Debug

/debug — Triage and Fix a Bug End-to-End

The disciplined debugging loop for Summer projects. Read the error before guessing. Form a hypothesis before editing. Verify the fix before declaring victory. Never grep the codebase before reading the actual error.

Core principle: The debugger already knows what's wrong. Your job is to listen to it before doing anything else.

When to use this skill

• The user says "it crashes", "it broke", "throws an error", "doesn't work", "freezes", "wrong behavior".
• The user invokes /summer debug, asks for a support report, or wants something they can send to Summer.
• The user pastes a stack trace.
• A previous build or test step failed and the user wants to fix it.

When NOT to use this skill

• The user is asking how to prevent a class of bug — that's a design question. Use the relevant discipline skill (gdscript-patterns, scene-composition).
• The user is asking for a feature with a known unfinished spec — that's make-game territory.
• The bug is in a file outside the game project that the host agent can debug natively (CI config, npm scripts) — use the host's debugger.

The Loop

Copy code
  Listen → Diagnose → Hypothesize → Propose → Fix → Verify

Do not skip steps. Do not loop back to "Hypothesize" without re-running the cheap diagnostic.

0. Support report mode

If the user asks for /summer debug, "send this to Summer", "make a report", or they are stuck in Codex, a cloud sandbox, or another agent environment, create a portable report first:

Copy code
summer_create_debug_report({
  issue: "<one-sentence user symptom>",
  include_play_session: <true only when the issue appears during runtime>
})

Tell the user where the Markdown file was written and remind them to review local paths and stack traces before sending. Then continue the normal debug loop only if they also asked you to fix the bug.

Fallback when the MCP tool is not available but shell is available:

Copy code
summer debug "<one-sentence user symptom>"

Add --play only when the issue appears after pressing Play.

1. Listen

Ask the user one focused question and wait for the answer:

What's happening, and when does it happen?

If they already gave a clear symptom in the request, skip the question. If they say "something's broken", that one question is your only call until they reply. Don't run tools yet.

2. Diagnose — cheapest tool first

In strict order. Stop at the first one that returns useful signal.

If summer_get_script_errors is clean and the user says it crashes only when running: it's a runtime bug. Go to step 2b.

2b. Runtime-only bugs

Reproduce it yourself first. Write a RunVerification probe that performs the repro steps (press / key), reports the state you expect to be wrong, and saves a frame. See "Drive the game yourself" below for the exact call. The returned results.errors_seen and reports are the diagnostic. This costs one call, touches nothing the user owns, and gives you a repro you can re-run after the fix.

Fall back to the editor session only when the repro genuinely needs a human (feel, hardware, "is this what you meant?"):

Copy code
  summer_clear_console
  summer_play
  ─▶ ASK USER: "Reproduce the bug now."
  (wait for confirmation)
  summer_get_debugger_errors
  summer_get_debugger_warnings   (if diagnostics showed warnings)
  summer_stop

If you go this route, do not skip the "reproduce now" prompt. Auto-running and grabbing whatever's in the buffer leads you to chase ghosts from previous sessions.

2c. MCP unavailable (engine not running, or no Summer install)

If summer_get_script_errors returns "Summer Engine is not running" or the tool isn't available:

1. Tell the user summer run starts it. One command restores every tool in this loop — try that before degrading to an interview.
2. If they cannot or will not, ask the user to copy-paste the Output panel and the Debugger panel from Summer Engine.
3. Reason over the pasted text exactly as you would over MCP output.
4. Continue with the rest of the loop unchanged.

Do NOT loop on MCP retry. Do NOT pretend the engine will come back. Use the fallback the moment it fails once.

3. Hypothesize — one specific theory

State exactly one hypothesis, in one sentence, naming the file and line.

Good: "Typo at scripts/player.gd:14 — GRAVTY should be GRAVITY." Bad: "Could be a typo, missing import, or wrong scope."

If you have multiple plausible theories, pick the highest-prior-probability one. The user can correct you if you're wrong; chasing all three at once wastes their time.

Verify the hypothesis without editing. If the error mentions a missing node, call summer_inspect_node to confirm. If it's a missing resource, summer_inspect_resource. If it's a scene-graph issue, summer_get_scene_tree. Confirm the world state matches the error before proposing a fix.

