English | 简体中文
Svif is a Project orchestration product coordinating durable Project continuity, execution surfaces, and capability providers.
The Project persists; Executors and execution environments may change.
Name. Svif is an Icelandic noun meaning flight, hovering, floating, or gliding. The same lexical element appears in svifryk, the Icelandic term used for suspended particulate matter. The name reflects Svif's product model: execution can move across Execution Surfaces and Capability Providers without making any one environment the owner of the Project. It also deliberately echoes Agnir's particle metaphor: Agnir names the durable pieces of Project truth; Svif names the motion around them as execution changes while Project continuity persists.
This section is for users. Pick the action that matches what you want to do.
Svif is not publicly listed yet. After publication, install it from the universal Plugins Directory shared by ChatGPT and Codex.
Install and enable Svif for this Project: https://github.com/iorLab/svif
The current self-distributed release is v0.2.0-preview.1. A compatible installer resolves the short intent above to that immutable Preview tag through the repository marketplace; it must not silently install moving main. Codex CLI and ChatGPT desktop/Codex are the supported Preview surfaces. If the active client cannot install repository Plugins, it must report the unsupported surface instead of claiming installation succeeded.
No recurring Svif installation prompt is required. Ask for the actual Project task.
Upgrade the Agnir used by this Project to the latest stable release: https://github.com/iorLab/agnir
A genuinely uninitialized Project does not need Agnir to be manually initialized first. On the founding repository/filesystem path, Svif's shared Skill establishes the required Agnir continuity and matching Svif Project Binding as part of first use.
Upgrading Agnir is not the same as upgrading Svif. Compatible Agnir upgrades preserve SVIF.yaml, Project identity, and durable Project continuity.
For Agents working in this repository. Users normally do not need to read this section.
- Discover. Treat this repository root as the authorized Project Entry Point. Read top-level
AGNIR.yamland validate the declared Agnir Core/profile compatibility and Project identity. - Load. Load Current State and Next Actions from the durable memory declared by
AGNIR.yaml. Load Decisions and Evidence when they materially constrain the current operation. Prefer durable Project truth over chat history or private executor memory unless superseded by a newer Principal instruction or a directly observed current Project fact. - Bind Svif. Read
SVIF.yamland the relevant current specifications before changing Svif product behavior. Preserve the configured Continuity Provider, execution, capability, authority, and Project-identity boundaries. - Work / checkpoint. Perform the actual Project work, then reconcile only material continuity changes at an intentional checkpoint, save-progress, finish, or repository commit boundary. Unchanged durable truth is a no-op. Material changes must form one coherent candidate; reject stale-base publication rather than overwriting newer truth, then verify the locator chain after publication.
- Commit / push. Authorized
commit,提交,提交代码, or equivalent intent means checkpoint before commit and preferably one revision for Project + Agnir changes.commit and push,提交推送, or equivalent adds push plus authoritative-ref verification. Merely observing an external commit triggers checkpoint evaluation, not an unconditional Agnir write.
Root AGENTS.md is only the activation locator to this section; it must not become a second copy of Project state or the Agnir procedure. The canonical activation route is:
Project root -> AGENTS.md -> README.md / Agnir Project Instructions -> AGNIR.yaml -> declared durable memory
The applied Agnir operational distribution is recorded in AGNIR.yaml under extensions.agnir/operations; this provenance does not replace Core/profile compatibility or Project identity.
If an activation locator, identity, required memory locator, or compatibility check fails, repair the earliest faulty layer when authorized. Do not invent Project state or silently fall back to chat history, sibling repositories, or retired layouts.
On first use of Svif in a genuinely uninitialized repository/filesystem Project, the shared Skill establishes the founding Agnir continuity plus a matching Svif Project Binding. Svif does not take over existing Project files. Existing AGENTS.md and README.md receive only the activation/instruction entry they need, while unrelated content is preserved.
Project/
├── AGENTS.md # [EDIT: add entry only] add Agnir activation locator; preserve existing instructions
├── README.md # [EDIT: add entry only] add ## Agnir Project Instructions; preserve existing content
├── AGNIR.yaml # [ADD] founding Agnir discovery anchor
├── brand/ # approved brand masters, exports, QA, references, and handoff
├── .agnir/ # [ADD] Project-owned durable continuity
│ ├── state.md # [ADD] current durable Project truth
│ ├── next-actions.md # [ADD] outstanding ordered work for the next Executor
│ ├── decisions.md # [ADD] durable decisions that constrain future work
│ └── evidence/ # [ADD] recovery/audit/material-claim evidence and checkpoints
└── SVIF.yaml # [ADD] Svif Project Binding: continuity, execution, capability and profile bindings
If compatible Agnir/Svif artifacts already exist, the Skill validates and reuses them rather than recreating them. Partial or contradictory artifacts are a repair case, not clean initialization. A Project intentionally bound to another Continuity Provider is not silently overwritten with Agnir.
These are founding repository-filesystem onboarding artifacts, not universal Svif kernel requirements. Svif coordinates replaceable providers and execution surfaces; it does not make Git, GitHub, Agnir, ChatGPT, or Cloudflare permanent kernel dependencies.
flowchart TB
P["Principal / User"] --> E["Execution Surface<br/>ChatGPT today"]
D["Installable Plugin<br/>Agent Plugins 1.0<br/>Skill-first MVP"] --> E
subgraph T["Target Project surface — first use"]
G["AGENTS.md<br/>EDIT: add activation locator only"]
H["README.md<br/>EDIT: add Agnir instructions only"]
Q["AGNIR.yaml + .agnir/<br/>ADD: founding continuity"]
B["SVIF.yaml<br/>ADD: Project binding"]
G --> H --> Q --> B
end
D -. "non-destructive first-use setup" .-> G
subgraph S["iorLab/svif"]
X["Execution integration<br/>src/svif/execution"]
O["Svif Orchestrator"]
K["Capability Providers<br/>src/svif/capabilities"]
R["Portable contracts<br/>Evidence · Authority · Profiles"]
X --> O
O --> K
O --- R
end
E --> X
B -. "configures" .-> O
Q -. "durable continuity" .-> C["Continuity Provider<br/>Agnir today"]
O --> C
K --> F["External systems<br/>Cloudflare today"]
C -. "independent protocol" .-> A["iorLab/agnir"]
The Project-surface nodes describe the first-use onboarding boundary; the product nodes describe Svif's replaceable runtime roles. Svif owns the coordination boundary. The Orchestrator does not permanently depend on Agnir, ChatGPT, or Cloudflare; they are the founding/current bindings for the Continuity Provider, Execution Surface, and Capability Provider roles.
The active canonical repository topology is deliberately small:
iorLab/svif— the complete Svif product: Orchestrator, integrations, capability providers, installable Plugin packaging, contracts, tests, and E2E fixtures;iorLab/agnir— the independent Agnir continuity protocol consumed through Svif's Continuity Provider interface.
Provider-specific Svif behavior stays in iorLab/svif unless it becomes an independently useful product or protocol in its own right.
flowchart TD
I["Principal intent"] --> P["Plugin / Execution Surface workflow<br/>discover Project and durable state"]
P --> B["Orchestrator.begin<br/>resolve Project binding"]
B --> A["Agnir Continuity Provider"]
A --> L["Load durable continuity"]
L --> M["Materialize Project-scoped execution context"]
M --> E["Execution Surface / Executor<br/>ChatGPT today"]
E --> W["Structured WorkResult<br/>subject + evidence + requested effect"]
W --> V{"Exact subject verified?"}
V -->|No| STOP["Stop / repair<br/>no false checkpoint"]
V -->|Yes| Q{"External effect requested?"}
Q -->|No| C["Reconcile + checkpoint"]
Q -->|Yes| U{"Required authority granted?"}
U -->|No| STOP
U -->|Yes| D["Capability Provider actuates<br/>Cloudflare today"]
D --> O["Independent observation"]
O --> R{"Observed subject / target match?"}
R -->|No| STOP
R -->|Yes| C
C --> A
C --> N["Durable Project truth ready<br/>for the next Executor"]
The default internal lifecycle is:
DISCOVER -> PLAN -> CHANGE -> VERIFY -> DELIVER -> OBSERVE -> CHECKPOINT
REPAIR returns to the earliest violated invariant. For externally driven surfaces such as ChatGPT, Orchestrator.begin() creates the bound operation/session and Orchestrator.complete() reconciles the returned result. Untrusted model/result payloads cannot self-grant protected authority.
