fix(claude): align beta assembly and Haiku helper detection with the 2.1.258 baseline - #5412
fix(claude): align beta assembly and Haiku helper detection with the 2.1.258 baseline#5412HYEONMIN94 wants to merge 5 commits into
Conversation
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 32b9d94eea
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".
| if requested[claudeAFKModeBeta] { | ||
| betas = append(betas, claudeAFKModeBeta) | ||
| } |
There was a problem hiding this comment.
When a caller-owned or confirmed-native request supplies afk-mode-2026-01-31 and the body declares an advisor tool without already supplying the advisor beta, withClaudeAdvisorToolBeta does not recognize AFK as a later-position beta and therefore produces ...,afk-mode-2026-01-31,advisor-tool-2026-03-01. This contradicts the newly documented Claude Code wire order, where advisor precedes AFK, and defeats the exact fingerprint reconstruction this path is intended to provide; add AFK to that helper's insertion boundaries.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 23528a4c37
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".
| if headerValue(headers, "X-Stainless-Async") != "" { | ||
| return false | ||
| } | ||
| compression := headerValue(headers, "Accept-Encoding") | ||
| if (shape == claudeCodeHelperShapeStructured && compression != "gzip, deflate, br, zstd") || | ||
| (shape == claudeCodeHelperShapeMinimal && compression != "gzip") { | ||
| if headerValue(headers, "Accept-Encoding") != "gzip, deflate, br, zstd" { | ||
| return false | ||
| } | ||
| requestID := headerValue(headers, "X-Client-Request-Id") | ||
| _, errRequestID := uuid.Parse(requestID) | ||
| return errRequestID == nil | ||
| return headerValue(headers, "X-Client-Request-Id") == "" |
There was a problem hiding this comment.
When an operator pins claude-header-defaults to the previously supported 2.1.220/0.94.0 baseline, the software tuple passes meetsClaudeDeviceProfileBaseline, but genuine 2.1.220 helpers still carry a request UUID and use the older shape-specific compression/async headers. These unconditional 2.1.258 checks therefore reject every such markerless helper; because helpers omit claude-code-20250219, they have no standard-signal fallback and are classified as unconfirmed and cloaked. Select the transport checks from the configured baseline or retain acceptance of the measured 2.1.220 envelope.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: cc79527e06
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".
| if requested[claudeAdvancedToolUseBeta] || claudeBodyUsesAdvancedToolUse(body) { | ||
| betas = append(betas, claudeAdvancedToolUseBeta) |
There was a problem hiding this comment.
When an operator pins claude-header-defaults to the supported 2.1.220/0.94.0 baseline, the outgoing User-Agent and package headers use that legacy tuple, but claudeCodeCLIBetas still unconditionally applies 2.1.258 behavior here and omits advanced-tool-use-2025-11-20 for ordinary non-empty tool lists. This creates a mixed fingerprint that no 2.1.220 client emitted; the new helper-transport branch preserves the configured legacy envelope, but the main request beta assembly has no equivalent version selection. Choose the beta policy from the configured baseline as well.
Useful? React with 👍 / 👎.
|
errors out with
|
|
chicken noodle soup chicken noodle |
|
Follow-up on the 400 I reported here: I traced it to base behaviour in the signature sanitizer, not to this PR. Claude Fable 5.1 emits a thinking-signature generation the CAIS validator does not recognise, so every thinking block is dropped, and a thinking-only assistant turn vanishes, which breaks the system-message placement rule. The fix is in a separate PR against dev: #5422. This PR is not at fault and is not blocked by it; the two were exercised together on a private build and work as expected. |
…e measured Claude Code 2.1.258 The default baseline now advertises claude-cli/2.1.258 with @anthropic-ai/sdk 0.112.1, but the per-request Anthropic-Beta assembly and the Haiku helper detection still follow the 2.1.220 measurements, so the outgoing fingerprint is one a real 2.1.258 never produces. Align the Anthropic-Beta assembly with 2.1.258: advanced-tool-use-2025-11-20 is sent only while tool search or another advanced tool-use feature is on the wire (or when the caller asks for it), and afk-mode-2026-01-31 is forwarded between fast-mode and extended-cache-ttl when the caller sends it. Re-measure the Haiku helper and count_tokens profiles on 2.1.258: add the advisor-tool + structured-outputs helper sequence, stop expecting X-Stainless-Async and x-client-request-id on helper requests (the SDK no longer sends the former, and the latter is attached only on a first-party base URL), and unify Accept-Encoding. count_tokens is unchanged. Refresh the config.example.yaml baseline comments and the test fixtures; add a registry test for claude-fable-5-1 and the captured 2.1.258 billing fingerprint.
withClaudeAdvisorToolBeta did not treat afk-mode-2026-01-31 as an insertion boundary, so a confirmed-native or caller-owned request that carried afk-mode and declared an advisor tool without the advisor beta ended up with "..., afk-mode, advisor-tool". Add afk-mode to the boundaries and pin the order in a test.
