The notrest plugin marketplace by Not Rest Inc..
/plugin marketplace add notrestai/notrest
/plugin install notrest@notrest
Renaming from oracle-suite? Same harness, new id — the plugin was renamed oracle-suite → notrest
(the marketplace is still notrest, so the install id is now notrest@notrest and skills invoke as
/notrest:<name>). Install notrest as above, then drop the old entry:
claude plugin uninstall oracle-suite
The session harness — the scaffolding a working session runs inside: discipline auto-anchored into
every session, delegation as an arrangement (agentswarm) with a spend ledger that makes the model-routing
policy checkable, an append-only estate that records every prompt and every agent, continuity + successor
escort across sessions, and an eval suite the harness runs against its own rules. Twenty-eight skills ride
on it — structured thinking from intake to handoff, built on three principles: token-lean (progressive disclosure, --quick modes, built-in search budgets), verified (real tiered sources, disconfirmation passes, honesty labels on every claim), and model-agnostic (the discipline lives in the skills, so any model runs them reliably):
| Skill | What it does |
|---|---|
| oracle | Session intake — six setup questions (Objective, Role, Architecture, Content, Leverage, Evaluation) + loads/creates the CLAUDE.md foundation. Invoke with "hey oracle" or /oracle. |
| researcher | Rigorous multi-pass research → a background doc + a decision dossier. |
| marketresearcher | Funnel-shaped market research → sizing, competitors, whitespace, an opportunity dossier. |
| stepbystep | Turns a goal/docs into a stress-tested, converged, ordered action plan. |
| actionplan | Expands a stepbystep plan into a copy-paste-ready runbook (exact commands per host). |
| critic | Adversarial red-team of any document — steelman, attack, alternatives, severity-tiered verdict. |
| explainer | Layered understanding of any topic — a correct mental model, plain → working → expert layers, the standard misconceptions, verify-it-yourself checks. |
| decider | Structures any decision — weighted criteria, evidence per option, sensitivity ("which assumption flips the winner"), pre-mortem, recommendation with reversibility framing. |
| factcheck | Claim-by-claim verification against real sources — ✅ CONFIRMED / 🟡 PLAUSIBLE / 🔴 REFUTED / 🔵 MISLEADING / ⚪ UNVERIFIABLE, with daisy-chained sources detected. |
| watch | Facts have shelf lives — the calendar-time sibling of factcheck. Pulls a dossier's load-bearing [cited] claims into watch/watchlist.md (verbatim claim · source · tier · first-verified · last-checked · status · cadence), re-verifies the due ones with the same rigor, and appends a dated DRIFT REPORT to watch/drift-log.md — HOLDS / DRIFTED / DEAD-SOURCE / UNVERIFIABLE, every run stamping the URLs it actually fetched. A dead source is never mistaken for a refuted claim, drift is never smoothed, and where a scheduled-tasks MCP exists it offers to run the cycle on a cadence — owner-confirmed, never faked. |
| introspect | Validated workspace self-reports — a black-box instrument inspired by Anthropic's J-space research: two-layer sealed snapshots (scored tokens + readable glosses) scored against subsequent behavior (verbalized / silent / turnover) with predictive lift vs a context-only control agent. Ships with a real example ledger. |
| game-forge | Complete playable games from a short request — fixed-timestep engine templates (browser + pygame), genre playbooks, juice/audio references, and a hard playtest gate: no game ships unrun. |
| gpt | Chat-first GPT lane (5.6 family, Codex CLI on your ChatGPT account): /gpt just talks to a persistent conversation after a 2-question guided setup — thinking level + how agentic. --once one-shots, --task background jobs with verified deliverables, --vs blind comparisons, --here repo-aware mode, and a no-questions --setup path for orchestrators. Opinions labeled, never treated as sources. |
| chatroom | Shared rooms for AI sessions — any Claude session joins via a tiny script (append-only room file, flock-atomic posts, armed watches as wakes), GPT joins through a persistent-memory bridge (gpt-bridge --once, mention-triggered, throttled). Local-machine v1; strict no-secrets rule. |
| archivist | Content continuity — scans every ORACLE output folder into one greppable oracle-index.md (title, date, path, each dossier's Read Me First), so "what do we already know about X?" costs one find instead of a re-spent search budget. Consult-before-spending; the index is a finding aid, never a source. |
| graph | An Obsidian-style file graph of any project — a python3-stdlib scanner walks the repo (.gitignore respected) and renders a self-contained, force-directed HTML view of every file and the references between them (wikilinks, markdown links, imports, sourced scripts, estate pointers, agent transcripts), at zero model tokens. register projects and all merges them into one cross-project map — the PM view: one session holding every project at once. |