Svif ships the Skill-first v0.2.0-preview.1 Plugin under plugin/, using the portable Agent Plugins 1.0.0 package layout and an additive OpenAI/Codex manifest:
svif/
├── .agents/plugins/marketplace.json
└── plugin/
├── plugin.json
├── .codex-plugin/plugin.json
├── README.md
└── skills/
└── svif/
└── SKILL.md
plugin/plugin.json remains the portable Agent Plugins manifest. plugin/.codex-plugin/plugin.json is the OpenAI/Codex manifest used for the shared Skill and public listing metadata. .agents/plugins/marketplace.json is the supported self-distributed Preview path for Codex CLI and ChatGPT desktop/Codex, while remaining separate from universal-directory publication.
The Repository Preview keeps the user request intentionally short. Version resolution, fixed-tag marketplace registration, client-capability checks, first-use Agnir bootstrap, and installation evidence belong to the installer/Plugin procedure rather than the prompt. See plugin/README.md for the exact fixed-tag route.
The mature personal-user path remains ChatGPT -> Plugins Directory -> discover Svif -> install -> invoke. Svif is not publicly listed yet, so this is a target consumer path rather than a currently available production install.
OpenAI's current public submission flow explicitly accepts a Skills-only Plugin. Svif's .codex-plugin/plugin.json has therefore been tightened to the current final-directory metadata limits, while the existing plugin/skills/svif/SKILL.md remains the single shared workflow implementation. MCP/App packaging is not a prerequisite for the initial public submission and must not be added merely to satisfy publication.
The Repository Preview and the future public personal-user release are different distribution layers. Codex CLI and ChatGPT desktop/Codex Preview installation must be validated independently; public review approval, directory publication, personal ChatGPT installation, invocation, Agnir activation, verification, and checkpoint remain separate evidence layers.
See plugin/README.md for public submission prerequisites, proposed listing metadata, review test cases, repository-marketplace development routes, and evidence boundaries.
This tree is the practical map of the repository. It is intentionally selective: it shows the directories and key files that explain where each product responsibility lives, rather than listing every fixture or evidence file.
svif/
├── .agents/plugins/ # OpenAI/Codex repository marketplace catalog
│ └── marketplace.json # maps the self-distributed Preview to the local ./plugin root
│
├── src/ # executable Svif product code
│ └── svif/
│ ├── runtime.py # Orchestrator kernel: begin/run/complete, verification, authority, reconciliation
│ ├── continuity/ # Continuity Provider implementations/adapters
│ │ └── agnir.py # founding Agnir repository/filesystem Continuity Provider
│ ├── execution/ # Execution Surface bridges
│ │ └── chatgpt.py # founding ChatGPT structured Execution Surface bridge
│ └── capabilities/ # Capability Providers that inspect or affect external systems
│ └── cloudflare.py # founding Cloudflare Workers Capability Provider
│
├── integrations/ # platform/provider integration boundaries
│ ├── chatgpt/ # ChatGPT app/MCP integration around the execution bridge
│ └── cloudflare/ # Cloudflare descriptor, transport boundary, and integration notes
│
├── plugin/ # installable Agent Plugins 1.0 distribution package
│ ├── plugin.json # portable Plugin manifest
│ ├── .codex-plugin/plugin.json # OpenAI/Codex + public-directory listing metadata
│ ├── README.md # public submission, package validation, installation and evidence guidance
│ └── skills/svif/SKILL.md # shared Svif Project-orchestration workflow Skill
│
├── spec/ # portable product contracts used by the Orchestrator and integrations
├── profiles/ # specialized behavior layered on portable contracts
├── schemas/ # machine-readable serializations of Svif contracts
├── tests/ # runtime, provider, surface, continuity, Plugin, and E2E tests
├── conformance/ # portable-contract conformance checks and fixtures
├── checks/ # repository/product-integrity checks
├── history/ # predecessor and retired-project evidence; not active runtime dependencies
│
├── .agnir/ # this Svif Project's canonical state, next actions, decisions, and evidence
├── .github/workflows/ # CI that runs repository, runtime, and conformance checks
├── AGENTS.md # minimal Agnir activation locator to this README's Project Instructions
├── AGNIR.yaml # locates this Project's Agnir continuity under the current filesystem profile
├── SVIF.yaml # repository/filesystem serialization of this Project's Svif binding
├── ARCHITECTURE.md # detailed product architecture and dependency boundaries
├── README.md # English project entry point and canonical Agnir Project Instructions
├── README.zh-CN.md # Simplified Chinese project entry point
└── VERSION # current Svif product/release version
For the fully expanded file-by-file map of the current main, including responsibility annotations for every tracked file, see REPOSITORY_TREE.md.
