night-market-change-control

Every change to this repo passes through a fixed law stack, a classification step, and a gauntlet of automated gates. This skill tells you which class your change is, which gates it must pass, and which rules are non-negotiable because a past incident made them so. Nothing here may be routed around: if a gate fails, fix the cause, never the gate.

Jargon used below: a "gate" is any automated check that can block a commit, PR, or release. An "ADR" is an Architecture Decision Record in docs/adr/. The "Iron Law" is constitution rule 3: no implementation without a failing test first.

The law stack

When two documents conflict, the higher one wins.

Rank	Source	Contents
1	CONSTITUTION.md	10 immutable rules, override and amendment process
2	.claude/rules/	8 project rules (markdown style, read budgets, slop gates, invariants)
3	docs/adr/	17 numbered decision records (0001 to 0017)
4	Guides in docs/	quality-gates, testing-guide, plugin-development-guide

A skill, hook, or agent instruction that says "skip rule N" without an explicit user grant or a merged amendment is itself a defect (CONSTITUTION.md, Override mechanism section).

The ten constitution rules

One line each. Where a rule was written in blood, the incident column names the blood.

#	Rule	Motivating incident
1	Disclose AI involvement in every PR; never strip real or add fake AI attribution	Enforced by hook plugins/imbue/hooks/vow_no_ai_attribution.py
2	AI commits over 200 changed lines need a spec, ADR, or plan doc first (lockfiles, fixtures, snapshots excluded)	Size = scrutiny principle; see the unbloat incident under non-negotiables
3	Iron Law TDD: failing test before implementation for plugin Python. Skills and prose need structural tests (test_skill_<name>.py)	Design principle, no single incident
4	One identity leak ("As a large language model") in any committed artifact is an automatic revert	Pattern catalog in scribe slop-detector
5	Quality claims ("fast", "production-ready") need in-repo evidence or deletion	Design principle
6	No bypassing gates: no --no-verify, no SKIP=hook, no unauthorized force push, no bare suppression comments without a stated reason	Neutered-mypy incident (CHANGELOG 1.9.12); bulk-ignore revert 06b9b1db
7	New dependencies need justification; 18 months unmaintained is presumed abandoned; verify AI-suggested package names against the registry	Slopsquatting defense; hook guard_package_hallucination.py (imbue)
8	Docs cost reader-time (audience x frequency x per-read time); writing effort must match	Design principle
9	Prefer deletion over rewriting; AI slop is overwhelmingly additive	Design principle
10	Errors are not optional: propagate by default, no bare except, safe-to-discard needs an inline "why" comment	Silent-failure sweep 666171c3; ecosystem bare-except campaign

Change classification

Classify before the first commit. The class decides the paperwork.

Change	Required process
AI-generated diff, 200 changed lines or fewer	Plain PR through the standard gates
AI-generated diff over 200 changed lines	Spec, ADR, or planning doc BEFORE the code (rule 2)
Load-bearing design decision	Numbered ADR in docs/adr/ (next number after 0017)
New project-wide rule	New file in .claude/rules/ via plain PR; existing rules cite an origin issue or discussion (#454, #457)
Constitution amendment	PR titled constitution: amend rule N, a summary of what changes and why, sign-off from the repo owner
Release	Version bump, changelog section, v-prefixed tag (lifecycle below)

The change lifecycle

1. Branch

Branch from master (the main branch) using <topic>-<version>:

git checkout master && git pull
git checkout -b my-topic-1.9.16

Live examples in git branch -a: ai-slop-1.9.4, bugfixes-1.9.5, discussions-fix-1.9.14. Never delete backup/unbloat-* branches. They are the recovery points for past deletion campaigns (backup/unbloat-20260328 is how the 2026-03-28 over-deletion was undone).

2. Local gates before committing

make lint          # ruff format + ruff check --fix + bandit
make typecheck     # per-plugin strict mypy (all plugins)
make test          # per-plugin pytest via scripts/run-plugin-tests.sh
make validate-all  # plugin structure validation

Plugin tests MUST run per plugin (make <plugin>-test or make -C plugins/<plugin> test). Root pytest excludes plugins/* to avoid import-path collisions.

3. The pre-commit gauntlet

git commit runs .pre-commit-config.yaml hooks in this order (verified 2026-07-02):

1. Suppression guards: check-noqa (blocks bare inline lint suppressions), check-docstring-quality, check-json-utils-drift, check-per-file-ignores.
2. Hook registration: check-hook-registrations, check-plugin-hooks.
3. Code gates: run-plugin-typecheck --all (strict mypy on every plugin, not just changed ones), run-plugin-tests --changed.
4. Skill and plugin validation: validate-skill-descriptions, one validate-<plugin>-plugin hook per plugin, check-context-optimization, validate-description-budget, check-markdown-links, capabilities-sync-check, check-skill-graph-drift, check-skill-exit-criteria-drift.
5. Standard file hygiene (trailing whitespace, YAML/TOML/JSON syntax, merge conflicts, debug statements).
6. bandit security scan.
7. ruff-format, ruff-fix, ruff-check.
8. Pin freshness: check-ruff-version, check-pinned-versions.

The typecheck gate runs --all deliberately. It used to run --changed, and a global mirrors-mypy hook silently disabled 13 error codes. Both were fixed in 1.9.12 (CHANGELOG). Do not weaken either setting.

4. PR flow

Run in this order (all are sanctum slash commands):

1. /sanctum:prepare-pr updates docs, runs tests, drafts the PR.
2. /sanctum:pr-review reviews scope, requirements, and code.
3. /sanctum:validate-pr builds and executes a diff-derived test plan, and proves revert-tests are genuine guards.
4. /sanctum:fix-pr and /sanctum:resolve-threads handle feedback.

