tzachbon/srulik-toolkit / plugins/srulik-toolkit/skills/agent-swarm
agent-swarm
VerifiedCoordinate delegated work through the child-agent capabilities available in the current harness. Use when the user asks to spawn agents, delegate, swarm, parallelize, race approaches, or when active instructions require an orchestrator and worker split. Does not require Traycer, a specific provider, or a fixed model.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 14.0 KB |
| agents/openai.yaml | config | 208 B |
| graphs/pr-review.yaml | config | 1.1 KB |
| graphs/task.yaml | config | 8.8 KB |
| references/graph-spec.md | markdown | 19.5 KB |
| references/host-usage.md | markdown | 246 B |
| references/live-view.md | markdown | 2.6 KB |
| references/t3-code.md | markdown | 3.4 KB |
references/t3-code.md
View on GitHub# Graph mode on T3 Code
This maps the generic steps of graph mode onto T3 Code tools.
| Step | T3 Code |
| --- | --- |
| List runnable models | `orchestrator_capabilities`. Write `[{provider, model, effort}]` to `<run-dir>/candidates.json` and pass it as `pick --candidates <run-dir>/candidates.json` for the providers that can run children. Use each provider instance ID as `provider`, and the effort you intend to use. |
| Read quota | Run `aa-models usage > <run-dir>/usage.json` and pass `--usage <run-dir>/usage.json` to `pick`. Where usage needs root, as on hosts whose secrets helper is root-only, use `sudo -n aa-models usage`. |
| Delegate a task | `delegate_task` with `target.providerInstanceId`, `target.model`, and the effort in `target.options` (the option ID that `orchestrator_capabilities` lists for that model), all taken from `pick`. Use `mode: "wait"` for one task, or `async` for a batch. Ask the child for JSON matching the node's `out`. |
| A child can't start | Record `done` with `failed` and `final: true`. |
| Open the live page | Start `graph.js view <run-dir> --port 47812` as a background process that outlives the turn, such as `nohup ... &`, and keep it running after the run ends. It listens on the machine's Tailscale address, so the printed link opens from the owner's other devices; on a machine without Tailscale it stays on loopback. Pass `--host <address>` only to pick another private address the owner reaches. Never use `0.0.0.0` or a public address. Give the owner the printed page URL. T3 lists the server under Servers, and clicking it opens the page. Also open it with `preview_open`: pass the printed page URL as `url` when it is not loopback, because the server then listens only on that address; on loopback, use `target: {kind: "environment-port", port}`. Every run uses port 47812, so a machine has at most one view server, showing the latest run. Before starting one, stop the `graph.js view` process that holds port 47812: take its PID from `ss -ltnp 'sport = :47812'`, confirm with `ps -o args= -p <pid>` that it runs `graph.js view`, then `kill` it. Avoid `pkill -f 'graph.js view'`, which also matches the shell running it. Reuse the same tab. Never stop a process on that port that isn't a graph view. If `preview_open` fails, report its exact error and the page URL instead. For example, a T3 server running as root makes Chrome refuse to start its sandbox. In that case, never change or restart the T3 server to open the page. |
| Daily refresh | `schedule_task`, daily: `node <plugin root>/scripts/graph.js refresh`. |
| Weekly thread scoring | `schedule_task`, weekly, read-only. Run `graph.js watermark` and list threads updated after its `since` (not since the last job run, so threads held back behind an undecided one are listed again) with `t3_thread_list` (`allProjects: true`) and read each with `t3_thread_read`. Have the cheapest suitable model label each thread as `{updated, model, kind, signals}`, then run `graph.js score <threads.json>` with the file of labels. Keep no thread text. |
Observed on T3 Code on 2026-11-10: `pr-review.yaml` ran on a real pull request with children on Claude (`claudeAgent`), Kimi (`oh-my-pi`) and Gemini (`antigravity`), picked by `pick` and delegated with `delegate_task`. Quota came from `sudo -n aa-models usage`. The live view was served by `graph.js view`; T3's own preview could not open it because the server runs as root, so a headless Chromium captured it instead.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 471k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |