跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
浏览器与自动化

blacksmith-orchestrate

Orchestrate many forge runs at once: analyze which tasks collide on the same files, schedule the colliding ones sequentially, and run the rest in parallel worktrees at the right model tier and review depth. Use when the user runs /blacksmith-orchestrate or asks to ship several issues, a milestone, or a written implementation plan in one go.

radimsem · 已有 AI 审核记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add radimsem/forge-skills --skill blacksmith-orchestrate

技能说明


name: blacksmith-orchestrate description: "Orchestrate many forge runs at once: analyze which tasks collide on the same files, schedule the colliding ones sequentially, and run the rest in parallel worktrees at the right model tier and review depth. Use when the user runs /blacksmith-orchestrate or asks to ship several issues, a milestone, or a written implementation plan in one go."

Blacksmith Orchestrate

Decide which tasks run, in what order, at what depth, by which model, in which worktree — then dispatch one forge run per task.

Overview

This is a wrapper around forge, not a fork of it. It never reimplements a forge step: every task it dispatches runs the ordinary twelve-step forge workflow, and everything this skill adds is a decision about those runs — which ones may run at the same time, how deep each one goes, and where each one's work lives.

Nine numbered steps, split by one gate, deliberately mirroring forge's shape:

Part 1 — Plan (1–5):     normalize → materialize → analyze → schedule → [GATE] battle plan
Part 2 — Execute (6–9):  provision → dispatch waves → relay → close-out

  gate held: no worktree is created and no forge Part 2 runs until "yes, forge them"
StepAction
1Parse the invocation; split orchestrator flags from pass-through forge flags; route the work source into a canonical task list
2Materialize refs — spec-file and free-form routes get real issue refs via /to-tickets, or synthetic T1..Tn IDs if the user declines filing
3Analyze — route each task to the plan-sourced path or the scout fan-out
4Build the collision graph, triage matrix and worktree grouping → scheduling waves
5[GATE] one battle plan, one approval
6Provision worktrees; write the run ledger
7Dispatch waves — one agent per task at its assigned model tier and forge depth
8Relay — prove a blocker's changes are on the dependent's base, rebase, release
9Close-out — aggregate report, PR list, worktree cleanup reminders, self-evolution

Forge's hard floors are inherited unchanged, and orchestrating many runs never relaxes them: every run's Step 12 /goal verification still has to pass before that run assembles a commit, and no dispatched run ever writes back to the source issue or ticket it was dispatched for — that stays with the user under every flag combination, afk included. That is narrower than "never writes to a tracker": Step 2 files new issues via /to-tickets for the plan and free-form routes, including under automode, exactly as forge's own Step 10 sanctions filing spin-off issues it did not author — see references/entry-routes.md. Whether a dispatched run may also commit and open a PR for its own task is a separate question, resolved three ways: by default, the Step 5 "yes, forge them" approval is the Step 12 commit-and-PR selection forge requires, granted once for every task the approved plan named (see Step 7); under automode without afk, nobody made that selection, so every dispatched run stops at its own Step 12 plan-only output, exactly as a standalone automode forge run does, and nothing is pushed; afk is the single sanctioned exception, authorizing unattended commit-and-PR as well as the blocking-PR merge it already governs — it is labelled that way everywhere it appears, so the exception stays auditable instead of becoming a quiet contradiction. Terminal PRs and worktree cleanup stay with the user under every flag combination, afk included.

"The agent" means whatever agent runs this skill, and "one agent per task" means whatever subagent or parallel-run mechanism the runtime exposes. Adapt every reference — config directory, agent guide, interview UI, host CLI — to your runtime, exactly as forge does. If the runtime cannot run work in parallel at all, say so and fall back to running the waves sequentially rather than pretending to fan out.

Runtime

This skill is Claude-Code-first, degrading elsewhere in the same way forge's codex and coderabbit flags do: each surface below is used where the host provides it, and the workflow still completes, one step slower, where it does not.

SurfaceUsed forWhy
Workflow toolStep 3b scout fan-outbounded parallel fan-out, schema-validated returns, keeps N proposals out of the orchestrator's context
Agent toolStep 7 implementation dispatchper-agent model selection, and the human gates and multi-hour parks live in the main loop
Ledger fileSteps 6–9 statesurvives /compact, crash and resume

This is the same runtime-parallelism degradation the Overview already states — "If the runtime cannot run work in parallel at all, say so and fall back to running [...] sequentially rather than pretending to fan out." This section does not restate that rule independently; it is the same rule applied to the three Claude-Code surfaces above. Concretely here: on a non-Claude-Code runtime the skill degrades with a one-line warning to sequential forge runs in dependency order, one worktree per component, no afk, and inline analysis instead of a scout fan-out.

Parameters

Parse the invocation as /blacksmith-orchestrate <work-source> [orchestrator-flags] [forge-flags]. Flag words are orthogonal and can appear anywhere in the request, exactly as in forge. The <work-source> is one of five routes — one or more issue refs, a tracker container such as a milestone or epic or label, plan <path>, a quoted free-form goal, or resume — and routes may combine in a single invocation, so every task carries the route it arrived by. Read references/entry-routes.md for the route table, how the routes normalize into one task list, and how containers expand.

Full flag matrix (effects, composition rules, conflicts): references/flags.md.

Orchestrator flags

These are consumed by t

阅读完整来源 ↗