For plugin-touching changes, add abstract:plugin-review at the matching tier:

Tier	When	Scope	Time
branch	Default during work	Affected and related plugins	~2 min
pr	Before merge	Affected and related plugins	~5 min
release	Before version bump	Every plugin	~15 min

5. Release

1. Bump the ecosystem version. .claude-plugin/marketplace.json is the source of truth; the bumper fans it out to every plugin's plugin.json, metadata.json, openpackage.yml, pyproject.toml, and __init__.py:

   python3 plugins/sanctum/scripts/update_versions.py 1.9.16
   

   Or use the sanctum:version-updates skill, which runs a git-workspace-review preflight first.

2. Move [Unreleased] entries in CHANGELOG.md into a dated version section (Keep a Changelog 1.1.0, SemVer). Never rewrite or de-slop historical entries.

3. Update the docs of record: docs/api-overview.md version reference, plugin READMEs, and the generated capabilities reference via /sanctum:sync-capabilities.

4. Run abstract:plugin-review --tier release.

5. /sanctum:create-tag pushes a v-prefixed tag. The tag matching v* triggers .github/workflows/cross-framework-publish.yml (semver validation, cross-framework build, tarballs, GitHub release). Pushes to master separately trigger trust-attestation.yml (full test run plus SLSA attestation).

Landing a change: checklist

Step	Action	Gate satisfied
Classify	Match the change against the classification table	Rule 2, amendment process
Branch	<topic>-<version> off master	Convention
Test first	Failing test before implementation	Rule 3 (Iron Law)
Local gates	make lint && make typecheck && make test	Rule 6
Commit	Plain commit, no bypass flags, honest attribution	Rules 1 and 6
PR	prepare-pr, pr-review, validate-pr; disclose AI involvement	Rule 1
Plugin review	abstract:plugin-review at branch or pr tier	Plugin quality
Merge	All CI checks green, review threads resolved	CI

Non-negotiables

Each row is a settled battle. Re-litigating one requires new evidence stronger than the incident that settled it.

Non-negotiable	Rationale	Incident	Evidence
Never bypass or weaken a gate	A green gate that skips checks is worse than no gate: it certifies broken code	mirrors-mypy silently disabled 13 error codes; typecheck ran changed-only; bulk ruff ignores hid 73 real violations	CHANGELOG 1.9.12; commit 06b9b1db
No swallowed errors, ever	catch-and-continue in a scanner drops findings silently and reports clean	Scanners B1-B4 returned clean results on malformed input for months	commit 666171c3 (PR #521, issue #575)
Deletion campaigns need a backup branch and a markdown-reference scan	Python import graphs miss scripts referenced only from skills and commands	Tier-3 unbloat deleted 182 files (66K lines); skill-referenced scripts had to be restored	commits a3f11323 (delete), 3f280334 (restore); branch backup/unbloat-20260328
No DRY consolidation across plugin boundaries	Plugins are independent deployables, and shared code couples their release cycles	tasks_manager consolidated to a shared script, reverted, then differentiated per plugin	commits 054e2679, 29961cd2, d89a55c7
No speculative infrastructure	Unused abstraction is pure carrying cost	LSP proxy landed without a consumer and was reverted in PR #193 review	commit bc318947
Identity leaks are an automatic revert	One leaked phrase proves unreviewed AI text shipped	Constitution rule 4	CONSTITUTION.md
Historical CHANGELOG entries are never edited	The changelog is a record, not prose to polish	Codified as an anti-goal in the slop rules	.claude/rules/slop-scan-for-docs.md (anti-goals)

When NOT to use

• A gate is failing and you need to diagnose why: use night-market-debugging-playbook.
• You want the full story behind an incident named above: use night-market-failure-archaeology.
• You need the mechanics of running tests, lint, or the release scripts (flags, artifacts, environments): use night-market-operations.
• You are judging whether a design fits the architecture: use night-market-architecture-contract.
• You are deciding what evidence a test must provide: use night-market-validation-and-qa.

Exit Criteria

• The change is classified against the classification table, and any over-200-line AI diff has a spec, ADR, or plan doc committed before the implementation.
• make lint, make typecheck, and make test pass locally, and no bypass flag (--no-verify, SKIP=, force push) appears in the branch history.
• The branch name matches <topic>-<version> and is based on master.
• The PR description discloses AI involvement (authored, co-authored, or reviewed).
• For a release: marketplace.json, the dated CHANGELOG.md section, and the pushed v* tag all carry the same version.
• Any constitution change is in a PR titled constitution: amend rule N with repo-owner sign-off.

Provenance and maintenance

Compiled 2026-07-02 against repo v1.9.15 (10 constitution rules, 8 project rules, 17 ADRs). Re-verify volatile facts before relying on them:

rg -c '^### [0-9]' CONSTITUTION.md          # expect 10
ls .claude/rules/ | wc -l                    # expect 8
ls docs/adr/ | wc -l                         # expect 17
rg -n '"version"' .claude-plugin/marketplace.json | head -1
rg -n '      - id:' .pre-commit-config.yaml  # current gate order
rg -n 'v\*' .github/workflows/cross-framework-publish.yml
rg -n 'amend rule' CONSTITUTION.md           # amendment title convention

Commit hashes cited above were verified with git log --oneline -1 <hash> on 2026-07-02. Unverified/candidate: the exact count of files restored after the unbloat incident (commit subjects confirm the delete and restore, not the restored-file count), and whether docs/api-overview.md holds a version table or a single version reference (one version mention verified).