← ClaudeAtlas

lilia-agent-debuglisted

Agent debugging workflow for the LiliaCode native desktop. Use when adding, changing, validating, or reviewing stable target IDs, the LILIA_AGENT_DEBUG TCP protocol, debug-only instrumentation, or cargo xtask agent-debug harness behavior.
sena-nana/LiliaCode · ★ 7 · AI & Automation · score 69
Install: claude install-skill sena-nana/LiliaCode
# Lilia Agent Debug Treat Agent debugging as a developer interface, not as visible product UI. Users see normal application state; Agents get stable hidden targets and a development-only protocol. Harness behavior, corpus, and evidence layout are defined in `docs/design/agent-debug-harness.md`. Run `cargo xtask agent-debug` from the repository root. ## Ownership - Protocol, debug commands, and release exclusion live in `apps/desktop/src/agent_debug.rs`. - Product target IDs live in `apps/desktop/src/target_ids.rs`. - xtask starts the debug desktop, drives observe/act, and writes `agent-debug-runs/lilia-*`. - Generic control instrumentation belongs with the control owner; see `$lilia-app-boundary`. ## Implementation Pattern 1. Gate the TCP listener with `LILIA_AGENT_DEBUG`. Release builds must not contain the listener, fixed debug markers, or test fixtures. 2. Expose stable target IDs for primary controls, important rows, retry/recover actions, filters, tabs, dialogs, and destructive confirmations. 3. Name targets by functional path, not translated text, layout position, or pixel coordinates: `lilia.settings.open`, `lilia.task-session.composer.input`. 4. Keep target IDs invisible and non-semantic to users. Do not add public technical instructions, automation labels, or debug-only copy. 5. `act` only accepts protocol-defined targets and typed actions. Do not drive the UI by screenshot matching or guessed coordinates. 6. Screenshots come from the real WGPU surface. A scree