lsm / HyperNeo / 31966757047

79%

Build Type

push

github

Committed by web-flow

Commit Message

chore: upgrade Claude Agent SDK to 0.3.233 (#2546)

* chore: upgrade Claude Agent SDK to 0.3.233

Bump @anthropic-ai/claude-agent-sdk from 0.3.179 to 0.3.233 and resync the
shared SDK type declarations so local consumers match the installed package.

Breaking API changes in 0.3.233:
- Query.interrupt() now resolves to SDKControlInterruptResponse | undefined
  instead of void; update QueryLike and the ACP adapter accordingly.
- CanUseTool returns PermissionResult | null and its options require a
  requestId; guard the null case and stamp a requestId in the ACP flow.

New message types emitted by the bundled Claude Code CLI (2.1.x):
- Filter (do not persist/broadcast) internal signals: command_lifecycle,
  conversation_reset, and active_goal, plus the background_tasks_changed and
  control_request_progress system subtypes.
- Render model_refusal_no_fallback so refusal content is shown instead of
  silently dropped.
- Update the compiled CLI fallback version to keep binary downloads aligned.

* fix(sdk): address 0.3.233 message-handling review findings

- Filter background_tasks_changed and control_request_progress before
  persistence/broadcast (not just from UI), so internal state pushes don't
  accumulate invisible rows or wake live subscribers.
- Strip terminal_slash_commands from the init slash command list so browser
  autocomplete never advertises terminal-only commands (exit, statusline, …).
- Route model_refusal_no_fallback through the nested subagent renderer so
  subagent refusals show content instead of a bare "System:" row.

* fix(sdk): address 0.3.233 review round-2 findings

- Honor local fallback scope: a subagent/side-question refusal only swaps that
  response's model, never the session model.
- Stamp browser-authored prompts with origin {kind:'human'} so strict
  human-trust gates don't classify composer input as unattributed.
- Fail closed on matchedAskRule instead of silently auto-allowing a tool the
  user explicitly gated with a pe... (continued)

Coverage Stats

43551 of 59164 branches covered (73.61%)

Branch coverage included in aggregate %.

36 of 44 new or added lines in 8 files covered. (81.82%)

12 existing lines in 7 files now uncovered.

59305 of 71252 relevant lines covered (83.23%)

3802.01 hits per line