[Feature]: Gemini provider missing image/multimodal input support

[rajnaveen344]

Summary

The Gemini provider's Part struct is text-only, preventing image inputs from being sent via the native inlineData API — even though Gemini models natively support multimodal image inputs.

Problem statement

Other providers (Anthropic, OpenAI-compatible) already parse [IMAGE:...] markers and convert them to their native image formats. The Gemini provider ignores these markers and sends them as raw text, making multimodal workflows broken for Gemini users.

Proposed solution

Replace the flat Part struct with a serde untagged enum supporting Part::Text and Part::InlineData variants. Add a build_user_parts() helper that reuses the existing multimodal::parse_image_markers() to extract image markers and produce Gemini-native inlineData entries. Supports multiple images per message.

Non-goals / out of scope

• Files API support for large images (>20MB) — can be a follow-up
• Video/audio multimodal inputs
• Changes to multimodal.rs or provider traits

Alternatives considered

• Adding image support via the Files API — more complex, requires upload step, not needed for inline use cases under 20MB

Acceptance criteria

• User messages with [IMAGE:data:...] markers produce correct inlineData parts in the Gemini API request
• Multiple images per message are supported
• Text-only messages serialize identically to current behavior (backward compatible)
• Tests cover text-only, single image, multiple images, image-only, and fallback paths

Architecture impact

None. Change is scoped to src/providers/gemini.rs. No trait changes, no new dependencies, no config changes.

Risk and rollback

Low risk. Single-file, additive change. git revert <commit> for rollback.

Breaking change?

No

Data hygiene checks

• No personal or sensitive data in tests — synthetic base64 stubs used
• Neutral project-scoped wording confirmed

[github-actions]

Thanks for opening this issue.

Before maintainers triage it, please confirm:

• Repro steps are complete and run on latest main
• Environment details are included (OS, Rust version, ZeroClaw version)
• Sensitive values are redacted

This helps us keep issue throughput high and response latency low.

[theonlyhennygod]

Implemented and opened PR: #2435

Root cause:

• Gemini provider request parts were modeled as text-only, so [IMAGE:data:...] markers never converted into Gemini-native image parts.

What changed:

• Switched Gemini request Part from a text struct to an enum supporting text and inlineData.
• Added build_user_parts() to parse user content with multimodal::parse_image_markers() and emit Gemini inlineData parts.
• Added focused tests for text-only, single-image, multiple-image, image-only, and invalid-marker fallback behavior.

Why this resolves it:

• User-role messages with inline image markers now serialize to Gemini’s expected inlineData request shape while preserving backward-compatible text-only behavior.

Validation highlights:

• cargo test providers::gemini::tests -- --nocapture
• cargo test build_user_parts_ -- --nocapture

[theonlyhennygod]

Merged via PR #2435

• merged at: 2026-03-01T20:19:39Z
• merge commit: 7350415

Branch rebased to latest main before merge.