fix: read the code execution tool's shell calls as code execution activity - #26
Closed
dtmeadows-ant wants to merge 2 commits into
Closed
fix: read the code execution tool's shell calls as code execution activity#26dtmeadows-ant wants to merge 2 commits into
dtmeadows-ant wants to merge 2 commits into
Conversation
…nt them Xcode 27 beta 5 removes `Transcript.CustomSegment`, which this package used to keep server-side tool calls and results (web search, web fetch, code execution) on the transcript. That storage is replaced, and replay of earlier assistant turns is rebuilt on top of the replacement so that everything the Messages API requires back verbatim actually goes back verbatim. How it works now: - As a turn streams, every completed content block is recorded, as the API sent it, in the metadata of the transcript entry that presents it: a thinking block on its reasoning entry, a tool_use block on its tool call, and text, server-tool traffic and anything unrecognized on the response entry. Each record carries the turn it belongs to and each block its position in that turn. - Building the next request, a turn's recorded blocks are replayed in position order, so thinking signatures, server tool results and citations return exactly as received however the framework's entries interleave. Entries with no record (an app-assembled transcript, an entry cut off before its first block completed) are rebuilt from their segments and signature. Tool calls and results the API would reject as unpaired, and turns left with nothing but thinking, are dropped instead of failing every later request. - Records are key-sorted JSON text under a reserved `claude.content` metadata key and survive `Transcript` persistence unchanged. Public API: - `ClaudeServerToolSegment` and its transcript accessors are removed. `ClaudeServerToolActivity` replaces them: typed web search / web fetch / code execution readings, or `.unrecognized` by tool name. - A response entry's segments run in the order the model produced them: one text segment per text block, and wherever the model called a server-side tool, an empty text segment whose id is the tool-use id holding that call's place. `Transcript.Response .claudeServerToolActivity(for:)` resolves such a segment to its activity so apps can render searches inline between paragraphs; `Transcript.claudeServerToolActivity` lists every round-trip in a conversation and pairs results that arrive in a later response. Also: - `pause_turn` responses are continued automatically, with usage accumulated across the continuation and trailing whitespace trimmed from the re-sent content as the API requires. - `ContentBlock` decodes any untyped block losslessly as `.raw`; `content_block_start` carries the block as sent; a `ContentAssembler` in ClaudeAPI rebuilds streamed blocks into their non-streaming form. - An App Attest token rejected on a continuation request is refreshed and retried like one rejected on the first request. Fixes #23
…ivity The code execution tool answers with `bash_code_execution` calls (input `command`, result `bash_code_execution_tool_result`), which the typed reading didn't match, so every code-execution activity surfaced as `.unrecognized`. Accept that name alongside plain `code_execution` and report whichever it was from `toolName`.
trilorez
force-pushed
the
trilorez/xcode-27-beta-5
branch
from
August 13, 2026 15:33
a96c6f8 to
28c8d2d
Compare
Collaborator
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.
Follow-up to #24, targeting its branch.
With
.codeExecutionenabled,code_execution_20260120answers withserver_tool_use name=bash_code_execution(inputcommand) →bash_code_execution_tool_result. The typed reading only matched plaincode_execution/{code}/code_execution_tool_result, so every code-execution activity surfaced as.unrecognized("bash_code_execution", …). Plaincode_executionblocks do still appear — other server tools (web search) emit them when they run code internally — so both names need to read as.codeExecution.This accepts either name, decodes
codeorcommand, expects<name>_tool_result, and hasactivity.toolNamereport whichever it was.text_editor_code_executionstays.unrecognized. No public API change.Verified on Xcode 27 beta 5 / iOS 27 beta 5 simulator: 212/212, and live a
.codeExecutionsession now yields.codeExecutionwithtoolName == "bash_code_execution"and.output(stdout:…), with the follow-up turn replaying fine.