Related Issues

• [Bug] Bedrock Converse sends Anthropic-specific cachePoint to Amazon Nova models #20815 / PR fix: disable prompt caching for non-Anthropic Bedrock models #20866 — fixes non-Anthropic Bedrock models (Nova, Mistral) receiving cachePoint. Does not address the TTL validation gap for Anthropic Claude models on Bedrock.
• [Feature]: Per-agent prompt cache control (cacheRetention config key) #17112 — requests per-agent cacheRetention config. Orthogonal scope.

Problem

When cacheRetention: "long" is configured for a Bedrock Claude model, the pi-ai provider unconditionally injects ttl: { ttl: "1h" } into cache points. However, only a subset of Claude models on Bedrock support 1-hour TTL. All others reject the request with an API error, breaking the agent turn entirely.

Additionally, the config auto-injection logic (which sets cacheRetention: "short" as default) only runs for provider === "anthropic", skipping amazon-bedrock. Users running Claude on Bedrock don't get prompt caching defaults unless they manually configure it.

PR #20866 correctly disables caching for non-Anthropic Bedrock models but doesn't address either of these issues.

Reproduction

1. Set config:

{
  "amazon-bedrock/us.anthropic.claude-opus-4-1-v1": {
    "params": { "cacheRetention": "long" }
  }
}

2. Send any message → API error:

"Extra inputs are not permitted"

Same error for Claude Opus 4.0, Sonnet 4.0, 3.7 Sonnet, 3.5 Haiku, 3.5 Sonnet v2.

For Nova models with cacheRetention: "long":

"Extended TTL prompt caching is only supported for Anthropic models"

Expected Behavior

cacheRetention: "long" should gracefully degrade to "short" (5-min TTL) for models that don't support 1-hour TTL, instead of causing an API error.

Models Supporting 1-Hour TTL (per AWS docs)

Only these Bedrock models accept ttl: "1h":

• Claude Opus 4.5+
• Claude Sonnet 4.5
• Claude Haiku 4.5

All other Claude models (Opus 4.0/4.1, Sonnet 4.0, 3.7 Sonnet, 3.5 Haiku/Sonnet) support only 5-min TTL.

Proposed Fix

These two issues belong in different layers:

1. supportsExtendedTTL guard → upstream pi-ai (amazon-bedrock.ts)

TTL support is a Bedrock API constraint, not openclaw business logic. pi-ai's streamBedrock already has supportsPromptCaching() as a model-level gate — supportsExtendedTTL is the same category of concern. Placing it in pi-ai means all consumers get the fix, rather than each one needing to wrap around the provider (as PR #20866 does for the Nova case).

function supportsExtendedTTL(modelId: string): boolean {
  const id = modelId.toLowerCase();
  if (!id.includes("anthropic") || !id.includes("claude")) return false;
  const m = id.match(/claude-(?:opus|sonnet|haiku)-4-(\d{1,2})(?:-\d{8}|-v)/);
  return m ? parseInt(m[1], 10) >= 5 : false;
}

In streamBedrock, downgrade "long" → "short" when !supportsExtendedTTL(model).

2. Config auto-injection → openclaw (config-*.ts)

This is openclaw's own config migration logic. The provider gate just needs to also accept "amazon-bedrock" (filtered to Anthropic Claude models via isAnthropicBedrockModel from PR #20866):

// Current:
if (!parsed || parsed.provider !== "anthropic") continue;

// Fix:
if (!parsed || (parsed.provider !== "anthropic" && parsed.provider !== "amazon-bedrock")) continue;

Environment

• OpenClaw: 2026.2.19-2
• Provider: amazon-bedrock
• Model: us.anthropic.claude-opus-4-6-v1 (and tested against Opus 4.1, Nova Micro)

Workaround

Local patch to amazon-bedrock.js adding supportsExtendedTTL() and effectiveCacheRetention() that downgrades "long" → "short" for unsupported models. Manual params.cacheRetention in config.