| spend | The token receipt — an append-only spend/ledger.md of every observed model spend (subagent counts, workflow totals, gpt tokens-used), graded observed/estimate; report prints the per-model split and exits 4 if Fable ever rode below the seat — the routing rule, made checkable. |
| doctor | The harness's self-check — one read-only pass, eight named PASS/WARN/FAIL checks, a fix command on every finding: SKILL.md front-matter a real YAML load would reject (the unquoted ": " that makes a skill invisible while the file sits on disk), manifest + tombstone version pins, skill-count drift across the four places the number is spelled as prose, hook scripts that exist and clear bash -n plus rename residue in the SessionStart echo, estate integrity (COORD header, agent ledger, spend.py report, compile candidates), repo-vs-clone-vs-installed drift, unanchored gitignore rules that silently un-track a whole skill dir, and stale version stamps on rendered surfaces. Every check is a scar from a real defect. Exits 0 / 5 warnings / 6 any fail, --json for machines, --plugin to inspect what a session is actually running. Reads only — never repairs, never bumps, never commits. |
| eval | The harness's law-conformance suite — one static pass over the shipped files asking whether every law the suite claims actually left a fingerprint in the text: the offload policy (explicit opus, no sonnet/haiku, fork ban wherever subagent_type is written, the rule echoed by SessionStart), honesty labels on every claim-making skill, scanners that exist and py_compile behind every zero-token claim, append-only estate ledgers, the worker contract (self-check + finishing up), front-matter YAML accepts with a real /slash trigger, the named safety laws (a draft is never sent · a dead source is never a refutation · never auto-install), and hooks that are silent-on-failure. Doctrine: a law that is well-encoded leaves a static fingerprint — check the fingerprint, not the behavior. ~0.1s, zero model tokens, exits 0 / 5 / 6, --json; every FAIL cites the file:line and a fix. behavior --case is a bounded, opt-in, code-graded (never LLM-judged) model path, deliberately outside the ship gate. doctor checks the install; eval checks the laws. |
| recap | The trail, turned into understanding — the read side of the estate. Walks COORD.md, COORD-AGENTS.md, git, spend/ledger.md and the dossiers in timestamp order and delivers the project's decision story: a narrated timeline, a who-was-consulted table (transcripts existence-checked), ships, costs, open threads — plus a self-contained clickable decision map (nodes = rulings / decisions / consultations / ships; edges = led-to / informed-by / reversed). Every claim carries its trail citation; conflicts between sources are surfaced, never smoothed. It derives — it never invents. |
| draft | The suite's outbound verb — every other skill ends at "you now know / you decided"; draft turns that dossier into the thing you send (email · exec memo · slack update · one-pager · status update, each a skeleton with a hard length budget). One law makes it trustworthy: [fact] vs [framing] — every factual sentence traces to a source line in a bottom-of-background source-map and keeps its honesty label, while framing choices (what to lead with, what to cut, what tone) are listed as choices. Persuasion never upgrades a label: [estimate] stays hedged in the final prose, [unverified] drops or carries its hedge — never becomes fact. Numbers never invented, quotes never manufactured, recipients never guessed. And a draft is never sent: sending is the owner's act in the owner's client, and a draft file is never evidence a message was delivered. |
| refuter | The harness's adversarial reviewer for executables — the contract that stops "have a lane check the work" from being re-improvised in every prompt. A lane other than the builder attacks ONE narrow target under a written brief (artifact inline, isolated scratch dir, hard prohibitions, a numbered attack ladder specialized per target, ~12 tool calls) and answers in a verdict grammar with teeth: CONFIRMED needs an exact command and pasted output, PLAUSIBLE needs inputs → state → wrong outcome and is explicitly not evidence, and a finding with neither is deleted before returning. The ladder runs irreversible-path safety → integrity invariants → claim honesty → partial-failure states → test honesty → environment assumptions; severity ranks breaks-irreversible-safety > breaks-claim-honesty > degrades > cosmetic. It also reports what SURVIVED, by name. Blast-radius tiered (full refuter for anything that pushes/installs/writes the estate, grep-and-read for docs, nothing for cosmetic), review-the-fix always goes to a different lane, repairs cap at two rounds — and the refuter never fixes what it finds and never grades its own prior findings. |
