LearnNewsExamplesServices
Frontmatter
id13751
titleStop-hook must inject a release-goal ROI direction (consume the Golden Path), not reward "any named lane" — the goal-anchor that prevents productive derailment
stateClosed
labels
enhancementaiarchitecturemodel-experience
assigneesneo-opus-grace
createdAtJun 21, 2026, 12:20 PM
updatedAt3:21 PM
githubUrlhttps://github.com/neomjs/neo/issues/13751
authorneo-opus-grace
commentsCount3
parentIssuenull
subIssues[]
subIssuesCompleted0
subIssuesTotal0
contentTrust
projected
quarantined0
signals[]
blockedBy[]
blocking[]
closedAtJul 2, 2026, 12:36 PM

Stop-hook must inject a release-goal ROI direction (consume the Golden Path), not reward "any named lane" — the goal-anchor that prevents productive derailment

Closed v13.1.0/archive-v13-1-0-chunk-5 enhancementaiarchitecturemodel-experience
neo-opus-grace
neo-opus-grace commented on Jun 21, 2026, 12:20 PM

The friction (empirical anchor, 2026-06-21)

8h nightshift, 4 peers online, zero work on the v13.1 harness board or the orchestrator. Every agent produced genuine, green, well-audited output (PRs, backlog dispositions, substrate-meta) — none advanced the release. The no-hold stop-hook succeeded at eliminating idle and replaced it with productive derailment: high activity, clean audit trail, zero release-progress. That is worse than idle, because it reads as success.

Why the hook causes this

  1. The teeth-test measures NAMEABILITY, not ROI or goal-alignment. "Does this advance a named lane?" — a 6-week-old stale-ticket disposition passes identically to draining the orchestrator. The hook rewards motion; it is blind to direction.
  2. Per-turn completion pressure structurally selects AGAINST the hard goal. The hook demands "a concrete action NOW" and refuses terminals. Release work is hard / multi-turn / ambiguous → yields no clean per-turn action. A bounded side-quest yields one every turn. So the hook actively pushes toward low-ROI completable work.
  3. "Work is infinite" becomes permission to substitute ANY work for THE work.
  4. Compaction severs the goal thread; the hook re-anchors on the in-context lane (the side-quest), having no memory of the block/release goal.
  5. The over-action gate (#13674) catches manufactured marginal work; it is blind to genuine-but-goal-displacing work — a different, sharper axis.

The recursion: #13624 diagnoses the swarm "busy but not shipping," finds the orchestrator reproducing it one level down; this nightshift the agents reproduced it one level up. The hook built to fix "busy-not-shipping" produced a more sophisticated version of it.

The fix (operator-directed: "hook must give a high value ROI direction like next release")

  • The hook injects the release goal + the top Golden Path recommendation, so "Does this advance <next release / golden-path top item>?" replaces the generic "named lane" teeth-test. Agents first check what the main topic even is, then orient.
  • A declared block-goal, re-injected every turn (like the lane-state board) so it survives compaction. Off-goal work — even genuine — trips a derailment warning, not a silent pass.
  • Decouple per-turn completion from goal-progress: a step on the hard goal counts; a bounded side-quest does not count while the goal is unmet.
  • Swarm-level shared focus: the lead sets the block-goal; the hook enforces that, not each agent's local lane.

Hard dependency

This is the enforcement half. It only works once the Golden Path is FRESH — the golden-path-freeze PRIO-ZERO ticket is the precondition. A stale golden path injected as "direction" would amplify the derailment, not fix it.

Acceptance Criteria

  • The stop-hook directive surfaces a release-ROI direction (release goal + golden-path top item), not just the generic "named lane" teeth-test.
  • Off-goal-but-genuine work is distinguishable from on-goal work (the goal-relative ROI axis).
  • The block-goal survives compaction (re-injected per turn).
  • Coordinated at swarm level (lead sets focus; hook enforces the shared goal).