Claude Code plugin for high-stakes multi-stage agentic work. Iron Rule: real-system evidence on disk, refuses verdicts without proof. Embedded sub-agent skills — sub-agents carry their skills, can't fail to load. Standalone — no required MCP servers.
Version 1.0.0 — first stable release. The five pillars are on disk and self-verified by /shannon:doctor.
# In Claude Code — install from the GitHub repo:
/plugin marketplace add krzemienski/shannon
/plugin install shannon@shannon
# Or from a local clone:
/plugin marketplace add /path/to/shannon
/plugin install shannon@shannon-localA Claude Code plugin that turns multi-stage agentic work — research, planning, implementation, validation, release — into a sequence of provable steps.
Five pillars:
- Embedded sub-agent skills. Sub-agents carry their skills inline; spawning is reliable.
- Orchestration. Single-message multi-Task dispatch. Sequential, parallel, competitive patterns.
- Iron Rule validation. Real-system evidence on disk. No mocks. No stubs. No test files.
- Meta-judge consensus. Rubric YAML generated before any judge runs. Hidden thresholds. Debate on disagreement.
- Self-instrumented.
/shannon:doctorand/shannon:auditwork — the plugin observes itself.
- Not a CLI tool. Plugin-only.
- Not an MCP framework. Uses MCP when present; degrades gracefully when absent.
- Not a replacement for Claude Code. Runs inside it.
- Not an IDE. Cursor/Cline/Continue.dev cover that surface.
Optional but pair well:
- Context7 — fresh library/API documentation
- sequential-thinking — systematic problem decomposition
Anything else: ship as research surfaces a need.
# 1. Install (from the published GitHub repo)
/plugin marketplace add krzemienski/shannon
/plugin install shannon@shannon
# 2. Activate in your project
/shannon:enforce on
# 3. Verify
/shannon:doctor
# 4. First workflow
/shannon:plan "Your feature here"
/shannon:cook plans/<date>-<slug>/Full guide: docs/QUICK_START.md.
| Command | Purpose |
|---|---|
/shannon:plan |
Plan author (linear / brownfield / converge / tournament / deep) |
/shannon:scope |
Brownfield reconnaissance — codebase + skill inventory + session context |
/shannon:cook |
End-to-end execution with iron-rule validation |
/shannon:autopilot |
Refusal-driven retry loop around cook |
/shannon:fix |
3-strike bug-fix runner |
/shannon:loop |
do/verify/reflect convergence loop |
/shannon:dispatch |
Sub-agent dispatch (sequential / parallel / competitive) |
/shannon:team |
Multi-teammate orchestration with file-ownership boundaries |
/shannon:validate |
Functional validation with cited evidence |
/shannon:audit |
Read-only audit (screen / app / session / drift / completion-evidence) |
/shannon:research |
Parallel researcher fan-out |
/shannon:prd |
Interview-driven PRD authoring |
/shannon:reflect |
Self-refinement (self / critique / memorize) |
/shannon:why |
Five-whys root-cause analysis |
/shannon:trace |
Session JSONL timeline |
/shannon:retro |
Mine sessions for retrospectives |
/shannon:resume |
Resume halted runs from evidence tree |
/shannon:enforce |
Toggle Shannon hooks (on/off) for current project |
/shannon:install |
Atomic install + activation |
/shannon:doctor |
Mechanical contract check |
Five layers:
- Skills (33) — canonical at
skills/<name>/SKILL.md+references/*.md(progressive disclosure) - Agents (9) —
agents/<name>/AGENT.mdwith embedded SKILL.md content inlined bybuild/embed-skills.py - Commands (20) —
/shannon:*entry points routing to agents/skills - Hooks (7) —
hooks/*.jswithrequired_hookscontract enforced byscripts/doctor.py - Validation harnesses (2) —
scripts/harness/sdk_runner.py+tmux_runner.py
Full breakdown: docs/ARCHITECTURE.md.
v1.0.0 — stable. All five layers are on disk and pass the mechanical contract in
scripts/doctor.py (10/10 checks). The build path that produced v1 is preserved below as
provenance.
| Milestone | Status | Output |
|---|---|---|
| PRD approval | ✅ | PRD-V1.md + DECISIONS.md |
| Skill curation (67 → 33) | ✅ | skills/ (33 dirs) |
| Agent embedding | ✅ | agents/ (9 manifests + AGENT.md + _built/) |
| Command consolidation (27 → 20) | ✅ | commands/ (20 dirs) |
| Hook curation (16 → 7) + contract | ✅ | hooks/ (7) + scripts/doctor.py |
| Validation harnesses | ✅ | scripts/harness/ (SDK + Tmux) |
| Documentation | ✅ | README + 7 docs/ files |
| Brownfield planning | ✅ | /shannon:scope + skill-inventory + --brownfield modes |
| v1.0.0 release | ✅ | doctor 10/10 PASS; counts reconciled to disk (33/9/20/7) |
Built on lessons from the v5/v6/v7 prototypes shipped in this repo and from Anthropic's ecosystem (anthropic-skills, claude-code-guide, sadd, oh-my-claudecode). The PRD-V1 reframe absorbed the productive ideas and dropped the rest.
MIT.
- docs/INSTALL.md — platform-specific install
- docs/QUICK_START.md — 10-minute path to first workflow
- docs/SKILLS_CATALOG.md — the 33 skills
- docs/ARCHITECTURE.md — 5 pillars + 5 layers + Architecture C internals
- docs/FUNCTIONAL_VALIDATION_GUIDE.md — Iron Rule + harness usage
- docs/CONTRIBUTING.md — how to add a skill/agent/hook/command
- docs/DEV_HARNESS.md — SDK + Tmux harness runbook