Local GPU Imagegen v0.9.1

Local GPU Imagegen v0.9.1

v0.9.1 completes the release-identity sync after the repository moved from
zc4578980-tech to ChevalGrand520: the packaged version, release-candidate
contracts, tests, and active documentation all agree on 0.9.1, and every
published artifact (PyPI wheel, MCP Registry descriptor, Git tag) points at the
same commit and wheel identity.

Highlights

• Release-identity sync: the packaged __version__, release-candidate
  expected constants, test contracts, and active docs were aligned to 0.9.1
  in one commit; historical v0.9.0 records remain untouched.
• Explicit safety-boundary tool definitions: confirmation, identity,
  idempotency, route, and plan fields now carry parameter descriptions without
  changing any name, type, or requirement.
• Machine-readable no-eligible-model result: recommend reports a hint
  field when no eligible model exists instead of an empty failure.
• Migrated repository identity: the git remote now points at
  https://github.com/ChevalGrand520/local-gpu-imagegen; historical
  zc4578980-tech references remain only in docs/evidence/ audit records.

Install

uvx --from local-gpu-imagegen==0.9.1 local-gpu-imagegen verify
uvx --from local-gpu-imagegen==0.9.1 local-gpu-imagegen setup codex --apply

Then restart Codex and ask it to run a supported ComfyUI API workflow from its
exact local path. The existing-backend path requires Python 3.11 or 3.12, an
already-running local ComfyUI instance, an installed model, and a supported
API-format workflow.

Verified Release Identity

• Source commit: 03bb59735e3075cb42830038c080483fa531126a
• Wheel: local_gpu_imagegen-0.9.1-py3-none-any.whl
• Wheel SHA-256:
  0b2f6b928f9d5cfe415b0d8559bd7dceb2cc7abaa2d8b70e54b56b9dc1ccd2c9
• Wheel bytes: 307669
• Python: 3.11 and 3.12
• MCP protocol: 2024-11-05
• MCP tools: exactly 17
• Exact-SHA CI: Windows and Ubuntu platform-neutral jobs passed on Python 3.11
  and 3.12 (Actions run 33361352325)
• Public distribution: the exact wheel is available from PyPI and the matching
  descriptor io.github.ChevalGrand520/local-gpu-imagegen version 0.9.1 is
  active in the official MCP Registry

Supported Scope And Limits

• The v0.9 guided bootstrap and managed-generation host scope is Windows 10/11
  x64 with NVIDIA. The universal wheel and Ubuntu CI do not imply Linux managed
  generation support.
• ComfyUI and the user-supplied model generate the pixels. Local GPU Imagegen
  controls workflow authority, reproducibility, review, recovery, and output
  evidence; it does not claim to improve the underlying model's image quality.
• No model weights are bundled. Runtime or model downloads are never silent and
  require a separately displayed plan and confirmation.
• Arbitrary custom-node graphs are not automatically repaired or installed.
• This release makes no production-readiness, latency, throughput, concurrency,
  VRAM-saving, or generalized character-quality claim.
• No generated image is attached to this Release.

See the
README,
Quickstart,
and Changelog
for setup details and the complete evidence boundary.

Local GPU Imagegen v0.9.0

Local GPU Imagegen v0.9.0

v0.9.0 makes the supported Windows/NVIDIA first-run path explicit while keeping
the project focused on one job: helping an AI coding agent run a supported
ComfyUI workflow with frozen inputs, bounded revisions, structured review, and
recoverable evidence.

Highlights

• Guided Windows bootstrap: inspect a complete plan before any transfer,
  including sources, licenses, estimated bytes, and the exact confirmation.
  Approved downloads are resumable and checksum-verified; extraction is
  bounded, installation is transactional, and readiness is checked before use.
• Bring your own ComfyUI workflow: inspect and register supported API-format
  workflows without silently rewriting the graph, switching the model, or
  installing custom nodes.
• Review-driven delivery: durable run manifests retain the exact route,
  workflow, model identity, prompts, settings, seed, round budget, reviews, and
  image hashes. Runs allow one to three successful rounds and can recover after
  interruption without silently spending another round.
• Auditable revisions: preserve/change requests create immutable child runs.
  Prompt refinement, low-strength img2img, and confirmed inpaint remain distinct
  routes with explicit boundaries.
• Byte-bound acceptance: a reviewed candidate is not silently accepted.
  Finalization requires a later user confirmation bound to the exact run,
  round, and PNG SHA-256.
• More reliable managed startup: concurrent MCP clients coordinate Windows
  portable ComfyUI startup and re-probe backend health before deciding whether
  an existing process can be reused.

