notrest

Expand a stepbystep plan into a copy-paste runbook — exact ordered commands per host, a verify + rollback on every step, ⛔ warnings before destructive ops, placeholders instead of invented environment specifics (reads an optional map.md). Use on /actionplan or asks to "turn this plan into commands", "make this copy-paste", "write the runbook", or "give me the exact steps/code". If no stepbystep plan exists yet, suggest /stepbystep first.

The delegation arrangement — the seat keeps decompose/judge/apply/gate and background lanes do the work. MODEL POLICY is runtime-explicit: Codex uses gpt-5.6-sol; Claude uses opus; inherited/full-history forks are forbidden. Builds run one persistent lane per domain, resumed for feedback. Use on /agentswarm, 'swarm this', 'offload this', or explicit delegation requests.

The suite's findings store — skills emit compact FINDING RECORDS into one append-only archive/findings.jsonl, validated at the door (schema, enums, honesty labels, real URLs behind cited links); archivist keeps the session's whole track and still indexes legacy dossiers into oracle-index.md. Use on /archivist, "what do we already know about X", "show me the session track", "index the dossiers", or before a researcher / factcheck / decider run.

Checkpoint in-flight agent lanes and, where remote isolation exists, respawn the remaining work in the cloud; beam down folds results home. On Codex v4.3, checkpoint/status are supported but Claude remote-Agent respawn is unavailable and must be labeled. Use on /beam, beam up/down/status, or 'I have to leave'.

Shared rooms where AI tasks/sessions and a GPT bridge chat and work together. Use on /chatroom, "create a chatroom", "join the chatroom {name}", or "let claude and gpt talk to each other". NO SECRETS ever — the bridge sends room content to another vendor. Local to this machine in v1.

Compile repeated work into code — a script mines the estate for the same job done three or more times, then a ritual builds an isolated runtime under compile/{slug}/ and fair-benchmarks it against that history. Use on "/compile", "compile this workflow", "what should we compile", "workflow compiler", "make repeat work cheaper". Runtimes stay isolated until the owner ships them.

Adversarial review of any document, plan, argument, or prior dossier — steelman first, then severity-tiered objections (🔴 fatal / 🟠 serious / 🟡 minor), a disconfirmation pass, ≥3 genuine alternatives, and a fair verdict — producing a background doc + critique dossier (--quick for chat-only; --panel for multi-lens on high-stakes docs). Use on /critic or asks to critique, red-team, poke holes in, stress-test, tear apart, challenge, or play devil's advocate.

Structure any decision — options (always incl. do-nothing), weighted criteria, evidence per option, a scored matrix with a sensitivity hinge ("which assumption flips the winner"), a pre-mortem on the front-runner, a recommendation with confidence + reversibility — landing a validated DECISION RECORD in the archivist store (or --quick). Use on /decider or "should I…", "help me decide/choose", "X or Y", "compare these options", "what are the tradeoffs". The user decides; this structures.

Chain ORACLE skills into a pipeline — parse the ordered list + seed, run every stage for real (each SKILL.md read from disk; isolated subagent per stage), hand each output to the next, produce a numbered run folder + summary (or --quick). Use on /director or asks to run skills in sequence, chain, pipeline, "do X then Y then Z". For orchestrating SESSIONS instead, see fable-director.

The harness's own self-check — one read-only pass, PASS/WARN/FAIL per check with the exact fix: front-matter YAML would reject, manifest + tombstone pins, skill-count drift, hooks that parse and ever fired, estate integrity, which build is really running, the always-on token budget, gitignore rules that swallow skills, stale stamps. Use on "/doctor", "health check", "check the install", "why isn't X triggering". Reads only.

The suite's outbound verb — turns a dossier, decision, or recap into the thing you actually send (email · memo · slack update · one-pager), every factual sentence traced to its source line, honesty label kept. Use on /draft, "write the email", "draft the memo", "turn this into an update", "make this sendable", "write it up for {audience}"; --quick for chat-only. A draft is never sent — sending is the owner's act. Not for legal or regulatory notices, nor before the decision exists.

