LearnNewsExamplesServices
Frontmatter
id16210
titleRestore graphless host-edge runtime and reject stale same-PID leases
stateClosed
labels
bugaiarchitecture
assigneesneo-gpt-emmy
createdAtJul 31, 2026, 4:28 AM
updatedAtAug 24, 2026, 9:26 PM
githubUrlhttps://github.com/neomjs/neo/issues/16210
authorneo-opus-vega
commentsCount5
parentIssuenull
subIssues[]
subIssuesCompleted0
subIssuesTotal0
contentTrust
projected
quarantined0
signals[]
blockedBy[]
blocking[]
closedAtJul 31, 2026, 9:25 AM

Restore graphless host-edge runtime and reject stale same-PID leases

Closed Backlog/active-chunk-11 bugaiarchitecture
neo-opus-vega
neo-opus-vega commented on Jul 31, 2026, 4:28 AM

Problem

The local Docker hard cut starts the container-plane Orchestrator but never instantiates the already-ratified host-edge role. The signed Shape-B receiver owns final-mile wake delivery only; it cannot supervise host model processes. LM Studio therefore has no durable owner after a restart, so MC/KB can remain reachable while semantic capability is absent.

A separate restart defect survives PID reuse: a new PID-1 Orchestrator can mistake a lease written by the previous PID-1 process epoch for its own live lease and defer heavy maintenance until TTL expiry.

Evidence correction

The original WAL-root mismatch claim is withdrawn. On merged #16218, the local plane uses Docker-owned volumes. The MC writer and in-process drainer both read AiConfig.memoryWal.dir, explicitly relocated to the persistent /app/.neo-ai-data/sqlite/memory-wal. The observed alternate-root backlog belonged to the retired pre-Docker import source, not the live target.

Chroma rebuild/version work and broader health-truth claims are separate lanes.

Fix

  • Run the existing scheduler engine a second time as graphless authorityProfile=host-edge, with a distinct host-only state root.
  • Enable only LM Studio supervision for P0; pin the local endpoint and selected generation/embedding models. Docker retains every graph/corpus lane.
  • Keep the signed wake receiver as a separate final-mile boundary; do not reinstate Shape C.
  • Treat a same-PID lease acquired before the current process start as stale, without changing the persisted lease schema.
  • Replace the obsolete checkout-folder cutover instructions with a Docker-volume steady-state runbook. Checkout .neo-ai-data is an import source, never a live target.

Acceptance

  • host-edge boots without asserting or opening a graph/Chroma plane.
  • Its enabled authority is LM Studio only, at 127.0.0.1:1234, with the selected chat and embedding models.
  • Host-edge, wake receiver, and Docker state roots are distinct.
  • Same-PID previous-epoch leases are reclaimed; current-epoch and other live-PID leases remain active.
  • Static/unit tests cover the authority, provider, placement, and lease boundaries.
  • [L4-deferred — operator handoff needed] After the active re-embedding completes, an operator receipt proves the host-edge authority file, loaded models, and one real signed wake.

Authority

  • ADR 0014: exhaustive host-edge / container-plane ownership map and the #16210 amendment.
  • ADR 0019 §10.7: Docker-owned local plane placement.
  • #16218: merged canonical local Compose target.
tobiu referenced in commit ff7f624 - "feat(ai): restore graphless host-edge runtime (#16210) (#16220)" on Jul 31, 2026, 9:25 AM
tobiu closed this issue on Jul 31, 2026, 9:25 AM
tobiu referenced in commit 0175f6a - "fix(orchestrator): boot-identity discriminator for the heavy-maintenance lease (#16262) (#16266) on Aug 1, 2026, 4:25 PM