Roblox Diagnostics And Tooling

Roblox House Stack — Diagnostics and Tooling

Overview

Every claim about a game's behavior should come from an instrument, not an impression: telemetry signals, analyzer counts, console captures, or one of the shipped scripts below. This skill catalogs each measurement instrument the house stack provides, the exact command to run it, and how to interpret what it prints. Commands are repo-root-relative and run in the current project repo.

When to use

• You need evidence for a balance/tuning discussion (spawn vs kill counts, cap hits, HP loss per cycle, deaths).
• You want the repo-wide type/lint/format status, or to know whether your change regressed it.
• You need to read the running game's console or screen from a Claude session.
• You want a one-command view of every feature flag, debug toggle, or stale plan file.

When NOT to use

Choosing an instrument

1. Balance telemetry (the evidence source for tuning)

House pattern: telemetry services live in src/Server/Features/Telemetry/ and are gated to the gameplay place — they check an environment/place predicate (typically driven by a StringValue that the per-place *.project.json sets) and no-op elsewhere. A lobby playtest produces zero balance telemetry. There are two kinds of instrument; find your project's concrete services with:

ls src/Server/Features/Telemetry/ 2>/dev/null
grep -rn "Signal\|Subscribe\|:Connect" src/Server/Features/Telemetry/ --include="*.luau" | head -30

Session-summary logger — per-cycle console block (Studio playtests)

Pattern: a service with a top-level local ENABLED: boolean toggle that subscribes to gameplay signals (cycle started/ended, NPC spawned/died, spawn-cap hit, player downed), accumulates counters across one gameplay cycle, and prints a single summary block at cycle end. Nothing is persisted; the console print IS the output. Search the console for the service's log prefix (e.g. [BalanceLog]).

To build the signal-wiring table for your project (which signals it subscribes to, where each is defined), grep the service for signal names and confirm each against the source service:

grep -n "= Signal" src/Server/Features/**/*.luau   # where signals are defined
grep -n "Connect\|Subscribe" src/Server/Features/Telemetry/*.luau

How to read it during a playtest: Studio's Output window, or from a Claude session via the MCP console tool (§3) after each cycle end. Generic interpretation heuristics:

• spawned − killed growing cycle-over-cycle = players falling behind the difficulty curve;
• spawn cap hit early in the cycle = spawn pressure is throttled and the intended difficulty isn't being delivered;
• defended-object HP delta per cycle is a primary difficulty signal.

One cycle is an anecdote — capture every cycle of a full run before proposing a tuning change (and even then, the change itself is gated by roblox-change-control hard rule: rationale + predicted metric + evidence path).

Case study (shipped firebit game, 2026): the night-summary block printed duration, zombies spawned/killed, whether the active-cap was hit and when, RV HP start/end, and players downed — five lines that turned "night 3 felt hard" into "night 3: 41 spawned, 38 killed, cap hit at +51s, RV lost 230 HP". Also learned there: the telemetry docstrings cited a docs path that had moved — verify any doc path a docstring cites before trusting it.

Analytics event service — post-launch events (live servers)

Pattern: a second service, gated only by the place predicate (no toggle), that emits design events via the Analytics wally package (lfg-studio/analytics — check wally.toml for the pinned version). Verified in the reference repo: Analytics.logPlayerEvent writes to both sinks — Roblox AnalyticsService:LogCustomEvent (read in Creator Hub → Analytics → Custom Events) and GameAnalytics addDesignEvent (read on the GameAnalytics dashboard). Re-verify the dual sink in your project's pinned copy:

grep -rn "LogCustomEvent\|addDesignEvent" Packages/_Index/lfg-studio_analytics*/analytics/src/init.luau

The layer is initialized by src/Server/Features/Core/AnalyticsService.luau (Analytics.start with the GameAnalytics gameKey/secretKey hardcoded in that file — don't copy them elsewhere). The build tag is version.txt plus a -dev suffix whenever game.GameId differs from the live experience ID recorded in the project's game data — so dev/Studio events are separable from live data by build. To list which events your project emits and what value they carry:

grep -n "event = \|logPlayerEvent" src/Server/Features/Telemetry/*.luau

Case study (shipped firebit game, 2026): three events — player death, defended-object destroyed (logged once per connected player), and paid-revive completed — each carrying days-survived as the value. UNVERIFIED at the time: nobody had opened either dashboard to confirm events were arriving; the emission path was verified in code only. Do that dashboard check in your own project before citing live numbers.

