--- name: fcis-owner-access description: Govern Fountain Coach owner passkeys, OAuth/Host access, machine enrollment, and reasoning-expressive FCIS OpenAPI 3.1 and MCP capability discovery. Use for access.fountain.coach fleet management, fcis.fountain.coach APIs, human- and LLM-readable operation IDs/descriptions, callable instrument inventories, per-client authorization, conformance, or API maintenance and release. Reuse the existing owner-access Skill, current MIDI2/FCIS-KIT identities, Host registry, FountainAuthKit, FountainStore and contract tests. Never treat API documentation as a grant or build a second authority. --- # FCIS Owner Access — Passkeys and Recovery ## Canonical Go stack and OpenAPI — owner-adopted 18 November 2026 **Peer canonical source:** `Fountain-Coach/.github/FCIS-STACK.md` defines the Go FCIS architecture and finite production DoD; `Fountain-Coach/.github/FCIS-OPENAPI.json` is its **same-level authority** for precise FCIS-owned HTTP operation IDs, schemas, security/refusal and target status. Read and maintain **both** together when HTTP boundaries change. Do not merely link to a domain-maintained OpenAPI or treat one as a generated authority of the other. The previous `midi2-gpu-fabric/apps/estate-publisher/contracts/fcis-owner.openapi.json` is **historical Swift handler compatibility input, not the Go-target writable source**. The new Go FCIS owns OAuth, WebAuthn owner identity, MCP and a fresh dedicated native Go FountainStore instance. The old Swift OAuth issuer, historical refresh-token state and shared old Store are not runtime prerequisites. This is **target specification**, not proof of a deployed Go release. Public `fcis.fountain.coach` remains the issuer and `access.fountain.coach/manage/` remains an owner UI consumer, never a grant writer. **Change route:** Edit the root `FCIS-OPENAPI.json` (OpenAPI 3.1) on reviewed organization Git, classify each operation `specified-only-not-deployed` until Go/HTTPS verification, enforce schema/$ref and unique operationId/summary/description, and update the root Stack if architecture changes. A domain copy may be verified as a pinned **derived projection** only; update the owner Skill's [OpenAPI maintenance protocol](references/openapi-maintenance.md) and existing Go migration Skill accordingly. Never invent an MCP tool, MIDI2 operation, Host grant or payment endpoint merely by writing OpenAPI. Priority-two Stripe remains an independent FCIS instrument/provider contract. The internal FCIS owner OpenAPI 3.1 contract now forms a bounded owner/Host subset within the full root OpenAPI. OpenAPI is a *protocol specification*, not an executable grant. **No hardcoded** `host.describe` ceiling is adopted; capability discovery must follow admitted Host instruments and the actual client grant. **A new operation is DOCUMENTED only after contract parity** with its source/handler shape; **CALLABLE or LIVE-ACCEPTED require separately witnessed execution**. For any proposed Go operation without a matching handler, mark `NOT_ESTABLISHED` rather than fabricating availability. Do not invent endpoints or permissions to fill an architecture diagram. ## Authority and observed baseline Keep the user's goal: **normal owner sign-in and independent recovery without SSH, passwords copied from an Ubuntu machine, or a second identity system**. First-device Golden Key and WebAuthn passkey enrollment, sign-out, and returning passwordless sign-in were live-accepted on a physical iPad on 12 November 2026. **Independent second-device account recovery remains NOT_ESTABLISHED**; inspect current issuer/owner evidence before making a fresh status claim. Separately classify DOCUMENTED, SOURCE-TESTED, STAGED and LIVE-ACCEPTED operations. Read published Governance [07](https://governance.fountain.coach/chapters/07-agent-operating-guide/), [08](https://governance.fountain.coach/chapters/08-validation-and-acceptance/), [142](https://governance.fountain.coach/chapters/142-fountainauthkit-governed-authorization-authority/), [152](https://governance.fountain.coach/chapters/152-the-host-is-the-instrument/), [161](https://governance.fountain.coach/chapters/161-host-enrollment-client-consent-and-access-management/), and [143](https://governance.fountain.coach/chapters/143-simplicity-is-key-one-bootstrap-one-trusted-session/) for trusted session and recovery. Check published `chapter-status.json`: Chapter 142 is a current *proposal*, not proof of deployed OIDC. Preserve authority boundaries: - `access.fountain.coach/manage/` is the **Fountain Press UI**, not a credential issuer, grant database, or authorization writer. - `fcis.fountain.coach` is the existing **OAuth issuer and MCP protected resource**; the Host already has a `FountainHostMCPOwnerAuthenticating` adapter for upstream human authentication. - FountainAuthKit owns scoped OAuth code/PKCE, client registry, refresh, consent and grant revocation; sign-in **does not itself authorize** client/Host execution. - The admitted Host owns protected operations; FountainStore provides durable evidence. A UI status is not an operation receipt. - The Ubuntu Mac mini's observed `fcis-access-staging` is a read-only snapshot, not the Hetzner production issuer. Do not infer their credential equality. The design anchor is `Fountain-Coach/fountain-press/docs/access-passkey-first-owner-signin-2026-11-12.md`. Read owning `AGENTS.md`, `PLANS.md`, production source provenance and exact scenario before editing. ## Repeatable procedure 1. **Inspect current identity and actual failure.** Inspect issuer metadata, running Host adapter, published Access JS, credential *metadata only*, current source SHA and current authorization; do not request, dump, paste, log, display or commit any password, Basic header, signing key, bearer/refresh token, passkey secret or SSH private key. Do not reset user credentials or disturb the existing ChatGPT grant. 2. **Bind the existing owner.** Register passkeys only following an independently verified owner action, associating the WebAuthn credential with the existing issuer subject; do not invent a second account, local permission store or login domain. 3. **Use vetted WebAuthn/FIDO2 verification.** On the issuer's chosen HTTPS origin, require fresh unpredictable challenge, correct Relying Party ID, origin, credential public key/signature, user verification, expiration, replay resistance, and rate limiting. Verify `access` vs `fcis` origins explicitly, default to an issuer-hosted top-level ceremony and avoid homemade authentication crypto. 4. **Preserve OAuth.** Apply [RFC 9700](https://www.rfc-editor.org/rfc/rfc9700.html), exact registered redirect URIs, authorization-code + PKCE S256, resource audiences, bounded consent, existing signing/grant/refresh continuity and denial after revocation. Never turn human login into automatic Host power. 5. **Select secure browser sessions.** Follow [RFC 10017](https://www.rfc-editor.org/rfc/rfc10017.html); evaluate a same-origin backend-for-frontend that holds OAuth tokens server-side and issues an opaque Secure/HttpOnly/SameSite cookie. Require CSRF defense, timeouts, logout, reauth/step-up, and negative browser-origin tests. A session cache is not another authority. Never put credentials or grants into static Press content or localStorage. 6. **Establish independent recovery.** Enroll a second usable authenticator with recent owner verification, revoke lost devices, prove second-device login and define a reviewed lost-all-device procedure. Keep existing Basic fallback controlled and audited during migration; retire it only after replacement and recovery are independently accepted. No unverified email/SMS shortcut. 7. **Implement in bounded source-owning slices.** Host/Kit own the verifier and authority; Press owns the user interface; Host/Store own effect receipts. Do not invent a second service, new `mcp.fountain.coach` front end or dead UI controls. 8. **Run the acceptance matrix.** Verify isolated WebAuthn wrong origin/RP/challenge/key/UV/replay denial, cookie/CSRF/logout/expiry tests, *physical* iPad Safari and Mac sign-in, accessibility, second-factor recovery, existing ChatGPT `host.describe`, and distinct disposable-client revoke -> refresh denied -> real MCP denied -> Store receipt. Simulated iPad viewports support but do not replace native testing. ## Internal FCIS OpenAPI maintenance For an internal FCIS owner/API change, OpenAPI drift, newly admitted Host HTTP route, grant/machine response or published fleet view, follow [the FCIS OpenAPI maintenance protocol](references/openapi-maintenance.md). It maintains **the one organization-root `FCIS-OPENAPI.json`** together with `FCIS-STACK.md`, verifies exact standalone Go handler/test parity when implemented, and uses the old Swift `midi2-gpu-fabric/apps/estate-publisher/contracts/fcis-owner.openapi.json` only as a pinned compatibility predecessor (including its Node source-drift gate and Swift tests where predecessor behavior is changed). Keep documentation, source tests, installed Go, signed Caddy release and physical iPad acceptance **distinct**. This internal OpenAPI 3.1 is FCIS-owned, never an external Stripe/provider schema and never a substitute for MIDI2 IDL/FCIS-KIT execution. No new API authority, runtime Kit, second Skill or plugin is created; the owner has explicitly promoted the already-existing OpenAPI to peer canonical Git source. ## Reasoning-expressive Host and instrument discovery — mandatory Apply Skill Creator's **name/description selection** principle to FCIS operations. The internal FCIS owner OpenAPI 3.1 contract is the **human- and machine-readable HTTP management projection**, not the FCIS-MIDI2 capability authority. Read [the FCIS OpenAPI maintenance protocol](references/openapi-maintenance.md) before any operation/capability change. 