ROST implementation methodCompass authoring guideInstall the ROST implementation SkillROST implementation evidence and groundingAICOS chat guideResponsibility Graph playbookCharter design playbookCharter authoring deep-diveAgent staffing playbookAdd agents to your Responsibility GraphDesign a custom agentAgent builder guideAgent definition import guideHow agents workTool access and vaultAvailable tools guideCLI and MCP installation guideSkill builder guideAgent Skill authoring guideAgent Skill setup guideAgent reference mapCascade guideSignal guideFriction guideSteward queue guideBilling and pricing guideConfirmations and human gates guideGoverned command capability matrixSettings guideMembers and invites guideNotifications guideLocal runner guideStock agents guideTroubleshooting guideAI model data handling guideSecurity model guideSync rhythm playbookMigration and legal notes
Docs

Install the ROST implementation Skill

Install the public implementation Skill through its signed, verified GitHub release so a general-purpose agent can run the ROST implementation method.

company setup

On this page

  • Install
  • What the installer verifies, in order
  • What gets installed
  • Update path
  • After install
  • Fallback
  • Public repository

The public ROST implementation Skill is one installable package that turns a general-purpose agent into a mode-aware ROST implementer. It covers regular-mode implementation today. Full Operator external-effect handoff activates only through live capability discovery in a future release; until then the Skill must not assume Full Operator authority is available.

Install

Run the installer from the public repository:

npx @rosthq/cli@latest skills install-implementation --client claude-code

Supported clients include claude-code, codex, and cursor. The command requires Node 22+ and a CLI version at or above the manifest's minimum. The installer works from the public repository alone; it does not require a ROST account first.

What the installer verifies, in order

This order is the security contract: the release signature is verified before any manifest or checksum is read.

1. Release-tag SSH signature — the annotated release tag's SSH signature is checked against the signer public key compiled into the rost. Unsigned, retagged, or wrong-signer releases are rejected here, before any manifest or checksum is read. 2. Repository owner — the repository owner must be agenticoperatingsystem. 3. Manifest checksum — the default tag is pinned to an exact manifest sha256 in the rost; the fetched manifest must match. 4. Per-file checksums — every installed file's sha256 is checked against the manifest. 5. Supported protocol range — the manifest protocol version must be compatible with the rost's minimum supported protocol.

A fetched release can never introduce its own trust root. Any key material inside release content is ignored. Signer rotation is an explicit rost update, never release-controlled metadata.

What gets installed

Files are placed under the client's skills directory as rost-implementation/, plus a .rost-skill-release.json provenance marker. The marker records source repository, tag, commit, manifest sha256, and signer fingerprint. The agent quotes this marker when reporting which implementation method ran.

For Claude Code the directory is $CLAUDE_CONFIG_DIR/skills/ when that variable is set, otherwise ~/.claude/skills/ — CLAUDE_CONFIG_DIR is the variable Claude Code itself reads. If CLAUDE_HOME is set without CLAUDE_CONFIG_DIR, the installer fails with a migration instruction (Set CLAUDE_CONFIG_DIR=<value> instead, or unset CLAUDE_HOME to install under the default ~/.claude); when both variables are set and name different directories, the installer uses CLAUDE_CONFIG_DIR and prints a notice that CLAUDE_HOME is ignored (an equivalent spelling prints no notice). For Codex the directory is $CODEX_HOME/skills/, otherwise ~/.codex/skills/. The installer prints the resolved directory and a line such as Claude Code will load this from <path> — check it matches the configuration the client launches with.

Update path

Re-run the installer to update. It re-verifies and replaces the pinned release. Use --tag <vX> to select a specific signed release. The stable install name rost-implementation is preserved across versions.

Re-running is also how you confirm an existing install is intact: the installer hashes the files ON DISK against the signed manifest, not just the recorded marker, and reinstalls if any file was edited, removed, or added. A matching provenance marker alone is not evidence about the installed bytes. Note the scope of what this proves — the files matched the manifest at the moment the installer completed, so re-run it immediately before a run that depends on the Skill being unmodified.

After install

The agent bootstraps per the Skill's SKILL.md:

1. Confirm rost presence and version. 2. Discover implementation-purpose device authentication live via rost docs / rost command list. Never use an ordinary login session for an implementation run. 3. If the implementation-bootstrap surface is absent in the installed rost, upgrade @rosthq/cli and, if still absent, stop with an honest prerequisite message. 4. Load doctrine via rost reference get rost-implementation-method.

Fallback

If the installer cannot run, the only supported step is installing or upgrading @rosthq/cli. Hand-copied skill files are unverified and unsupported.

Public repository

Canonical source: github.com/agenticoperatingsystem/rost-implementation-skill.

Version

2026-07-30.2

Used by

  • human
  • cli
  • mcp
  • in app agent

API

/api/reference/files/implementation-skill-install