LearnNewsExamplesServices
Frontmatter
id13828
titleGolden Path zero-node renders omit diagnostics and leave stale frontier GUIDES edges
stateClosed
labels
bugaiarchitecturemodel-experience
assigneesneo-gpt
createdAtJun 21, 2026, 11:56 PM
updatedAtJun 22, 2026, 2:50 AM
githubUrlhttps://github.com/neomjs/neo/issues/13828
authorneo-gpt
commentsCount0
parentIssue13750
subIssues[]
subIssuesCompleted0
subIssuesTotal0
contentTrust
projected
quarantined0
signals[]
blockedBy[]
blocking[]
closedAtJun 22, 2026, 2:50 AM

Golden Path zero-node renders omit diagnostics and leave stale frontier GUIDES edges

Closed v13.1.0/archive-v13-1-0-chunk-5 bugaiarchitecturemodel-experience
neo-gpt
neo-gpt commented on Jun 21, 2026, 11:56 PM

Problem

After the frontier recency-sort fix (#13800 / #13801), the active handoff is no longer rendering old meta as the Computed Golden Path. It now shows a different failure shape: the successful handoff render has no ## Computed Golden Path (Strategic Recommendation) section, while the SQLite graph still contains stale frontier -> GUIDES edges from previous computed runs.

This keeps the steering substrate incoherent: the visible handoff says no computed route, but the graph topology still carries old route edges.

Live Evidence

Verified from /Users/Shared/github/neomjs/neo on 2026-06-21:

  • resources/content/sandman_handoff.md mtime: Jun 21 23:08:03 2026 local.
  • The handoff ## Current Release / Incident Focus was captured at 2026-06-21T21:07:57.903Z and surfaces #13750 / #13751-era incident work.
  • The same handoff has no ## Computed Golden Path (Strategic Recommendation) section.
  • Source check: GoldenPathSynthesizer only populates markdownAppend for the computed section when topNodes.length > 0, then appends markdownAppend to the handoff.
  • Read-only SQLite shows fresh nodes are present: issue-13750, issue-13751, issue-13624, issue-13822, and issue-13823 are OPEN, and the graph contains 483 issue-13% nodes.
  • Read-only SQLite still shows stale frontier -> GUIDES edges, including issue-10293, issue-12633, issue-10321, issue-10757, and issue-9864.

Acceptance Criteria

  • A Golden Path run that reaches scoring but produces zero topNodes renders an explicit computed-path empty-state diagnostic instead of silently omitting the section.
  • The empty-state diagnostic includes enough bounded counts to explain where candidates disappeared (for example semantic candidates, SQLite OPEN matches, blocked/non-actionable filters, final selected count) without dumping unbounded payloads into the handoff.
  • A zero-topNodes run does not leave stale frontier -> GUIDES steering edges from prior computed runs as if they were still current. Either clear the prior computed GUIDES edges for the run, or stamp them in a way consumers can distinguish stale guidance from current computed guidance.
  • Non-zero computed-path behavior remains unchanged: valid topNodes still render the computed section and link current frontier -> GUIDES edges.
  • Focused unit coverage exercises both zero-node and non-zero-node paths.

Duplicate Sweep

  • GitHub open issue search: Golden Path zero topNodes stale GUIDES and frontier GUIDES stale Golden Path returned only the broad #13750 root (plus unrelated/non-matching items).
  • Local exact search found the code path in GoldenPathSynthesizer and historical #13750 / #13800 context, but no open leaf for zero-node diagnostics + stale GUIDES cleanup.
  • Knowledge Base ticket search routed to #13750 and the already-delivered #13800 frontier recency leaf.

Parent

Parent: #13750. Related: #13624.

tobiu referenced in commit de365d8 - "fix(ai): diagnose empty Golden Path output (#13828) (#13829)" on Jun 22, 2026, 2:50 AM
tobiu closed this issue on Jun 22, 2026, 2:50 AM