This skill records the design decisions that hold claude-night-market
together, the invariants that enforcement code keeps true, and the weak
points that are known and accepted. Read it before proposing a change
that crosses a plugin boundary, touches a hook, adds a skill, or bumps
a version. Every claim cites in-repo evidence: an ADR (Architecture
Decision Record, in docs/adr/), a commit hash, or a checked-in
enforcement file. Verify a citation before relying on it:
git log --oneline -1 <hash>
rg -n "Status" docs/adr/<file>.md
Load-bearing decisions
1. Plugins are self-contained deployables
Each of the 23 plugins under plugins/ must install and run alone.
There is no shared registry and no root-level shared library that
plugins import at runtime (ADR-0001, Accepted). Plugins detect each
other at runtime via filesystem checks and must degrade gracefully when
a sibling is absent.
Cross-plugin DRY is an anti-pattern here, and that is settled by
experiment, not taste. Commit 054e2679 consolidated 1164 lines of
duplicated tasks_manager.py from attune, sanctum, and spec-kit into a
shared root script. It broke plugin self-containment and was reverted
in 29961cd2. The durable fix, d89a55c7, made the copies
per-plugin and intentionally different. docs/dependency-audit.md
records the per-plugin copies as the approved state. Do not re-propose
the consolidation.
Practical test before you extract shared code: if a user installs only one plugin from the marketplace, does your change still work? If not, duplicate the code into each plugin instead.
2. Hooks run under host Python from a cache directory
Claude Code copies an installed plugin into a cache directory and runs
its hooks under the host system Python, which can be as old as 3.9.
The repo itself is Python 3.12 (root pyproject.toml,
requires-python >= 3.12). Only hook scripts and their transitive
import chains carry the 3.9 constraint. Five contract rules follow,
each purchased with an outage:
| Rule | Why (evidence) |
|---|---|
| Hook code and every transitive import must be Python 3.9 compatible | datetime.UTC (a 3.11+ alias) broke the whole hook import chain repeatedly. ruff kept auto-reverting the fix until UP017 was globally ignored (pyproject.toml line 165) and an AST scan test held the line (plugins/leyline/tests/test_python39_compat.py) |
Read the hook payload as JSON on stdin, never from CLAUDE_TOOL_* env vars |
Claude Code does not set those env vars, so env-reading hooks were silent no-ops for months (full record: night-market-failure-archaeology SB9). Canonical reader: read_hook_payload() in plugins/abstract/hooks/shared/hook_io.py |
| No relative paths in hooks | The cache directory is not the repo checkout. conserve's session-start hook broke on a relative path and now inlines its JSON utilities (CHANGELOG) |
| Hook entrypoints must import safely under a bare interpreter | An eager import yaml in gauntlet made every git commit emit ModuleNotFoundError. Guarded in 45dd77ef (#518), anthropic deferred in 9bfc0a7a |
Subprocess timeouts must sit below the budget registered in hooks.json |
herald's LLM call once outlived its registered Stop-hook budget, so the harness killed the hook with no verdict at all (full record: night-market-failure-archaeology SB7). Guard test: plugins/herald/tests/unit/test_double_shot_latte.py |
CI enforcement: .github/workflows/python39-compat.yml runs two
gates with uneven coverage. Gate 1 (ruff UP007, flags 3.10+ union
syntax) covers 12 plugins' hooks/ dirs. Gate 2 (hook test suites
inside a real Python 3.9 venv) covers only 7 plugins: abstract,
conserve, egregore, imbue, leyline, memory-palace, sanctum. herald
ships a Stop hook yet appears in neither gate.
3. One ecosystem version, fanned out to every manifest
.claude-plugin/marketplace.json is the version source of truth
(1.9.15 as of 2026-07-02). Each plugin carries three manifests that
must stay in sync with it and with each other:
.claude-plugin/plugin.json(name, version, component arrays).claude-plugin/metadata.json(version plus dependency hints)openpackage.yml(cross-framework manifest)
Plus pyproject.toml and any __init__.py carrying __version__.
Never hand-edit versions across files. Use the bumper, which finds and
rewrites all of them:
uv run python plugins/sanctum/scripts/update_versions.py <version>
One trap: metadata.json dependencies (for example imbue declaring
"abstract": ">=2.0.0") is a separate semver namespace for capability
compatibility. It does not track the marketplace version and a 2.0.0
there does not mean marketplace 2.0.0 exists.
