consensus-rnd

English canonical public documentation. Chinese companion.

consensus-rnd publishes Agent Skills for Claude Code, Codex, Cursor, and Gemini. Its sshx skill brings isolated worker perspectives to an authorized decision or implementation task, converges on a concrete plan, and checks the result through independent review. The deliverable is a prompt contract and its supporting files, shared across platform manifests.

Activate it only for an explicit request for consensus, multi-angle thinking or delegated review, or an applicable host/repository mandate. Risk alone and a generic architecture review request do not activate it; within authorization, weigh whether separate perspectives justify the cost. See the trigger and exclusions.

What It Provides

skill	What it is for	Runtime shape
sshx	Worker-delegated inline consensus for a decision, plan, or finished change.	Prompt contract with worker runners; no skill-owned daemon, GitHub, git, label, release, or lifecycle orchestration authority; optional heca MCP delegation inside heca.

Core

The engine is not "run the same prompt several times and vote." It is biased, independent, multi-angle convergence:

• Biased independent perspectives: thinking and review workers start from different priors and must not read same-round peer output before sealing their own verdicts.
• Meta-judge convergence: disagreement is reduced into fixed exits: consensus, a compatible concrete plan, or a bounded stall that is surfaced honestly.
• Concrete plans pass the gate: an obvious direction still needs independent validation before implementation.
• Symmetric verification: finished work passes independent reviewers and the fixed review truth table before it is declared done.
• Pure orchestration: the caller dispatches workers and aggregates bounded conclusions; implementation, verification, repair, and design solving remain delegated.

Protocol at a Glance

The protocol has seven ordered stages. The caller handles intake, dispatch, convergence, and reporting; workers own thinking, implementation, and review.

Stage	Purpose
intake	Record the user's goal, constraints, success criteria, and host capabilities.
choose_worker_mode	Check available worker carriers before dispatch.
thinking_panel_workers	Test the plan through six isolated perspectives.
meta_judge	Converge conclusions using the thinking truth table.
implementation_worker	Delegate the approved outcomes and scope.
review_triplet_workers	Review the completed candidate independently.
fix_or_done	Route repairs, further review, or completion under the existing gates.

The thinking panel assigns a distinct objective to each seat:

Thinking seat	Lens
teleology	Purpose and the form it requires.
parsimony	How much mechanism is necessary.
fidelity	Source-verified premises and representative evidence.
natural-ownership	The layer that naturally owns the duty.
proportional-containment	The justified scope, authority, and duration.
worth	Whether the benefit warrants the cost.

The review triplet applies the review truth table:

Review seat	Focus
architecture	Boundaries, contracts, coupling, and maintainability.
quality	Behavior, edge cases, failures, and user impact.
tests	Coverage, determinism, and verification strength.

A termination gate applies inside the final stage only when the host has positively declared a goal-driven continuation mechanism in the goal contract, and an affirmative satisfaction claim is being considered. It is not an eighth stage and grants no control over the host mechanism.

Only confirmed defects against the current goal or an existing contract enter automatic repairs. Unverified objections alone authorize no repair or blocking goal-gap/abstention; bounded checks investigate unknowns only when the decision depends on them. Findings retain evidence_status and user_goal_basis; uncertain changes to product or permission semantics go to their owner for clarification. A user correction cancels dependent stale plan, review and repair authority while preserving history and active-job protection. See goal corrections and admission.

Risks and Limits

• Broad execution access: the Codex runner defaults to danger-full-access. Read the invocation contract and threat model before supplying a work target.
• Heca execution: seats run unattended with full permissions; heca owns their agents, so host job time limits and teardown do not reach them. On the ordinary route, the caller collects each attempt's successful or failed result and then archives that attempt's own agent. On the stopped-failure route, the confirmed stop archive is the archive and comes before recording the failure. An attempt that was neither collected nor stopped, such as a caller interrupted before collection, remains unarchived; an archive call that fails after collection is reported as unresolved housekeeping. Archived history stays readable in heca. Support is limited to recorded evidence. See worker delegation.
• Cost and duration: a full initial run dispatches at least 10 worker turns, each potentially making multiple model calls: six thinking seats, one implementation worker, and three reviewers. Repair, focused, repeated review, and applicable termination passes add work. There is no pass_budget cap by default; see the protocol, panel, review, and pass accounting.
• External content transfer: the oracle seat sends the inlined brief through NyxID to an external ChatGPT Pro account. Consider what the brief contains before dispatch; see worker delegation and the oracle request contract.
• Artifact retention: runner briefs, logs, and results accumulate under ${SSHX_HOME:-$HOME/.sshx}/<flight-id>/attempt-<n> until cleaned or pruned. See the run directory and inactive-flight sweep; a short retention window can remove a still-running silent flight.
• Host responsibility: runner time limits and teardown of the whole job tree belong to the host harness. Teardown behavior for Codex, Cursor, and Gemini hosts is unverified in the teardown specification; also see delegation.
• Changing contract: this is a prerelease, as indicated by the package metadata. The contract may change; check the skill and contribution guidance when updating.
• Untrusted output: consensus remains AI output requiring independent verification (repository invariant FI-001). Context isolation does not prove independent priors or better results (delegation limits). The Lean model checks formal objects, not the truth of evidence or compliance of real runs; correspondence with the prose remains a human judgment.

Quick Start

Prerequisites and carriers

The host must provide background execution with completion notification, time limits, and whole-job teardown for runner dispatch. Worker delegation owns capability checks, seat assignment, bounded retries, and fallback.