1. **Describe intended use, not just implementation.** For each existing HTTP operation use a stable, meaningfully named `operationId`, a short action-oriented `summary`, and a `description` explaining *when to select it*, *when not to*, required subject/client/target and scope, read versus mutation consequences, expected typed result and failure/evidence semantics. Explain parameter meaning and real constraints in OpenAPI 3.1 JSON Schema. Give safe examples when they aid selection. Do not rename shipped IDs or wire fields without compatibility/versioning. 2. **Discover the actual capabilities.** Resolve the signed/typed MIDI2 IDL, FCIS-KIT instrument contract, installed Host registry, target routing, issuer admission and per-client grant. The HTTP inventory and MCP `tools/list` are distinct projections of that authority, not handwritten substitutes for it. Only expose operations actually implemented and registered on the selected Host; preserve their canonical identities and versions in both projections. A broad Host description or an `OpenAPI` operation is not an executable grant. 3. **No hardcoded `host.describe` capability ceiling.** Treat today's `OwnerFleetHost.capabilities.items.enum = ["host.describe"]` and the MCP adapter's special-case list as **current implementation limitations**, never as constitutional FCIS capability policy. A new admitted, authorized operation must not require writing a second inventory by hand. Reconcile registry-backed HTTP capability metadata with MCP advertised tools, typed arguments and runtime dispatch **through existing native adapters**. Until supported and tested, label any extension NOT_ESTABLISHED; do not invent endpoints, auto-grant permissions, or present unimplemented tools as callable. 4. **Keep capability states independent.** Distinguish `declared`, `installed`, `admitted`, `routable`, `authorized-for-this-client`, `callable`, and `live-accepted`; attach source revision, issuer/Host observation time and evidence identity to positive claims. A machine can be enrolled yet offline; an installed instrument can lack a grant; a documented operation can lack a deployed route. Unknown state is explicitly NOT_ESTABLISHED, never silently coerced to true/false or advertised as a success. 5. **Select productively, execute narrowly.** The LLM may choose among semantic descriptions and inspect a detailed contract progressively, but cannot infer permissions from names, suggest an unchecked generic terminal fallback, or derive a usable credential. The existing Host revalidates exact target, owner/client relationship, scope, operation, parameters, bounded effects and authorization **at invocation**. FountainStore and independent witnesses—not the chat transcript, OpenAPI document, MCP listing or HTTP status alone—establish terminal effects. 6. **Enforce drift gates before release.** Extend existing Node OpenAPI source-drift and native Swift handler tests, plus existing MCP/Host focused tests, to compare current issuer inventory, admitted operation IDs, schema/version mapping, `tools/list`, real dispatch and per-client refusal. Reject stale or hardcoded capability enumerations presented as complete, orphaned/invented operations, ambiguous semantic descriptions, unsupported state transitions and false callability/health claims. A new operation is DOCUMENTED only after contract parity; CALLABLE or LIVE-ACCEPTED require separately witnessed execution and Store evidence. Preserve the current production service and grants until the separately governed release passes. ## Reporting and maintenance Use [the acceptance matrix](references/acceptance.md) to report source revision, owner/issuer/Host binding, exact tested behavior, negative tests, real-device evidence and next missing gate. Unknown is **NOT_ESTABLISHED**, not a manufactured outage. A Git commit, HTTPS 200, or Press badge never proves deployed passkeys. This is an **agent procedure**, not a new Kit, OIDC provider or account system. Maintain only in `Fountain-Coach/.github/.codex/skills/fcis-owner-access`; release via the single GitHub-backed Fountain Coach FCIS plugin under [Governance 160](https://governance.fountain.coach/chapters/160-skill-sovereignty-and-governed-distribution/), protected PR, SHA-256 file-tree verification, and independent hosted version/selection readback. **No skill.zip, duplicate plugin or standalone personal Skill.**