v0.8.14 / v0.11.2

New features and improvements

1. 01

   Skill library cleanup: routers at the top, references at the leaves

   The most visible work is not a command-line feature, but the shape of the manuals agents load. Long utility skills now expose a smaller routing surface and move heavy details into named references or assets.

   • lingtai-issue-report now routes to an evidence checklist, report template, and filing flow instead of mixing the whole process into one document.
   • listen, academic-research, vision, dj, and find-something-to-do moved under swiss-knife/reference/, making the top-level skill catalog read more like entry points than a flat toolbox.
   • Portal guide, tutorial guide, recipe, dev guide, daily reflection, and web-browsing utilities continue the nested-reference pattern introduced in earlier releases.

   Why it matters:Resident prompts should remember where to go, not carry every procedure. Smaller routers make the system prompt lighter, while leaf references keep detailed workflows available when the task actually needs them.

2. 02

   Molt and procedures: the full handoff scaffold moves to the right layer

   The old lingtai-molt-template utility is folded into the kernel-side operating model. The resident procedure now points clearly to psyche-manual, and the full nine-section consequential-molt template lives as a packaged asset.

   • Kernel procedures now make the “read psyche-manual before a consequential molt” route explicit.
   • psyche-manual became a lighter router with assets/molt-template.md as the full scaffold.
   • Changed docs replace stale “Codex Cheat Sheet” wording with durable-memory and successor-briefing language.

   Why it matters:A molt summary is not a transcript; it is the next self’s operating brief. Keeping the full scaffold in the intrinsic psyche manual makes that rule available to every agent without bloating the always-on prompt.

3. 03

   Release-window rescue: Codex/gpt-5.5 tool schema scrub and scheduler removal

   The final kernel rescue commit is included in this release because it directly affects whether Codex-backed gpt-5.5 agents can boot and call tools reliably.

   • Merge 70e7c02 carries commits 3086312 and adea979: tool schemas are scrubbed before being sent to the OpenAI/Codex Responses backend, removing JSON Schema shapes that the backend rejects.
   • The built-in email recurring-send scheduler was removed from schema/i18n/manager code; recurring work now routes back to host cron / launchd / systemd as documented in bash-manual.
   • The cleanup touched seven kernel files (+83 / -460) and the release tests were aligned to assert that the scheduler surface is absent rather than accidentally still advertised.

   Why it matters:The fix reduces a hard availability failure mode: agents should not be blocked before their first real action because the model backend rejects the tool definitions themselves. Removing the duplicate scheduler also leaves one clearer scheduling path.

4. 04

   Reliability polish across setup, presets, validation, and release hygiene

   Alongside the documentation and skill refactors, this window includes smaller fixes that make setup, testing, and release verification less surprising.

   • TUI setup/preset/API-key handling and MiniMax-M3 default updates were validated through Go tests and release builds.
   • Portal release validation now explicitly builds portal/web before portal Go tests so embedded assets exist.
   • The release gate normalized docs/stars/stars.csv line endings so git diff --check stays clean.

   Why it matters:Small release-gate fixes are easy to dismiss, but they are exactly what keeps the next release from spending time rediscovering the same failure modes.