Imported from 1122-gggggg/wulai-sfm-localization (AGENTS.md). Install upstream with npx skills add 1122-gggggg/wulai-sfm-localization. Copyright stays with the author.

Project agent instructions

Scope and decisions

• Follow system and developer instructions, then the user's current request. Use these rules and applicable skills as project defaults. Preserve explicit safety boundaries below; prompt cleanup does not authorize changing them.
• Complete the requested work through implementation and relevant verification. State material assumptions. For routine, reversible choices, use repository conventions and proceed. Ask when missing information materially changes the outcome, scope, or authorization; continue independent work while waiting.
• If an instruction blocks progress, identify its file and exact rule, explain the blocked action, and state what input is needed to continue.

Changes and completion

• Read the files relevant to the change. Keep edits minimal and match existing style; preserve unrelated work. Remove only dead code introduced by your edits.
• Do not add speculative features, abstractions, or unrelated refactors.
• For multi-step work, briefly state the outcome and how it will be verified. Fix failures caused by the change and rerun affected checks. Once relevant checks pass, stop unless new evidence warrants broader verification.
• Use a regression test for a behavioral bug when practical. For documentation or other low-impact edits, inspect the diff and relevant links or syntax; do not add tests that merely repeat the implementation.
• Report the result, verification, and any remaining limitation concisely. Distinguish checks actually run from checks that require other environments.

Context and tools

• Load only skills whose workflow matches the task, and only the references needed for that workflow. Avoid reading a full repository map before each edit.
• When CodeGraph is available and indexed, prefer it for symbols, call graphs, impact, and focused source context. Batch related symbols and use its results without redundant searches. Check source when results are stale or incomplete.
• Use rg for literal text and rg --files for file discovery. If CodeGraph is unavailable or uninitialized, use these tools to continue; offer index setup when useful rather than making initialization a prerequisite for the task.
• Avoid rereading unchanged files or loading files over 100 KB unless needed. Verify APIs, flags, versions, and commit IDs from source or documentation.
• When delegation is authorized and the parent model is gpt-5.6-sol, use gpt-5.6-luna with max reasoning only for bounded, independently verifiable tasks. Specify inputs, ownership, output, and checks. Keep ambiguous decisions and final integration in the parent agent.

Project references and flight boundaries

• For test placement and collection, use tests/README.md. Run only checks known to be offline or simulated without further approval; inspect unfamiliar launchers before execution.
• For flight behavior, live hardware, or operator controls, read 控制介面程式/SAFETY.md and follow 定位演算法/AGENTS.md.
• Only the operator may initiate real takeoff or autonomous flight through the operator interface. Agents must not arm motors, send takeoff commands, or run scripts that cause real flight, even on a conversational request to do so. If verification requires flight, prepare the offline work and leave the physical operation to the operator.
• Do not change takeoff, landing, forced landing on close, hold-to-move PCMD, or Esc freeze semantics without an explicit request for that behavior change. Such a request does not authorize an agent to operate the aircraft.

Output

• Be concise, use the user's language, and avoid flattery, emojis, em dashes, and closing fluff. Include technical detail only when it supports the result.