debug-godot-game

Debug a Godot game

Preserve the failure, isolate one cause, apply the smallest repair, and rerun the same reproduction. Godot Agent Loop supports Godot 4.7 or later; report older engine behavior as outside the supported boundary.

Control contract

• Validate projectPath against effective MCP roots and allowed directories before mutation. Preserve a reproducible baseline and changed-file snapshot.
• Record whether the user requested watched or unattended work. For watched work, call editor_session with ensure and launch enabled; stop if a usable editor cannot be established instead of silently continuing detached. Unattended repairs use the host's normal file tools on .tscn, .gd, and project.godot.
• Distinguish persistent scene/resource/script/settings repair from runtime-ephemeral observation or mutation. Stop the runtime before persistent repair unless an editor-native operation is explicitly safe during play and record why.
• Use canonical core tools directly (compact is only the compatibility alias). Resolve hidden game_get_property, game_set_property, game_call_method, and game_eval through godot_catalog detail, then invoke them with godot_call; never call a hidden tool directly.
• If Pause Agent blocks a mutation, do not retry or bypass it. Continue only observation or safe teardown and report the effective blocked tool.
• Use privileged reflection or evaluation only when its group is already enabled, the hypothesis requires it, and safer observations cannot distinguish the cause.

Workflow

1. Reproduce before editing. Use realtime run_project for visual or timing complaints and repeat the user's exact input or a bounded game_scenario.
2. Capture the minimum distinguishing evidence: cursor-bounded game_get_logs, game_get_errors, concise game_get_scene_tree/game_get_ui/game_get_node_info state, and game_screenshot only when rendering matters.
3. Classify the failing boundary: parse/startup, persistent scene/resource, import, runtime state, input/timing, rendering, audio, export, or platform/toolchain.
4. State one falsifiable hypothesis and one observation that distinguishes it from the nearest alternative. Change one independent variable per trial; do not disable multiple systems and infer a single cause.
5. Drive input through game_scenario. Use a game_key_hold step for continuous movement with a bounded wait or observation and a release in the same scenario; do not hand-assemble repeated game_key_press taps. Never leave input held across a separate MCP call or while reasoning. Use bounded game_wait_until, never manual sleeps.
6. Stop the project before persistent repair. Apply the smallest matching scene, script, resource, or setting change as one coherent undoable change through editor_transaction (watched) or the host's file tools.
7. Run run_project_tests discovery or headless checks before runtime proof. Repeat the exact baseline, stress/recovery input, and observation; then run adjacent regression checks with verify_project or run_project_tests.
8. Separate measured FPS/frame time/process/render data from unavailable GPU metrics and subjective reports of feel. Do not substitute simulation counters for displayed frame pacing.
9. Release held input, call stop_project, remove only identified probes and MCP-owned transient artifacts, and independently check cleanup.