$ the-wire · showcase
Hook failures can now block tool calls
By RepoJournal · Filed · About Anthropic · Composed from the cited sources · methodology
3 people shipped this
The Python Agent SDK lets hook failures fail closed, so a crashing PreToolUse guard no longer silently permits the tool call it was meant to police.
Add on_failure to HookMatcher anthropics/claude-agent-sdk-python
A Python hook callback that raised, returned None, or returned a value the CLI rejected was treated as no decision, so a crashed PreToolUse guard let the tool run. Setting HookMatcher(on_failure="block") now sends onFailure on the matcher in initialize and blocks the action; hooks that leave it unset send nothing new, so existing behavior is unchanged. The CLI-side and TypeScript work lands sep...
test(e2e): accept either shell tool in shell-dependent tests anthropics/claude-agent-sdk-python
Seven end-to-end tests that require a shell command now accept either Bash or PowerShell instead of matching Bash alone, which had kept a PowerShell call from reaching the hook and failed with "PostToolUse hook should have been invoked, got: []". The failure hit test_hook_with_continue_and_stop_reason on main on 2026-09-29, 10-06, 10-08 and 10-30.
Merge pull request #106583 from anthropics/claude/gh-triage-workflows-5-5-m7kir1 anthropics/claude-code
The repo's issue triage and dedupe workflows now run on Claude Sonnet 5.5 instead of their previous model.
chore: bump Claude Code to 2.1.317 and Agent SDK to 0.3.317 anthropics/claude-code-action
The action tracks the release train: Claude Code moves to 2.1.317 and the Agent SDK to 0.3.317, mirrored into claude-code-base-action and the bundled CLI in claude-agent-sdk-python. Routine bumps; the changelog and mods housekeeping round out the day.