4. Skill discovery budget is finite and overflow is silent
Claude Code loads every installed skill's description into context at 2% of the context window, with a 16,000-character fallback (ADR-0004, Accepted, updated 2026-05-21). Descriptions that exceed the budget make skills invisible with no error anywhere. That is why:
- Each description is capped at 160 characters
(
docs/skill-description-guide.md), enforced by thevalidate-description-budgetpre-commit hook backed byplugins/abstract/scripts/validate_budget.py. - The ecosystem-wide validator ceiling is 90,000 characters
(
DEFAULT_BUDGETinplugins/abstract/scripts/validate_budget.py, overridable viaSLASH_COMMAND_TOOL_CHAR_BUDGET). ADR-0004 anddocs/skill-description-guide.mdstill cite a stale 60,000 figure; the script is the enforcer. Flag the ADR for an update through change control.
When adding a skill, spend the 160 characters on trigger phrases, not on restating the name.
5. Collective memory and trust are external services
Decisions, learnings, and audit syntheses are posted to GitHub
Discussions, which act as agent collective memory across sessions
(ADR-0007, Accepted). There is no gh discussion subcommand: all
Discussions access goes through gh api graphql. Release trust is
established by SLSA attestation (Supply-chain Levels for Software
Artifacts) of trust-report.json on master pushes
(.github/workflows/trust-attestation.yml). The original blockchain
design (ERC-8004) was dropped for cost. ADR-0008 is marked Superseded
2026-03-15 and now points at GitHub Attestations. See
night-market-collective-memory for the Discussions workflow.
6. Every skill has one of three roles
docs/skill-integration-guide.md defines the role taxonomy used when
judging whether a skill is an orphan or a hub:
| Role | Inbound refs | Invoked directly | Example |
|---|---|---|---|
entrypoint |
low (0-3) | yes | sanctum:do-issue |
library |
high (4+) | rarely | imbue:proof-of-work |
hook-target |
varies | no | imbue:vow-enforcement |
A skill with zero inbound Skill() references is only a problem if it
also has no command path and no hook that loads it. Check the role
before archiving anything.
Invariants
Each row names the enforcement that keeps the invariant true. If you weaken the enforcement, you own the failure mode in the third column.
| Invariant | Enforced by | What breaks if violated |
|---|---|---|
| Hook import chains are Python 3.9 safe | .github/workflows/python39-compat.yml (2 gates) plus AST scan test plugins/leyline/tests/test_python39_compat.py |
Total hook outage: one 3.11-only import kills every hook that transitively loads the module |
No bare except: and errors propagate by default |
ruff E ruleset (root pyproject.toml) plus CONSTITUTION.md rule 10 |
Scanners silently drop files and report success on garbage input, and failures become invisible |
| All manifests carry the same ecosystem version | plugins/sanctum/scripts/update_versions.py rewrites pyproject/plugin.json/metadata.json/openpackage.yml/__init__.py in one pass |
Version drift across ~100 files, and marketplace.json stops being the source of truth |
book/src/reference/capabilities-reference.md matches plugin registrations |
scripts/capabilities-sync-check.sh via make docs-sync-check and .github/workflows/capabilities-sync.yml |
Published docs advertise components that do not exist, or hide ones that do |
No new dangling Skill(plugin:name) references |
scripts/check_skill_graph_drift.py ratchet against scripts/skill_graph_baseline.json |
A model tries to load the referenced skill and silently gets nothing |
Every SKILL.md has an ## Exit Criteria section |
scripts/check_skill_exit_criteria_drift.py ratchet against scripts/skill_exit_criteria_baseline.json |
Skills the model cannot tell when to stop executing (vague success criteria at scale) |
| Hook subprocess timeout < registered hook budget | Guard test in plugins/herald/tests/unit/test_double_shot_latte.py asserts LLM_TIMEOUT_SECONDS fits inside the hooks.json timeout |
Harness kills the hook mid-flight and no verdict is emitted at all |
| Hook payloads are read stdin-first | plugins/abstract/hooks/shared/hook_io.py read_hook_payload() shared by hook scripts |
Hooks become silent no-ops (this happened, and the only symptom was a starved learning digest) |
The two ratchet scripts share one mechanic: they fail a commit only
when the violation count rises above the committed baseline, and they
ask you to lower the baseline when the count drops. Never raise a
baseline to get a commit through. That is routing around change
control. One documented exception: a brand-new library skill
legitimately starts uncalled, and scripts/check_skill_graph_drift.py
(plus the _comment in scripts/skill_graph_baseline.json) instructs
you to raise max_uncalled_libraries to record the 30-day consumer
grace period that .claude/rules/shared-utility-consumer-rule.md
grants. See night-market-change-control.
