Night market config catalog

Night Market Config Catalog

Every configuration axis in this repo: where it lives, what the options and defaults are, whether it is production policy or an experimental opt-in, and what guards it. Flags drift, so every axis ends with a one-line re-verification command. Run it before trusting a value stated here.

Terms used once, defined once:

• Axis: one independently tunable configuration surface (a file, a key, or an environment variable).
• Guard: the mechanism that notices when the axis breaks or drifts (a test, a pre-commit hook, a CI job, or a script).
• Shadow mode: a hook that evaluates and warns but does not block. The blocking posture is opt-in.

Axis table

Surface	File	Key options and defaults	Status	Guarded by
Quality gate thresholds	.claude/quality_gates.json	enforce_blocking true, max_critical_issues 3, max_warnings_per_dimension 5; per-dimension keys below	Production policy	Documented in docs/quality-gates.md; consumed as policy prose by sanctum PR workflows (see caveat below)
Context governance	.claude/context_governance.json	enforce_strict_limits false, require_progressive_disclosure true, require_modular_structure true, require_optimization_level "standard", block_on_critical_violations false, max_violations_per_file 5	Production, advisory	docs/quality-gates.md
Repo Claude settings	.claude/settings.json	Contains only a description key. NO hooks registered at repo level; all active hooks ship from plugins	Production invariant	Convention; see night-market-architecture-contract
Behavioral rules	.claude/rules/*.md (8 files)	bounded-discovery, markdown-formatting, plan-before-large-dispatch, prefer-invariants-over-fallbacks, prefer-rg-over-grep, shared-utility-consumer-rule, skill-exit-criteria, slop-scan-for-docs	Production	slop CI, pre-commit, review practice
Python toolchain	root pyproject.toml	See "Root pyproject tool tables" below	Production	Pre-commit hooks, typecheck.yml, security.yml
Per-plugin coverage	plugins/*/pyproject.toml [tool.nightmarket]	coverage_threshold = 90 in 19 plugins, 85 in gauntlet (20 files total)	Production	Read by scripts/run-plugin-tests.sh
Pre-commit pins	.pre-commit-config.yaml	Remote pins: pre-commit-hooks v6.0.0, bandit 1.8.6 (last release supporting the 3.9 system interpreter). Ruff runs from a local repo via the uv-managed binary so hook and make format share one version	Production	scripts/check_pinned_versions.py, python39-compat.yml
Plugin manifest trio	plugins/<p>/.claude-plugin/plugin.json, .claude-plugin/metadata.json, openpackage.yml	All three carry version in lockstep with the marketplace (1.9.15 at compile time)	Production	plugins/sanctum/scripts/update_versions.py, plugins/abstract/scripts/validate_plugin.py
Marketplace version	.claude-plugin/marketplace.json top-level version	Single ecosystem version, source of truth for the fan-out	Production	update_versions.py <version> bumper
MCP servers	.mcp.json	One stdio server: markitdown via uvx markitdown-mcp	Production	Manual
LSP config	.cclsp.json	pylsp for py/pyi; typescript-language-server for js/ts and markdown	Production	Manual
Feature-review scoring	.feature-review.yaml	version: 1; weight tables for value and cost; thresholds high_priority 2.8, medium_priority 1.8, confidence_warning 0.6	Production	Consumed by imbue:feature-review
Egregore runtime config	.egregore/config.json (runtime file, created in the target repo, not committed here)	Nested dataclasses in plugins/egregore/scripts/config.py: overseer, alerts, pipeline, budget, discussions. pipeline.completion_integrity = False	Experimental opt-in flag inside production config	Unit tests in plugins/egregore/tests/test_config.py (default, roundtrip, and raw-JSON opt-in paths)
Herald Stop-hook judge	plugins/herald/hooks/hooks.json + env vars	Stop hook double_shot_latte.py registered with timeout: 10; internal LLM_TIMEOUT_SECONDS = 8	Hook production, LLM path experimental	Guard test asserts LLM timeout stays under the registered hook budget

Caveat on the two .claude/*.json policy files: no Python script in the repo reads them directly. They are policy inputs referenced by docs/quality-gates.md and by sanctum command prose (prepare-pr, pr-review configuration). Treat them as contract documents for review workflows, not as runtime config a daemon loads. Direct programmatic consumers: none found at compile time (candidate: wire one before tightening values).

Quality gate dimensions (.claude/quality_gates.json)

Per-dimension keys and defaults:

Dimension	Keys	Blocks?
performance	max_file_size_kb 20, max_tokens_per_file 5000, max_function_lines 60, max_complexity_score 12	No (block_on_violation false)
security	block_hardcoded_secrets true, block_insecure_functions true, require_input_validation true	Yes (block_on_violation true)
maintainability	max_technical_debt_ratio 0.3, max_nesting_depth 5	No
compliance	require_plugin_structure true, require_proper_metadata true	No

Security is the only blocking dimension. Everything else is advisory, capped globally by max_critical_issues: 3.

