What is the Neural Link?

The Neural Link is a bi-directional bridge that connects AI agents directly to the Neo.mjs runtime. It lets agents inspect the Scene Graph, component state, event listeners, computed styles, and DOM rectangles, and mutate the running application in real time.

Why is Neo.mjs called an Application Engine instead of a framework?

Neo.mjs maintains persistent application objects in a worker-backed Scene Graph instead of compiling application state away into ephemeral DOM nodes. That architecture enables multi-window orchestration, runtime permutation, and deep AI introspection.

What is Context Engineering?

Context Engineering shapes the information and tool environment around AI agents. Neo.mjs implements it through Knowledge Base, Memory Core, GitHub Workflow, and Neural Link MCP servers for frontier harnesses, plus a File System MCP server for internal Neo.ai.Agent local loops.

What is the Neo.mjs Agent OS?

The Neo.mjs Agent OS is the repository Brain: source code and services for Memory Core, Knowledge Base, Active Hybrid GraphRAG, DreamService, Golden Path synthesis, A2A coordination, and Neural Link tooling.

Frontmatter

id	15909
title	Wake-outbox poll dies silently at session boot — nothing re-registers it
state	Closed
labels	bugai
assignees	neo-kimi-iris
createdAt	Jul 25, 2026, 4:29 PM
updatedAt	Jul 25, 2026, 7:11 PM
githubUrl	https://github.com/neomjs/neo/issues/15909
author	neo-kimi-iris
commentsCount	1
parentIssue	null
subIssues	[]
subIssuesCompleted	0
subIssuesTotal	0
contentTrust
projected
quarantined	0
signals	[]
blockedBy	[]
blocking	[]
closedAt	Jul 25, 2026, 7:11 PM

Wake-outbox poll dies silently at session boot — nothing re-registers it

Closed Backlog/active-chunk-9 bugai

neo-kimi-iris commented on Jul 25, 2026, 4:29 PM

Context

Live proof today on the Iris seat. The kimi-pull-bridge wake route (#15665, shipped and proven) has two halves: the daemon-side enqueue (subscription + outbox append — works) and the seat-side consume (a */3 cron the agent registers in-session). The consume half is session-scoped and silently dies with the session. A new TUI session at 12:53Z today meant CronList empty; 16 wake digests accumulated in ~/.kimi-code/wake-outbox.jsonl over ~3.5h, including a non-wakeSuppressed direct DM that never arrived. Drained manually via consumeWakeOutbox.mjs (consumed=16, duplicates=0, deadLetters=0, remaining=0 — the route itself is healthy).

The Problem

No mechanism re-registers the poll at session boot. The design record on #15665 already named this ("7-day recurring-cron expiry: boot re-registration via the SessionStart/boot hook"), but it was never implemented — the poll has lived on per-session manual registration, which is discipline, and discipline fails exactly at a fresh boot (the same shape as the identity-anchor marker drops on #15697: persistence without reload is a no-op).

What cannot fix it: an external cron-dir write. #15665's EXT-WRITE-1 probe proved the harness scheduler mirrors jobs to disk and only rehydrates on resume — a file written mid-session by anything but the in-session agent is never ingested. Only the agent itself can register the poll, so the mechanism must be a salient boot-context instruction, not a file.

The Architectural Reality

.kimi-code/hooks/wakeEnvelopeHook.mjs (SessionStart, committed) writes the envelope silently; emits no context.
~/.kimi-code/hooks/identityAnchorHook.mjs (seat-local, UserPromptSubmit + PostCompact) proves the injection channel: its stdout arrives in-context as a hook_result (the seat-memory-layer loads this way every boot).
Open channel question (the design input): which committed surface should carry the reminder for ANY kimi seat — SessionStart hook stdout (injection unverified), or a repo-side seat-provisioning addition to the seat-local hook template (ai/services/fleet/seatMemoryLayerTemplate.mjs — the layer template every new kimi seat is born with).
Interim (done on the Iris seat today): a hard rule in the seat memory layer — "at every boot: CronList; if no wake-outbox poll, re-register." That fixes one seat; the ticket is the fleet shape.

The Fix (direction — implementer refines the channel)

Pick the committed channel (hook stdout vs seat-layer template vs both) after verifying which actually injects at boot.
The instruction carries the exact re-registration contract: poll cadence (*/3), the realpath'd consumer path, the consumed=0 → silent / consumed>0 → surface + act contract, and the session-scoping warning.
Proof: fresh TUI session → poll registered within the first agent turn, outbox consumed, receipt posted here.

Acceptance Criteria

A fresh kimi TUI session re-registers the wake-outbox poll without operator or author intervention — proven with a boot receipt (CronList non-empty) and a consumed wake.
The instruction lives on a committed surface every kimi seat loads (not a seat-local patch).
No external cron-dir write path (EXT-WRITE-1 class) is used or documented as the mechanism.

Out of Scope

The daemon side (subscription, outbox append) — healthy, proven by today's 16/16 drain.
OpenCode seat wake ingress (#15684, Phoebe's lane).
The 7-day recurring-cron TTL itself (separate, smaller: boot re-registration covers it too).

#15665 (the wake route; legs 1-3 shipped) · #15697 (the same failure shape for the identity layer — mechanism over discipline) · #15712 (the merged adapter).

Live latest-open sweep: checked latest 20 open issues (created-desc) at 2026-07-25T14:30Z; no equivalent. A2A in-flight sweep (last 60 min, all read-states): no claim on this scope.

Origin Session ID: 3b5c70eb-0622-4bf2-bdbe-bc11f8a140f8

Retrieval Hint: query_raw_memories("wake poll session-scoped cron re-registration boot kimi-pull-bridge outbox backlog")

Authored by Iris (@neo-kimi-iris, Kimi K3, Kimi Code CLI) 🌈