Install

uvx --from local-gpu-imagegen==0.9.0 local-gpu-imagegen verify
uvx --from local-gpu-imagegen==0.9.0 local-gpu-imagegen setup codex --apply

Then restart Codex and ask it to run a supported ComfyUI API workflow from its
exact local path. The existing-backend path requires Python 3.11 or 3.12, an
already-running local ComfyUI instance, an installed model, and a supported
API-format workflow.

Verified Release Identity

• Source commit: b9018f7a72bd20752740b3d87b977120851208d4
• Wheel: local_gpu_imagegen-0.9.0-py3-none-any.whl
• Wheel SHA-256:
  99c02f963e86d3867165bf16981f6445cb41cf9c8f208bc360578db483bcd179
• Python: 3.11 and 3.12
• MCP protocol: 2024-11-05
• MCP tools: exactly 17
• Exact-SHA CI: Windows and Ubuntu platform-neutral jobs passed on Python 3.11
  and 3.12
• Windows model-free suite: 1163 passed with 12 documented platform or
  privilege skips
• Offline candidate verifier: 28/28 passed
• Public distribution: the exact wheel is available from PyPI and the matching
  descriptor is active in the official MCP Registry

Supported Scope And Limits

• The v0.9 guided bootstrap and managed-generation host scope is Windows 10/11
  x64 with NVIDIA. The universal wheel and Ubuntu CI do not imply Linux managed
  generation support.
• ComfyUI and the user-supplied model generate the pixels. Local GPU Imagegen
  controls workflow authority, reproducibility, review, recovery, and output
  evidence; it does not claim to improve the underlying model's image quality.
• No model weights are bundled. Runtime or model downloads are never silent and
  require a separately displayed plan and confirmation.
• Arbitrary custom-node graphs are not automatically repaired or installed.
• This release makes no production-readiness, latency, throughput, concurrency,
  VRAM-saving, or generalized character-quality claim.
• No generated image is attached to this Release.

See the README,
Quickstart,
and Changelog
for setup details and the complete evidence boundary.

Local GPU Imagegen v0.8.3

Local GPU Imagegen v0.8.3 adds an opt-in supervisor for an existing Windows portable ComfyUI installation while preserving the default bring-your-own-backend behavior.

What's new

• serve and the official Codex and Claude Code setup flows now accept --auto-start-comfyui with an exact --comfyui-root.
• Managed mode validates the fixed portable layout and launches python_embeded\python.exe -s ComfyUI\main.py on loopback only.
• An already-running endpoint is reused without being owned or stopped. A child started by the MCP process is stopped at MCP exit only when its queue is known to be empty; non-empty or unknown queues are retained and reported.
• First-readiness waiting is bounded and configurable from 1 to 300 seconds, with a 120-second default. doctor remains read-only.
• Model-free tests cover existing-endpoint reuse, exact portable commands, non-loopback rejection, paths with spaces, bounded readiness, and queue-aware cleanup.

Install and verify

uvx local-gpu-imagegen verify
uvx local-gpu-imagegen setup codex --apply
# or
uvx local-gpu-imagegen setup claude-code --apply

To opt into managed startup for one existing Windows portable installation:

uvx local-gpu-imagegen setup codex --apply `
  --auto-start-comfyui `
  --comfyui-root "<ComfyUI_windows_portable>"

Verified release evidence

• Annotated tag v0.8.3 resolves to exact commit 66278380d2bb87f585148297777427b40ec8aa54.
• The exact-SHA GitHub Actions matrix passed on Windows and Ubuntu with Python 3.11 and 3.12.
• Public PyPI clean-cache acquisition returned version 0.8.3, MCP protocol 2024-11-05, and exactly 17 tools.
• The public wheel is local_gpu_imagegen-0.8.3-py3-none-any.whl, 260,937 bytes, SHA-256 215ea2a4544fd2fb03fb76002f54ff81bb645347e7f32be76abea65b938d548c, and is not yanked.
• Exact public 0.8.3 Claude Code setup and tool-catalog ingestion completed with Connected and no schema errors.

Boundaries

• Managed mode starts only an existing, explicitly selected Windows portable ComfyUI installation. It does not install ComfyUI, models, or custom nodes.
• There are no silent model downloads or model switches. ComfyUI and other integrated open-source components generate the pixels; this project does not train a model or modify diffusion algorithms.
• Model-free protocol and lifecycle tests do not establish image quality, performance, VRAM usage, backend readiness on another host, or production readiness.
• Claude Code tool-catalog ingestion is verified; a Claude Code image-generation session remains pending.
• Regional and two-stage composition routes remain experimental and outside the ordinary golden path.

Full 0.8.2 to 0.8.3 source change:
a0e08c5...6627838