OpenCode-StoryForge is the editorial swarm plugin for OpenCode that orchestrates a curated team of AI agents to draft, research, revise, and polish multi-stage writing projects. Designed with GitHub and npm users in mind, it markets a catchy collaborative workflow, pro-level polishing crew, and clear configuration knobs while still linking down to the technical usage sections below.
- Editor-in-Chief: Orchestrator, direction setter.
- Writer: Draft creator.
- Researcher: Fact gatherer.
- Section Editor: Structure reviewer.
- Copy Editor: Language polisher and AI slop remover.
- Fact Checker: Verification expert.
- Reader Advocate: Audience representative.
OpenCode surfaces editor_in_chief as the primary role inside the UI, so you can pick it from the agent selector just like any other OpenCode plugin. When you kick off the Editor-in-Chief, StoryForge inherits whichever model you selected for that session (the same behavior as opencode-swarm), and you can still override the model via opencode-writer-swarm config if you need a different fallback.
- Add the plugin to your
opencode.json. - Configure models in
~/.config/opencode/opencode-writer-swarm.jsonor.opencode/opencode-writer-swarm.json. - Invoke
@editor_in_chiefwith a writing request.
- Brief: Editor-in-Chief creates a brief.
- Research: Researcher gathers facts.
- Plan: Editor-in-Chief creates a content plan.
- Draft: Writer produces the first draft.
- Review: Section Editor, Copy Editor, Fact Checker, and Reader Advocate review the draft.
- Polish: Copy Editor does a final polish.
- Delivery: Final output saved to
.writer/final/.
| Setting | Description |
|---|---|
config_validation_enabled (CONFIG_VALIDATION_ENABLED) |
Enables the prototype-pollution guard in deepMerge. Set to false to fall back to the previous behavior. |
FILE_VALIDATION_ENABLED |
Controls symlink/size/depth validation for .writer files. Disable to revert to the legacy permissive reader. |
FILE_RETRY_ENABLED / WRITER_MAX_RETRIES |
Toggles exponential backoff retries for writer file writes and limits the number of retries. |
LOG_REDACTION_ENABLED |
When true (default), startup logs redact keys ending in _KEY, _SECRET, or _TOKEN. Setting to false temporarily disables redaction for debugging. |
VERBOSE_INIT / LOG_LEVEL=debug |
Emit detailed initialization metadata (agent count, sanitized config keys) during plugin startup. |
The plugin exposes a /swarm command namespace for inspecting and managing the swarm state:
/swarm diagnose– Runs health checks on.swarm/plan.md,.swarm/context.md, and the plugin config./swarm export– Emits the current plan/context bundle as a JSON snapshot for backups or migration./swarm reset --confirm– Securely deletes.swarm/plan.mdand.swarm/context.mdafter a confirmation warning, leaving the workspace in a clean state.
Each slash command validates all arguments and file paths before performing I/O, matching the guardrail philosophy outlined below.
Guardrails ensure no single session exceeds configured limits. The defaults (exposed under guardrails in your config) are:
max_tool_calls: 200max_duration_minutes: 30max_repetitions: 10max_consecutive_errors: 5warning_threshold: 0.5
Warnings fire when you cross 50% of a limit (logged as Guardrail warning: Approaching tool call limit). When limits are exceeded the guardrail hook throws to halt further tool execution and protect the agent loop.
Context-budget warnings continue to execute through experimental.chat.system.transform, ensuring the architect agent receives system-level alerts when plan/context token budgets reach 0.7 (warning) or 0.9 (critical) of the configured window.
- Slash commands for
/swarm diagnose,/swarm export, and/swarm reset --confirm. - Guardrail engine enforcing tool-call/duration/repetition/error limits with configurable warning thresholds.
- Evidence store keeping
.swarm/evidenceshrink-wrapped with retention and guardrail-aware reads. - Documentation, tests, and build/typecheck scripts locked in for a full verification phase.