Summary

Two changes to make cacheRetention work end-to-end for Anthropic Claude models on the amazon-bedrock provider.

Changes

1. src/config/defaults.ts — config auto-injection

The config auto-injection in applyContextPruningDefaults only sets cacheRetention: "short" for provider === "anthropic", skipping Bedrock Claude models. Extended the provider gate to also match amazon-bedrock models whose ID contains anthropic.claude.

2. src/agents/pi-embedded-runner/extra-params.ts — runtime pass-through

resolveCacheRetention hard-gates on provider !== "anthropic" and returns undefined for all other providers. This means cacheRetention configured in model params is never consumed at runtime for Bedrock models. Extended the gate to also accept amazon-bedrock when cacheRetention is explicitly set in extraParams. No default is applied for Bedrock since pi-ai's own provider already defaults to "short" internally.

Related

• Fixes the config + runtime parts of [Bug]: Bedrock cacheRetention: "long" sends unsupported ttl: "1h" to models that only support 5-min TTL #21986
• Complements PR fix: disable prompt caching for non-Anthropic Bedrock models #20866 (disables caching for non-Anthropic Bedrock models)
• The TTL validation gap (cacheRetention: "long" → ttl: "1h" on unsupported models) is a separate upstream pi-ai concern, tracked in [Bug]: Bedrock cacheRetention: "long" sends unsupported ttl: "1h" to models that only support 5-min TTL #21986 and pi-mono discussion #1571

Testing

• AI-assisted (Kiro CLI)
• Lightly tested against local openclaw instance with amazon-bedrock/us.anthropic.claude-opus-4-6-v1
• Confirmed cacheRetention is auto-injected and consumed at runtime for Bedrock Claude models
• Confirmed non-Anthropic Bedrock models are correctly skipped

Greptile Summary

Extends cacheRetention support to Anthropic Claude models running on Amazon Bedrock by adding two complementary changes:

• Config auto-injection in applyContextPruningDefaults now recognizes Bedrock Claude models (matching amazon-bedrock provider with model IDs containing anthropic.claude) and applies the default cacheRetention: "short" setting
• Runtime pass-through in resolveCacheRetention now accepts Bedrock when cacheRetention is explicitly configured, allowing the auto-injected or user-specified values to flow through to the streaming API

The changes correctly preserve the existing behavior for direct Anthropic models while extending support to Bedrock's Anthropic Claude variants. The string matching logic handles both region-prefixed (us.anthropic.claude-*) and non-prefixed (anthropic.claude-*) model ID formats.

Confidence Score: 5/5

• This PR is safe to merge with minimal risk
• The changes are well-scoped, logically sound, and follow the existing patterns in the codebase. The two-part approach (config auto-injection + runtime pass-through) correctly implements the feature. The string matching logic using .toLowerCase().includes("anthropic.claude") appropriately handles all documented Bedrock Claude model ID formats, and the guard condition extraParams?.cacheRetention !== undefined ensures Bedrock only gets cacheRetention when explicitly configured (via auto-injection or user config). The changes preserve backward compatibility with direct Anthropic models and existing cacheRetention behavior.
• No files require special attention

_{Last reviewed commit: fb85a4d}

_{(3/5) Reply to the agent's comments like "Can you suggest a fix for this @greptileai?" or ask follow-up questions!}