[chris7iu]

Summary

Follows the official codex CLI's stateful WebSocket previous_response_id delta continuation and chained /responses/compact encrypted_content consolidation to eliminate the ~95% redundant session history re-upload in upstream traffic:

• Synthetic subagent continuation key when x-claude-code-agent-id is absent (src/providers/codex/continuation.rs, src/providers/codex/mod.rs, tests/codex_agent_continuation.rs): When Claude Code spawns subagents that share x-claude-code-session-id without sending x-claude-code-agent-id, resolve_continuation_owner derives a deterministic synthetic owner key (ConversationIdentity::Agent(session_id, "synthetic,<sha256>")) from prompt_signature + the leading developer envelope (AdditionalTools and leading developer instructions in Responses Lite) + the first role == "user" message in body.input. This prevents concurrent subagents from colliding with the main thread or each other, avoiding PromptChanged/NonPrefix continuation evictions and full-history re-uploads.
• Unbounded WebSocket continuation transcripts (src/providers/codex/continuation.rs): Removes MAX_OWNER_TRANSCRIPT_BYTES (2 MB) and MAX_TOTAL_TRANSCRIPT_BYTES (20 MB) while keeping MAX_STATES and TTL_MS bounding. Long reasoning-heavy sessions with high-entropy reasoning.encrypted_content blobs no longer abort previous_response_id delta continuation after ~15–25 turns and fall back to re-uploading the full multi-megabyte transcript on every turn.
• Unbounded post-compaction state & multi-compaction chaining (src/providers/codex/compaction.rs, src/providers/codex/mod.rs):
  • Removes MAX_STATE_BYTES (4 MB) and MAX_TOTAL_STATE_BYTES (20 MB) and extends STATE_TTL_MS to 24 hours so compacted sessions never drop their consolidated Compaction { encrypted_content } item.
  • Applies apply_compaction_replay before begin_compaction / request_compaction on subsequent compaction boundaries so consecutive compactions (both automatic and manual /compact) consolidate the prior Compaction { encrypted_content } item alongside new turns via /responses/compact, matching official codex CLI behavior without any arbitrary reasoning retention cap $K$.
• Hook-tolerant compaction detection (src/providers/codex/translate/request.rs, src/providers/codex/compaction.rs): Ignores trailing system / developer messages injected by Claude Code UserPromptSubmit hooks when detecting compaction boundary requests (is_compact_messages_request) and when stripping the compaction instruction (without_compaction_instruction).
• Coexistence of requestCompression and previousResponseId (src/config.rs): Keeps CodexTransport::Auto when both codex.requestCompression and codex.previousResponseId are enabled so normal turns use WebSocket previous_response_id delta continuation (~2–20 KB/turn) while HTTP requests (/responses/compact and HTTP fallbacks) use gzip request body compression.

Test Plan

• cargo fmt --all -- --check
• cargo clippy --all-targets -- -D warnings
• cargo test --lib config::tests::codex_request_compression
• cargo test --lib providers::codex::continuation::tests
• cargo test --lib providers::codex::compaction::tests
• cargo test --lib providers::codex::translate::request::tests::compact
• cargo test --test codex_agent_continuation
• cargo test --test smoke_cutover smoke_codex_http_server_compaction