This repository houses the global configuration, landing page profile, central installation scripts, and shared developer doctrines for TheColliery organization.
- CoalMine: Quality-canary skills suite for AI coding agents — nine canaries plus the auto-cadence hooks that run them unprompted.
- CoalTipple: Model/effort router — delegate-down to save, escalate-up for quality, with a fail-safe model-ranking Lock.
- CoalBoard: Consensus & debate board — diverse lenses verify error-not-allowed work before it ships, or audit existing work and report findings.
- CoalHearth: Session warm-resume — journals state so an interrupted session resumes from a recovery block instead of a manual rebuild.
- CoalFace: Fan-out discipline — scout → waves → QC → one writer, the whole swarm bounded by the estimated solo $-cost, never a token bound.
- CoalWash (rc): Memory washer/defragmenter, two lanes: class-B (session-loaded memory + governance) cleans the FAT never the MEAT, under a code-enforced gate proving zero structured-token loss, deletes plan-sourced and undo-backed (snapshot + whole-run rollback), apply transactional; class-A (the at-rest transcript estate) is never semantic-edited — byte-identity only, copy-verify-then-delete with a round-trip restore, and wizard-invoked only.
- CoalLedger (beta): Docs-health canary suite — CoalMine for documentation (wrong · incomplete · stale · contradictory · broken · unreadable · config-gated leaked),
doc-structuredetecting through a CommonMark+GFM AST engine, plus a quiet docs memory-drift reminder; cross-agent + multilingual.
You can install any or all tools in the series (CoalMine, CoalTipple, CoalBoard, CoalHearth, CoalFace, CoalWash (beta), CoalLedger (beta)) using the central installer located in this repository:
curl -fsSL https://raw.githubusercontent.com/TheColliery/.github/main/install.mjs -o colliery-install.mjs
node colliery-install.mjs all # the whole suite (all 7)
node colliery-install.mjs 3 # just CoalBoard
node colliery-install.mjs 1 2 3 # CoalMine + CoalTipple + CoalBoard
node colliery-install.mjs 6 7 # CoalWash + CoalLedger (beta)Periodic blind output-quality evaluations and regression tests. To keep them honest, this page links to each tool's RESULTS.md (the single source of truth) and never copies the figures -- so a number here cannot drift from, or be invented apart from, the measured run.
- CoalMine -- defect-detection recall/precision across 7 of the 9 canaries, multiple engines (Claude tiers + Antigravity cross-vendor). Results: benchmarks/CoalMine/RESULTS.md.
- CoalTipple -- ON-vs-OFF paired benchmark (4 tasks × 3 tiers): routing ON = 4/4 task quality on both baselines vs OFF = 3/4, ~23% cheaper from an Opus main / cost-neutral from a Sonnet main; plus a 5-domain (crypto, proof, research, legal, voice) output-correctness matrix, Claude Code. Results: benchmarks/CoalTipple/RESULTS.md; routing-cost savings: benchmarks/CoalTipple/ROUTING-SAVINGS.md.
- CoalBoard -- with-the-board-vs-without on error-not-allowed tasks, two platforms (Claude Code reliability + Antigravity cross-vendor), including the honest correlated-blind-spot ceiling. Results: benchmarks/CoalBoard/RESULTS.md.
- CoalHearth -- interruption damage: cold restart vs warm resume on a 10-file mid-refactor (token delta + the categorical state-fidelity recovery), Claude Code. Results: benchmarks/CoalHearth/RESULTS.md.
- CoalWash -- durability campaign CLOSED: 53/53 loss-taxonomy classes covered (26 forced-wear-measured, 27 adversarial-verified), wear arm retired on 2-dry with every detected loss restored, Thai arm 33/33 tripwires, adversarial 0/33 traps leaked, workability 10/10 parity; a controlled equal-size fidelity test lost 0 of 448 structured tokens where a fair naive compress lost 218.8. The safety floor is CODE-held; model quality moved yield, never the loss floor. The headline sawtooth-vs-bloat saving is still unmeasured (protocol ready). Results: benchmarks/CoalWash/RESULTS.md.
- CoalLedger -- not yet measured (launched unbenchmarked rather than with an invented number); the engine's mechanical layer is fixture-gated in-repo, and the first published benchmark will follow the flock shape here.
- CoalFace -- fan-out cost: solo vs ad-hoc vs scout-digest on a shared-context worksite (tokens + dollars + the coarse-pack lever), Claude Code + Antigravity cross-vendor. Results: benchmarks/CoalFace/RESULTS.md.
This specific repository (.github) manages the following internal assets:
profile/README.md: The main landing page displayed on the TheColliery Organization Profile.install.mjs: The unified cross-platform DLC selector installation script for all active tools in the suite.DESIGN-PRINCIPLES.md: The Quantum 11 performance & design guidelines governing all series projects.hooks-safety.md: The Phoenix 13 safety hook commandments (sandboxing, fail-silent execution).scripts-quality.md: The CLI and script quality rules for release-gating.DOC-PATTERN.md: The shared public-doc pattern (README / SECURITY / CONTRIBUTING / PRIVACY / CHANGELOG) that keeps every repo's docs sibling-consistent.SKILL-REPO-PATTERN.md: The shared repository layout and machinery — manifests, source-vs-dist, config system, gates, workflows.RELEASE-PATTERN.md: The shape of a GitHub Release — title, notes body, and which tags get one.PERMISSION-MATRIX.md: The least-power capability each tool requests, and the rights it deliberately never asks for.SWEEP-MARKS.md: The event → doc-spot registry: which surfaces each recurring event rots, and who owns sweeping each one.CLAUDE-AI-INSTALL.md: Which skills run on claude.ai, and how to package one for it.benchmarks/: The per-tool benchmark records the Series Benchmarks section above links to.
For the public-facing organization profile, please visit github.com/TheColliery.