4. Propose — ask before writing

Surface the proposed fix in plain language and ask permission. Two patterns:

Code fix:

May I edit scripts/player.gd to rename GRAVTY → GRAVITY on line 14?

Scene fix vs code fix:

The script calls audio.play() but the AudioStreamPlayer child is gone. Two options:

1. Re-add the AudioStreamPlayer to ./Coin (preserves the original behavior).
2. Null-check in coin.gd (defensive, but the audio is silent).

Which one do you want?

Never unilaterally pick when there are two equally valid fixes (scene vs code, defensive vs strict, fast vs correct). Ask.

5. Fix — minimal, focused

• For GDScript edits: Read the 20–40 lines around the error, Edit the exact change. Don't read the whole file. Don't reformat. Don't rename other things.
• For scene edits: use the appropriate summer_* tool (summer_add_node, summer_set_prop, summer_replace_node). Group multi-step changes in summer_batch for one undo step.
• Nested resource properties: summer_set_resource_property works against inline sub_resource targets too — pass nodePath, resourceProperty, and subProperty (there is no dotted "mesh.size" form). Structural failures (node not found, property is not a resource, resource is null) are explicit errors; a bad value shape is not — a JSON object instead of an engine literal string, a misspelled key/subProperty, or a wrong-typed value returns ok:true yet silently no-ops or coerces destructively on current engines (dict → material cleared; dict → Color(0,0,0,1); newer engines reject these with unknown_property / bad_value_shape / type_mismatch). Pass class names and Color(...)/Vector3(...) strings, and confirm the fix in the saved .tscn or a snapshot diff, never from ok alone. See summer_docs_read id reference/mcp-tools-reference.

6. Verify — re-run the diagnostic that found it

The fix is not done until the same diagnostic that found the bug returns clean.

• Script error → re-run summer_get_script_errors.
• Runtime error → re-run the play/reproduce/check-debugger loop.
• Console warning → summer_clear_console, then summer_play and confirm clean output.

If the diagnostic is still red after the fix, you formed the wrong hypothesis. Go back to step 3 — do not make a second edit on top of the first. Revert if the first edit didn't help, then re-hypothesize.

Anti-Patterns

Quick reference — common Summer Engine bug families

Closing

A debug session is done when:

1. The user-reported symptom no longer reproduces.
2. The diagnostic that originally flagged the bug returns clean.
3. No new errors or warnings have been introduced.

Tell the user one short sentence: "Fixed <file>:<line> — <change>. Diagnostic clean." Then stop.

What the MCP debug tools CAN and CAN'T see (read this before claiming "clean")

The static and boot-time tools (summer_get_script_errors, summer_get_console, summer_get_diagnostics) can't substitute for play-testing — but you have a play-testing route, so a false-clean report is a choice, not a limitation. Know which tool answers which question.

What the MCP CAN see

• Script parse errors (summer_get_script_errors) — full text + file:line. Reliable.
• Editor console output (summer_get_console) — print statements, editor-side warnings, std startup messages. Full text.
• Runtime debugger error count + full text (summer_get_debugger_errors) — returns errors_data with error, error_descr, callstack, file, function, line. Reliable for errors.
• Runtime debugger warning text (summer_get_debugger_warnings) — same structured shape as the errors tool, filtered to severity warning. Use it whenever summer_get_diagnostics shows a non-zero debugger.warnings; do not report a warning count you never read.
• Diagnostics summary (summer_get_diagnostics) — counts of console errors, debugger errors, debugger warnings, script errors. Tells you where to drill. Counts only — it carries no FPS, frame time, draw calls, or physics-body numbers.
• Scene tree + node properties (summer_get_scene_tree, summer_inspect_node) — only the edited scene, not the running game's live tree.
• Live runtime state (summer_get_runtime_tree, summer_inspect_runtime_node, summer_game_probe) — the running game's tree, properties and frames, on engine builds with the runtime-control ops. See agent-playtesting.
• Rendered pixels (summer_screenshot) — target:"viewport" for the editor's current view, target:"scene" for an offscreen render of a scene file, target:"game" for a frame of the running game (that one needs the Summer desktop app bridge and fails cleanly over a plain local connection). You look at the actual image, not a description of it.
• Gameplay behaviour, driven by you (RunVerification via summer_batch) — see the next section. Input, live tree, and frames from a real running instance.
• Whether the game is running (summer_is_running) — and on which scene.

