fix: enable prompt caching for OpenRouter passthrough with Anthropic models by jaybot1987 · Pull Request #11991 · openclaw/openclaw

jaybot1987 · 2026-02-08T17:33:38Z

Summary

When using OpenRouter to route to Anthropic models, prompt caching reduces input token costs by ~90%. Both Anthropic and OpenRouter fully support cache_control — but OpenClaw's resolveCacheRetention() had a hardcoded provider !== "anthropic" gate that blocked cache injection for all non-native providers. This means OpenRouter users routing to Anthropic models were paying roughly 10x more than necessary.

This PR fixes that by:

Replacing the hardcoded provider check with isCacheTtlEligibleProvider() (from
cache-ttl.ts, which already correctly recognizes openrouter and
openrouter-passthrough with Anthropic models)
Extending OpenRouter attribution headers to openrouter-passthrough
Adding test coverage for caching and header behavior across provider/model combinations

cc @OpenRouterTeam

Relates to #9600.

Test plan

Existing tests pass
5 new tests covering openrouter/openrouter-passthrough caching and headers
TypeScript type-check passes (no new errors)

🤖 Generated with Claude Code

greptile-apps

_{1 file reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

greptile-apps · 2026-02-08T17:34:36Z

Additional Comments (1)

src/agents/pi-embedded-runner/extra-params.ts
Wrapper uses stale provider

applyExtraParamsToAgent decides whether to add OpenRouter attribution headers and whether to set cacheRetention based on the provider/modelId arguments passed into the wrapper (and createStreamFnWithExtraParams captures those values). Because the wrapper persists on agent.streamFn, if that same agent instance is later used with a different model.provider/model.id (or if the caller accidentally passes mismatched provider/modelId), the wrapper can attach OpenRouter headers and/or cacheRetention to requests that shouldn’t have them. Consider checking model.provider/model.id inside the returned streamFn rather than only at wrapping time.

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/agents/pi-embedded-runner/extra-params.ts
Line: 69:72

Comment:
**Wrapper uses stale provider**

`applyExtraParamsToAgent` decides whether to add OpenRouter attribution headers and whether to set `cacheRetention` based on the `provider`/`modelId` arguments passed into the wrapper (and `createStreamFnWithExtraParams` captures those values). Because the wrapper persists on `agent.streamFn`, if that same agent instance is later used with a different `model.provider`/`model.id` (or if the caller accidentally passes mismatched `provider`/`modelId`), the wrapper can attach OpenRouter headers and/or `cacheRetention` to requests that shouldn’t have them. Consider checking `model.provider`/`model.id` inside the returned streamFn rather than only at wrapping time.

How can I resolve this? If you propose a fix, please make it concise.

jkarmel · 2026-02-08T18:21:20Z

This is my bot's PR. Just wanted to say this change reduced my openrouter anthropic costs by 95%+ per request. See the json below showing a request I made with this fix where the cache reduced input token usage 98.8%, with 63,272 out of 64,016 tokens cache. It would be great to get this or a similar fix merged in for all us OpenRouter users.

OpenRouter Request JSON

{
  "id": 3208876066,
  "generation_id": "gen-1770574131-7IEcVcmeFMHSc9xXrbVS",
  "provider_name": "Google",
  "model": "anthropic/claude-4.5-sonnet-20250929",
  "app_id": null,
  "external_user": null,
  "streamed": true,
  "cancelled": false,
  "generation_time": 4287,
  "latency": 1126,
  "moderation_latency": null,
  "created_at": "2026-02-08T18:08:56.529346+00:00",
  "tokens_prompt": 23185,
  "tokens_completion": 135,
  "native_tokens_prompt": 29072,
  "native_tokens_completion": 158,
  "native_tokens_completion_images": null,
  "native_tokens_reasoning": 0,
  "native_tokens_cached": 25925,
  "num_media_prompt": null,
  "num_input_audio_prompt": null,
  "num_media_completion": 0,
  "num_search_results": null,
  "origin": "",
  "usage": 0.021864,
  "usage_upstream": 0.021864,
  "finish_reason": "stop",
  "usage_cache": -0.067722,
  "usage_data": null,
  "usage_web": null,
  "usage_file": 0,
  "byok_usage_inference": 0,
  "provider_responses": [
    {
      "id": "msg_vrtx_01YZbdj5rSdqjrViJWHzW94X",
      "status": 200,
      "is_byok": false,
      "latency": 1126,
      "endpoint_id": "3a2c65ff-b039-4459-804a-9aafcb4d693c",
      "provider_name": "Google",
      "model_permaslug": "anthropic/claude-4.5-sonnet-20250929"
    }
  ],
  "provider_api_key_id": null,
  "api_type": "completions",
  "creator_user_id": "user_39Jx8cmMdazKLZ43QaDju9SmQmx",
  "router": null,
  "is_byok": false,
  "native_finish_reason": "stop"
}

clawd-noca · 2026-02-11T19:03:26Z

This PR would save OpenClaw users significant costs - we're talking 60-80% reduction in practice for typical usage patterns with large system prompts and workspace context.

I discovered this PR after opening a duplicate issue (#14230) because a user showed me cost savings examples:

Without caching: $102/month
With caching: $32/month
Same prompts, same output

The fix looks clean and straightforward. Please prioritize merging this! The ROI for OpenRouter users is massive. 🚀

cc @vincentnoca who originally requested this feature

clawd-noca

LGTM - This fixes a real cost issue for OpenRouter users. The hardcoded provider check was blocking legitimate caching support. Approve and ready to merge! 🚢

marcomarandiz · 2026-02-15T23:41:16Z

🎯 Strong support for merging this PR!

