MY Plugin marketplace for Claude Code. The agents here were created from something I stumbled across by accident, wrapped as a solution.
The Spector sub-agent came from me trying to articulate to Kimi K2.5 how a dashboard with an endless combination of settings was manipulated by an end user, why they'd want certain settings, and what rules governed whether or not the combination of settings was valid. It was one of those things where you'd look at the area of the UI and know instantly how it worked, but couldn't explain it. Kimi looked at my screenshots and the bug in the code and was able to describe it with comically dense narrative, later improved by Claude. This agent tries to extend my experience with that to an entire code base. (this one needs a little work.)
The Paranoid E2E Architect sub-agent was borne from a frustration that Claude could generate a lot of common sense unit tests that covered 65-70% of how a complex application was supposed to work, but completely whiffed on the need for tests many developers with an eye toward excellent UX would catch. Language models are good at validating a spec was implemented and functions with the spec as a guide. If it wasn't defined in the spec, the model isn't going to manufacture it out of thin air. That's on you, the human. This sub-agent attempts to close the coverage gap between what the model knows needs to be tested and the mutations in an application that aren't in the spec, but a human would ID.
High-density functional spec generation from code. Three-layer pipeline:
- Scope discovery — identifies modules, groups files, estimates scale
- Parallel capture (Sonnet 4.5) — structured behavioral inventories per module
- Sequential compression (Opus 4.6) — dense functional specs at 20:1–50:1 compression
claude /plugins add-marketplace briankelley/claude-marketplace
claude /plugins install spector# Spec a single file
/spector path/to/module.py
# Spec an entire directory
/spector src/
# Spec current working directory
/spector
# Custom output path
/spector src/ docs/my-spec.mdDense functional specifications where every token carries semantic load. Behavior-first, state-machine-oriented, zero filler. A 1000-line module compresses to 20–50 lines of spec.
See spector/examples/config-page-iconography.md for the gold-standard reference output.
| Component | Model | Role |
|---|---|---|
/spector skill |
— | Orchestrator: scope discovery, dispatch, assembly |
spec-scanner agent |
Sonnet 4.5 | Broad capture: reads code, emits structured inventory |
spec-compressor agent |
Opus 4.6 | Precision compression: inventory → dense spec |
Adversarial E2E test layer that hunts for unguarded destructive paths in web applications. Builds a registry of every state-mutating operation, fuzzes every permutation of empty/default/boundary input against prHomee-existing data, and enforces that no write path exists without a corresponding guard. Does not verify happy paths — other tests handle those.
claude /plugins install paranoid-e2e-architect# Launch against a project
/paranoid /path/to/project
# Or invoke the agent directly via Task tool- Destructive Action Inventory — crawls every page, catalogs every write trigger, cross-references against the Write Registry
- State Snapshot Diffing — hashes files before/after write actions, flags any mutation from no-op submissions
- Adversarial Prompt Templates — optional LLM-driven threat analysis of each page's interactive elements
- Permutation Fuzzing — systematically tests every registry entry × 10 input strategies (empty, boundary, type confusion, etc.)
- Loss Annotation Policy Tests — enforces that irreversible/unbackupable operations require confirmation
A complete paranoid test suite: paranoid_helpers.py (registry + utilities), test_paranoid_inventory.py, test_paranoid_snapshots.py, test_paranoid_fuzzing.py, test_paranoid_loss_annotations.py, test_paranoid_adversarial.py, plus conftest fixtures with pre-seeded state.
See paranoid-e2e-architect/references/dispatch-template.md for the structured dispatch format.
MIT