Kernel v0.14.1

New features and improvements

1. 01

   Codex Responses can continue incrementally

   LingTai now preserves the WebSocket-side state needed to continue a Codex Responses turn through previous_response_id when the local history still matches the remote baseline.

   • ws_incremental means LingTai continued the existing remote previous_response_id chain instead of replaying the full request.
   • ws_full means LingTai rebuilt a complete request from local history and started a fresh remote state chain.
   • The runtime keeps fallback paths so prefix mismatch, missing baseline, or explicit reset can safely choose a full request.

   Why it matters:Codex cache affinity depends on the remote Responses state chain, not only on repeating the same prompt bytes.

2. 02

   Tool-output baselines and fresh epochs are explicit

   Tool outputs are frozen against the request baseline, and LingTai can intentionally leave stale remote metadata behind by starting a fresh epoch.

   • Frozen tool-output cache prevents later local mutation from changing what an incremental Codex request is replying to.
   • Periodic epoch reset defaults to 20 turns and sends one complete request rebuilt from current local chat_history.
   • A successful local system(action="summarize") also triggers a fresh Codex epoch on the next request.

   Why it matters:A fresh epoch is the safe way to discard stale remote-state baggage without deleting LingTai local history.

3. 03

   Summarize guidance now protects Codex cache

   Jason noticed that repeated one-by-one summarize calls were repeatedly breaking the Codex cache chain; the runtime comment now makes that cost visible.

   • For Codex, each summarize forces the next request to start as ws_full instead of continuing as ws_incremental.
   • The adapter comment strongly discourages consecutive summarize calls within about five turns.
   • Ordinary long tool results should be read first and summarized together once their raw payloads are no longer needed.

   Why it matters:Other providers are less sensitive here, but Codex uses `previous_response_id` state, so unnecessary fresh epochs are real cache misses.

4. 04

   Default Codex identity is honest LingTai

   The release removes stale comments that described the old official-CLI-shaped experiment as the default.

   • Default requests use originator=lingtai and User-Agent=LingTai/<version>.
   • The official Codex CLI-shaped identity remains only as an explicit local comparison switch.
   • Comments, tests, and src/lingtai/llm/openai/ANATOMY.md were updated to match the shipped behavior.

   Why it matters:Protocol experiments are valuable, but the shipped runtime should be clear about who it is.

5. 05

   Validation and release hygiene

   The release also repairs stale kernel ANATOMY citations and records exact artifact hashes from the published files.

   • Full kernel pytest passed: 2715 passed, 4 skipped in 304.41s.
   • Targeted Codex identity/comment tests passed: 42 passed.
   • Custom all-ANATOMY citation check passed: 599 citations checked, 0 issues.
   • Wheel and sdist passed twine check, contained no __pycache__ / .pyc, and were verified on PyPI after upload.

   Why it matters:This was a small version bump, but the state-machine behavior deserved full validation and durable release evidence.