Kernel v0.14.0 · TUI/Portal v0.9.5

New features and improvements

1. 01

   Honest Codex metadata and forwarded account identity

   Kernel v0.14.0 makes what the runtime tells the Codex backend about itself accurate rather than approximate.

   • Agents send an honest metadata envelope to the Codex backend so backend-side provenance reflects reality (#461).
   • The operator's own ChatGPT-Account-ID header is forwarded on Codex requests, so attribution and account routing are correct (#454).
   • Codex thinking config and the ChatGPT account-id path were repaired across the window (#457).

   Why it matters:Backend trust and cache affinity should rest on the truth about who is calling, not on guesswork or a stand-in identity.

2. 02

   Measured latency and tool-result visibility

   A slow turn becomes a measurable fact, and long tool results become visible without flooding the context budget.

   • New LLM latency telemetry records provider-wait and per-phase timings (#366).
   • Long tool results are surfaced in agent meta with a top-10 preview list and char counts (#451, #452, #453, #465).
   • Tool-result metadata is nested under _meta and excluded from summarize sizing so it no longer distorts the context budget.

   Why it matters:Operators can see where a turn's time goes and what evidence exists, without paying for it in working context.

3. 03

   Turn recovery and calmer molt discipline

   The window hardens recovery so a single stuck turn cannot take down a long session, and retires legacy molt-pressure config.

   • Durable tool results are replayed before heal, hardening post-tool continuation recovery (#297).
   • Turns poisoned by a stuck worker (WorkerStillRunning) now recover instead of wedging (#463).
   • Legacy molt-pressure configuration is ignored; molt pressure moved into agent meta (#464, #467).

   Why it matters:Long-running agents need to survive a single bad turn and shed context on intent, not on a configured pressure number.

4. 04

   A faster cockpit with fuller notifications

   TUI/Portal v0.9.5 trims the launch path and makes notifications more readable while work is in motion.

   • The authoritative session rebuild is deferred off the launch path and the session cache is concurrency-safe, so startup is faster (#397).
   • Notification snapshots render their full meta envelope and markdown blocks, are scrollable, and persist (#391, #392, #393).
   • /daemons lazy-loads run detail, token-ledger reads stream, and the initial mail view shows a loading banner (#395, #390, #398).

   Why it matters:Less time waiting on launch, and more of what an agent saw visible exactly where operators look.

5. 05

   Contributors and the fate of every proposal

   A release window is more than its merged commits. Below is everyone who opened a pull request or a substantive issue in this window — including work that was not merged but still shaped the result.

   • @huangzesen (maintainer) — authored the bulk of the window (86 merged PRs across both repos) covering Codex honesty, latency telemetry, tool-result metadata, summarize/molt discipline, and the release itself; two PRs (#369, #340) were closed unmerged after being superseded.
   • @TZZheng — merged durable tool-result replay (#297), LLM latency telemetry (#366), the dev-guide cache-hit-rate skill (#396), and a duplicate-process fix (#459); the earlier Codex reasoning-effort preset PR (#280) was closed unmerged, then the idea landed via #394.
   • @BatalloLu — fixed setup so it preserves agent identity / legacy preset fields (#372, merged).
   • @9s5bz2jvd2-lang (Wang Runyuan) — proposed structured daemon task capsules (#388) and terminal async daemon dispatch (#390), both closed unmerged, plus an open daemon-first policy PR (#389); the accompanying runtime issues (#409 molt keep_last, #422 large-result cleanup latency, #423 post-molt codename retrieval) are included because they framed problems this and future windows act on.

   Why it matters:Recording closed-unmerged PRs and the issues behind them keeps the contributor history honest: ideas often arrive before the merge that implements them.