Dynamic Workflow Forge

Forge or review Claude Code dynamic workflows: JavaScript orchestration scripts that coordinate many subagents in the background, keep intermediate state in script variables, and return one consolidated result. Use this skill for large-scale audits, migrations, cross-checked research, and adversarial review patterns where the orchestration itself must be readable, repeatable, and resumable inside the same Claude Code session. [DOC]

Do not use this skill for static slash-command workflow definitions. If the target is a Markdown command, .claude/commands/*.md, or a phase-map prompt without the Agent SDK Workflow tool or dynamic JavaScript runtime, route to workflow-forge instead. [EXPLICIT]

Required Inputs

• Task objective, target paths or corpus, and expected final report or change.
• Scale signal: approximate file count, issue count, source count, or agent fan-out need.
• Runtime gate: Claude Code version, surface, provider route, and whether dynamic workflows are enabled.
• Risk controls: approval mode, tool allowlist, max spend or slice-first budget, model routing, and file-edit boundaries.
• Reuse target: one-off run, saved project workflow, or saved personal workflow.

Core Decision

Choose a dynamic workflow only when code-held orchestration is the advantage.

Forge Procedure

1. Gate the Runtime

• Require Claude Code v2.1.154 or later for dynamic workflows.
• Require Agent SDK v0.3.149 or later when authoring or invoking the Workflow tool from SDK code. coverage_gap [INFERENCIA]: este número de versión y los helpers agent()/parallel()/pipeline() NO están documentados en code.claude.com/docs/en/workflows (2026-06-20); confirmar contra release notes o el repo del Agent SDK antes de depender de ellos.
• Confirm the provider or plan route: paid Claude Code plan, Anthropic API, Amazon Bedrock, Google Cloud Vertex AI, or Microsoft Foundry.
• Confirm workflows are not disabled by /config, disableWorkflows, managed settings, or CLAUDE_CODE_DISABLE_WORKFLOWS=1.
• Confirm ultracode only when the selected model exposes xhigh effort.

2. Classify the Run Shape

• Use direct opt-in when the user says "use a workflow", "run a workflow", or includes ultracode.
• Use /effort ultracode only when every substantive task in the session should be eligible for workflow planning.
• Use saved workflows when the same orchestration should recur; save from /workflows to .claude/workflows/ for project reuse or ~/.claude/workflows/ for personal reuse.
• Use args for saved workflow input such as paths, issue IDs, source lists, or review criteria.

3. Design the Script Contract

• Make the script coordinate agents; do not put direct shell or filesystem work in the workflow script itself.
• Keep intermediate evidence, votes, and retry state in script variables.
• Cap concurrency at 16 agents or lower and cap total agents well below 1,000.
• Split stages that need human sign-off into separate workflows because mid-run user input is not available except permission prompts.
• Add /workflows management notes: view phase progress, inspect agents, pause or resume with p, stop with x, restart an agent with r, and save with s.

4. Add Cross-Checks

• Use independent agents for discovery slices, then separate verifier agents for adversarial review.
• Require each high-impact claim or proposed change to survive at least one independent check.
• For research workflows, require source-backed claim voting before synthesis.
• For code workflows, require targeted tests, static checks, or reviewer votes before final edits are reported as ready.

5. Control Cost and Failure

• Start with a narrow slice before whole-repo or whole-corpus runs.
• Route lower-risk phases to smaller models when acceptable; reserve the strongest model for synthesis, architecture, or adversarial judgment.
• Pre-approve known-safe tools before long background runs; unallowlisted shell, web, or MCP calls can still prompt or fail depending on surface.
• Make resume scope explicit: paused runs can resume in the same session, but a new Claude Code session starts the workflow fresh.

Deterministic Validator

Use scripts/compile-dynamic-workflow-forge.py when a structured workflow spec must be validated or rendered reproducibly.

Run desde la raíz del skill (las rutas son relativas a ese directorio):

cd "$CLAUDE_PLUGIN_ROOT/skills/dynamic-workflow-forge"
python3 scripts/compile-dynamic-workflow-forge.py \
  --input scripts/fixtures/dynamic-security-audit.json \
  --output /tmp/dynamic-security-audit.md

Load assets/dynamic-workflow-policy.json for local gate rules, assets/dynamic-workflow-schema.json for the structured spec shape, and references/domain-knowledge.md for detailed routing guidance.

Output Requirements

• State whether the result is a dynamic workflow, a saved workflow, or a validation report.
• Include runtime/provider gates, orchestration phases, subagent roles, cross-check pattern, /workflows controls, save location, limits, cost controls, and static-workflow boundary.
• Mark missing runtime facts as coverage_gap instead of assuming availability.
• Keep any produced workflow script reviewable before launch and explicit about permissions.

Validation Gate

• Dynamic-workflow route is justified over subagents, skills, teams, and workflow-forge.
• Claude Code and SDK version gates are stated when relevant.
• Provider, plan, and disable-setting gates are checked or marked coverage_gap.
• Script contract coordinates subagents and excludes direct shell/filesystem work from the script body.
• /workflows management, saved workflow locations, args, resume scope, and background behavior are covered.
• Agent counts respect 16 concurrent and 1,000 total limits.
• Cross-check or adversarial review pattern is present for high-impact work.
• Cost controls include slice-first execution and model-routing guidance.
• Static slash-command workflows are routed to workflow-forge.

Reference Files

Related katas (toolkit)

• katas-deterministic-agent-loop
• katas-prefix-caching
• katas-session-resume-fork

Contract

• Aceptación: workflow JS determinista, resumible, con estado en script y fan-out acotado. [EXPLICIT]
• Límites: orquesta; la lógica de cada paso vive en agentes/skills. [EXPLICIT]
• Casos borde: gate de versión Agent SDK; helpers agent()/parallel()/pipeline() no documentados en workflows.md [coverage_gap]. [EXPLICIT]
• Supuestos: [SUPUESTO] Workflow tool disponible (requiere probe; ver §1 Gate the Runtime). [SUPUESTO]
• Trade-off: workflow vs equipo/subagente — determinismo y resume a cambio de setup. [EXPLICIT]

Packet

Capas del packet, cargables bajo demanda (disciplina ICM: una capa por vez, nunca todas juntas): references/ guías de profundidad (cargar UNA por etapa) · knowledge/ cuerpo de conocimiento · prompts/ prompts listos · examples/ salida de ejemplo · agents/ subagentes del packet · templates/ plantilla de output · scripts/ automatización local · assets/ recursos estáticos.