| director | Chains the other skills into a pipeline (e.g. marketresearcher → critic → stepbystep). |
| fable-director | Seats and runs the "3 DEVS AND A RELAY" multi-session dev arrangement — a metered Fable/Opus director that applies all code, flat Opus dev/QC lanes handing EDIT SPECS, per-lane blackboard files with token-watch wakes, a QC refuter relay, and a self-carrying rotation ritual. Bundles the V4 protocol of record + a repo scaffolder. Invoke with /fable-director. |
| compile | The estate's fourth verb — move repeated work into code. A python3-stdlib scanner mines COORD.md, COORD-AGENTS.md and spend/ledger.md (optionally session transcripts) for the same job done three or more times and ranks the candidates at zero model tokens; a seat-run ritual then reconstructs that job's functional contract from the trail with recap-grade citations, an Opus builder lane emits an isolated runtime under compile/<slug>/ (runner · typed schemas for the model calls that survive · validators · replay fixtures), an independent refuter attacks it, and the seat runs a fair benchmark against the history it came from. Cheaper-but-worse is a failed compile; every metric carries observed/estimate/unavailable provenance; the one-time compilation cost is reported separately and never amortized; nothing installs itself — shipping a compiled runtime is a normal versioned release. |
| agentswarm | The fast delegation arrangement — the seat (the main session's model: Fable or Opus alike) keeps decompose, judge, apply all edits, gate ships, and offloads everything else to concurrent in-session Opus lanes; the harness is the wire (no blackboards, no watches, no rotation — deaf lanes and split-brain deleted by construction). Owner model policy: every offloaded job runs on Opus. QC refuter runs as code after every finding; spend-ledger receipts per lane; archivist consult before fan-outs. Use fable-director only when lanes must outlive the machine or span accounts. |
| sessionend | Captures session state into START-HERE / HANDOFF / STATE / CLAUDE.md so the next session resumes seamlessly. |
| fable-mode | The Fable behavioral contract: the ORIENT → PROBE → ACT → PROVE → BANK loop, hard rules — including the offload model policy (Fable never rides in a subagent: every offloaded job runs on explicit Opus, never inherited Fable credit; delegation via agentswarm) — THE FABLE DIFFERENCE (instinct → fable move), a consumer-side verification cookbook, the outage playbook (reroute, stage, never stall), tool-graph craft, and situational profiles. Auto-anchored every session by the SessionStart hook; full contract via /fable-mode. |
oracle + sessionend are bookends: ORACLE loads the foundation at the start, sessionend updates it at the end — together they make sessions continuous. fable-mode is the posture in between: the SessionStart hook injects its discipline anchor every session, so reliable working habits are on by default without anyone typing /fable-mode.
- SessionStart — self-updates the plugin from git (fire-and-forget
--ff-only; note: marketplace-cache installs aren't git clones — there the update path isclaude plugin update notrest@notrest); injects the fable-discipline anchor AND the offload model policy hard rule (Fable sessions offload every spawned job on explicit Opus — never sonnet/haiku, never inherited Fable — with agentswarm as the default delegation arrangement) every session; auto-createsCOORD.md— the session coordination ledger (the fable-coord behavior, generalized) — at any git-repo root and nudges reading its tail when it exists; detectsSTART-HERE.md/HANDOFF.mdand adds a resume nudge; detects per-laneCOORD-<LANE>.md(or legacyFABLE-COORD*.md) blackboards and nudges seating the fable-director. - UserPromptSubmit — when
COORD.mdexists, injects a one-line reminder on every prompt: append one honest ledger line when this prompt's work lands (ask → landed | evidence). - PreCompact — reminds that a deliberate
/sessionendhandoff preserves more than automatic compaction. - SubagentStop — when a spawned agent finishes inside a git repo, auto-writes one machine-written line (id · model · last conclusion · transcript path) to
COORD-AGENTS.mdat the repo root, so the session's decision pattern — which agents were consulted and what each concluded — lands in the repo at zero model-token cost (the archivist indexes it; entries point at the full transcripts).
docs/TUTORIAL.md — install → the shape of every skill → your first full loop → which-skill-when → chains worth knowing.
docs/oracle-skill-flow.html — a one-page, self-contained diagram of the oracle intake (foundation & resume → the six questions → routing → the continuity loop). GitHub shows HTML as source — open the file locally in a browser, or via GitHub Pages if enabled for this repo.
MIT © 2026 Not Rest Inc.