudflare_documentation
…s endpoint Claude Code sends /v1/messages/count_tokens without max_tokens since it only needs token counting, not generation. The required max_tokens field caused 422 validation errors, breaking context usage tracking and preventing conversation compaction.
…s response - Make AnthropicTool.input_schema optional to accept Anthropic built-in server tools (web_search, code_execution, bash, text_editor) that don't have input_schema. These were causing 422 validation errors. - Silently strip server tools in converter since Kiro API can't handle them, while keeping custom tools working as before. - Add context_management.original_input_tokens to count_tokens response to match Anthropic API spec.
The gateway was reporting input_tokens ~2000 tokens too low because Kiro API adds an internal system prompt that the gateway can't see. This caused Claude Code to trigger auto-compression too late. - Add KIRO_PROMPT_OVERHEAD_TOKENS (default 2000, configurable via env) to local token estimates in both messages and count_tokens endpoints - Use context_usage_percentage from Kiro API to derive accurate input_tokens at end of stream, included in message_delta.usage - Apply same fix to non-streaming collect_anthropic_response path Before: message_start input_tokens=464 (actual: 2448) — 5.3x undercount After: message_start input_tokens=2463, message_delta input_tokens=2448
|
Thanks for the PR! 🎉 Before merge, we need a one-time CLA confirmation. Full CLA text: Please reply once with: You need to write once, all further messages from me can be ignored. |
- Fix server tool detection: check 'input_schema is None' not 'not input_schema'
(empty dict {} is valid for tools with no parameters)
- Update anthropic_to_kiro tests to use .payload on KiroPayloadResult
- Update build_kiro_payload tests to use .payload on KiroPayloadResult
- Replace validate_tool_names tests with truncate_tool_names tests
- Update AnthropicTool tests for optional name/input_schema
- Update max_tokens test for optional default (4096)
All 1412 tests pass.
|
Thanks for the PR! 🎉 Before merge, we need a one-time CLA confirmation. Full CLA text: Please reply once with: You need to write once, all further messages from me can be ignored. |
Author
|
I have read the CLA and I accept its terms |
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
3 participants
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.
Problem
Claude Code (and other clients) can't trigger auto-compression on time because the gateway reports
input_tokens~2,000 tokens too low.The Kiro API adds an internal system prompt (~2,000 tokens) that the gateway can't see when counting tokens locally:
Clients see the low
input_tokens, think there's plenty of context room, and don't compress until it's too late.Solution
1.
KIRO_PROMPT_OVERHEAD_TOKENS(config.py)Configurable constant (default 2000, env var
KIRO_PROMPT_OVERHEAD_TOKENS) added to all local token estimates. Accounts for Kiro's internal system prompt.2. Accurate
input_tokensinmessage_delta(streaming_anthropic.py)At end of stream, uses
context_usage_percentagefrom Kiro API to derive realinput_tokensand include it inmessage_delta.usage.3. Same fix for non-streaming path
After fix
Slight overcount is intentional — better to compress early than too late.
Depends on