feat: transfer task list across agent sessions via CLAUDE_CODE_TASK_LIST_ID

[Wirasm]

Summary

• Set CLAUDE_CODE_TASK_LIST_ID env var when spawning Claude agents so task lists persist across session stop/resume cycles
• ID is generated deterministically from the session ID (kild-{session_id} with / sanitized to -), stored in session JSON, and injected into both terminal and daemon spawn paths
• On resume: reuses existing task list ID so tasks survive restarts
• On fresh open: generates new task list ID for a clean start
• On destroy: cleans up ~/.claude/tasks/{task_list_id}/ directory

Post-review fixes

• Fixed shell injection in terminal export commands (env var values now single-quoted)
• Sanitized / to - in task list IDs for flat directory names
• Replaced full session re-save in agent_status with patch_session_json_field() to prevent older binaries from dropping new fields
• Extracted cleanup_task_list() for testability
• Added tests for daemon env var injection, destroy cleanup, slash sanitization, and JSON field preservation
• Updated docs (save-tasks.md, CLAUDE.md)

Related issues

• Daemon PTY sessions exit immediately after open/resume #320 — daemon PTY sessions exit immediately (found during e2e testing, separate issue)
• agent-status full session re-save can drop fields from newer binary versions #321 — remaining save_session_to_file call sites that could drop fields from newer versions

Test plan

• cargo fmt --check passes
• cargo clippy --all -- -D warnings passes
• cargo test --all passes (1,506 tests, 0 failures)
• cargo build --all passes
• Unit tests for generate_task_list_id() and task_list_env_vars()
• Test for slash sanitization in task list ID
• Tests for build_daemon_create_request with Some(task_list_id), non-Claude agents, and None
• Tests for cleanup_task_list (removal, nonexistent dir, None guard)
• Test for patch_session_json_field preserving unknown fields
• Non-Claude agents get no task list env var
• Old session JSON without task_list_id deserializes to None (serde default)
• E2E: task list created in kild, visible after stop/reopen via status bar and TaskList tool

Closes #296

[Wirasm]

PR Review Summary

Critical Issues (1 found)

Agent	Issue	Location
silent-failure-hunter	Shell injection risk: Unquoted env var values in export command construction. If task_list_id contains shell metacharacters (spaces, semicolons, backticks), arbitrary command execution is possible. Values should be single-quoted.	handler.rs:252, handler.rs:881

Important Issues (3 found)

Agent	Issue	Location
pr-test-analyzer	Task list cleanup on destroy has zero test coverage — path construction, removal, error handling all untested	destroy.rs:333-358
pr-test-analyzer	Daemon env var injection untested — all existing build_daemon_create_request tests pass None for task_list_id, so the new branch is never exercised	handler.rs:1101-1108
pr-test-analyzer	create_session integration not tested — no test verifies Claude agent gets a task_list_id generated and persisted in session JSON	handler.rs:178-187

Suggestions (6 found)

Agent	Suggestion	Location
pr-test-analyzer	Terminal command prepending with export syntax not tested	handler.rs:244-258
pr-test-analyzer	Task list ID reuse on resume vs fresh generation not integration tested	handler.rs:801-812
pr-test-analyzer	Session serialization round-trip doesn't exercise task_list_id field	types.rs
comment-analyzer	task_list_id field doc could mention None for non-Claude agents	types.rs:176-184
comment-analyzer	Terminal export comment could explain architectural difference from daemon path	handler.rs:244
type-design-analyzer	Consider extracting CLAUDE_CODE_TASK_LIST_ID as a named constant	resume.rs:48

Strengths

• Clean separation of concerns: ID generation in agents/resume.rs, lifecycle in handler.rs, cleanup in destroy.rs
• Well-designed types: pragmatic Option<String> with #[serde(default)] for backward compatibility (7.75/10 type design score)
• Accurate, concise documentation on all new functions and fields
• Unit tests for helper functions (generate_task_list_id, task_list_env_vars) cover both positive and negative cases
• All existing test constructors updated for new Session::new() parameter

Documentation Updates

• .claude/commands/save-tasks.md — Rewrote for automatic task list transfer (removed stale manual instructions)
• CLAUDE.md — Added task list cleanup to destroy.rs integration points

Verdict

NEEDS FIXES — 1 critical shell injection issue must be resolved before merge. Test coverage gaps should be addressed for key code paths (destroy cleanup, daemon env var injection).

Recommended Actions

1. Fix shell injection: single-quote env var values in terminal export commands, or validate task_list_id format
2. Add test for build_daemon_create_request with Some(task_list_id) to verify daemon env var injection
3. Add test for task list cleanup on destroy
4. Consider adding integration test for create_session with Claude agent verifying task_list_id generation

[Wirasm]

PR Review Summary

Critical Issues (1 found)

Agent	Issue	Location
silent-failure-hunter	patch_session_json_field silently succeeds when session JSON is not an object — field update is silently skipped, no error returned, caller thinks heartbeat was updated	persistence.rs:179-181

Details: The if let Some(obj) = json.as_object_mut() pattern silently does nothing when the JSON root is not an object (e.g., corrupted file with [], null, or a primitive). The function returns Ok(()) even though no field was updated. This means agent_status.rs thinks last_activity was patched when it wasn't, violating the "No Silent Failures" project principle.

Fix: Replace with:

let obj = json.as_object_mut().ok_or_else(|| SessionError::IoError {
    source: std::io::Error::new(
        std::io::ErrorKind::InvalidData,
        "session JSON root is not an object",
    ),
})?;
obj.insert(field.to_string(), value);

Important Issues (2 found)

Agent	Issue	Location
silent-failure-hunter	cleanup_task_list logs at error! level but treats failure as non-fatal — inconsistent severity (should be warn!)	destroy.rs:90
comment-analyzer	"Existing path" comment is misleading — should be "Terminal path" to match daemon path naming convention	handler.rs:243, 872

Suggestions (5 found)

Agent	Suggestion	Location
type-design-analyzer	Consider SessionField enum for patch_session_json_field field parameter to prevent typos	persistence.rs:168
pr-test-analyzer	Add test for patch_session_json_field with non-object JSON (validates the critical fix)	persistence.rs
pr-test-analyzer	Add test for terminal export format string construction	handler.rs:250-254
comment-analyzer	Remove "(existing behavior)" from terminal path comment in open_session	handler.rs:872
comment-analyzer	Add atomic write note to patch_session_json_field docstring	persistence.rs:161

Strengths

• Clean shell escaping in terminal export commands (single-quote wrapping with proper '\'' escaping)
• Path traversal prevention: branch validation rejects .., / sanitized to - in task list IDs
• Atomic file operations via temp file + rename in JSON patching
• Excellent backward-compatibility story via patch_session_json_field preserving unknown fields
• Strong test coverage: 12 new tests covering ID generation, env var injection, cleanup, and JSON patching
• Well-structured docstrings on all new public functions with lifecycle documentation

Type Design Assessment

Type	Score	Status
Session.task_list_id	7/10	Adequate
generate_task_list_id()	6/10	Adequate
task_list_env_vars()	7/10	Adequate
patch_session_json_field()	5/10	Needs improvement (stringly-typed field, silent failure)
Session::new() (15 params)	6/10	Watch for future growth

Verdict

NEEDS ONE FIX — The silent failure in patch_session_json_field must be addressed. Everything else is ready.