2. Static analysis — exact commands and the regression bar

All tools come from aftman.toml (in the reference repo: luau-lsp 1.63.0, selene 0.29.0, stylua 2.3.1 — check your project's aftman.toml for its pins); run aftman install if missing. Run everything from the repo root.

Type check (the one that matters)

.claude/skills/roblox-diagnostics-and-tooling/scripts/analyze-all.sh          # src/ (default)
.claude/skills/roblox-diagnostics-and-tooling/scripts/analyze-all.sh src/Server

which is equivalent to:

luau-lsp analyze \
  --defs=globalTypes.d.luau \
  --sourcemap=sourcemap.json \
  --no-strict-dm-types \
  --ignore="Packages/**" \
  --ignore="DevPackages/**" \
  --ignore="*.spec.lua" \
  --ignore="*.spec.luau" \
  src/

This is the same flag set as the per-file PostToolUse hook (scripts/luau-analyze-hook.sh in the house boilerplate) plus --ignore="DevPackages/**": without it, directory-wide runs drown in dozens of Jest-internal type errors from DevPackages/_Index/jsdotlua_jest-core.... The script also dedupes luau-lsp's habit of printing some diagnostics twice (once with a [game/...] DataModel path, once relative) and, unlike the hook, fails loudly when sourcemap.json / globalTypes.d.luau are missing (the hook silently exits 0 — see roblox-build-and-env for the sourcemap traps).

Interpretation / regression bar. The intended baseline is zero diagnostics. If your project carries known debt, the baseline is a recorded list: run the script, save its output, and treat any diagnostic not in that list as a regression to fix before moving on. The per-file hook will block you on edited files anyway, but new-file and cross-file breakage only shows up in the directory-wide run. Establish/refresh your baseline:

.claude/skills/roblox-diagnostics-and-tooling/scripts/analyze-all.sh | tee /tmp/luau-baseline.txt

Case study (shipped firebit game, 2026): a dedicated type-cleanup pass took src/ from 80 luau-lsp errors to 0; the working tree later carried 5 known diagnostics (a DevPackages typing gap, one open type-debt pair, two DeprecatedApi warnings). The team's rule was exactly the above: those 5 are the list; a 6th is a regression.

Lint

selene src

Config: selene.toml (house convention: std = "seleneCustom+roblox", excludes build.luau, test/**/*, *.d.luau — verify with cat selene.toml). Record your project's current error/warning counts the same way as the type baseline (selene src 2>&1 | tail -4); the regression bar is: don't add to them.

Format

stylua --check src          # report only
stylua path/to/File.luau    # format ONLY files you touched

If stylua --check src reports non-conforming files you didn't touch, the codebase is not stylua-clean — never run stylua src repo-wide in that state: it would produce a mass diff unrelated to your task (the project CLAUDE.md's surgical-changes rule). Use --check to see whether your files conform, and format only those. Case study (shipped firebit game, 2026): 114 files failed --check; a repo-wide format would have buried any real change in a 114-file diff.

3. Studio-side measurement (MCP)

When Roblox Studio is open with the MCP server connected, a Claude session can measure the running game directly. The core tools (names as documented in the roblox-testing plugin; they may arrive namespaced, e.g. mcp__roblox-studio__start_stop_play — discover the exact names with ToolSearch):

Pattern for a measured playtest: start play → let a full gameplay cycle elapse → console_output and extract the summary block → repeat per cycle → stop play. Keep the raw blocks as the evidence artifact for any tuning proposal.

For frame time, MicroProfiler, and memory analysis use the skill roblox-performance; for breakpoint-level inspection of a live bug use the roblox-debugging plugin skill.

4. /test and /playtest as instruments

Reach for /roblox-testing:test when the question is "did I break a pure function/module" (Jest Lua over __tests__/*.spec.luau; its power is proportional to how many specs your project actually has — count them with find src -name "*.spec.luau" | wc -l). Reach for /roblox-testing:playtest <scenario> when the question is "does the game still boot / does this flow still work end-to-end" — scenarios live in tests/scenarios/ (house baseline: at least smoke.md). Both require Studio + rojo serve on default.project.json. Conventions, evidence bar, and how to certify a test as trustworthy live in roblox-validation-and-qa — treat any harness whose own end-to-end run hasn't been verified in your project as pending-verification.

