Skip to content

Configuration Reference (LeafFuzz)

Configuration files are generated on first server startup under plugins/LeafFuzz/ on Paper, Spigot, Purpur, and Folia (or config/leaffuzz/ on Fabric and NeoForge).


File Overview

File Purpose Key Sections
config.yml Default episode budget, concurrent player roster, slow-tick MSPT threshold, disposable-server guard, extra command denylist, multi-boot campaign settings, and optional AST-grounded LLM audit check, campaign, llmAudit
messages.yml All operator-facing chat/console prefix, permission, parameter, busy, and reload strings prefix, noPermission, badParameter, invalidCommand, busy, reloadSuccess, reloadFailed
personas.jsonl (optional) Custom sleuth_persona_v1 JSONL behavioral profiles (Int8[16x16] policy genome + biometric cadence) loaded alongside the 8 built-in archetypes One JSON object per line (schema: "sleuth_persona_v1")
reports/<target>-qc-report.md Self-contained 5-stage Markdown QC scorecard written automatically after every /fuzz check run Stage 1..5 verdicts, LuckPerms / commands.yml fixes, ddmin reproducers, exc:/stall:/hazard: stacks, Bonferroni OLS retention slopes, and Order-1/Order-2 bytecode causal findings
reports/<target>-qc-data.yml & .json Machine-readable quality and safety moderation evidence leading on line 1 with status: <PASS\|WARN\|FAIL> status, moderation_ready, stage_verdicts, quality_metrics, shrunk_reproducers, fault_signatures, retention_findings, causal_findings

config.yml Settings

Key Type Default Description
check.defaultEpisodes Integer 250 Default number of UCB1 synthetic persona bot episodes executed when episodes=<n> is omitted from /fuzz check. Valid range: 1..100000.
check.defaultPlayers Integer | String 1 Default concurrent player survival roster size (1..500, or "auto" / "ramp" to progressively ramp concurrent players until instability) dispersed across an Archimedean spiral when players= is omitted from /fuzz check.
check.slowTickThresholdMs Number 100.0 Main-thread tick duration threshold in milliseconds (>= 10.0) above which LogAndStallSignatureOracle captures a slow-tick stack trace (stall:) outside JVM GC pause windows.
check.requireDisposableForAll Boolean true Blocks full target=all checks unless .bench-disposable exists in the LeafFuzz data directory or server root; this marker is an explicit opt-in, not ownership verification.
check.extraDenylist List\<String> [] Additional command patterns merged with the shared denylist and blocked before synthetic dispatch across Paper/Folia, Fabric, and NeoForge; dispatch also blocks save-all, save-off, timings, and spark.
campaign.restartStrategy String "AUTO" Multi-boot campaign restart strategy ("AUTO", "NONE", "SPIGOT_RESTART", "SUPERVISOR_STOP").
campaign.maxRestartRetries Integer 3 Maximum crash/restart retries (1..20) before FuzzCampaignStateStore halts or isolates a crashing archive.
llmAudit.llmAuditEnabled Boolean false Enables optional AST-grounded root-cause explanations and advisory Jminor causal chain auditing via LlmCausalChainAuditor. Makes zero network calls when false.
llmAudit.llmProvider String "OFF" LLM provider identifier ("OFF", "GEMINI", "OPENAI_COMPATIBLE", "ANTHROPIC", or "OLLAMA").
llmAudit.llmModel String "gemini-2.5-flash" Model identifier sent to the configured provider.
llmAudit.llmEndpoint String "" Custom HTTP endpoint URL, or empty to use the provider's default endpoint (http://127.0.0.1:11434/v1/chat/completions for OLLAMA).
llmAudit.llmApiKeyEnv String "GEMINI_API_KEY" Environment variable name holding the provider API key (never stored in plaintext in reports/).
llmAudit.explainFuzzFaults Boolean true When llmAudit.llmAuditEnabled is true, attaches closed-world verified Class#method root-cause explanations to proven Stage 2/3/5 findings without altering pass/fail verdicts.
llmAudit.maxChainsPerTarget Integer 6 Maximum number of Jminor causal chains or findings sent to the LLM per target (1..32).

messages.yml Settings

Key Default Description
prefix "&8[&2LeafFuzz&8] &r" Chat and console prefix prepended to /fuzz replies.
noPermission "&cYou do not have permission to run this LeafFuzz command." Sent when a caller lacks leaffuzz.admin.
badParameter "&cInvalid parameter '[param]': '[value]'." Sent when a name=value parameter fails validation.
invalidCommand "&cUnknown subcommand '[arg]'. Use /fuzz help." Sent when an unknown subcommand or bare positional token is supplied.
busy "&eA diagnostic snapshot or fuzz pass is already running; please wait." Sent when /fuzz check or /fuzz leaks snapshot is invoked while another pass is already active.
reloadSuccess "&aLeafFuzz configuration and messages reloaded." Sent after /fuzz reload validates and reloads config.yml and messages.yml.
reloadFailed "&cLeafFuzz reload failed: [errors]" Sent when config.yml fails validation during /fuzz reload.