deep-security-review skill: Pass 0 lieutenant Task dispatch consistently fails with "Tool execution was cancelled"

[kumamaki]

The deep-security-review skill spawns parallel Task subagents for Pass 0 lieutenant enumeration. Send 4 Task calls in one assistant message (the skill's BATCH_SIZE_LIEUTENANTS default) and all 4 return Tool execution was cancelled. Tasks cancel before they read a single file.

Environment

• Droid CLI 0.164.0
• macOS (darwin 25.5.0)
• Orchestrator model: GLM-5.2 (Droid Core)
• Skill: builtin:deep-security-review (via builtin:security-review, "Deep" mode)
• Target: ~720 source files (Rust, Swift, TypeScript, Vue)

Reproduction

1. Run security-review, pick "Full-project" then "Deep"
2. Accept the consent gates (defaults)
3. The skill reaches Pass 0 and dispatches Task calls for area lieutenants
4. Every Task call returns Error: Tool execution was cancelled

Three attempts

• 6 Task calls at once: all cancelled
• 4 Task calls (clean, deduplicated, full prompts): all cancelled
• 13 Task calls (accidental duplication): all cancelled

Tasks cancel before producing output. No files read, no findings generated. The error surfaced to the user: The AI model timed out. Please retry or switch models with /model.

The gap

The skill sets BATCH_SIZE_LIEUTENANTS = 4 and MAX_CONCURRENT_TASKS = 8. Four parallel Tasks should work. They do not. The skill has a fallback for nested lieutenants that lose their Task tool, but the orchestrator itself has no fallback when dispatch fails. The pipeline cannot pass Pass 0.

Note

The skill's SKILL.md runs ~3000 lines. The model timeout may stem from the orchestrator processing that prompt while spawning multiple subagents, not from a runtime concurrency limit.