BlendSmith v1.0.0 — Adaptive Production Loop

BlendSmith v1.0.0 — Adaptive Production Loop

v1.0.0 makes AI-assisted Blender production less repetitive without losing evidence or authority, especially for reference-guided work.

Highlights

• Reference-image continuity: SHA-256 pinning for supplied reference images, retained in candidate dependencies.
• Mandatory reference-to-render comparison: visual ACCEPT requires comparison with every registered reference, with the source identity, viewed renders and concrete observations.
• Saved-scene integrity: independently reopen the pinned Blender file, rejecting missing used images and required packed references that are absent.
• Lighter agent workflow: sparse method-family defaults, compact next output, identity-bound contract templates, and Quick Tune for bounded parameter edits.
• Adaptive reviews: record all defects, reassess recurring issues and major regressions, and preserve authority and owner approval gates.

Install / upgrade

Download and install the attached blendsmith-1.0.0-py3-none-any.whl using:

python -m pip install --upgrade blendsmith-1.0.0-py3-none-any.whl
blendsmith skill-install --force
blendsmith doctor

Restart Codex to load the upgraded Skill. Both CLI/Core and Skill must be version 1.0.0.

Verification

Full regression tests and Ruff passed; GitHub Actions CI covers Windows and Ubuntu. A real Blender 5.2.1 reopen check passed with a packed reference and correctly rejected an unpacked reference. No causal A/B quality or performance gain is claimed; private benchmark fixtures are not distributed.

AI acceptance is not owner acceptance. Publication of a produced 3D asset still requires explicit owner approval bound to the exact candidate SHA.

Changelog · Reference-guided workflow

BlendSmith v0.0.3 — Domain Knowledge

BlendSmith v0.0.3 — Domain Knowledge

v0.0.3 moves the production loop one level further upstream: research before the production decision when geometry depends on real-world knowledge.

Highlights

• Domain Research Gate with seven explicit risk families
• SHA-bound Domain Knowledge and Domain Practice receipts
• domain_constraints carried into Method Plan work units
• confidence propagation that prevents evidence from becoming stronger downstream
• bounded Domain Knowledge Cache with scope/freshness/epoch checks
• cache reuses facts, never production authority
• STRUCTURAL / CONTRACT changes can return to Domain Research
• explicit upstream reconsideration requires fresh knowledge instead of immediately reusing the questioned cache
• checkpoints, resume, blendsmith next, CLI schemas, and the bundled Skill preserve the full Research → Knowledge → Practice → Method Plan chain
• v0.0.2 run compatibility retained

Validated with repo-source tests, adversarial cache/receipt tamper tests, three-domain virtual dogfood, package/wheel smoke tests, and self-security review.

BlendSmith v0.0.2 - Production Structure

Production Structure turns BlendSmith's original review/repair loop into a dependency-aware production control loop.

Highlights:

• Work Unit Decomposition Gate and method-family accounting
• Production Graph with upstream revision and downstream invalidation
• Change Impact: LOCAL / METHOD / STRUCTURAL / CONTRACT
• Method Continuity and Global Reassessment
• Exploratory Live GUI Review
• Method Selection v2 and bounded discovery cache
• Fresh same-run recovery, evidence profiles, and next-action introspection

AI_ACCEPTED remains distinct from HUMAN_ACCEPTED.

BlendSmith v0.0.1

Initial public release of BlendSmith.\n\nHighlights:\n- Model-neutral CLI/Core for AI-assisted Blender production loops\n- Specialized-first Method Selection Gate\n- Evidence-backed visual review and bounded repair\n- Method reconsideration for wrong production approaches\n- Live-GUI verification when available\n- Exact SHA-bound human approval\n- Checkpoint/resume with method-authority restoration\n- TTL retention and verified GC boundaries\n- Bundled Codex Skill with CLI/Skill SHA + version pin checks\n\nPrimary runtime target: Blender 5.2 LTS.