菜单
Drive the obtnet fleet (distributed build/run/sync across machines) via obt.net.py — node routing, async jobs, artifact fetch, git-free tree sync. Use whenever work should run on another machine or files must move between machines without git.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Orchestrate multi-agent engine work on orkid — decomposing milestone specs into lanes, assigning machines/branches, briefing implementer and gate-runner subagents, reviewing, adjudicating gates, and integrating. Use when coordinating subagent-based implementation of a spec/milestone rather than implementing directly.
How to author procedural content in orkid using HyperSyn (Reality Synthesizer) — a Python-as-DSL system for composing particle systems, 2D/3D procedural textures, mesh/SDF operations, and (eventually) audio/sequencer graphs into compact runtime experiences. Use when the user asks about HyperSyn, ptex2d, ptex3d, hypermesh, dataflow DSL, FXV2 codegen from dataflow, subprocess graph validation, or the hypergraph composition layer.
Operate THIS machine as an obtnet fleet node — launching the node daemon correctly, what the controller may do to this machine (jobs, builds into the staging, checkout gitsync/sync), node-side etiquette and troubleshooting. Use when setting up, running, or debugging the node side of the fleet (the controller-side skill is `obtnet`).
Answer questions about orkid's OBT build system, ork.build.py, CMake structure, module layout, ork.python environment, staging directories, pyext builds, and environment variables. Use when the user asks about building, compiling, CMake, or project structure.
Answer questions about orkid's ComputeInterface (CI), compute shader dispatch, indirect dispatch, SSBO binding for compute, memory barriers, image/sampler binding, dispatch phases. Use when the user asks about compute shaders, GPU compute, or dispatch.
Answer questions about orkid's compositor system, CompositingData/Scene/Technique, render nodes (Forward/Unlit/Picking), post-FX nodes (ACES/HSVG/Bloom/User), output nodes (Screen/RtGroup/VR/File), presets (ForwardPBR/Unlit/Picking), RtGroup render targets, and Python compositor bindings. Use when the user asks about compositing, post-processing, render targets, or rendering presets.
| name | obtnet |
| description | Drive the obtnet fleet (distributed build/run/sync across machines) via obt.net.py — node routing, async jobs, artifact fetch, git-free tree sync. Use whenever work should run on another machine or files must move between machines without git. |
Every command prints ONE greppable verdict line last: [obtnet] ok|FAIL <verb> <node> ....
Trust it. Do not fetch or dump logs on success. --json on any verb for tool parsing.
Controller resolution: --controller > $OBTNET_CONTROLLER > ~/.obt-global/obtnet.json
(set once with obt.net.py config tcp://HOST) > localhost. Bare HOST or HOST:PORT are
accepted anywhere an address is (port defaults to 7461).
obt.net.py list shows the fleet. Any <node> argument accepts a literal name or a
@ selector, routed to the least-busy live match:
@any · @linux / @mac · @gpu · @gpu=5090 (capability substring) · @5090
(substring across name/os/arch/gpu) · comma-composed: @linux,gpu.
FAN-OUT on run/build/test/scene/sync/gitsync: @each (every live node), @each:linux
(every match), or a,b lists — execution verbs submit to ALL then wait (simultaneous
cross-machine builds/tests), working-copy verbs loop; per-node verdicts + a fleet summary.
obt.net.py list # fleet + load + clock offsets
obt.net.py run <node> [--timeout S] [--cwd D] -- CMD... # short sync command (<~60s)
obt.net.py submit <node> [--out GLOB] -- CMD... # long work -> job id
obt.net.py build <node> [--cwd D] # ork.build.py (or -- CMD);
# verdict: errors=N warnings=N
# (error-grep beats lying rc0)
obt.net.py test <node> -- CMD... # verdict: passed=N failed=N
obt.net.py scene <node> [--out G] [--env K=V] -- CMD... # OFFSCREEN forced unless
# --windowed (owner consent);
# cook/settle/fps lines surfaced
obt.net.py wait <node> j-XXXX # verdict; auto stderr-tail on fail
obt.net.py log <node> j-XXXX [--stderr --tail N --grep P] # bounded peek at a LIVE job
obt.net.py jobs # fleet-wide job table
obt.net.py cancel <node> j-XXXX # SIGTERM the job's pgid
obt.net.py fetch <node> <sha256> <dest> # artifact by sha (from verdicts)
obt.net.py sync <node> <local_dir> <remote_dir> # make remote == local (git-free)
obt.net.py sync <node> L R --pull # make local == remote
obt.net.py gitsync <node> <local_repo> <remote_repo> # align git base (branch+HEAD)
# via bundle, NO push; --tree
# chains the uncommitted delta
obt.net.py diff <node> <local_dir> <remote_dir> # +/-/M/L classified; rc0=equal
obt.net.py watch [--node N] [--grep P] # live fleet event tail (ctrl-c)
Nodes run inside OBT shells, so staged tools (ork.build.py, obt.dep.*, ork.cpp.db.*)
work directly: obt.net.py run @linux,gpu -- ork.cpp.db.search.py MySymbol --porcelain.
submit (never run) and give an honest --timeout.gitsync it (aligned bases = diffs/patches
line up; it fails loudly on divergence or tracked dirt), then --tree/sync for the
uncommitted delta. Never assume the node is current.sync dry-runs first when the target matters: --dry (counts only); --delete is opt-in.
Verdict tree=MATCH is cryptographic proof both sides are equal — nothing else to check.jobs before heavy work on a busy node.log --grep, never full dumps.