Skip to content

Latest commit

 

History

History
112 lines (89 loc) · 4.64 KB

File metadata and controls

112 lines (89 loc) · 4.64 KB

Remis for Codex: Agent API quickstart

Remis is the reliable execution plane for game localization. Codex is the natural-language control plane that can inspect a workspace, explain a plan, call the localhost API, and stop at safety gates.

The Skill is an operator manual. The product remains the Remis desktop application, FastAPI service, workflow engine, validator, repair system, project store, and review/export interface.

Install the operator Skill

Give this prompt to your Agent:

Clone and install Remis from https://github.com/Drlinglong/Remis, read the bundled Agent Skill, start it locally, and verify that it is ready to use. When installation succeeds, confirm that Remis is ready. Then briefly explain that external providers such as OpenAI or Google require an API key to access their services. An API key is a private credential that may be linked to billing; enter it only in Remis Settings → API Settings, never in chat. Local providers such as LM Studio or Ollama do not need one.

Codex discovers the repository Skill at:

.agents/skills/remis-agent/SKILL.md

Verify the local service

Start the installed desktop application or run scripts\developer_tools\windows\run-dev.bat from the repository root. The default port is 1453; respect REMIS_BACKEND_PORT if the launcher reports an override. Then check:

Invoke-RestMethod http://127.0.0.1:1453/api/health
Invoke-RestMethod http://127.0.0.1:1453/api/agent/preflight
Invoke-RestMethod http://127.0.0.1:1453/api/agent/capabilities

The capability response lists public game, language, provider, and workflow information. It does not expose provider credentials.

Run preflight before every workflow. It performs a live latest-release check against the official GitHub repository and reports whether provider setup is missing. On a first installation, configure a cloud provider key in Remis Settings > API Settings, or deliberately select and test a keyless local provider. An API key is a secret credential issued by the model provider for authentication and often billing; it belongs in Remis, never in Agent chat.

Use the API safely

The governed sequence is:

flowchart TD
    C[Codex understands intent] --> S[Remis Skill applies rules]
    S --> I[Inspect mod and create plan]
    I --> A{User approval required?}
    A -- No, dry run --> R[Remis localhost API]
    A -- Yes, approved --> R
    A -- Not approved --> X[Stop without side effects]
    R --> W[Workflow engine]
    W --> V[Validation and bounded repair]
    V --> H{Human review needed?}
    H -- Yes --> Q[Review in Remis]
    H -- No --> E{Approve export or overwrite?}
    E -- Approved --> O[Installable localized mod]
    E -- Not approved --> X
Loading

Use the detailed payload and status reference in .agents/skills/remis-agent/references/api-workflow.md.

Trust boundaries

  • Localhost only: Remis binds to the local machine.
  • Keys stay in Remis: Agent responses and capability discovery never include provider API keys.
  • Release check before work: every Agent workflow begins by checking the installed version against the latest official GitHub Release.
  • Approval before spending: real model-backed translation requires a plan-specific approval.
  • Approval before repair: model-backed repair is a separate write/cost gate.
  • Approval before export: target paths and overwrite state are previewed before deployment.
  • Validated outputs: game variables, syntax, encoding, and structure remain under deterministic checks.
  • Audit trail: plans, tasks, repair attempts, approvals, and recoverable job snapshots persist in local Remis data.

Current boundary

In the 3.2.0 stable and Agent Preview channels, the Agent API exposes approval-gated translation start, checkpoint resume, repair, export, and task cancel operations. Cancellation is cooperative and applies only to supported translation task kinds; it is not an immediate process kill. Pause remains unsupported because the runner has no safe cooperative pause boundary.

Published context releases can be read through the Agent context endpoints. Direct Agent context-analysis planning/start is not exposed yet; use the normal Remis archive workflow and its Task Center approval gates. Archive removal is a separate approval-gated operation.

Checkpoint resume is valid only when the persisted checkpoint is compatible with the current source/configuration snapshot. Repeating the same resume request with its idempotency key must return the existing child task rather than starting another translation.

Interactive API documentation is available at http://127.0.0.1:1453/docs while the service is running.