Known weak points
Stated plainly so nobody rediscovers them the hard way. These are open by decision or by neglect, not secrets.
-
Skill-discovery overflow is silent. If total description overhead exceeds the context budget, skills vanish with no error (ADR-0004). The 160-char cap and validator ceiling are preventive guards. There is no runtime detection of overflow.
-
The quality-gate federation has one real orchestrator. The gate skills (karpathy-principles, scope-guard, proof-of-work, justify, rigorous-reasoning, vow-enforcement) are designed to compose, but only
egregore:quality-gatesequences the full pipeline today (docs/quality-gates.md, "Who currently orchestrates"). Every other consumer picks gates ad hoc. -
docs/project-brief.md,docs/specification.md, anddocs/implementation-plan.mdare overwritten each feature cycle. The current contents describe only the latest cycle (insight-palace bridge as of 2026-07-02). Past cycles survive only in git history. Do not cite these files as a permanent record. -
Duplicate module names across plugins force per-plugin mypy. Running
mypy plugins/from the root collides on duplicate module names (comment in.github/workflows/typecheck.yml). The same constraint makes root pytest excludeplugins/*(norecursedirsin rootpyproject.toml, and rootconftest.pydocuments theImportPathMismatchError). Always run plugin tests and typechecks per plugin. -
mdbook is unpinned in the deploy workflow.
.github/workflows/deploy-book.ymlinstallsmdbook-version: 'latest', so an upstream mdbook release can break the book deploy with no repo change. Candidate fix: pin a version. Not done as of 2026-07-02. -
The autonomous loop rides an undocumented harness behavior. Egregore continues across turns because its Stop hook returns
{"decision": "block", "reason": ...}and the harness feeds the reason back as the next instruction. Upstream documents Stop hooks as a gate on stopping, not as a continuation primitive, and the sanctioned alternatives (/loop,CronCreate) are session-scoped. ralph-wiggum rides the same behavior, bounded by an iteration count. Egregore's cost is bounded as of9f31a878(stall detection, default 3), and the reliance is not. No test can cover the harness end of it. Full record:docs/adr/0022-stop-hook-reinjection-as-continuation.md(2026-08-23).The reliance stands. What changed is that its failure is now observable rather than silent.
plugins/egregore/scripts/continuation_baton.pyhas the session record each handoff with the deadline by which the next turn should have started, so reinjection quietly ceasing strands a baton with the process still alive. That is a distinct signal, where before an upstream change and a finished run looked the same from outside. Stranded means stalled rather than old, so a long healthy run never reads as one. The watchdog does not consume the baton yet, which is accepted debt indocs/adr/0023-continuation-baton-makes-a-dropped-turn-observable.md(2026-08-25).
ADR digest
All 23 records live in docs/adr/. Statuses for 0001 to 0017 were
read from the files on 2026-07-02, and 0018 to 0022 on 2026-08-23.