Root pyproject tool tables

Root pyproject.toml (version 1.9.15, requires-python >=3.12) carries these tables:

• [tool.ruff]: line-length 88, target-version py312.
• [tool.ruff.lint] select: E, W, F, I, B, C4, UP, PLC0415, PLR0913. extend-ignore: E501, B007, B008, B905, E402, and UP017. UP017 is load-bearing: it keeps timezone.utc instead of the 3.11+ datetime.UTC alias because hook scripts must import under the system Python 3.9. Removing it re-breaks hooks (this has happened three or more times, and an AST test now holds the line).
• [tool.ruff.lint.per-file-ignores]: tests get S-series and assert allowances; **/__init__.py gets F401; **/hooks/*.py gets BLE001, PLC0415, E402 (resilience lazy imports); scripts/*.py gets E402.
• [tool.pytest.ini_options]: norecursedirs includes plugins/*, so plugin tests MUST run per-plugin. Markers: unit, integration, e2e, slow, network, plugin, skill, hook, command, bdd, benchmark. --strict-markers is on.
• [tool.coverage.report]: fail_under = 85 at root.
• [tool.mypy]: python_version = "3.12" with strict-leaning flags (disallow_untyped_defs, warn_unused_ignores, etc.).
• [tool.uv.workspace] members: plugins/abstract, plugins/leyline, plugins/conserve, plugins/core (a minimal set for linting, not all 26 plugin dirs).
• [tool.bandit] skips: B101, B103, B105, B108, B110, B112, B310, B311, B404, B603, B607, B608. Each skip has an inline rationale comment in the file. Keep that pattern when adding one.

Manifest trio and the two semver namespaces

Each plugin ships three manifests that must stay in sync:

• .claude-plugin/plugin.json: name, version, component arrays.
• .claude-plugin/metadata.json: version, dependencies, provides, claude integration block.
• openpackage.yml: cross-framework manifest (name, version, description, keywords).

Two distinct semver namespaces appear in metadata.json and are easy to confuse:

1. The ecosystem version ("version": "1.9.15"), bumped in lockstep from marketplace.json by plugins/sanctum/scripts/update_versions.py <version>.
2. Capability dependency ranges, for example imbue declares "dependencies": {"abstract": ">=2.0.0"}. That 2.0.0 is a capability-contract version, NOT the marketplace version. Do not "fix" it to 1.9.x during a bump.

Environment variables

Variable	Consumer	Values and default	Effect	Status
VOW_SHADOW_MODE	imbue hooks via plugins/imbue/hooks/shared/vow_utils.py	Default "1" (shadow). Set 0 to block	Vow hooks (bounded reads, scope ramp, package hallucination, no-AI-attribution, no-emoji-commits) warn instead of blocking	Shadow default is production; blocking posture experimental
UV_TOOL_DIR	root Makefile (line 9)	Default ./.uv-tools	Where uv installs tools; ruff resolves from $(UV_TOOL_DIR)/ruff/bin first on PATH	Production
SKRILLS_REPO	root Makefile (skrills-build)	Default $HOME/skrills	Source checkout used to cargo-build the skrills binary	Production (skrills itself optional; Python fallbacks exist)
CONSERVATION_MODE	plugins/conserve/hooks/session-start.sh	quick, deep, normal; default normal; unknown values warn and fall back to normal	How much conservation guidance loads at session start	Production
DOUBLE_SHOT_LATTE_LLM	plugins/herald/hooks/double_shot_latte.py	Set "1" to enable	Enables the LLM "second shot" tiebreaker, only on the single ambiguous Stop-hook outcome	Experimental opt-in, default off
DOUBLE_SHOT_LATTE_MODEL	same hook	Default "haiku"	Model used for the second shot	Experimental
DOUBLE_SHOT_LATTE_MAX_CONTINUATIONS	same hook	Integer (read from env; see hook source for cap semantics)	Caps forced continuations	Production guard
CLAUDE_HOOK_JUDGE_MODE	same hook	"true" disables the LLM path	Prevents the judge from invoking itself recursively	Production guard

Experimental and opt-in flags

Both current experiments follow the same pattern: default off, guard tests, blocking behavior only on explicit opt-in.

egregore completion_integrity

• Key: pipeline.completion_integrity, a bool defaulting to False in the PipelineConfig dataclass at plugins/egregore/scripts/config.py (line 49 at compile time).
• Stored in the target repo's .egregore/config.json, read by the summon skill alongside .egregore/manifest.json.
• When True: a quality verdict of "fix-required" becomes a step failure (cannot advance to ship) and merge is held for human review regardless of auto_merge.
• Guards: plugins/egregore/tests/test_config.py covers the False default, the opt-in roundtrip, and loading from a hand-written raw JSON file.

herald LLM second shot

• The Stop hook itself is production (registered in plugins/herald/hooks/hooks.json with timeout: 10).
• The LLM tiebreaker is gated three ways: DOUBLE_SHOT_LATTE_LLM must be "1", judge mode must not be active, and the turn must land on the single ambiguous outcome (every confident verdict skips the LLM entirely).
• LLM_TIMEOUT_SECONDS = 8 is deliberately under the 10-second registered budget. A value above the budget once made the harness kill the hook before any verdict (full record: night-market-failure-archaeology SB7). A guard test now asserts the ordering. Never raise the LLM timeout without checking the registered budget first.