5. Shipped scripts (this skill's scripts/, all read-only)

All three assume the house repo layout and locate the repo root themselves (git rev-parse), so they can be invoked from any cwd inside the project. All three were test-run against the reference repo on 2026-07-05.

analyze-all.sh

Covered in §2. Exit 0 clean / 1 diagnostics / 2 missing preconditions.

flags-report.sh

.claude/skills/roblox-diagnostics-and-tooling/scripts/flags-report.sh

Prints four sections: the full FeatureFlags table (from the house-standard src/Shared/Class/FeatureFlags.luau) with line numbers, every local *DEBUG* constant in src/ with file:line: declaration, telemetry ENABLED toggles, and a drift section listing every DEBUG value that is not false. Output shape:

=== FeatureFlags (src/Shared/Class/FeatureFlags.luau) ===
  :15  SaveInStudio = true
  ...
=== DEBUG constants in src/ ===
  src/Server/Features/Xxx.luau:27:local DEBUG_PRINTS = false
  ...
=== Telemetry toggles ===
  ...
=== DEBUG values that are NOT false (check these before a release) ===
  ! src/.../Xxx.luau:79:local DEBUG_FORCE_STATE: number? = 1

Interpretation: the ! section mixes true drift with tuning parameters that merely share the DEBUG prefix (e.g. a visual constant only used when a master DEBUG_MODE is true) — read each hit's context before acting. The dangerous class is a non-false DEBUG constant that lives in a production code path. Case study (shipped firebit game, 2026): a DEBUG_FORCE_STRUCTURE_CHUNK = 1 constant was live in production paths, forcing a specific structure onto the first map chunk for everyone. Whether to change any flag is roblox-config-and-flags / roblox-change-control territory — and remember hard rule 3: flag-gated dormant code is deliberate, never refactor or delete it. This script only makes drift visible.

plan-board-drift.sh

.claude/skills/roblox-diagnostics-and-tooling/scripts/plan-board-drift.sh

Tables every docs/plans/*.md with frontmatter status + open/checked box counts, counts docs/plans/archive/, and flags drift:

PLAN                                          STATUS    OPEN CHECKED
some-feature.md                               active       1       8
...
Archive: N plan(s) in docs/plans/archive/
=== Drift flags ===
  ! done-plan.md: status done but not moved to docs/plans/archive/
  ? old-plan.md: all 7 boxes checked but status is '(none)' (finished but never closed out?)
  i N plan(s) have no status frontmatter (predate the workflow)

Interpretation: ! = a status: done plan sitting outside archive/ (a direct workflow violation — the project's CLAUDE.md owns the lifecycle rules); ? = probably-finished plans never closed out; i = plans predating the workflow. Known doc debt and how to pay it down is roblox-docs-and-writing territory; this script is how you re-measure it.

Common mistakes

• Running stylua src repo-wide on a non-clean codebase. You'd bury your change in an unrelated mega-diff. Check with stylua --check src first; format only files you touched.
• Directory-wide luau-lsp analyze without --ignore="DevPackages/**". Jest-internal errors flood the output and hide real regressions.
• Treating the per-file hook as full coverage. It analyzes only the edited file and silently passes when sourcemap.json/globalTypes.d.luau are missing. After creating/renaming files, regenerate the sourcemap (roblox-build-and-env) and run analyze-all.sh.
• Proposing tuning changes from one cycle's feel. Capture the summary blocks for a full run; the change itself still needs rationale + predicted metric + evidence path per roblox-change-control.
• Expecting balance telemetry outside the gameplay place. Place-gated telemetry services return early elsewhere (e.g. the lobby); playtest the gameplay place (npm run dev:<place>).
• Expecting session-summary data anywhere but the console. It only prints; if you didn't capture console_output (or copy the Output window) before stopping play, the data is gone.
• Reading live-game numbers from Studio analytics. Studio/dev sessions emit with a -dev build tag; filter by build when reading dashboards.
• Trusting "0 errors" from a pipe. luau-lsp/selene/stylua exit codes are eaten by | tail; use the shipped scripts or check pipestatus.

Provenance and maintenance

Derived on 2026-07-05 from the firebit house-stack reference project; all commands and paths were verified there on that date. Everything project-specific is volatile — re-verify against the CURRENT project before relying on it: