[Feat]: Add enterprise cloud model adapters (Bedrock Converse, Azure OpenAI, etc.)

[rosspeili]

Feature Description

Add first-class SkillLoader adapters for enterprise / cloud-hosted model APIs whose tool-calling wire format differs from the existing five adapters (OpenAI, DeepSeek-compatible, Claude, Gemini, Ollama prompt).

Phase 1 (this issue’s MVP — pick one provider in first PR):

• to_bedrock_tool() (or equivalent) for AWS Bedrock Converse / tool-use API — map manifest.yaml parameters to Bedrock’s tool spec (name sanitization, schema shape per AWS docs).
• Tests in tests/test_loader.py (no live AWS calls; fixture manifests only).
• docs/usage/bedrock.md — install (boto3 or skillware[bedrock] extra if added), IAM/credentials, minimal agent loop snippet.
• Optional runnable example under examples/ + examples/README.md row.
• CHANGELOG.md [Unreleased].

Phase 2+ (follow-up issues/PRs — do not bundle in MVP):

• Azure OpenAI (often OpenAI-compatible — document first; adapter only if schema diverges)
• Google Vertex AI (may share Gemini shapes — evaluate reuse vs to_gemini_tool())
• Cohere tool API (if schema differs materially)
• IBM watsonx / other enterprise endpoints — only with clear demand

Explicit non-goals for this track:

• Per-vendor adapters for OpenAI-compatible APIs (Kimi, GLM, Groq, etc.) — covered by docs issue for to_openai_tool() + base_url
• New adapters for every local Ollama model name (Hermes, Qwen, …) — use to_ollama_prompt()
• Running inference inside Skillware — adapters are schema conversion only, same as today

Rationale

• Enterprise adoption: Teams on AWS/Azure/GCP often cannot use direct OpenAI/Anthropic/Google API keys; they route through Bedrock, Vertex, or Azure OpenAI with different SDKs and IAM.
• Honest provider matrix: Today we document five consumer paths; cloud consoles are a gap.
• Complements trust/secrets work ([RFC]: Pluggable Secret Management Architecture & Context Binding #39, [Feat]: Scoped environment injection for skill execution (sub-issue of #17; complements #39) #111): Enterprise hosts often pair with IAM roles and secret managers; adapters are the integration surface, not secret storage.
• Avoid adapter sprawl: Only add methods where wire format differs; document OpenAI-compatible cloud endpoints under the docs issue instead of duplicating to_openai_tool().

Affected paths (optional)

skillware/core/loader.py, tests/test_loader.py, pyproject.toml (optional [bedrock] extra), docs/usage/README.md, docs/usage/bedrock.md, examples/bedrock_*.py, CHANGELOG.md, docs/contributing/ai_native_workflow.md

Implementation Idea

Design principles:

1. Mirror existing adapter pattern in loader.py — static method on SkillLoader, input skill_bundle, output provider-native tool object or dict.
2. Reuse _sanitize_function_tool_name() where provider allows; add _sanitize_bedrock_tool_name() only if AWS rules differ.
3. Keep optional SDK deps out of core install — try/import with clear ImportError message (same as to_gemini_tool()).
4. Phased delivery: Bedrock first PR; Azure/Vertex/Cohere as separate issues after validating schema delta.
5. No breaking changes to execute(params) or existing adapters.

Bedrock sketch (illustrative — verify against current AWS Converse API):

# skillware/core/loader.py
@staticmethod
def to_bedrock_tool(skill_bundle: Dict[str, Any]) -> Dict[str, Any]:
    manifest = skill_bundle.get("manifest", {})
    return {
        "toolSpec": {
            "name": SkillLoader._sanitize_function_tool_name(manifest.get("name", "")),
            "description": manifest.get("description", ""),
            "inputSchema": {"json": manifest.get("parameters", {})},
        }
    }

Tests: Parametrize manifest fixtures (same as existing loader adapter tests); assert key fields present; no network.

Docs: Bedrock guide covers IAM role vs access keys, region, model ID, and that skills still use env_vars for skill keys (Etherscan, etc.) separate from Bedrock credentials.

Risks / how to avoid breaking the repo:

Risk	Mitigation
Adapter explosion	One cloud provider per PR; RFC comment on #39-sized changes
Optional dep bloat	[bedrock] extra; not in [all] unless justified
Schema drift vs AWS	Link official AWS docs in bedrock.md; version note in CHANGELOG
Duplicate OpenAI-compatible cloud	Document “use to_openai_tool() + Azure endpoint” before new adapter
Example/doc drift	Run pytest tests/test_registry_docs.py; update usage index

Acceptance criteria (MVP / Bedrock PR):

• SkillLoader.to_bedrock_tool(bundle) implemented and tested offline
• Usage guide + README index row
• Optional extra documented in install_extras.md if new dependency
• Existing five adapters unchanged; full pytest tests/test_loader.py green
• Follow-up issues linked for Phase 2 providers