World-Forge: Instructions file for Codex

AGENTS.md — standing instructions for agentic tools

Read this first. It applies to Kilo Code, Cline, and any other agentic IDE tool opening this workspace. It is deliberately short — it routes you to the right document for your session type instead of duplicating them.

First: which kind of session is this?

A. Running the pipeline (most sessions)

The user typed /worldforge brainstorm, /worldforge start, /worldforge revise, /worldforge resync-preset, /worldforge convert, or a resume/skip/status variant.

• Open workflows/world-forge.md and follow it. It is the router and the source of truth for what runs when: it dispatches build runs to four stage files (world-forge-discovery.md → world-forge-drafting.md → world-forge-validation.md → world-forge-construction.md). Load the router plus the active stage file only — never all of them. Revise runs live in workflows/world-forge-revise.md, convert runs in workflows/world-forge-convert.md, and preset resync / audition / post-launch routing in workflows/world-forge-postlaunch.md.
• Kilo Code: run commands through WorldForge-Orchestrator. The user selects it in the agent picker and types /worldforge … to it. Its system prompt is the router, and its permission block lets it write only the ledger and dispatch only WorldForge-* agents. If you are Kilo's general-purpose Code agent and receive a /worldforge command, tell the user to switch to WorldForge-Orchestrator — if they want you to proceed anyway, follow the same dispatch rules below (you are not permission-restricted, so the rules are on you).
• Interactive agents need a relay. Dispatched agents cannot ask the user questions directly; their questions come back as their result. Relay them verbatim, never answer on the user's behalf, and pass the user's reply back verbatim to the same child session (router → DISPATCH PROTOCOL rule 5).
• Kilo Code Subagent Delegation: Whether you are the orchestrator or the Code agent, do not run the pipeline commands inline. Use your subagent delegation tool (the task tool) to route the command to the correct custom subagent:
  • /worldforge brainstorm or /world-forge brainstorm → Dispatch to WorldForge-Brainstormer (optional ideation upstream of Phase 0; writes informal Brainstorm_Notes.md, no World Seed)
  • /worldforge start or /world-forge start → Dispatch to WorldForge-Interviewer
  • /worldforge revise or /world-forge revise (with --freeform, --target, etc.) → Dispatch to WorldForge-Reviser
  • /worldforge convert or /world-forge convert → Dispatch to WorldForge-Converter
  • /worldforge resume phase[N] → Dispatch to the custom agent defined for Phase N (e.g., WorldForge-Editor for Phase 3)
• If a dispatch or model call fails, never do the phase's work inline. Retry the dispatch; if it stays unavailable, set the Pipeline State Ledger status to BLOCKED, halt, and surface it to the user (router → DISPATCH PROTOCOL rule 1). An inline-produced phase has no agent spec, no Context Manifest, and no valid sign-off — treat its output as if the phase never ran.
• Pipeline files are READ-ONLY at runtime: everything under agent_roles/, templates/, workflows/, plus Notes_On_functionality.md and Notes_Quick_Reference.md. You write only to the world project's Drafts/, Export/, World_Seed.md, and report files. If you find yourself about to edit a pipeline file mid-run, stop and surface it to the user — an agent went off-script.
• Load only what each phase needs. Every agent spec begins with a 📂 CONTEXT MANIFEST listing exactly what to load. Honor it. In particular, never load Samples/, wiki/, CLAUDE.md, CHANGELOG.md, or tutorial.md during a run — they are human-facing or maintenance material and only burn context.
• For SillyTavern runtime questions (position values, lorebook flags, token budget, prompt assembly order): consult Notes_Quick_Reference.md first. Open the full Notes_On_functionality.md only when the quick reference does not settle the question or the agent spec explicitly requires a section of it.