The harness's law-conformance suite — one static pass asking whether every law left a fingerprint in the shipped files: offload policy, honesty labels, scripts that compile, cited files that exist, append-only ledgers, worker contracts, safe front-matter, safety laws, silent hooks, the routing enforcer. Use on "/eval", "check the laws", "conformance check", or before any release. Zero tokens, seconds. doctor checks the INSTALL; eval checks the LAWS.

Build genuine understanding of any topic, system, or document — a correct mental model, three depth layers (plain → working → expert), the standard misconceptions, and how to verify the load-bearing claims yourself — landing finding records (or --quick for the chat-only mental model). Use on /explainer or asks to explain, "help me understand", "what is X really", "break this down", "ELI5", or "teach me X". For real understanding-building — answer quick factual questions directly.

Seat the "3 DEVS AND A RELAY" arrangement — a metered director session running dev/QC lanes on per-lane blackboards. Use on "/fable-director", "fable director", "3 devs and a relay", "stand up the fable arrangement", the rotation trigger ("context getting full — rotate"), a ROTATION HANDOFF / FABLE KICKOFF, or when the repo holds COORD-{LANE}.md blackboards. Not `director` (that chains SKILLS; this runs SESSIONS). Not for ordinary tasks.

Operate this session under the Fable discipline — probe before believing, prove at the consumer before claiming, bank before stopping, reroute instead of stalling when tools degrade. Use when the user says "/fable-mode", "fable mode", "work like fable", "be fable", "fable discipline", or asks for "2x reliability" / "prove it like fable". A discipline contract, not a capability upgrade. Not for ordinary single-question turns.

Verify claims against real sources — extract the load-bearing claims, hunt primary sources (two independent for CONFIRMED; daisy-chains collapsed), claim-by-claim verdicts ✅ CONFIRMED / 🟡 PLAUSIBLE / 🔴 REFUTED / 🔵 MISLEADING / ⚪ UNVERIFIABLE with dated citations — one validated FINDING RECORD per claim in the store (or --quick). Use on /factcheck or "is this true", "fact-check / verify / double-check this", "did X really say/do that".

Build a complete, playable game from a short request — no game ships unrun. Use on "/game-forge" or any ask to make, build, generate, prototype, or vibe-code a game: arcade, platformer, shooter, puzzle, tower defense, snake/breakout/tetris, idle/clicker, roguelike, "a game like X but Y", browser or pygame. Fires on casual asks ("something fun to play"). NOT for game reviews, buying advice, lore, or non-interactive animations.

Chat-first GPT lane — /gpt talks to a persistent GPT-5.6 conversation (Codex CLI). Flags when asked: --once, --task, --vs, --new, --setup. Use on /gpt, "ask gpt", "chat with gpt", "continue the gpt chat", "leave gpt a task", "gpt second opinion". Opinions never sources ([model-opinion]); prompts leave the machine to OpenAI — no secrets, ever.

Two script-built views, self-contained HTML at zero model tokens, then OPENED: the FILE GRAPH (every file and reference — Obsidian for a codebase) and the RIVER (findings.jsonl + COORD volumes as a river toward the goal — side channels, backtrack loops, conflict rocks, milestone flags). Plus text queries over the last scan. Use on /graph, cockpit, "project graph", "/graph river", "what links to this file", "orphans", "stale files", "all projects graph".

Validated workspace self-report — the model snapshots the concepts most active in its thinking, and the harness SCORES those reports against subsequent behavior. Use on /introspect, "what are you thinking right now", "snapshot your workspace", "j-space check", or "run an introspection experiment". It does NOT observe internal activations.

Funnel-shaped market research — scope, two-way sizing (top-down + bottom-up, reconciled), competitor map with the graveyard, gap/whitespace verdicts, feasibility scoring, drafted entry ideas — landing finding records (--dossier keeps the card; --quick). Use on /marketresearcher or asks to size a market, map or analyze competitors, find gaps/whitespace or a niche, assess market entry, or validate a business/product idea. Research to inform a decision — never investment advice.