… drop unused helper parameters Place claude-fable-5-1 after claude-fable-5 and use the description from the router-for-me/models catalog so the embedded fallback matches the live one. Remove the shape parameter from measuredClaudeCodeHelperHeadersMatch and the transport parameters from the test header builders; they stopped varying once the 2.1.258 helper envelope became a single shape.
…aseline When an operator pins claude-header-defaults to 2.1.220, genuine 2.1.220 helpers still carry a UUID request id and the shape-specific async and compression headers, so applying the 2.1.258 checks unconditionally rejected every such helper and cloaked it. Check the measured 2.1.220 envelope when the baseline version is below 2.1.258 and the 2.1.258 envelope otherwise, and pin both cases in tests.
An operator who pins claude-header-defaults to 2.1.220 already gets the 2.1.220 helper transport checks, but the main-request beta assembly still applied the 2.1.258 policy and dropped advanced-tool-use from ordinary tool lists, producing a fingerprint no 2.1.220 client emitted. Move the version split into a shared helper, let claudeCodeCLIBetas emit the 2.1.220 policy (advanced-tool-use with any non-empty tool list, no afk-mode) below 2.1.258, and use the same split for the helper Accept-Encoding fallback. The helper beta sequence first observed on 2.1.258 is not accepted under an older baseline, and config.example.yaml documents the behaviour.
c6ba1aa to
1fbbda8
Compare
HYEONMIN94
changed the title
3 participants
devnow advertises Claude Code 2.1.258 as the default baseline, which clears the version gate forclaude-fable-5-1. The per-requestAnthropic-Betaassembly, the Haiku helper detection and the count_tokens profile still follow the 2.1.220 measurements, so the outgoing fingerprint isUser-Agent2.1.258 with a 2.1.220 body. This PR aligns those with a measured 2.1.258, and keeps the 2.1.220 policy for an operator who pins the baseline below 2.1.258.Related to #5402.
How 2.1.258 was measured
ANTHROPIC_BASE_URLwas pointed at a local capture server and the installedclaude2.1.258 binary was run on the interactiveclientrypoint and the-p(sdk-cli) entrypoint with an OAuth login. A Read of a large file and a Bash call were used to trigger the count_tokens pre-check and the Haiku helpers. The beta assembly order, the billing-header hash and the header conditions were cross-checked against the code inside the 2.1.258 bundle.What changed compared to 2.1.220
advanced-tool-use-2025-11-20: previously sent whenever the request had tools; now sent only while tool search is active. A request carrying 158 tools without tool search had no such beta.afk-mode-2026-01-31: new. Sits betweenfast-modeandextended-cache-ttlon auto-mode sessions.advisor-tool, structured-outputswhen the account's advisor feature flag is on and cache diagnostics are off.X-Stainless-Asyncis never sent.Accept-Encodingisgzip, deflate, br, zstdfor both the minimal and the structured helper.x-client-request-idis attached only when the base URL is api.anthropic.com, so a request that reaches the proxy never carries one.x-app,anthropic-version, the billing-header hash algorithm (the captured suffixes are reproduced),context_management,cache_controlttl, themetadata.user_idshape, the constant beta set and order, and count_tokens (five betas, body of model/messages/tools, noX-Stainless-Timeout).Changes
internal/runtime/executor/claude_executor_request.go:advanced-tool-useis sent only when the body has atool_search_tool_*tool,defer_loading,input_examplesorallowed_callers, or when the caller requested it.afk-modeis forwarded at its measured position when the caller sends it, andadvisor-toolstays ahead of it. HelperAccept-Encodingdefault and comments updated.internal/runtime/executor/helps/claude_client_detection.go: new sequence in the helper beta allowlist. Helper transport checks follow 2.1.258 (no async header, one compression set, no request id on requests reaching the proxy).internal/runtime/executor/helps/claude_device_profile.go:ClaudeBaselineUsesLegacyWirereports whether the configuredclaude-header-defaultsbaseline predates 2.1.258. The beta assembly and the helper transport checks use it to fall back to the measured 2.1.220 shape, so a pinned legacy baseline keeps a consistent fingerprint.internal/runtime/executor/claude_executor_tokens.go,config.example.yaml: comments.claude-fable-5-1, captured billing fingerprint suffix pinned.Verification
gofmt -l .clean,go build ./cmd/server,go test ./...pass on top of currentdev.--local-modeland a Claude OAuth credential:claude-fable-5-1request returns200and the response model isclaude-fable-5-1;claude-fable-5still returns200ANTHROPIC_BASE_URL+ANTHROPIC_AUTH_TOKENresponds normally in both-pand interactive mode, every request200claude-header-defaultspinned to 2.1.220 / 0.94.0, the beta assembly and the helper transport checks return to the 2.1.220 shape (covered by tests)Notes
advisor-tool-2026-03-01on every main-thread request when the account's advisor feature flag is on. That is account dependent, so the existing rule (only when requested or when the body declares an advisor tool) is kept.cc_prev_req=/cc_prompt_id=to the billing header, and a few headers, only when the base URL is api.anthropic.com. They cannot be observed through a custom base URL and are not synthesised here.