Repository navigation
fix(client): capture and replay Anthropic redacted thinking blocks - #7019
Merged
Hmbown merged 3 commits intoOct 18, 2026
Merged
Conversation
Anthropic sometimes replaces a whole thinking block with an encrypted `redacted_thinking` content block. Main had no representation for it: the SSE `content_block_start` failed to decode and the block was dropped, the non-stream response decode failed wholesale, and nothing was ever replayed. A tool loop that replays the bare `tool_use` without its thinking block then fails with "Expected `thinking` or `redacted_thinking`, but found `tool_use`" - an unretryable 400 on every retry of that turn. Fix, ported from fork PR codewhale-hq#91 (redacted-thinking slice only): - models: add a tolerant `ContentBlockStart::RedactedThinking` stream variant (missing-key and explicit-null `data` decode to an empty payload instead of failing the event) and a serde-default `redacted_data` field on the neutral `ContentBlock::Thinking`, so persisted sessions round-trip and older sessions load unchanged - stream capture: park redacted blocks byte-exact the moment they arrive (and finished thinking blocks in wire order), persist them, keep them in the mid-stream retry prefix, and treat a stream that captured one as received content instead of silently re-issuing the request - non-stream capture: rewrite `redacted_thinking` blocks into the neutral shape before decoding so the response (and its tool calls) survives - replay: send the payload back byte-exact on native `*.anthropic.com` routes; withhold it from every non-native Messages host, dropping redacted-only assistant messages instead of emitting `content: []` - other wires (Chat Completions, Responses) pick explicit fields only, so the payload never leaves the Anthropic dialect, and the transcript renders redacted blocks like any textless thinking block - header only, payload bytes never displayed - engine: redacted starts count as actionable content and their payload counts toward the stream content cap Out of scope, still withheld from the fork commit: multi-turn thinking strip rules, Responses pairing integrity, Gemini signature-gate scoping, encrypted-reasoning capture widening, OpenRouter reasoning details, and the compaction-estimator byte accounting. Co-authored-by: zhuowp <9426526+zhuowp@users.noreply.github.com> 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
#7016, #7018 and #7019 from @asto18089 carry a Co-authored-by trailer for @zhuowp (the work was ported from a fork they contributed to). scripts/check-contributor-credit.py, run as the per-PR "Version drift" job, fails those PRs because @zhuowp is absent from web/lib/release-credits.ts. docs/CONTRIBUTORS.md and CHANGELOG.md already credit them. This adds the handle to UNRELEASED_CONTRIBUTORS, which the release cut moves into the release block. Refs #7016 Refs #7018 Refs #7019 Verified: python3 scripts/check-contributor-credit.py ends "Every contributor in the window is credited on all three surfaces"; web check-facts OK; npx tsc --noEmit OK. 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
Review follow-up on #7019 (@asto18089). The PR made any thinking block flushed into completed_thinking count as received content in stream_died_with_nothing. flush_thinking_block keeps blocks whose text is whitespace or that carry only opaque state. On Responses/Codex routes a reasoning item with encrypted content but an empty summary is common, so such a block blocked the silent NoContentStreamDeath retry where main retried. A whitespace-only chat reasoning chunk did the same. The guard now counts a completed block only when it is redacted or has non-blank thinking: the rule current_thinking already follows. Redacted blocks still count, which is what the PR needed. The PR's comment above the guard is updated to say so. Refs #7019 Verified on the merged tree (merge b6373c2 plus this change): scripts/dev-test.sh tui redacted thinking anthropic stream_: 442 run, 442 passed. scripts/dev-test.sh protocol redacted: 3 passed. cargo test -p codewhale-models --lib redacted: 2 passed. 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 b6373c2. Thank you @asto18089. Before this, a I merged it locally because it conflicted with #7014 in one hunk of One follow-up on Not in a release yet. — 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
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 sometimes returns a
redacted_thinkingcontent block instead of a decodable thinking block (safety/policy on extended-thinking models). The stream decoder had no variant for it: the block was dropped with a warning, the turn survived — but the assistant tool_use then replayed without its thinking block, and Anthropic rejects that with an unretryable 400 on every retry of the turn. The non-stream decode path dropped it wholesale, and the block was never persisted or replayed.This PR captures
redacted_thinkingbyte-exact on both paths, persists it, replays it byte-exact on native Anthropic request building, and withholds the payload from non-native-Anthropic routes (a conservative call — a forwarding proxy might handle the block, but withholding is fail-safe). Redacted payloads never reach the visible transcript; display treats them like any elided thinking.Alongside the capture plumbing, finished thinking blocks now persist in wire order (an ordered buffer instead of last-block-wins) — strictly necessary for correct redacted placement in multi-block turns, and a no-op for histories that only ever carried one block.
Scope
Deliberately excluded from the same-titled fork work: Responses pairing integrity, final-tool-turn strip rules, Gemini signature-gate scoping, encrypted-reasoning capture widening (upstream already captures on all Responses routes that send
include), signature-only-block persistence, and token-estimator accounting for redacted bytes (disclosed here as a known approximation).Tests
"data": nullgateway shapes), non-stream decode shim, native byte-exact replay, non-native strip, redacted-only assistant message drop, serde round-trip with field-absent session compat, and an engine-level pin that the block persists into the follow-up request in wire order.cargo test -p codewhale-models --locked44 passed;-p codewhale-tui --lib redacted34 passed; targetedanthropic/thinking/flushfilters green; clippy-D warningsclean (models, protocol, tui lib);check --workspacegreen.Ported from the Pinvou/CodeWhale fork (#91), redacted-thinking slice only.
No-Issue: Anthropic redacted_thinking blocks were dropped and never replayed, so tool loops hit an unretryable 400