Index of every page on this site: llms.txt
All skills
vercel-labs avatar

/vgpu-agent-flow

@773351f
by Vercel Labsvercel-labs/vgpu2.7k stars
142
Run it yourself

Lead workflow for developing vgpu features and Blender assets with the repository's subharness specialists. Use for agent-based research, API design, implementation, shaders, materials, simulations, or ImageGen-guided asset iteration, and when the human mentions the agent flow or subharness team.

Use this Skill: https://skilld.dev/gh/vercel-labs/vgpu/vgpu-agent-flow

Nothing lands on disk. Nothing to clean up.

Fork this Skill

Edit a local copy. It keeps the author and licence.

referencesasset-iteration.md

≈2.6k tokens on demand. Your agent reads this file only when SKILL.md points to it.

Blender asset iteration

Use this lane when the user requests an authored Blender asset, screenshot-driven refinement, or an ImageGen concept translated into a runtime example. Follow the selected repository workflow; reuse the user's creative delegation instead of inventing a public-API approval stage.

repo:asset-author runs Astra (gpt-6-astra) with declared Opus 5.5 children: subagent:producer handles LOW meshes, bakes and packaging when production is assigned; subagent:runtime specializes in shaders and disjoint browser rendering files; subagent:critic reviews actual captures and costs. These roles intentionally have no model fallback: report unavailable requested models. Reading a skill does not grant a CLI child access to the lead's image-generation tools.

Astra owns modeling, Blender renders, authored materials, environment design, camera composition and intended lighting. For web delivery or optimization, read asset production: the producer derives LOW and its bakes from a pinned HIGH source. The author can handle a small combined assignment itself when requested; avoid overlapping ownership. It uses actual Blender renders to refine concepts and directs the next visual change. The runtime specialist owns runtime PBR and baked-lighting composition, sky/water shaders, reflections, shadows and color-management parity with that scene. If native Blender execution requires the lead, Astra still specifies the render and inspects the returned images; the lead provides execution support without transferring artistic ownership to the runtime child.

Before launch, put the scope, file ownership, target renderer/cameras, output paths, budget and iteration cap in .context/work/<topic>/brief.md and decisions.md. Numeric budgets belong to the task. If omitted, use three revisions with at most two critic passes per revision. Inspect the topic's recorded lesson evidence when available; don't blindly copy past aesthetics. Read skills/vgpu/SKILL.md and its Blender references, verify the installed Blender version, and resolve runtime APIs through the target package's installed docs.

pnpm exec subharness check repo:asset-author
pnpm exec subharness check repo:asset-runtime
pnpm exec subharness check repo:asset-critic
pnpm exec subharness run repo:asset-author --prompt-file .context/work/<topic>/asset-task.md

Use native background commands or detached tasks as described in the parent skill. Check every terminal outcome, including declared children. A readiness check is not proof of tool execution. Keep the author handoff compact: delegate detailed runtime diagnostics and full capture review to the assigned children, then read their result summaries and selected decisive images. Preserve comprehensive evidence on disk instead of repeatedly loading every image and manifest into the parent. If the harness fails with an input/context limit, retain completed artifacts and account for cancelled children's partial writes before starting a fresh session with a concise handoff; retrying the same oversized session does not recover it.

The lead generates a concept with the built-in ImageGen tool when requested, saves it in the topic, and passes its path. The author, or the lead after a reported native block, supplies actual captures; record who produced them. Inspect a capture before using it as an ImageGen edit target; label its role and preserve silhouette, camera, scale and attachments unless intentionally changing them. Save the reimagined image separately as a concept. Translate selected changes into real geometry/materials, then capture the implementation again. Never substitute a generated image for proof of a working model.

Use a file handoff while the author is active: iteration-status.md records revision, child IDs, actual output paths and readiness/blockers; image-request.md records the edit target, prompt, preserved invariants and desired result path. The lead reads these and sends the generated path with subharness send <author-session> --delivery steer --prompt-file <response-file>. Use queued follow-ups when idle. A final response also works when no independent work remains. Do not assume the author can message the lead without this file/CLI handoff.

Give each shared handoff one writer. When the lead owns preview servers, the runtime child writes its build request and the lead writes a separate current-service record with build hashes, asset paths and ports. Read that record before captures and final reporting; do not append replies to a request another agent will regenerate. Keep superseded build manifests and failed captures intact. A queued child follow-up is not evidence that the running child has received a newer server state.

Use queue for follow-ups to the Opus runtime and critic children. In the current Subharness Claude adapter, steer means interrupt: active work stops and a replacement task starts with partial file effects preserved. Reserve it for deliberate interruption and inspect the reported delivery mode and terminal outcomes. Lead-to-Codex-author steering supports in-place delivery.

