feat: apply 10 community PRs as batch

- anthropics#806: setting_sources=[] truthiness fix
- anthropics#803: betas=[]/plugins=[] truthiness fix
- anthropics#786: ThinkingBlock missing signature crash fix
- anthropics#790: suppress ProcessError when result already received
- anthropics#658: capture real stderr in ProcessError
- anthropics#791: suppress stale task notifications between turns
- anthropics#763: guard malformed CLAUDE_CODE_STREAM_CLOSE_TIMEOUT env var
- anthropics#805: delete_session() cascades subagent transcript dir
- anthropics#804: top-level skills option on ClaudeAgentOptions
- anthropics#691: PostCompact hook event type support

479 tests passing, mypy clean, ruff clean.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: IgorTavcar

COMMUNITY_BATCH_2026-04-10.md

@@ -0,0 +1,304 @@
+# Community Batch — 2026-04-10
+
+Upstream state at time of application: `64e8dee` (SDK v0.1.58, CLI 2.1.100).
+Branch: `community-fixes` (reset to upstream/main, then patched).
+
+This document describes every change cherry-picked from open community PRs
+on `anthropics/claude-agent-sdk-python` and applied as a single batch.
+
+---
+
+## 1. PR #806 — `setting_sources=[]` truthiness fix
+
+**Issue:** #794
+**File:** `src/claude_agent_sdk/_internal/transport/subprocess_cli.py`
+
+### Problem
+
+`_build_command()` used `if self._options.setting_sources:` — a truthiness
+check. In Python, `[]` is falsy, so passing `setting_sources=[]` (meaning
+"load no setting sources") silently did nothing. The CLI fell back to loading
+all sources (user, project, local).
+
+### Fix
+
+Changed to `if self._options.setting_sources is not None:`. When the list is
+empty, the SDK now sends `--setting-sources ""` to the CLI, which correctly
+disables all setting sources.
+
+### Test changes
+
+- `test_build_command_setting_sources_omitted_when_empty` renamed to
+  `test_build_command_setting_sources_empty_list_passes_empty_string` and
+  updated to assert that `--setting-sources` IS present with value `""`.
+
+---
+
+## 2. PR #803 — `betas=[]` and `plugins=[]` truthiness fix
+
+**File:** `src/claude_agent_sdk/_internal/transport/subprocess_cli.py`
+
+### Problem
+
+Same class of bug as #806. `betas` and `plugins` used truthiness checks,
+making empty lists indistinguishable from `None` (unset).
+
+### Fix
+
+- `betas`: changed to `if self._options.betas is not None:`. Empty list sends
+  `--betas ""`.
+- `plugins`: changed to `if self._options.plugins is not None:`. Empty list
+  simply iterates zero times (no behavioral change, but semantically correct).
+
+---
+
+## 3. PR #786 — ThinkingBlock missing `signature` field
+
+**File:** `src/claude_agent_sdk/_internal/message_parser.py`
+
+### Problem
+
+`message_parser.py:110` used `block["signature"]` — a direct dict lookup.
+The Anthropic API can return thinking blocks without a `signature` in certain
+conditions (redacted thinking blocks, streaming edge cases, older cached
+responses), causing a `KeyError` crash.
+
+### Fix
+
+One-line change: `block["signature"]` -> `block.get("signature", "")`.
+
+### Test added
+
+- `test_parse_assistant_message_with_thinking_missing_signature` — verifies
+  parsing succeeds with an empty string default for the signature field.
+
+---
+
+## 4. PR #790 — Suppress `ProcessError` when result already received
+
+**File:** `src/claude_agent_sdk/_internal/query.py`
+
+### Problem
+
+When using StructuredOutput (or any tool_use stop reason), the CLI exits with
+code 1 because it called a tool but received no tool_result response. The
+transport raises `ProcessError` after the process exits, even though a valid
+`ResultMessage` was already streamed to the consumer.
+
+### Fix
+
+Added a `except ProcessError` handler in `_read_messages()` that checks
+`_first_result_event.is_set()`. If a result was already received, the error
+is logged as a warning and suppressed. If no result was received, the error
+propagates normally.
+
+### Import added
+
+`from .._errors import ProcessError` — needed in `query.py` to catch the
+specific exception type.
+
+---
+
+## 5. PR #658 — Capture real stderr in `ProcessError`
+
+**File:** `src/claude_agent_sdk/_internal/transport/subprocess_cli.py`
+
+### Problem
+
+When the CLI subprocess exits non-zero, `ProcessError` was raised with
+`stderr="Check stderr output for details"` — a hardcoded string. The actual
+stderr was only piped when the user explicitly provided a `stderr` callback
+or enabled debug mode. This made debugging `Command failed with exit code 1`
+failures impossible without forking the SDK.
+
+### Fix
+
+Three changes:
+
+1. **Always pipe stderr** — `stderr_dest` is now unconditionally `PIPE`,
+   removing the conditional `should_pipe_stderr` check. The stderr stream
+   setup is also unconditional.
+
+2. **Buffer stderr lines** — Added `self._stderr_buffer: list[str] = []`.
+   Every line read from stderr is appended to this buffer (in addition to
+   being forwarded to the user callback if one exists).
+
+3. **Use buffer in ProcessError** — When creating `ProcessError` on non-zero
+   exit, `stderr` is set to `"\n".join(self._stderr_buffer)` instead of the
+   hardcoded placeholder.
+
+The buffer is cleared on `close()` to prevent memory leaks across reuse.
+
+---
+
+## 6. PR #791 — Suppress stale task notifications between turns