Drive the game yourself — RunVerification

For any bug that only fires from gameplay (movement, weapon firing, level transitions, button clicks), do not hand the repro to the user first. Spawn a hidden, disposable game instance that runs a GDScript probe and dies. It never touches the user's editor session.

Copy code
summer_batch ops:[{
  "op": "RunVerification",
  "probe_source": "extends SummerProbeBase\nfunc _ready() -> void:\n\tawait super._ready()\n\tawait get_tree().process_frame\n\tvar p := get_tree().root.find_child(\"Player\", true, false)\n\treport(\"y0\", p.global_position.y)\n\tawait press(\"jump\", 120)\n\tawait get_tree().create_timer(0.4).timeout\n\treport(\"y1\", p.global_position.y)\n\tsave_frame(\"after\")\n\tdump_tree()\n\tfinish()",
  "max_seconds": 20
}]

Probe API: report(key, value), save_frame(name), dump_tree(max_depth), press(action, hold_ms), key(keycode, hold_ms), finish(). Returns {ok, results, frames, out_dir}; results.errors_seen carries error-level engine/script messages from that run, so a probe is also a clean way to catch a runtime error without polluting the editor's debugger buffer.

This covers what static diagnostics cannot: pressed input, live scene tree (dump_tree reads the running tree, unlike summer_get_scene_tree), state sampled over several frames, and real rendered frames.

summer_batch forwards unknown ops verbatim, which is how you reach RunVerification with no dedicated tool. Note that the sibling SimulateInput op is not reachable this way — non-bridge callers get {ok:false, failure_reason:"unsupported_transport"}; it only works from the in-editor chat bridge. Drive input with press() / key() inside a probe instead.

What still genuinely needs the user

• Whether it feels right. Frame counts and reports cannot tell you the jump feels floaty. See debugging-game-feel.
• Hardware-specific and non-deterministic behaviour. A bug that only fires on their GPU, their controller, or one run in twenty.
• A judgement call on intent. "Is this the behaviour you wanted?" is a question, not a measurement.

Everything else — press the button, read the state, look at the frame — is yours to do.

The "static-only is not testing" rule

A clean summer_get_diagnostics after summer_play only proves: the game booted without parse errors or @implicit_ready null-derefs. It does NOT prove gameplay works. Auto-fire weapons, level transitions, spell casting, boss attacks, UI interactions all happen after boot and won't surface in the diagnostics until they fire.

When you've made code changes and want to declare them "verified":

1. Required: summer_get_script_errors clean on every modified file.
2. Required: summer_play boot returns 0 errors (use the scene most likely to exercise the change).
3. Required: exercise the specific scenario the change touches with a RunVerification probe — press the input, assert the state, save a frame. Quote the reports you got back.
4. Ask the user to play only for what a probe cannot judge: feel, hardware-specific behaviour, or whether the result is what they wanted.
5. Never claim "verified" from static analysis alone. Say "compiles and boots clean — needs play-test to confirm ."

The cost of skipping it

A typical cautionary scenario: many parallel agents write thousands of lines of game code in one batch with verification static-only. The build is "diagnostic-clean" yet the running game has telegraph meshes that render the wrong axis, parse-valid scripts that crash on first autoload because of a guard pattern that doesn't compile, projectile tracers sized so they read as straight lines instead of projectiles, transform-leaked colliders from a copied scene, and @onready paths that throw on every boot when the script is loaded into a sibling scene that lacks one of the referenced children. None of these surface in summer_get_diagnostics until something actually plays the game — which is exactly what a RunVerification probe does, and why "I had no way to test it" is not an available excuse.

Static analysis is necessary but never sufficient.

From a shell

No Summer MCP tools in this session? Run each one as a command; when summer --version is missing or below 4.0, use npx -y summer-engine@latest in place of summer. Fields are flags in kebab-case (--scene-path res://main.tscn) or one --input '<json>', and summer ops describe <op> prints the exact input. The summer-cli skill has the rest.