Repository navigation
fix(client): drop sampling params for the Claude 5 generation and Haiku 5.5 - #7014
Merged
Hmbown merged 2 commits intoOct 18, 2026
Merged
Conversation
…ku 5.5 The Anthropic sampling restriction blankets every model released after Opus 4.6: the Messages API reference states those models "do not support" temperature/top_p beyond their backwards-compatible defaults (1.0 / >= 0.99), and the migration guides restate it per family — "Setting temperature, top_p, or top_k to any non-default value on Claude Opus 4.7 and later models, including Claude Opus 5.5, returns a 400 error" — covering the Claude 5 generation, Haiku 5.5 (defaults only: temperature 1, a lone top_p 0.99), and the Fable/Mythos lines. The gate only knew opus-4-7/4-8 and fable/mythos, so user-configured temperature/top_p reached Claude 5 and Haiku 5.5 models and 400'd the request. Dropping the params entirely is always safe; the gate extends its own family list and deliberately stays independent of model_supports_reasoning's exact-id list, which answers a different question (thinking capability, not sampling rejection). Pinned across opus-5(-5) plus a datestamped spelling, sonnet-5(-5), haiku-5-5, fable-5, and mythos-5, with the older opus-4-6/sonnet-4-6/ haiku-4-5 passthrough asserted. Docs re-checked 2026-10-08. Signed-off-by: asto <asto18089@126.com>
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
github-actions
Bot
added
the
contribution-gate
Hmbown
pushed a commit
that referenced
this pull request
Oct 18, 2026
Maintainer follow-ups from the review of three merged contributor PRs from @asto18089. #7014 (Claude 5 sampling gate): the "never send both" assertion sat after the older-model loop, so it checked the first claude-opus-4-8 body, which never carries top_p, and could not catch a regression. It now runs inside the loop for each 4.x model, and a top_p-only request is checked to pass through, as the PR body described. #7015 (Kimi empty `items` repair): once an empty `items` is dropped, a required array with minItems >= 1 still failed mfjs_schema_can_terminate with InvalidSchemaNode, so that tool was still excluded. MFJS allows `items` to be absent, and an unconstrained array can always be satisfied, so absent `items` now counts as terminating. #7017 (era-scoped checkpoint recognition): the docs said an unstamped history "predates stamping". A session that has never been compacted has no stamped carrier either, so a pasted legacy header still reads as a checkpoint at a session's first compaction. SURVIVAL_CONTRACT.md and the history_is_post_provenance doc now state that limit. Refs #7014 Refs #7015 Refs #7017 Verified: scripts/dev-test.sh tui body_drops_sampling schema_sanitize moonshot kimi compaction: 355 run, 355 passed (includes body_drops_sampling_params_for_models_that_reject_them and 141 kimi/moonshot tests). validate_survival_contract.mjs: ok 17 fixtures. cargo fmt --check clean. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NBKbhx19BRbZtvWK8rVKjz
Collaborator
|
🤖 This reply is from Claude Opus 5.5, an AI agent working on behalf of the Codewhale maintainers. It is posted through @Hmbown's account, but Hunter did not write it. Merged in a3c3a8e. Thank you @asto18089. For routes without exact catalog facts, which covers custom or proxy base URLs, dated snapshot ids and Messages-wire gateways, this heuristic is the only thing that keeps One test follow-up on — Written by Claude Opus 5.5 (AI agent) on behalf of the maintainers. |
Hmbown
pushed a commit
that referenced
this pull request
Oct 18, 2026
fix(client): capture and replay Anthropic redacted thinking blocks Merged locally because the PR conflicted with #7014 in one hunk of crates/tui/src/client/anthropic.rs (build_anthropic_body): this PR's non-native redacted_thinking withhold block sits directly above the sampling-parameters comment that #7014 rewrote. Resolution keeps both: the withhold block, then #7014's updated sampling comment. PR checks on its last head (5851b95): Lint, Test (ubuntu/macos/windows), npm wrapper smoke all passing. No-Issue: Anthropic redacted_thinking blocks were dropped and never replayed, so tool loops hit an unretryable 400. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NBKbhx19BRbZtvWK8rVKjz
Hmbown
pushed a commit
that referenced
this pull request
Oct 18, 2026
Adds [Unreleased] entries for #7014, #7015, #7017, #7018, #7019 and #7020, crediting @asto18089 (and @zhuowp for the fork ports). Each merged without a changelog line. The entry for #7020 names its behavior change: the app-server's legacy [providers.custom] table now refuses a non-Chat wire with provider_wire_format_unsupported instead of forwarding a Chat body. Refs #7014 Refs #7015 Refs #7017 Refs #7018 Refs #7019 Refs #7020 Refs #6997 Verified: sync-changelog --check up to date; web check-facts OK; check-contributor-credit "Every contributor in the window is credited on all three surfaces". Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NBKbhx19BRbZtvWK8rVKjz
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
2 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.
What
anthropic_model_rejects_samplinggatestemperature/top_pout of Anthropic Messages bodies. The gate only knewopus-4-7,opus-4-8,fable, andmythos. Every model released after Opus 4.6 carries the same restriction — the Messages API reference states those models "do not support" settingtemperature/top_pbeyond their backwards-compatible defaults (temperature: 1.0,top_p >= 0.99), and the migration guides restate it per family: "Settingtemperature,top_p, ortop_kto any non-default value on Claude Opus 4.7 and later models, including Claude Opus 5.5, returns a 400 error". So user-configured sampling params reached the Claude 5 generation and Haiku 5.5 and returned a 400.The gate extends its own family list with
opus-5,sonnet-5, andhaiku-5substring arms. Dropping the params entirely is always safe; sending a forbidden one bricks the request. The gate deliberately stays independent ofmodel_supports_reasoning's exact-id list, which answers a different question (thinking capability, not sampling rejection).Pinned across opus-5(-5) plus a datestamped spelling, sonnet-5(-5), haiku-5-5, fable-5, and mythos-5, with the older opus-4-6 / sonnet-4-6 / haiku-4-5 passthrough asserted so a future over-broad arm like
contains("opus-4-")cannot slip through. Docs re-checked 2026-10-08.Tests
body_drops_sampling_params_for_models_that_reject_them: the eight Claude-5/Haiku-5.5/Fable/Mythos spellings must omit both params; older models keep temperature (or top_p when alone).cargo test -p codewhale-tui --lib body_drops_samplinggreen; clippy-D warningsclean (lib).Ported from the Pinvou/CodeWhale fork (#93, commit 3118c36); the Gemini and DeepSeek hunks of that fork commit are already covered upstream and are not included.
No-Issue: the heuristic Anthropic sampling gate missed Claude 5 / Haiku 5.5 on routes without exact catalog facts (port of fork PR #93)