Probe native execution before a long run. Blender may fail during Metal initialization in a macOS child sandbox even though --version succeeds. Report the log and exact command instead of retrying unchanged or broadening permissions. The lead can run Blender within its existing authorization and return captures. Use --python-exit-code 1, check completion logs and validate serialized assets: a process exit alone is insufficient evidence.

When agents continue editing during a native run, execute a verified source snapshot. Copy the entry point and its local imports into a new directory, compare every copied hash with the author's handoff, and keep that directory unchanged until completion. Record external inputs separately. Release the editable sources after verification; a queued steering message alone does not guarantee that an agent's in-flight edits have stopped. Check snapshot hashes again at completion, and retain failed scene checkpoints before retrying a revised source. When a copied helper imports installed packages, verify resolution from the snapshot before a long bake. Copying the helper can change Node's package lookup path. Preserve the installed dependency identity and use an explicit process-scoped lookup path when needed; a successful import from the original directory does not test the copied execution context.

The runtime child may capture through a browser tool or an installed agent-browser CLI using its Bash access. If unavailable, it reports browser capture: not performed (<reason>) and the exact server/capture steps; the lead owns the outstanding check. Record the actual renderer backend and test controls as well as saving the screenshot. A WebGL fallback does not prove WebGPU output.

For each iteration retain:

  • source revision or parameters that reconstruct the previous asset;
  • actual hero, rear and relevant close-up captures under recorded camera/light/pose settings;
  • export triangles, vertices, primitive/material counts, file bytes and texture residency;
  • runtime validation and measured draw counters, separately from offline estimates;
  • critic findings, chosen fixes and the next measured result.

Each critic handoff includes both the previous and current captures/metrics, or explicitly says the first revision has no implementation baseline. Generated references remain separately labeled.

Stop when the caller's visual requirements and measured budgets pass, or at the agreed iteration cap with explicit remaining findings. Do not add an unattended infinite improvement loop. Keep reproducible source and runtime assets outside scratch; save screenshots under .context/.

A cap is an execution limit, not a visual acceptance criterion. Keep a REVISE verdict when reference fidelity remains unmet; never convert it to PASS merely because topology checks pass or the revisions are exhausted. A user's rejection or changed brief reopens earlier acceptance. For close-up or architectural work, choose texel density from the required view before baking: tiled material detail and unique lighting UVs can preserve detail more economically than one low-resolution unique atlas. Requested baked AO/lightmaps need actual bake provenance and matched runtime captures with each contribution disabled independently.

For atlas repairs, compare the full matched views as well as the original defect crops. A new layout may repair one edge while making another exposed face black. Minimum chart-box width, positive triangle area and zero overlaps do not establish adequate raster coverage of visible subregions. Use AO-off comparisons to locate candidates, then inspect the actual texture and surface visibility to distinguish a bad bake from legitimate contact occlusion. Keep the candidate out of delivery while confirmed regressions remain.

When changing terrain beside water, verify actual mesh contacts and its intersection with the water plane. Manifold closure and source height samples are useful structural checks, but can pass while the rendered coast contains unintended submerged pockets. Inspect elevated and shoreline views and measure the resulting footprint before regenerating a shoreline field.

Improve the team from observed results

Apply the repository agent handoff contract to author/runtime/critic coordination. Keep coordination lessons in this pipeline and its agent definitions; put transferable modeling, baking and runtime-validation guidance in skills/vgpu/blender/. Do not make the portable Blender skill depend on Subharness roles, handoff files or repository-specific services.

Write candidate lessons in .context/work/<topic>/lessons.md: evidence path, observed symptom, cause, intervention, matched before/after result and scope of applicability. The lead verifies them and makes a narrow change to the relevant agent or this guide when the evidence justifies it. Promote durable lessons with a reproducible procedure or a source fixture included in the same delivery, so a later agent can recheck them without private task artifacts. Keep example-specific case studies with their source evidence; do not make this lane depend on an unpublished example. Reject unverified claims and task-specific style preferences masquerading as general rules. This improves instructions and examples, not model weights.

Source: SKILL.md on GitHub

skilld matched fixed text patterns in SKILL.md and file names. Patterns miss obfuscated code.

skilld run checks every file with the same patterns. It asks for approval before it loads a Skill with a behavior marked Needs approval.

No alerts22d3 checks · Risk SAFE
  • Gen Agent Trust Hub22d

    This skill provides a comprehensive workflow for graphics development and Blender asset creation by coordinating specialized sub-agents. It follows secure practices for managing environment configuration and includes clear boundaries for task delegation.

  • Socket22d

    No alerts

  • Snyk22d

    Risk: LOW · No issues

Signed by skilld at 773351f. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub last month.

Steadyupdated last month

README badge

README badge for vercel-labs/vgpu/vgpu-agent-flow