Adding a config axis

Checklist for any new flag, threshold, or env var:

• Default preserves current behavior (default off, or the value the system already exhibits).
• A guard test asserts the default AND the opt-in path, including loading from raw serialized config (model: plugins/egregore/tests/test_config.py).
• Env vars validate their input and fall back to the default on unknown values with a warning (model: plugins/conserve/hooks/session-start.sh).
• Any timeout inside a hook is asserted against the hook's registered budget in hooks.json.
• Add a row to the axis or env var table in this file, with status (production or experimental) and guard.
• Add a one-line re-verification command under Provenance.
• If the axis affects gating or review policy, update docs/quality-gates.md and route the change through night-market-change-control.

When NOT to use

• Running tests, lint, gates, or releases: use night-market-operations.
• Deciding whether a config change is allowed at all, or how to review it: use night-market-change-control.
• Plugin/skill/hook mechanics (what hooks.json fields mean, how manifests load): use claude-code-plugin-reference.
• Recreating the toolchain (uv, Python versions, skrills): use night-market-build-and-env.
• Diagnosing a broken gate or hook: use night-market-debugging-playbook.

Exit Criteria

• Every re-verification command under Provenance exits 0 and its output matches the value stated in this catalog.
• jq -r '.version' .claude-plugin/marketplace.json equals the version in each plugin's plugin.json, metadata.json, and openpackage.yml (spot-check at least one plugin).
• rg -l "coverage_threshold" plugins/*/pyproject.toml | wc -l returns 20. Any other count means this catalog is stale and must be updated before use.
• rg -n "completion_integrity: bool = False" plugins/egregore/scripts/config.py matches. If the default changed, the Experimental section here is wrong and must be rewritten.
• A new config axis added by the current change appears in the axis or env var table with a guard and a re-verification one-liner.

Provenance and maintenance

Compiled 2026-07-02 against repo v1.9.15 (branch discussions-fix-1.9.14). Line numbers and counts are volatile; re-verify before quoting.

Re-verification one-liners per axis:

# Quality gates and context governance
jq . .claude/quality_gates.json
jq . .claude/context_governance.json

# Repo settings must stay hook-free
jq 'keys' .claude/settings.json   # expect ["description"]

# Behavioral rules roster
ls -1 .claude/rules/

# Root toolchain tables
rg -n "fail_under|extend-ignore|^select|target-version" pyproject.toml
rg -n "\[tool.bandit\]" -A 20 pyproject.toml
rg -n "\[tool.uv.workspace\]" -A 8 pyproject.toml

# Per-plugin coverage thresholds
rg -n "coverage_threshold" plugins/*/pyproject.toml

# Pre-commit pins
rg -n "rev:" .pre-commit-config.yaml

# Marketplace version and manifest fan-out (imbue as sample)
jq -r '.version' .claude-plugin/marketplace.json
jq -r '.version' plugins/imbue/.claude-plugin/plugin.json
jq -r '.version, .dependencies' plugins/imbue/.claude-plugin/metadata.json
head -3 plugins/imbue/openpackage.yml

# Root dotfiles
jq . .mcp.json
jq '.servers[].command' .cclsp.json
rg -n "high_priority|medium_priority|confidence_warning" .feature-review.yaml

# Env vars
rg -n "VOW_SHADOW_MODE" plugins/imbue/hooks/shared/vow_utils.py
rg -n "UV_TOOL_DIR|SKRILLS_REPO" Makefile
rg -n "CONSERVATION_MODE" plugins/conserve/hooks/session-start.sh

# Experimental flags
rg -n "completion_integrity" plugins/egregore/scripts/config.py
rg -n "DOUBLE_SHOT_LATTE|LLM_TIMEOUT_SECONDS|JUDGE_MODE" plugins/herald/hooks/double_shot_latte.py
jq . plugins/herald/hooks/hooks.json

Known drift risks: pre-commit rev: pins (guarded by scripts/check_pinned_versions.py), the coverage-threshold file count (20 at compile time), the egregore config.py line number, and the promotion status of the two experimental flags (see night-market-completion-integrity-campaign for the promotion question).