| ADR | Title | Status | One-line takeaway |
|---|---|---|---|
| 0001 | Plugin Dependency Isolation | Accepted | No shared registry. Runtime filesystem detection with graceful degradation |
| 0002 | Extract QuotaTracker to Leyline | Accepted | Quota tracking moved from conjure into leyline for reuse |
| 0003 | Command Description Refactoring | Accepted | Two-part descriptions: short display line, rich identification text |
| 0004 | Skill Description Budget Optimization | Accepted (updated 2026-05-21) | 160-char cap. Budget is 2% of context, 16k fallback. Its 60k validator ceiling is stale: the enforcing script uses 90k |
| 0005 | Attune Plugin Discoverability Enhancement | Accepted (2026-02-05) | Superpowers-style trigger phrasing for attune components |
| 0006 | Self-Adapting Skill Health | Accepted (2026-02-15) | Homeostatic monitoring detects degrading skills |
| 0007 | GitHub Discussions as Agent Collective Memory | Accepted (2026-02-19) | Decisions and learnings posted to Discussions (GraphQL only) |
| 0008 | Behavioral Contract Verification Framework | Superseded (2026-03-15) | ERC-8004 blockchain dropped for cost. GitHub Attestations (SLSA) instead |
| 0009 | Sidecar Service Discovery via Port Files | Accepted | Sidecar daemons publish their ports through port files |
| 0010 | Stacked Diff Workflows | Accepted | git --update-refs plus gh pr create --base as zero-dependency stacking |
| 0011 | Shared Session-Capture Envelope | Accepted | One JSON envelope shape for friction and trace payloads, separate files |
| 0012 | Confidence-Tagged Agent Claims | Superseded by 0017 | VERIFIED/INFERRED/ASSUMPTION tags: no enforcement built |
| 0013 | Operationalizing Naur Theory-Building | Superseded by 0017 | Theory-building rituals: folded into 0017's lighter form |
| 0014 | Pensive Review-Skill Consolidation | Accepted (sequencing only) | 9 review skills, 5 sharing a verdict scaffold. Consolidation sequenced, not done |
| 0015 | Over-Built Orchestrator Skill Simplification | Accepted (data-collection phase) | Collect 30 days of usage data before simplifying over-built skills |
| 0016 | Wire-or-Archive for Three Orphan Skills | Accepted (decisions recorded) | Orphans judged on demand signal, wiring cost, and reference value |
| 0017 | Decisions on Confidence-Tagging and Theory-Building | Accepted (2026-06-18) | No enforcement mechanisms and voluntary use. Supersedes 0012 and 0013 |
| 0018 | Reuse memory-palace's Graph for the Tome Research Engine | Accepted | Tome builds on the existing graph instead of a second one |
| 0019 | Retire docs/superpowers/ and Record Its Shipped Design Decisions | Accepted | The design folder became ADR text; the folder is gone |
| 0020 | Detect a Thin Field with Positive Controls, Not Overlap Estimation | Accepted | Frontier detection uses positive controls, not overlap math |
| 0021 | PR Descriptions Carry Six Dimensions in Two Registers | Accepted (2026-08-12) | Six dimensions, but only some earn a heading |
| 0022 | The Autonomous Loop Continues by Riding the Stop Hook | Accepted, with a named open problem (2026-08-23) | Stop-hook re-injection is the loop. Cost bounded by stall detection; the reliance stays open |
When NOT to use
| You actually need | Use instead |
|---|---|
| How a change is classified, gated, and reviewed | night-market-change-control |
| The full story of a past incident or revert | night-market-failure-archaeology |
| Commands to run tests, lint, release, publish | night-market-operations |
| Every config file, default, and env var | night-market-config-catalog |
| Plugin/skill/hook mechanics in the abstract | claude-code-plugin-reference |
| Triage for a live failure | night-market-debugging-playbook |
Exit Criteria
- You can state why cross-plugin shared code is rejected, citing
ADR-0001 and commits
054e2679and29961cd2, and both commits resolve viagit log --oneline -1 <hash>. - Every enforcement file named in the Invariants table exists:
lseach path in the second column from the repo root with zero errors. - The ADR digest row count matches
ls docs/adr/*.md | wc -l(17 as of 2026-07-02). - For a proposed change, you can name which invariant row (if any) it touches and which weak point (if any) it worsens, before writing code.
- The version printed by
rg -m1 '"version"' .claude-plugin/marketplace.jsonmatches the version in every plugin'splugin.jsonyou spot-check.
Provenance and maintenance
Compiled 2026-07-02 against repo v1.9.15 on branch
discussions-fix-1.9.14. Skill/plugin counts drift: 198 SKILL.md files
under plugins/ and 23 marketplace plugins at compile time.
Re-verify volatile facts:
# ADR count and statuses
rg -m1 -n "Status" docs/adr/*.md
# Ecosystem version source of truth
rg -m1 '"version"' .claude-plugin/marketplace.json
# Discovery budget numbers (2%, 16k fallback; the ADR still says 60k)
rg -n "16,000|60,000" docs/adr/0004-skill-description-budget-optimization.md
# Enforced validator ceiling (90,000 as of 2026-07-03)
rg -n "DEFAULT_BUDGET" plugins/abstract/scripts/validate_budget.py
# Who orchestrates the gate federation today
rg -n "orchestrat" docs/quality-gates.md
# mdbook still unpinned?
rg -n "mdbook-version" .github/workflows/deploy-book.yml
# Ratchet baselines still present
ls scripts/skill_graph_baseline.json scripts/skill_exit_criteria_baseline.json
# Plugin and skill counts
python3 -c "import json;print(len(json.load(open('.claude-plugin/marketplace.json'))['plugins']))"
find plugins -name SKILL.md | wc -l