+
+**File:** `src/claude_agent_sdk/client.py`
+
+### Problem (Issue #788)
+
+When a background task (spawned via `run_in_background=True`) completed
+between turns, its `TaskNotificationMessage` sat in the shared message
+buffer. The next `receive_response()` call would yield it as the first
+message — before any response to the new query, causing the model to respond
+to stale task context.
+
+### Fix
+
+Added turn-tracking state to `ClaudeSDKClient`:
+- `_current_turn: int` — incremented each time a `ResultMessage` is yielded.
+- `_task_turn_map: dict[str, int]` — maps task IDs to the turn they started.
+
+`receive_response()` now defers task-lifecycle events that arrive before the
+first non-task message of the current turn. Once a non-task message arrives
+(proving the CLI is processing the latest query), deferred events are
+re-evaluated:
+- Notifications for tasks started in a **previous turn** are **dropped**.
+- Notifications for tasks started in the **current turn** (or unknown IDs)
+  are **yielded**.
+
+### Tests added (4)
+
+- `test_stale_notification_before_turn2_is_suppressed`
+- `test_notification_arriving_mid_turn_is_yielded`
+- `test_turn_counter_increments_and_cleans_map`
+- `test_unknown_task_id_notification_is_yielded`
+
+---
+
+## 7. PR #763 — Guard against malformed `CLAUDE_CODE_STREAM_CLOSE_TIMEOUT`
+
+**File:** `src/claude_agent_sdk/client.py`
+
+### Problem
+
+`ClaudeSDKClient.connect()` used `int(os.environ[...])` to parse the timeout
+env var. If the value was malformed (e.g., `"60s"`), connection setup raised
+`ValueError` before the SDK could initialize.
+
+### Fix
+
+Added `_parse_timeout_ms_from_env()` helper that:
+- Returns the default (60000ms) when the env var is unset.
+- Uses `float()` parsing with fallback on `TypeError`/`ValueError`.
+- Rejects non-finite values (`inf`, `nan`).
+
+### Test added
+
+- `test_connect_with_invalid_timeout_env_falls_back_to_default`
+
+---
+
+## 8. PR #805 — `delete_session()` cascades subagent transcript directory
+
+**File:** `src/claude_agent_sdk/_internal/session_mutations.py`
+
+### Problem
+
+`delete_session()` only removed the `{session_id}.jsonl` file. The sibling
+`{session_id}/` subdirectory (which holds subagent transcripts) was left
+behind, leaking disk space.
+
+### Fix
+
+After removing the `.jsonl` file, `shutil.rmtree(path.parent / session_id,
+ignore_errors=True)` cleans up the subagent directory. `ignore_errors=True`
+because most sessions never spawn subagents, so the directory usually doesn't
+exist.
+
+### Import added
+
+`import shutil`
+
+### Test added
+
+- `test_removes_subagent_transcript_dir` — creates a session with a sibling
+  subagent dir, deletes the session, asserts both are gone.
+
+---
+
+## 9. PR #804 — Top-level `skills` option on `ClaudeAgentOptions`
+
+**Files:**
+- `src/claude_agent_sdk/types.py`
+- `src/claude_agent_sdk/_internal/transport/subprocess_cli.py`
+- `src/claude_agent_sdk/_internal/query.py`
+- `src/claude_agent_sdk/_internal/client.py`
+- `src/claude_agent_sdk/client.py`
+
+### Problem
+
+Enabling Skills required two non-obvious steps in unrelated fields:
+
+```python
+options = ClaudeAgentOptions(
+    allowed_tools=["Skill"],
+    setting_sources=["user", "project"],
+)
+```
+
+### Fix
+
+Added `skills: list[str] | Literal["all"] | None = None` to
+`ClaudeAgentOptions`. The SDK now handles the wiring automatically:
+
+- `skills="all"` -> injects `"Skill"` into `allowed_tools`, defaults
+  `setting_sources` to `["user", "project"]`.
+- `skills=["pdf", "docx"]` -> injects `"Skill(pdf)"`, `"Skill(docx)"` into
+  `allowed_tools`, defaults `setting_sources`.
+- `skills=None` -> no-op (default).
+
+The `_apply_skills_defaults()` method in `SubprocessCLITransport` computes
+effective values without mutating the original options object.
+
+The skills list is also sent on the `initialize` control request so
+supporting CLIs can filter which skills appear in the system prompt.
+
+### Tests added (10)
+
+- Transport tests for all skills scenarios (none, all, empty, named, merge,
+  preserve user setting_sources, no mutation, no duplicates).
+- Query initialize tests for skills list vs. all/None serialization.
+
+---
+
+## 10. PR #691 — `PostCompact` hook event support
+
+**Files:**
+- `src/claude_agent_sdk/types.py`
+- `src/claude_agent_sdk/__init__.py`
+
+### Problem
+
+The CLI (v2.1.76+) fires a `PostCompact` event after context compaction
+completes, but the Python SDK had no type for it.
+
+### Fix
+
+Added to `types.py`:
+- `PostCompactHookInput` — with `trigger` (`"manual"` | `"auto"`) and
+  `compact_summary` fields.
+- `PostCompactHookSpecificOutput` — with `additionalContext` field.
+- Added `"PostCompact"` to the `HookEvent` union.
+- Added both types to the `HookInput` and `HookSpecificOutput` unions.
+
+Exported both new types from `__init__.py` and added them to `__all__`.
+
+### Tests added (3)
+
+- `test_post_compact_hook_input` (auto trigger)
+- `test_post_compact_hook_input_manual_trigger`
+- `test_post_compact_hook_specific_output`
+
+---
+
+## Verification
+
+After applying all changes:
+
+| Check | Result |
+|-------|--------|
+| `ruff check` | Clean |
+| `ruff format` | Clean |
+| `mypy src/` | Success: no issues found in 15 source files |
+| `pytest tests/` | **479 passed** (20 new tests) |

