Release 3.0.0: replace the mandatory method corpus with a microkernel and eight modules

[mzored]

Implements skiphow-unified-implementation-spec.md end to end.

What changes for a project that installs SkipHow

The agent stops running a process on your project. Installing SkipHow no longer gives a repository a tracker convention, no longer puts an item in front of every review branch, and no longer commits as a condition of finishing. The always-loaded kernel keeps authority, trust, preservation of work the run does not own, delegate safety, and what may be claimed as done. Everything else became guidance the agent consults when the work makes it worth reading.

Projects installed on 2.x will see fewer records written and fewer side effects they did not ask for, which is why this is a major version.

Architecture

Twenty-three method files became eight: product, technical design, diagnosis, verification, delegation, tracked work, integration, and writing for agents. Fourteen were consolidated, five deleted outright. The changelog says where each went.

	2.16.1	3.0.0
kernel SKILL.md	16,797 B	13,286 B	−20.9%
references/	82,199 B	46,251 B	−43.7%
normative corpus	98,996 B	59,537 B	−39.9%
method files	23	8	−65%

The kernel fell by a fifth while gaining the explicit trust hierarchy, the delegate brief contract, and credential handling that used to sit in a method, so the procedure removed from it is larger than the net figure shows.

Review

Five review lanes ran over the change. Two findings were regressions of rules earlier releases had added to fix observed failures, and both are fixed:

• The tracker claim rule was carried across the consolidation without the guard that made it safe, so a request only to diagnose one item would again have demanded a tracker write — the defect 2.8.0 fixed. It is now gated on an authorized write and carries its guard.
• The kernel paragraph on how a protected or human-only ask is worded had been deleted. A receipt had produced it, after a run put a credential ask to an owner as a menu of two technical mechanisms. It is back in the kernel.

Rules stated in two or three places were cut to one home each, with the kernel keeping every invariant. Sentences lost with deleted methods were restored where nothing covered them, including aborting or discarding work in an active operation and rewriting shared history.

Seven claims in docs/prior-art.md and five in the changelog entry described a package that no longer shipped. Every claim on both was checked against the files and corrected.

Evidence

evals/ records thirteen synthetic cases for the behaviors this release changed, across eight fixtures and three comparison arms, each naming one observable with the events it expects and forbids. Nothing in it has been run. It gates nothing, no pull request needs one, and every case stands not_run / UNVERIFIED. It makes a later receipt reproducible and comparable; it is not evidence.

Whether the compact package produces better model behavior than the one it replaces remains UNVERIFIED. No receipt compares them.

Verification

• python scripts/check.py — passed
• 184 tests — passed
• git diff --check — clean
• Claude package validation and isolated install — PASS
• Codex package validation — PASS (against the pinned plugin-creator validator)
• Codex isolated install — UNVERIFIED; forcing it fails on a local /etc/codex/requirements.toml marketplace-source policy, and CI skips isolated installation entirely

🤖 Generated with Claude Code