Dispatches GitHub issue queue tasks to parallel subagents with memory caps and a pinned status panel.
npx skills add https://github.com/theitush/yajna --skill orchestrate-theitush
Please help me install this Agent Skill. Skill: orchestrate Source: https://github.com/theitush/yajna/tree/main/.claude/skills/orchestrate Command: npx skills add https://github.com/theitush/yajna --skill orchestrate-theitush
systemd-run with MemoryMax and MemorySwapMax=0 cgroup caps sized from measured peaks, so an overrun kills only that job instead of the whole session tree. - Live status panel and report footer: The orchestrate-status tool renders a pinned Running / Planned / Done table with per-worker token counts, countdown ETAs, and clickable transcript links, then prints the same table as the final report footer. - Use Case: Given five open issues on the yajna repo, run a pass that works three of them simultaneously in separate subagents — each with its own path ownership, model, and memory cap — verifies and commits their results by path, and closes with a priced backlog and a full timing report. ## Quick Start Invoke /orchestrate to work this repo's issue queue as a dispatcher with parallel subagents and a pinned status panel.High-intent search queries and answers about installing and using this skill.
Split the task queue by the file paths each task touches: tasks with disjoint paths run simultaneously in separate subagents, while tasks sharing paths run serially in one lane. Each spawn must name the paths the worker owns and the paths it must not touch.
Run every heavy job under systemd-run with a cgroup memory cap: `systemd-run --user --scope -p MemoryMax=<N>G -p MemorySwapMax=0 -- <command>`. A capped job that overruns dies alone with exit 137 instead of triggering the OOM killer against the whole session tree.
Run the command under `/usr/bin/time -v` and read the `Maximum resident set size` field from its output. Record that number in the task result so future runs size their cgroup cap from a measured peak rather than an estimate.
Only when their file paths are disjoint, and commits must be made by path, never with `git add -A`, since the tree is shared. Where paths cannot be separated, spawn the worker with worktree isolation and merge its branch when it reports.
Restating creates a second copy that can drift from the original ask, and transcription errors in the copy send workers building the wrong thing. The brief should point the worker at the issue body's Ask block and add the dispatcher's own interpretation only as clearly marked separate text.
AutumnsGroveOrganize GitHub project board issues by assigning sizes, priorities, and milestones.
solo-istAudit GitHub backlogs and project boards for stale items and drift.
ByronWilliamsCPATrack Claude Code token usage across local JSONL transcripts.
tusososDispatch fresh subagents per task to isolate context and run two-stage reviews.