TheColliery
OrganizationMemory washer/defragmenter for AI agent memory: class-B semantic wash (fidelity-gated, blocks any structured-token drop, snapshot+rollback) and class-A byte-identity session estate (ULTRA archive/restore, RE-TIER) - never semantic-edited. Cleans the FAT, never the MEAT. Slows memory-overhead growth; does not eliminate it. Zero-dependency.
Categories
Indexed Skills (4)
coalwash
Memory washer/defragmenter for agent memory — two lanes: class-B (memory+governance) cleans the FAT, never the MEAT; class-A (transcripts) is byte-identity-only. Fidelity-first: a free mechanical Quick pass + a CODE gate blocking any STRUCTURED-token drop by diff; the paid Full pass is a separate consent; every DELETE/MERGE is plan-sourced + snapshot-backed. Session-start gauge; Fat hysteresis arms OBESE (auto-Quick, standing config, never asks). FULL = the economic cut-point (break-even proven, numbers shown): force-runs Quick; still over → one wizard ask, re-armed on growth. A capacity wall also forces FULL (wash-if-fat, else externalize). NO calendar cadence — never loop it. A manual `/coalwash` runs a fat-only, muscle-reorg, or estate pass. localOnly = Quick-only, no sub sees memory content. Honest: slows memory-overhead growth, does NOT eliminate it. Triggers: "/coalwash", "clean memory", "defrag memory", a [CoalWash] band nudge. Cross-agent (Claude Code validated). Zero-dep, offline, no API keys.
coaltipple
Model/effort router, ALWAYS-ON via its conductor hook: every prompt is routed, and routing returns exactly one of THREE outcomes — SELF (do it yourself at a pinned tier/effort; the routine result — routing that lands on self is still routing), delegate-DOWN (a large cheap mechanical task goes to a cheaper tier to save tokens), or escalate-UP (a task beyond the current model's competence goes to a stronger tier for quality). Sensitive/security work is NEVER delegated down, whatever its size. Invoke manually when weighing where a task should run. Claude Code only -- the platform's spawn tool must accept a worker model parameter (Antigravity etc. cannot actuate this); support reviewed monthly.
coalface
Fan-out discipline for swarm work. When a task decomposes into many units (a bulk refactor, a repo-wide sweep, a corpus batch), CoalFace runs it as a disciplined factory: mandatory SCOUT surveys the worksite, deterministic PARTITION merges overlapping/tiny spots, workers return anchor-edit orders as TEXT (propose-not-execute), QC checks scope+spec at collection, main = SINGLE WRITER (pre-swarm snapshot + domain gate), RECEIPT shows tokens-vs-solo. Wallet caps DOLLAR cost at ~solo via cheap tiers (raw tokens run HIGHER — fan-out ×N the per-sub baseline), not raw tokens. Modes: coalfaceMode auto (default, rides the contract at/above autoFanoutFloor units) | on (scout every prompt) | off. Manual "/coalface" or "swarm this" convenes it in any mode except off. Cross-agent (native subagent tool; no fan-out → sequential-pipeline degrade). Disciplines fan-outs that would happen anyway — does not make models smarter or guarantee correctness. Zero-dependency, offline, no API keys.
coalboard
Consensus & debate board. AUTO-trigger = the error-not-allowed slice (security/crypto, DB/financial migrations, high-precision math, anything catastrophic-on-error); convened MANUALLY ("/coalboard") it generalizes to any hard problem in ANY domain worth several diverse perspectives. OPINION lane: about to ask the user to settle a decision → ADD an "ask CB" option (~4 lenses + judge); the pick IS the consent, never auto-convened. With the user's consent it convenes diverse epistemic lenses (empirical/source-grounded, formal, show-me skeptic) in PARALLEL; a judge synthesizes on VERIFIED inputs; an independent out-of-frame solver breaks ties; the human signs off. Bounded cost (no whack-a-mole) + zero-breakage (staging) — improves correctness, claims no reliability number. Off ~90% of the time. Triggers: "/coalboard", "convene the board", a critical-task signal, a CoalTipple hand-off. Cross-agent (verified: Claude Code + Antigravity; others designed-for, unverified). Zero-dependency, offline, no API keys.
Bio shown is the top-scored skill's repo description as a fallback — real GitHub bios land in a future update.