Shannon turns multi-stage agentic work — research, planning, implementation, validation, release — into a sequence of provable steps. Its Iron Rule refuses any PASS that can’t cite a real artifact on disk: no mocks, no stubs, no test files. Sub-agents carry their skills inline, so spawning never silently fails. The plugin verifies its own integrity with /shannon:doctor before it asks you to trust a run.
Shannon treats a multi-stage run as a chain of claims, and refuses to let any claim stand without cited evidence. The discipline is structural — enforced by hooks and a mechanical self-check — not a prompt you hope the model remembers.
scripts/doctor.py enforces the contract mechanically.skills/<name>/SKILL.md, but each of the 11 agents ships with its skills inlined at build time by build/embed-skills.py. A spawned sub-agent can’t fail to load its own skill — the silent-invocation-failure mode is designed out, not hoped against./shannon:doctor runs a 10-check mechanical contract over the manifest, counts, hook registration, embed state, and body references. /shannon:audit reviews screens, apps, sessions, and drift. Shannon is self-instrumented — it proves its own integrity before it asks you to trust a run.Shannon ships as a Claude Code plugin. Add the marketplace, activate the hooks in your project, verify with /shannon:doctor, then drive a plan-then-execute run that captures evidence at every gate. Two examples below.
# 1. Add the marketplace and install the plugin. $ claude plugin marketplace add krzemienski/shannon $ claude plugin install shannon@shannon # 2. Activate Shannon's hooks for this project, then verify. $ /shannon:enforce on $ /shannon:doctor # plugin-manifest ... skills-count ... agents-count ... → 10/10 checks PASS · 0 mismatches
# Author a validation-gated plan, then execute it under the Iron Rule. # --mode picks the strategy: linear (default) · converge · tournament · deep $ /shannon:plan "Add rate limiting to the login route" --mode deep # → plans/<date>-<slug>/plan.md + phase-NN.md (gate per phase) $ /shannon:cook plans/<date>-<slug>/ # Each phase: build → exercise the real system → capture evidence → verdict → evidence/ on disk · cited PASS/FAIL · refusal if proof is missing
Every phase of a Shannon run passes through the same four-step functional-validation gate. A phase that can’t produce a cited artifact does not advance — the run refuses rather than fabricates a green check.
The pillars are not aspirations — each is a mechanism on disk. Embedded skills are built by a script, the Iron Rule is enforced by hooks, the meta-judge writes its rubric before it judges, and the self-check is a Python contract you can run yourself.
Grounded in the on-disk plugin tree at github.com/krzemienski/shannon and the live /shannon:doctor 10-check mechanical contract.
Install as a Claude Code plugin from GitHub or a local clone. Shannon is standalone — no required MCP servers. Context7 and sequential-thinking pair well but are optional.
# Option A — from the GitHub repo $ claude plugin marketplace add krzemienski/shannon $ claude plugin install shannon@shannon # Option B — from a local clone $ claude plugin marketplace add /path/to/shannon $ claude plugin install shannon@shannon-local # Activate + verify, then run your first gated workflow. $ /shannon:enforce on $ /shannon:doctor → 10/10 checks PASS · plan with /shannon:plan, execute with /shannon:cook
No SaaS, no telemetry, no closed source. The skills, agents, hooks, harnesses, and the doctor contract all live in the public repository. Eight pages worth your time on day one.
A green check that cites nothing is a guess. Shannon refuses the guess — and refusal is a feature.
Shannon’s v1.2.0 grows the surface to 36 skills, 11 agents, 22 commands, and 7 hooks while still passing its own mechanical contract (/shannon:doctor 10/10). The v5/v6/v7 prototypes are preserved as provenance.
/shannon:doctor still 10/10. Adds DeepPlan + wave-execution (dependency-ordered parallel waves with between-wave synthesis gates), /shannon:plan-deep as a true alias of plan --mode deep, the evidence-gated forge pipeline (analysis → docs → plan → oracle review → execute → validation → indexing → 3-reviewer consensus → 3-oracle quorum → completion-gate), and the skills10x activation harness with transitive registration./shannon:doctor (10/10 mechanical checks PASS). Skill set curated to 33, agents to 9, commands to 20, hooks to 7 with a required_hooks contract. Embedded sub-agent skills via build/embed-skills.py — sub-agents can’t fail to load. Standalone: no required MCP servers.