Description

Two related prompt-caching defects on the Bedrock / aggregating-provider paths. invokeModel runs injectPromptCaching(body, provider) before dispatch, which rewrites string system fields and message content into Anthropic cache_control blocks. Two downstream problems follow from that:

1. Bedrock Converse mapping is broken (src/clients/databricks.js, invokeBedrock). The message map (content: ... c.text ...) copies only c.text, dropping cache_control so Bedrock never actually caches. Worse, the system field is wrapped as converseBody.system = [{ text: bedrockBody.system }] while bedrockBody.system is now an array of blocks — producing system: [{ text: [ ... ] }], which violates the Converse contract (system: [{ text: "<string>" }]) and triggers a ValidationException.

2. Injection is not gated by model capability (src/clients/prompt-cache-injection.js, needsCacheInjection / injectPromptCaching). needsCacheInjection only checks provider membership in EXPLICIT_CACHE_PROVIDERS. But bedrock and openrouter route to many model families, and most (Llama, Mistral, Titan, Cohere, GPT, ...) don't support cache_control. Markers were injected regardless, yielding request shapes those models reject or silently ignore.

Type of Change

• Bug fix
• New feature
• Documentation update
• Performance improvement
• Refactoring

Changes Made

• Bedrock Converse shape fix (src/clients/databricks.js): added normalizeBodyForConverse(), called at the top of invokeBedrock before the Converse request is built. It strips cache_control and flattens any array-shaped system / message content back to plain strings (the Converse API has no cache_control concept). It works on a shallow copy, so the caller's body is untouched.
• Model-capability gate (src/clients/prompt-cache-injection.js): added modelSupportsCacheControl(body, provider) and call it inside injectPromptCaching. It inspects the resolved model id (body._tierModel / body.model) and skips injection for families known not to support cache_control. It is intentionally fail-open: Anthropic-only providers and unknown/absent model ids still inject, so the existing provider-only needsCacheInjection(provider) signature and behavior are preserved.

Testing

• Existing tests pass (test/prompt-cache-injection.test.js, unchanged assertions still green)
• Added new tests
  • test/databricks-bedrock-converse.test.js — asserts that after injectPromptCaching(body, 'bedrock') the normalized Converse payload has a string system and { text: string } content blocks.
  • Extended test/prompt-cache-injection.test.js with capability-gate cases (Claude allowed, Llama/Titan/Mistral/Cohere/GPT skipped, unknown model id still injects).
• Manual testing performed (ran the affected suites locally with the project's node --test runner)

node --test test/prompt-cache-injection.test.js test/databricks-bedrock-converse.test.js
# tests 32 pass 32 fail 0

Checklist

• Code follows style guidelines (2-space, CommonJS, existing conventions)
• Documentation updated (no user-facing behavior/config change)
• Tests added/updated
• No new warnings (eslint over the touched files reports no new errors)

Notes for reviewers

npm install against the default registry fails in my environment (the lockfile resolves to a private registry I can't reach), so I installed dev deps from the public npm registry to run eslint and the tests; the committed package-lock.json is unchanged. Linting the two touched source files adds zero new errors over the current main baseline.