Night Market Config Catalog

Environment variables

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).