OpenRouter compaction/summary requests drop x-session-id when cacheRetention is none

[Shoshkin]

What happened?

Compaction and branch-summary requests to OpenRouter are sent without x-session-id, so OpenRouter logs them as session_id: null. They don't show up under the session in Logs/Analytics/Sessions, and they get no sticky routing.

Cause: these calls set cacheRetention: "none" (#6618). openai-completions.ts and anthropic-messages.ts then drop the session ID completely:

const cacheSessionId = cacheRetention === "none" ? undefined : options?.sessionId;

x-session-id is OpenRouter's routing/observability key and does not write to the cache. The uuidv7() that createSummaryRequestOptions() generates for these calls never reaches the provider.

Steps to reproduce

1. Use an OpenRouter model and trigger compaction (/compact).
2. Look up the compaction generation: GET /api/v1/generation?id=<id> returns session_id: null. Normal turns in the same session have a session ID.

Expected behavior

OpenRouter summary calls keep x-session-id. prompt_cache_key and cache_control stay omitted under cacheRetention: "none".

I'd like to submit the fix. It's OpenRouter-only, so other session-affinity formats are unchanged: Shoshkin/pi@main...fix/openai-completions-session-affinity-cache-none

This reverses the cache-disabled part of bbb61e3 (#9102) for OpenRouter. Happy to adjust if that was intentional.

Version

0.85.1, still present on main (0.87.1)

---

AI disclosure: the investigation, fix and this text were drafted with Claude Code and reviewed by me.

[github-actions]

This issue was auto-closed. All issues from new contributors are auto-closed by default.

Maintainers review auto-closed issues daily and reopen worthwhile ones. Issues that do not meet the quality bar in CONTRIBUTING.md will not be reopened or receive a reply.

If a maintainer replies lgtmi on one of your issues, your future issues will stay open. If a maintainer replies lgtm, your future issues and PRs will stay open. The command must be at the start of the reply (optionally after one or more @username mentions) or at the end.

See CONTRIBUTING.md.

[holny]

AI-assisted content (per CONTRIBUTING.md): drafted by my local agent, reviewed and posted by me.

Confirmed on main (5fd446c): both adapters gate the session id on cache retention —

• openai-completions.ts:343 / anthropic-messages.ts:564: const cacheSessionId = cacheRetention === "none" ? undefined : options?.sessionId;

and that value is what createClient(...) receives, where its only use is the session-affinity headers (headers["x-session-id"] = sessionId behind compat.sendSessionAffinityHeaders; anthropic-messages.ts:974-976 the same). So with cacheRetention: "none" the routing header disappears entirely, even though x-session-id is OpenRouter's routing/observability key and never participates in the cache — exactly the session_id: null compaction entries you measured. The cache-key payload keeps its own separate gating, so the fix is to feed the affinity header the raw options.sessionId regardless of cacheRetention. (Adjacent heads-up: PR #9442 touches the prompt_cache_key gating, not this header.)

I'd like to implement this myself and can open a PR. If it meets the quality bar, a maintainer lgtm reply would keep my future issues and PRs open.