Python is the current executable reference vehicle; it does not freeze the eventual distribution technology. The installable Plugin is now an active product artifact rather than a future-only target.
- Agnir repository/filesystem continuity adapter exists.
- ChatGPT structured execution bridge supports externally driven
Orchestrator.begin()/Orchestrator.complete()handoff. - Cloudflare provider logic is owned by Svif and uses an injected transport boundary, so tests do not require live credentials.
tests/test_founding_e2e.pycomposes all three through the real Orchestrator boundary.plugin/plugin.json+plugin/skills/svif/SKILL.mdremain the portable Plugin MVP package.plugin/.codex-plugin/plugin.jsonnow also satisfies the public-directory listing limits currently guarded by repository tests..agents/plugins/marketplace.jsonis the supported self-distributed Preview route for Codex CLI and ChatGPT desktop/Codex; it is not public-directory publication or the primary personal ChatGPT Web onboarding path.- Protected authority remains outside untrusted model/result payloads.
- External success requires exact verified-subject delivery plus independent observation before checkpoint.
The founding E2E is intentionally credential-free. It proves the Svif product loop and provider boundaries, not live Cloudflare production delivery.
SVIF.yaml is the repository/filesystem serialization of project-binding/0.2 for this Project. It also registers the current product-owned Plugin artifacts while keeping continuity, execution, and capability bindings replaceable.
README.md and README.zh-CN.md are maintained as parallel entry points. Any change to product architecture, component ownership, dependency direction, authority/provenance boundaries, runtime flow, distribution status, or documented repository structure must update both language versions in the same change set.
Before the Architecture Diagram, README content is deliberately limited to Start Here for users, the canonical Agnir Project Instructions for Agents, and What Svif Adds to a Project as the concrete first-use Project-surface explanation. Install and Agnir-upgrade prompts are shown as standalone text code blocks so GitHub exposes copy controls. The Architecture Diagram mirrors that non-destructive first-use boundary; Runtime / Operation Flow remains a post-bootstrap runtime view and does not carry installation-mutation labels. Both Mermaid diagrams use conservative GitHub-compatible syntax: explicit nodes, quoted labels with <br/>, ordinary directed edges, and no direct edge to a subgraph. Publication workflow, Plugin packaging rationale, compatibility detail, and implementation explanation belong after the architecture entry point or in dedicated documents.
The exhaustive companion REPOSITORY_TREE.md is the file-level map of the active repository and must be updated whenever tracked files are added, removed, moved, or materially change responsibility.
python checks/check_repository.py
python conformance/check_contracts.py
PYTHONPATH=src python -m unittest discover -s tests -vFirst, complete the self-distributed v0.2.0-preview.1 acceptance path in Codex CLI and ChatGPT desktop/Codex using the exact immutable candidate. Then publish the verified tag as a GitHub Prerelease and retain the same Skills-only package for the later OpenAI Platform submission. Universal-directory publication still requires the applicable publisher identity and review flow; MCP/App packaging remains a later capability increment, not a release gate. Live Cloudflare actuation remains separately gated.