aibridge-workflow-orchestration

name: aibridge-workflow-orchestration description: AIBridge workflow and multi-agent orchestration guidance. Use when Codex needs to design, review, or execute a multi-agent workflow plan, split Unity work into parallel or pipeline agent roles, define structured workflow artifacts, choose between batch/multi automation and agent orchestration, run adversarial verification, investigate Runtime debug evidence, sweep multiple Runtime targets, or prepare AIBridge workflow recipes

AIBridge Workflow Orchestration

Focus on workflow recipes, multi-agent orchestration, parallel review, pipeline validation, adversarial verification, Runtime debug investigations, Runtime target sweeps, or structured workflow artifacts

Leave routine single-file edits, simple CLI command lookup, and ordinary Unity validation to aibridge-development-workflow and aibridge

Core Rules

• Keep orchestration explicit: phases, roles, dependencies, gates, artifacts, and expected outputs.
• Keep execution visible: when running a workflow, use short status blocks with explicit labels, 【模式：...】 for the active business mode and -> ... for phase/step progress.
• Prefer parallel read and serial write. Use parallel write only when worktree isolation, file ownership, merge strategy, and validation gates are explicit.
• Use pipeline for staged per-item work. Use parallel barriers only when a downstream step needs all upstream results.
• Use structured outputs for intermediate results: findings, verdicts, plans, patch proposals, validation results, artifact references, and Runtime target references.
• Treat Skill routing as preflight; scope phase-specific Skills to the phase that needs them, then pass compact handoff summaries and artifact references across phase boundaries.
• Separate claims from evidence. Treat AIBridge CLI, Runtime, screenshots, logs, tests, and Code Index output as evidence.
• Do not describe workflow as a generic AI agent scheduler. Current CLI support covers recipe list/validate/plan/init, active run begin/attach/finish, deterministic run-cli steps, ordinary-command artifact attach, external result import, adapter export, gates, and reports; agent and manual steps require an external executor.

Reference Loading

• Read references/orchestration-patterns.md before selecting parallel, pipeline, barrier, adversarial, or Runtime sweep patterns
• Read references/recipe-schema.md before writing or reviewing a workflow recipe document
• Read references/evidence-schema.md when defining or importing structured external results
• Read references/builtin-recipes.md before drafting common Unity implementation, review, Runtime validation, Prefab sweep, or bug-hunter workflows
• For cross-turn, resumed, or external-agent work, follow the active-run, evidence schema, and import rules in references/orchestration-patterns.md and references/recipe-schema.md