lingtai v0.12.4

lingtai v0.12.4

Patch release focused on making the Codex backend cheaper and more cache-friendly for LingTai agents and daemons.

Highlights

• Stabilizes Codex prompt-cache identity for root agents by deriving one 8-character session/thread/cache key from the resolved init.json path.
• Gives LingTai-backend daemon runs their own daemon-scoped LLM service and per-run Codex cache anchor, derived from the daemon run's daemon.json path.
• Keeps session-id, thread-id, and prompt_cache_key aligned so repeated turns in the same agent/run can hit the same Codex-side cache bucket, while separate daemon runs do not collide with or poison each other.
• Fixes related Codex thread/session handling, including last-call-id salt behavior.
• Includes daemon watchdog isolation, embedded IMAP reply attachment handling, wheel sidecar smoke-test coverage, and repository cleanup.

Why this matters

Before this patch, LingTai's Codex path could miss cache opportunities because the identity used for session/thread/cache routing was either not stable enough for the same logical agent, or too shared across daemon runs. That meant similar prompts from the same long-lived agent could look unrelated to the backend, while independent daemon runs could accidentally share cache identity.

v0.12.4 makes the identity explicit:

• root agents use a stable hash of the agent's resolved init.json path;
• daemon runs use their own daemon.json path as a cache anchor;
• each run uses the same value for session-id, thread-id, and prompt-cache key.

The result is a better chance of cache hits for repeated work within the same agent/run, and cleaner isolation between daemons.

Validation

Release gates were run in a clean release worktree before publishing:

• git diff --check
• python -m pytest -q
• python -m build
• python -m twine check dist/*