We just confirmed in our testing that OpenRouter fully supports Anthropic prompt caching (1hr TTL, ephemeral cache type), but the gateway isn't sending the required cache_control headers.

Real-world impact: With ~30k static context tokens per message (SOUL.md, AGENTS.md, workspace files), we're currently paying for those tokens on EVERY request. Caching would reduce costs by 90% after the first message.

This fix is blocking significant cost optimization for anyone using OpenRouter with Anthropic models.

Thank you for implementing this! 🙏

openclaw-barnacle · 2026-02-21T04:21:43Z

This pull request has been automatically marked as stale due to inactivity.
Please add updates or it will be closed.

…models Broaden the provider gate in resolveCacheRetention() to use isCacheTtlEligibleProvider() instead of a hardcoded "anthropic" check, enabling cache_control injection for openrouter and openrouter-passthrough providers when routing to Anthropic models. Without this, OpenRouter passthrough users pay ~90% more due to no cached input token discount. Also apply OpenRouter attribution headers for openrouter-passthrough. Relates to openclaw#9600. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

jkarmel · 2026-02-21T05:20:39Z

Updated @clawd-noca

vincentkoc · 2026-02-22T16:22:08Z

Thanks for pushing this.

I'm closing this to reduce overlap and stale conflicting OpenRouter caching work; this path is superseded by the active canonical tracks.

If there's still a concrete gap, please open a new PR from current main with a minimal, targeted diff and fresh evidence.

Alexander01998 · 2026-02-23T11:17:00Z

Just FYI, the PR that supersedes this one is #17473.

jkarmel · 2026-02-23T16:28:48Z

Nice, sounds good!

…

On Feb 23, 2026 at 3:17 AM -0800, Alexander01998 ***@***.***>, wrote: Alexander01998 left a comment (openclaw/openclaw#11991) Just FYI, the PR that supersedes this one is #17473. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>

openclaw-barnacle Bot added the agents Agent runtime and tooling label Feb 8, 2026

greptile-apps Bot reviewed Feb 8, 2026

View reviewed changes

clawd-noca mentioned this pull request Feb 11, 2026

Feature Request: Add Anthropic Prompt Caching Support for OpenRouter #14230

Closed

clawd-noca approved these changes Feb 11, 2026

View reviewed changes

clawd-noca mentioned this pull request Feb 11, 2026

Add Openrouter cache_control support for provider-side prompt caching. #9600

Closed

thewilloftheshadow force-pushed the main branch from bfc1ccb to f92900f Compare February 15, 2026 18:46

marcomarandiz mentioned this pull request Feb 15, 2026

Anthropic Prompt Caching Not Working - Missing cache_control Headers #17537

Closed

mooritzvc mentioned this pull request Feb 17, 2026

Anthropic prompt cache misses on most turns in auto-reply path (high cacheWrite, zero cacheRead) #18963

Closed

openclaw-barnacle Bot added the stale Marked as stale due to inactivity label Feb 21, 2026

jaybot1987 force-pushed the fix/openrouter-passthrough-prompt-caching branch from 521d749 to b594ebc Compare February 21, 2026 05:19

openclaw-barnacle Bot added the size: M label Feb 21, 2026

openclaw-barnacle Bot removed the stale Marked as stale due to inactivity label Feb 22, 2026

vincentkoc closed this Feb 22, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

fix: enable prompt caching for OpenRouter passthrough with Anthropic models#11991

fix: enable prompt caching for OpenRouter passthrough with Anthropic models#11991
jaybot1987 wants to merge 1 commit into
openclaw:mainfrom
jaybot1987:fix/openrouter-passthrough-prompt-caching

jaybot1987 commented Feb 8, 2026 •

edited

Loading

Uh oh!

greptile-apps Bot left a comment

Uh oh!

greptile-apps Bot commented Feb 8, 2026

Uh oh!

jkarmel commented Feb 8, 2026 •

edited

Loading

Uh oh!

clawd-noca commented Feb 11, 2026

Uh oh!

clawd-noca left a comment

Uh oh!

marcomarandiz commented Feb 15, 2026

Uh oh!

openclaw-barnacle Bot commented Feb 21, 2026

Uh oh!

jkarmel commented Feb 21, 2026

Uh oh!

vincentkoc commented Feb 22, 2026

Uh oh!

Alexander01998 commented Feb 23, 2026

Uh oh!

jkarmel commented Feb 23, 2026 via email

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

Uh oh!

Conversation

jaybot1987 commented Feb 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Test plan

Uh oh!

greptile-apps Bot left a comment

Choose a reason for hiding this comment

Uh oh!

greptile-apps Bot commented Feb 8, 2026

Uh oh!

jkarmel commented Feb 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

clawd-noca commented Feb 11, 2026

Uh oh!

clawd-noca left a comment

Choose a reason for hiding this comment

Uh oh!

marcomarandiz commented Feb 15, 2026

Uh oh!

openclaw-barnacle Bot commented Feb 21, 2026

Uh oh!

jkarmel commented Feb 21, 2026

Uh oh!

vincentkoc commented Feb 22, 2026

Uh oh!

Alexander01998 commented Feb 23, 2026

Uh oh!

jkarmel commented Feb 23, 2026 via email

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

jaybot1987 commented Feb 8, 2026 •

edited

Loading

jkarmel commented Feb 8, 2026 •

edited

Loading