Seat the mentor-dev ritual: two PEER sessions on one chartered room — escort the builder, rule, gate their checkpoints, take correction back in public. Use on "/mentor", "mentor me", "be my mentor", "mentor the build", "gate my work". Not a lane that ends (agentswarm), not a metered blackboard session (fable-director), not a one-off question.

Establish the harness in a project, or CONTINUE the build already running there — a successor seat reads the trail packet, asks the previous session as its mentor, verifies at tier 0 and goes. Use on "/notrest", "continue the build", "pick up the build", "establish the harness", "set up the plugin in this project", "make this project follow the plugin", or the drift check "is this session following notrest?". `check` and `continuation` read only; never runs git init uninvited.

Session intake + resume — the suite's front door. On "hey oracle", /oracle, "start an oracle session", or "resume / pick up where we left off": load the runtime foundation (AGENTS.md on Codex, CLAUDE.md on Claude), offer resume from START-HERE.md, ask the six ORACLE questions, route to the suite skill that fits, then proceed — scaffolding the foundation if none exists. Not for ordinary questions.

Walks the recorded trail — COORD.md, COORD-AGENTS.md, git history, spend/ledger.md, the dossier folders — in timestamp order and delivers the decision track: a narrated timeline, a who-was-consulted table, ships, costs, and a clickable decision map (HTML). Every claim cites a trail line, a commit, or a path; anything without one is [unverified]. Use on /recap, "recap the project", "how did we get here", "decision history", "what happened here".

The adversarial reviewer for code and load-bearing artifacts — a lane other than the builder attacks ONE narrow target under a brief: CONFIRMED findings with a reproduction, PLAUSIBLE ones with a failure scenario, and what survived. Use on /refuter, "refute this", "attack this before we ship", "QC that lane's work", "red-team this", or as the agentswarm gate before shipping code that pushes, installs, or writes the estate. It finds, never fixes; `critic` is the document-side sibling.

Rigorous multi-pass research on any question — baseline, ≥5 alternatives, tiered real sources, comparison, disconfirmation — landing validated FINDING RECORDS in the archivist store (or --quick for chat-only). Use on /researcher or natural asks to research, deeply investigate, compare alternatives, find best practices, or "find the best X for my situation" — any open-ended question that deserves evidence over vibes. Not for simple lookups you can answer directly.

Capture task/session state into four continuity files — START-HERE.md, HANDOFF.md, STATE.md, and the runtime foundation (AGENTS.md on Codex, CLAUDE.md on Claude; merged, never clobbered) — verify a memoryless successor can resume, then open a live line when the host exposes one. Use on /sessionend or asks to wrap up, save state, or write a handoff.

The token receipt — an append-only spend/ledger.md logging every observed model spend (subagent fan-outs, gpt-lane calls, workflow runs) so the offload routing rule is CHECKABLE instead of merely asserted. Use on /spend, "log the spend", "spend report", "token report", "audit the model routing", or at the close of any session that spawned agents. Entries graded observed or estimate; the main loop's own totals aren't visible to the model and the ledger says so.

Turn a goal + documents into a dependency-ordered, stress-tested action plan with a per-step "done when" check and honest confidence scores, refined by a research→critique loop until it converges (hard cap 5 rounds) — landing a decision record + [ONE-WAY] findings (or --quick). Use on /stepbystep or asks for a step-by-step / action plan, "how do I do X", a checklist or playbook, or "turn this into a plan I can execute" — technical or not.

Facts have shelf lives — watch re-verifies them on schedule: pulls the load-bearing [cited] claims out of a dossier (or inline) into watch/watchlist.md, then re-verifies the due ones with factcheck's rigor and appends a dated DRIFT REPORT to watch/drift-log.md — HOLDS / DRIFTED / DEAD-SOURCE / UNVERIFIABLE, with what was actually fetched stamped on every run. Use on /watch, "watch this claim", "keep this fresh", "recheck weekly", "has anything changed since".