Carrier	Prerequisites and role
codex-cli	Bash, jq, and an executable, usable Codex CLI; a writable run root and supplied work target. Initial seats when heca is unavailable and fallback seats inside heca use the Codex runner.
heca	Explicit user or host opt-in for this sshx run, with heca's MCP agent tools exposed to the session and an existing placement matching the work target. Disabled by default; takes all initial Codex seats when available, including implementation and repair. See Using Heca Workers and worker delegation.
nyxid-oracle	Python 3, Bash and jq for path projection, and a usable NyxID CLI with an active broker pool containing an online worker. The oracle runner owns dispatch; see pool selection and login diagnostics.
isolated-token-subagent	A host subagent facility with isolated token context; same-round peers cannot read one another's outputs before returning their verdicts.
abstain	Stop honestly when no carrier is available or a required flight exhausts retries and fallback; this is an outcome, not an execution carrier.

Every multi-seat stage initially assigns one seat to the isolated subagent, one to the oracle, and the rest to heca when explicitly enabled and available, otherwise Codex CLI, subject to seat constraints and fallback. The tests reviewer needs a carrier that can execute verification in the work target. Carrier assignment and isolation do not by themselves establish model diversity; see the owning policy.

Using Heca Workers

sshx supports heca alongside codex-cli and nyxid-oracle as a worker carrier. Heca is disabled by default. When the user or host explicitly enables it for this sshx run and its non-mutating capability check finds the exposed MCP agent tools, it takes every seat otherwise assigned to Codex CLI: the remainder of each multi-seat stage and every single-worker stage, including implementation and repair. When heca is unavailable, existing carrier composition stays unchanged. The fallback priority stays unchanged, so Codex CLI remains reachable inside heca.

Enable it explicitly, for example: “Use heca workers for this sshx run.” A host instruction explicitly enabling heca for the run also qualifies. Silence, merely mentioning heca, running inside heca, or exposed tools alone do not enable it. The caller records enablement, its source (or absence), and tool exposure once per run in the existing run/gate evidence. Both opt-in and exposed MCP agent tools are required for initial composition, single-worker dispatch, and fallback; opt-in cannot supply missing tools. Seats use those tools directly. You or the host may select a provider and model per seat; by default, seats use heca's codex provider when reported enabled, otherwise the caller's own heca provider, with heca's default model. The work target must have a matching existing heca placement. Worker delegation owns the contract and its evidence limits.

Claude Code

Install the marketplace and plugin, then invoke the skill with your task:

/plugin marketplace add ChronoAIProject/consensus-rnd
/plugin install consensus-rnd@consensus-rnd
/consensus-rnd:sshx <task>

The plugin identity is declared in the marketplace and plugin manifest.

Codex / Cursor

The Codex and Cursor manifests both declare "skills": "./skills/". Use the platform's plugin support to load the repository and read the skill contract; installation and invocation details depend on the host.

Verify the effective copy and refresh manually

Before dispatch, worker delegation records the actually read skill path, SHA-256, available checkout revision/dirty state, and current Heca policy in existing gate evidence. A manifest version or upstream release cannot identify a copied or locally edited skill, or prove an open session adopted it. Explicit user disable wins opt-in on every route, including retry, fallback, repair and resume; tools alone never opt in.

Inspect the path reported by the host, following any symlink, rather than assuming an installation directory. For a checkout, these read-only commands distinguish bytes from revision and manifest metadata:

shasum -a 256 <effective-skill-root>/SKILL.md
git -C <source-checkout> rev-parse HEAD
git -C <source-checkout> status --short

Compare the effective file with the intended checkout's skills/sshx/SKILL.md using diff; also compare supporting files because the whole skill directory is the distribution unit. A copied installation may have no Git revision: report that limitation and keep the content hash. None of these observations proves current-session adoption. Confirm the newly read path/hash and policy in the next run; if the host cannot expose that association, label adoption ASSUMED-UNVERIFIED.

Manual refresh is an owner action, never automatic sshx behavior. For a direct-folder installation, refresh the entire resolved skill directory from a verified checkout using the host's supported skill installation procedure, then compare again. Official Codex skills guidance documents discovery locations, symlink support and automatic change detection; restart Codex if an update does not appear. When multiple copies exist, verify the selected path. This does not assert that a legacy ~/.codex/skills copy is the selected source or that an existing session reloaded it.

For the Claude marketplace installation above, official plugin guidance documents claude plugin update consensus-rnd@consensus-rnd, then /reload-plugins to apply updated disk content to the session. Use the installed host's supported UI if that command is unavailable. For Cursor and Gemini, use their installed plugin/extension manager and confirm the effective source again; no reload command or adoption behavior is claimed here. These official owner procedures were not exercised on this host: local refresh/reload remains ASSUMED-UNVERIFIED. The skill executes no refresh or host/global setting change.

Gemini CLI

The extension manifest selects GEMINI.md as its context file. GEMINI.md lists the available skill and directs the agent to read its contract on demand.

Architecture

Surface	Owning source
Public identity and introduction	This README and its Chinese companion.
Repository rules and structure	CLAUDE.md, also exposed through the tracked AGENTS.md symlink.
Protocol and behavioral contract	SKILL.md.
Codex execution and artifact maintenance	CODEX_WORKER_SPEC.md and scripts.
Heca MCP delegation	SKILL.md.
NyxID broker dispatch	ORACLE_WORKER_SPEC.md.
Formal evidence and its limits	Formal model README.
Platform entrypoints	Claude marketplace, Claude plugin, Codex, Cursor, and Gemini.
Manifest version mapping	.version-bump.json and bump_version.py.
Repository verification	CI workflow, repository tests, and skill tests.

Contributing

See CONTRIBUTING.md for development prerequisites, local checks, and maintainer-owned release steps. Repository maintenance guidance is in Chinese.

License

MIT © ChronoAIProject