src/claude_agent_sdk/__init__.py

@@ -70,6 +70,8 @@
     PermissionResultAllow,
     PermissionResultDeny,
     PermissionUpdate,
+    PostCompactHookInput,
+    PostCompactHookSpecificOutput,
     PostToolUseFailureHookInput,
     PostToolUseFailureHookSpecificOutput,
     PostToolUseHookInput,
@@ -554,6 +556,8 @@ async def call_tool(name: str, arguments: dict[str, Any]) -> Any:
     "StopHookInput",
     "SubagentStopHookInput",
     "PreCompactHookInput",
+    "PostCompactHookInput",
+    "PostCompactHookSpecificOutput",
     "NotificationHookInput",
     "SubagentStartHookInput",
     "PermissionRequestHookInput",

src/claude_agent_sdk/_internal/client.py

@@ -128,6 +128,7 @@ async def process_query(
             initialize_timeout=initialize_timeout,
             agents=agents_dict,
             exclude_dynamic_sections=exclude_dynamic_sections,
+            skills=configured_options.skills,
         )
 
         try:

src/claude_agent_sdk/_internal/message_parser.py

@@ -107,7 +107,7 @@ def parse_message(data: dict[str, Any]) -> Message | None:
                             content_blocks.append(
                                 ThinkingBlock(
                                     thinking=block["thinking"],
-                                    signature=block["signature"],
+                                    signature=block.get("signature", ""),
                                 )
                             )
                         case "tool_use":