v0.2.3 - description cap conform + doctrine link

Fixed

• SKILL frontmatter description trimmed 1049 -> 985 chars (the 1024 cross-platform cap; the one-flock conform CoalBoard + CoalWash already did) - every auto-trigger word and both honesty clauses survive verbatim.
• README doctrine link now points at the .github repo where the doctrine docs live.
• The [0.2.2] CHANGELOG heading an in-progress edit had overwritten is restored.

Credit: agnix + the user's CoalBoard nasa audit. Gate: 19/19.

v0.2.2 — wallet honesty (DOLLAR bound, not token bound)

Fixed

Honesty: the "wallet ≈ solo cost" claim was FALSE in raw tokens on the always-loaded / front-door surfaces (SKILL description + body, README table + prose, the conductor hook messages, marketplace/plugin descriptions). The shipped benchmark refutes it: fan-out multiplies the fixed ~per-sub baseline by N, so raw tokens run higher than solo (~5.3× on a small 6-spot job). The wallet is a dollar bound (cheap worker tiers, −15% vs solo-on-an-expensive-main) + a wall-time bound, never a token bound. Every surface reworded; the token-budget keyword → cost-budget. The CHANGELOG [0.2.0] note + the benchmark already told the truth — this conforms the lagging headline surfaces. (Board-2 dogfood finding; the "false claim worse than none" class.)

v0.2.1 - description trim + sibling links

PATCH: SKILL description trimmed to 1017 chars (cross-platform-safe cap; always-loaded per session = a standing token saving), contract-template code spans un-wrapped (external-linter FP fix), README names CoalWash + CoalLedger and gains the claude.ai sequential-degrade compatibility row. No behavior change. Full history: CHANGELOG.md

v0.2.0 - stable: graduated on the first real full-pipeline run

CoalFace leaves beta. The graduation gate - held honestly since the first public beta ("graduates only when a REAL /coalface swarm runs on a REAL flat worksite") - is satisfied by a genuine job: a 15-spot doc-conform sweep across 5 sibling repositories, running the full contract end-to-end.

The run: scout (spot list + shared-digest + invariants) ? deterministic partition (the min-unit floor merged one repo's spots into the conductor) ? 1 wave, width 4 (slow-start) ? 4 workers returning anchor-edit orders as text ? QC at collection caught 4 real issues before any apply (a scout anchor drift, 2 template path errors, a memory-vs-code contradiction) ? single-writer sequential apply ? domain gates green ? receipt.

Honest wallet note: on this job the swarm did not beat the solo estimate (?220k worker tokens vs a ~150-250k solo estimate) - consistent with the shipped benchmark: the wallet is a bound and a right-sizing discipline, not a token saving. The graduation evidence is the discipline working - QC catches before apply, no worker ever writing the tree, gates green - not a token win.

Code is identical to 0.2.0-beta.3; this release changes the version and status labels only. Full history: CHANGELOG.md.

v0.1.0-beta.1 - initial public beta

Initial public BETA of CoalFace — the fan-out discipline for AI coding agents.

CoalFace does not create fan-outs and does not make models smarter — it disciplines the fan-outs an agent would already do, so the same promises come with structural guarantees:

	ad-hoc fan-out	via CoalFace
tokens	unbounded	bounded ~ the solo-cost estimate (wallet + shared digest + min-unit floor + no self-retry)
speed	under-fanned in practice	full wave width, bandwidth 1-100% with TCP-style AIMD backoff
quality	a stray worker writes your real tree	workers return anchor-edit orders as text; ONE writer applies behind a snapshot + QC + domain gate

Flow: mandatory scout -> deterministic partition -> waves of structural-leaf workers -> mechanical QC (one rework round) -> single-writer apply (skip-and-flag, per-class atomic rollback) -> receipt. Activation: coalfaceMode auto (factory) / on / off + manual /coalface. Composes with the family: CoalBoard wins error-not-allowed work, CoalTipple tiering optional, subagent-safety inherited.

Zero-dependency, offline, no API keys. 19 hermetic node:test cases. Beta: Claude Code verified; other concurrent-subagent platforms design-supported (reports welcome via the platform-report template).

Install:

claude plugin marketplace add TheColliery/CoalFace
claude plugin install coalface@coalface