spike(compaction): evaluate agentgateway compression for stateless routes

[robinbraemer]

Outcome

Determine whether agentgateway request-local context compression improves stateless AgentOS routes without replacing Pi's native OpenAI compaction lifecycle.

Scope

• Confirm the feature's released version, support status, exact /v1/compress contract, and upgrade/rollback behavior.
• Benchmark no compression, current Pi native server compaction, and gateway compression on disposable/stateless workloads.
• Measure total-session cost, task success, turns, latency, provider cache reads, tokens saved, and failure behavior.
• Prove /v1/responses/compact passes through agentgateway byte- and artifact-correctly.
• Keep persistent Pi sessions, opaque encrypted_content persistence, replay compatibility, and portable fallback owned by the existing AgentOS extension.

Acceptance criteria

• The spike records whether the advertised feature is released and pins the tested build.
• Gateway compression is never enabled for persistent Pi routes by default.
• OpenAI compact artifacts survive gateway routing and later compatible replay.
• Fail-open/fail-closed behavior is explicit and cannot silently corrupt tool-call pairing.
• Adoption requires better task success and total trajectory cost, not only smaller individual requests.

References

• https://agentgateway.dev/blog/2026-07-27-optimize-token-cost-with-context-compression/
• https://developers.openai.com/cookbook/examples/gpt-5/codex_prompting_guide#compaction