Kernel v0.16.0 · TUI/Portal v0.10.3

New features and improvements

1. 01

   Cache-miss budgets and honest context-pressure reminders

   The kernel gives every session and molt a spendable cache-miss budget, and moves the current context-pressure reminders to durable metadata so a molt now cue survives instead of flickering. The TUI surfaces the same number inline.

   • manifest.cache_miss_budget and AgentConfig.cache_miss_budget add a per-session, per-molt cache-miss soft cap defaulting to 1,000,000 tokens; when it is reached the runtime restamps cache miss budget {budget} reached, molt now under _meta.tool_meta.context.molt, alongside cache_miss_budget and cache_miss_tokens (kernel #641).
   • Current context-pressure reminders now live in durable _meta.tool_meta.context.molt, while reconstruction one-shots stay under _meta.tool_meta.reconstruction.molt, so a standing molt cue is no longer confused with a one-time rebuild (kernel #640, #638, #607).
   • The TUI mail/home telemetry shows current-session cache-miss tokens immediately after the token total, e.g. tok 1.1M (miss 8.6k) cache … (TUI #498).

   Why it matters:A cache miss is where a long run quietly burns money and context; making it a budget with a visible running count and a durable `molt now` cue turns a hidden cost into a spendable one the operator and the agent can both see and act on.

2. 02

   Soul flow becomes an explicit opt-in

   Soul flow no longer runs by default. It is gated behind an environment flag, and the TUI setup and kanban teach that model so operators know exactly when it is on.

   • Soul flow is env opt-in by default via LINGTAI_SOUL_FLOW_ENABLED; soul(action="flow") returns a stable disabled result unless the flag is set (kernel #639).
   • The TUI setup and kanban reflect the opt-in model: soul flow is disabled by default unless LINGTAI_SOUL_FLOW_ENABLED is set, so the cockpit and the runtime agree on the same default (TUI #497).

   Why it matters:A powerful behavior that runs silently by default is a surprise waiting to happen; making soul flow an explicit flag with a stable disabled result means it is on only when someone chose it, and the cockpit shows that choice rather than guessing at it.

3. 03

   Leaner, more inspectable runtime metadata

   Progressive disclosure gets sharper on both sides: the kernel flattens and thins its token/meta surfaces, and the TUI exposes the model-visible summaries those surfaces produce.

   • Kernel token/meta surfaces are tightened: compact flattened token diagnostics, current time under tool_meta, sparse agent_meta, sparse notifications, and large-result ranking in agent_meta (kernel #608, #618, #620, #609).
   • A-priori summary=true support lands for bash, read, and grep, so a tool call can declare up front that its result should be summarized for the model (kernel #586).
   • The TUI presents summary=true tool results as a model-visible summary after the raw tool output, and Ctrl+O / soul-mode tool-call previews expose summarized model-visible results with longer summary previews (TUI #477, #481, #480, #479).

   Why it matters:An agent reasons better on a lean, ranked view than on a wall of raw meta, and an operator trusts the cockpit more when it shows exactly what the model saw; flattening the surfaces and surfacing their summaries keeps both looking at the same honest, inspectable picture.

4. 04

   Daemon and runtime backend safety and breadth

   The runtime widens which backends it can drive and hardens the seams where long-lived daemons wait, send mail, and replay provider state.

   • A Kimi Code backend is added, and a claude-p background wait guard keeps a backgrounded run from racing ahead of its own completion (kernel #634, #611).
   • Daemon email is surfaced as an explicit tool, common daemon MCP completion is filled in, and HTTP identity headers are attached to outbound runtime calls (kernel #584, #636).
   • Codex encrypted-reasoning replay now self-heals, so a provider-side reasoning-replay hiccup no longer strands the run; backend, runtime, and helper refactors round out the window (kernel #633).

   Why it matters:A fleet that runs unattended for hours fails at the seams — a backgrounded wait that races, a daemon that cannot send mail, a provider replay that stalls; closing those seams and adding a backend is how the runtime stays reliable across more of the ways people actually run it.

5. 05

   Cockpit polish, maintenance, and release hygiene

   Around the headline work, the cockpit gains a small everyday convenience, a dead surface is removed, and the paired versions were validated from clean release worktrees before publication.

   • Clipboard image paste support landed for chat input, and the stamina surfaces were removed, with bundled skill/anatomy/frontmatter metadata refreshed (TUI #467, #478, #464–#459).
   • Kernel-side documentation, anatomy, and frontmatter maintenance continued alongside the feature work (many TZZheng-led docs/helper refactors; nudge-storm fix #576).
   • Strict ranges: kernel v0.15.3..v0.16.0 (88 commits, 283 files changed, +15,005/-6,365) and TUI/Portal v0.10.2..v0.10.3 (32 commits, 117 files changed, +2,904/-450).

   Why it matters:A release earns trust in the small things too — a paste that just works, a dead readout removed rather than left to mislead, and full gate evidence cut from the exact validated commits; recording that hygiene is part of shipping responsibly.