feat: add xAI image edit and refresh Imagine provider

[jovijovi]

Summary

Updates the existing xAI image PR into a current upstream-clean patch on top of NousResearch/hermes-agent:main.

This PR now adds:

• a dedicated image_edit tool surface for edit-capable image providers;
• optional provider-level supports_edit() / edit(...) contract with unsupported_capability fallback for generate-only providers;
• xAI Grok Imagine refresh aligned with the current Imagine API:
  • default model: grok-imagine-image-quality;
  • safe fallback for deprecated/unknown model ids;
  • official xAI aspect-ratio wire values plus Hermes aliases;
  • resolution honored for both generation and edit paths;
  • response_format: b64_json to avoid depending on ephemeral imgen.x.ai URLs;
  • single-image edit, xAI Files file_id, multi-image edit up to 3 inputs;
  • optional direct-provider n and storage_options support with privacy-default validation;
  • cached outputs plus metadata echo for debugging.

Compatibility notes:

• Preserves the current upstream image_generate(image_url=..., reference_image_urls=...) contract by routing xAI image inputs through the same edit implementation.
• image_edit is exposed in the image_gen toolset and gated at runtime by provider support.
• Non-edit providers keep the default unsupported-capability response; they are not forced to implement editing.

Verification

Hermes local gates on the rebuilt branch:

uv run --frozen pytest tests/plugins/image_gen/test_xai_provider.py tests/tools/test_image_edit_tool.py -q -o 'addopts='
# 100 passed

uv run --frozen pytest \
  tests/agent/test_image_gen_provider_edit.py \
  tests/plugins/image_gen/test_xai_provider.py \
  tests/tools/test_image_edit_tool.py \
  tests/tools/test_image_generation_plugin_dispatch.py \
  tests/test_toolsets.py \
  tests/tools/test_image_generation_image_to_image.py \
  -q -o 'addopts='
# 165 passed

python3 -m py_compile \
  agent/image_gen_provider.py \
  plugins/image_gen/xai/__init__.py \
  tests/agent/test_image_gen_provider_edit.py \
  tests/plugins/image_gen/test_xai_provider.py \
  tests/tools/test_image_edit_tool.py \
  tools/image_edit_tool.py \
  toolsets.py
# OK

git diff --cached --check
# OK before commit

Independent read-only review:

• Codex blocker review: PASS
• Previous blocker found and fixed: edit successes now return modality="image"; direct edit and image_generate(image_url=...) have regression coverage.

Scope hygiene

This update intentionally excludes downstream-only Sachima roadmap/dev-log/runtime-rollout evidence, local media/cache artifacts, private profiles, credentials, and local workspace paths.

[teknium1]

Thanks for the xAI API refresh work. The useful provider changes need a targeted salvage onto current main rather than a direct merge.

Problems

• agent/image_gen_provider.py:479 adds a local-file read path without raise_if_read_blocked; plugins/image_gen/xai/__init__.py:234 reaches it. Current main guards xAI reads at plugins/image_gen/xai/__init__.py:140-145 (commit c1826e269) and tests that blocked credentials are never opened.
• tools/image_edit_tool.py:57-60 checks any available edit provider, but dispatch at :82-85 uses only the active provider. This can expose a tool that necessarily returns unsupported_capability.
• Current main already deliberately unifies image editing under image_generate (agent/image_gen_provider.py:16-24, commit c02192ff6); adding image_edit to _HERMES_CORE_TOOLS duplicates that model-facing surface.
• The provider rewrite must retain current image_gen.xai.storage behavior at plugins/image_gen/xai/__init__.py:289-333.

Suggested changes

• Rebase the provider changes conceptually onto the unified image_generate path, preserve current storage and file-safety behavior, and add the active-provider availability test.

Automated hermes-sweeper review.

[teknium1]

This reads a model-supplied path without agent.file_safety.raise_if_read_blocked. Current main guards xAI local-image reads before opening them (plugins/image_gen/xai/__init__.py:140-145, c1826e269); put that guard in this shared helper before any filesystem access and add a no-read credential-store regression test.

[teknium1]

Availability is based on any discovered provider, but dispatch later uses only get_active_provider(). If xAI is available while the selected provider cannot edit, this exposes image_edit for a guaranteed unsupported-capability response. Resolve and check the active provider here instead.

[teknium1]

Please preserve the current config-driven image_gen.xai.storage path when salvaging this rewrite. Current main builds and sends configured storage options by default at plugins/image_gen/xai/__init__.py:289-333; accepting only direct kwargs would silently drop that existing behavior.

[teknium1]

Thanks for the thorough work here, @jovijovi — closing after review, with two reasons:

1. The core outcome (xAI image editing) has since landed on main via a different route: the xAI provider now routes generate(image_url=...) to /v1/images/edits (with multi-source support, storage options, and — as of Image picker no longer crashes on stale cross-provider models; xAI Grok catalog goes live-driven (salvage #77238) #89841/Picked image model now reaches xAI generation; OpenRouter picker lists all live image models (salvage #55893) #89890 — edit-capable model selection incl. grok-imagine-image-2.0). Your PR was the earlier articulation of that need and the JSON-body-not-multipart insight matches what shipped.

2. The image_edit core tool + supports_edit() ABC half doesn't fit our tool-surface policy: every core model tool ships on every API call, and image editing already flows through the existing image_generate tool (image_url / reference_image_urls params), so a second tool would duplicate that surface. The capability signaling your ABC change aimed at exists today via capabilities().

Appreciate the contribution — the editing capability is live and your PR helped map the requirements.