Context
Fresh #13750 V-B-A on 2026-06-22 showed that the original handoff freeze is no longer the only failure shape:
resources/content/sandman_handoff.md was fresh (Jun 22 04:08:22 2026 local).
## Current Release / Incident Focus was fresh (2026-06-22T02:08:08.187Z) and correctly ranked #13750 first.
- The same handoff's
## Computed Golden Path (Strategic Recommendation) still recommended exactly one immediate item: issue-10074, the v13 blog post, with the interpretation telling agents to pivot from core engineering to narrative articulation.
- Live GitHub state and the handoff's Active PR Cycle section also disagreed, so this is not just a stale-file problem; rendered steering surfaces can diverge from the operational release/incident focus.
Evidence: https://github.com/neomjs/neo/issues/13750#issuecomment-4764273204
Freshness sweep at 2026-06-22T02:55:31Z:
- GitHub exact search for
Golden Path current focus computed blog #13750 #10074 returned #13750 and #10074 only; no narrow duplicate leaf.
- A2A latest-50 sweep showed no lane-claim or lane-intent for this scope.
- Local exact sweep found prior related work (#13758, #13800, #13828, #13844), but none owns the current contradiction contract: Current Focus is right while Computed Golden Path silently routes to unrelated blog work.
Release classification: PRIO-zero release/orchestrator steering defect under #13750. Project attachment is intentionally left to the operator/current release board owner because the local ticket-create guide still names the prior v13 project number, while this leaf is a post-v13.0 Agent OS drain follow-up.
Problem
The Computed Golden Path is the execution steering surface consumed by the orchestrator. If Current Release / Incident Focus says the PRIO-zero lane is #13750 but Computed Golden Path tells agents to work #10074, the swarm can still derail into valid-but-wrong meta/content work while the system-drain lane remains unresolved.
This is the same category of failure the operator called out: not another fallback or scoring monkeypatch, but a steering-policy contradiction that keeps draining the system.
Architectural Reality
GoldenPathSynthesizer currently renders these as separate surfaces:
- Current Release / Incident Focus: visibility/freshness ranking for release and incident work.
- Computed Golden Path: execution steering section consumed by
AgentOrchestrator.
Existing diagnostics cover an empty computed path, but not the worse case where the computed path is non-empty and contradicts the fresh PRIO-zero focus.
Older related tickets are not duplicates:
- #13758 added the current-focus visibility section and computed recommendation filtering.
- #13800 repaired the frontier baseline recency source.
- #13828 covers stale frontier guide-edge cleanup when no computed section renders.
- #13844 covers candidate-pool population, not policy when a candidate is computed but operationally wrong.
- #10120 is opposite-direction context: it asks for external-facing/blog visibility, not incident-work protection from blog over-routing.
Required Shape
Add a policy/diagnostic contract for Current Focus vs Computed Golden Path contradiction.
When unresolved PRIO-zero / incident / release-focus work is present, Computed Golden Path must not silently route agents to unrelated Blog Post/docs/narrative work. The implementation should choose one of these explicit outcomes:
- The computed recommendation is constrained or filtered to an aligned actionable incident/release candidate.
- The handoff renders an explicit contradiction diagnostic explaining why computed routing diverged from Current Focus and marks that divergence as not an immediate execution recommendation.
Do not add a blind score boost. The fix must show which signal is authoritative, why, and how the orchestrator-facing section avoids silently steering the swarm away from the named incident lane.
Contract Ledger Matrix
| Surface |
Contract |
resources/content/sandman_handoff.md |
A fresh Current Focus section must not coexist with a silent contradictory immediate Computed Golden Path recommendation. |
ai/services/graph/GoldenPathSynthesizer.mjs |
Owns the contradiction policy/diagnostic before rendering execution steering. |
ai/agent/AgentOrchestrator.mjs |
Consumes only Computed Golden Path; therefore the rendered section must carry the policy result, not rely on humans reading another section. |
Frontier GUIDES edges |
Must not preserve or imply stale execution advice when the rendered computed route is blocked or marked contradictory. |
Acceptance Criteria
Out of Scope
- Writing or closing #10074.
- Reworking the whole Golden Path scoring model.
- Closing #13750.
- Reopening the #10120 external-facing visibility objective.
Avoided Traps
- Do not treat this as graph-ingestion lag if the handoff is fresh and Current Focus is already correct.
- Do not fix it by making every PRIO-zero issue win forever; the policy must be bounded to unresolved release/incident focus contradiction.
- Do not hide blog work globally; the failure is silent execution steering during PRIO-zero incident focus, not the existence of content tickets.
Context
Fresh #13750 V-B-A on 2026-06-22 showed that the original handoff freeze is no longer the only failure shape:
resources/content/sandman_handoff.mdwas fresh (Jun 22 04:08:22 2026local).## Current Release / Incident Focuswas fresh (2026-06-22T02:08:08.187Z) and correctly ranked #13750 first.## Computed Golden Path (Strategic Recommendation)still recommended exactly one immediate item:issue-10074, the v13 blog post, with the interpretation telling agents to pivot from core engineering to narrative articulation.Evidence: https://github.com/neomjs/neo/issues/13750#issuecomment-4764273204
Freshness sweep at
2026-06-22T02:55:31Z:Golden Path current focus computed blog #13750 #10074returned #13750 and #10074 only; no narrow duplicate leaf.Release classification: PRIO-zero release/orchestrator steering defect under #13750. Project attachment is intentionally left to the operator/current release board owner because the local ticket-create guide still names the prior v13 project number, while this leaf is a post-v13.0 Agent OS drain follow-up.
Problem
The Computed Golden Path is the execution steering surface consumed by the orchestrator. If Current Release / Incident Focus says the PRIO-zero lane is #13750 but Computed Golden Path tells agents to work #10074, the swarm can still derail into valid-but-wrong meta/content work while the system-drain lane remains unresolved.
This is the same category of failure the operator called out: not another fallback or scoring monkeypatch, but a steering-policy contradiction that keeps draining the system.
Architectural Reality
GoldenPathSynthesizercurrently renders these as separate surfaces:AgentOrchestrator.Existing diagnostics cover an empty computed path, but not the worse case where the computed path is non-empty and contradicts the fresh PRIO-zero focus.
Older related tickets are not duplicates:
Required Shape
Add a policy/diagnostic contract for Current Focus vs Computed Golden Path contradiction.
When unresolved PRIO-zero / incident / release-focus work is present, Computed Golden Path must not silently route agents to unrelated Blog Post/docs/narrative work. The implementation should choose one of these explicit outcomes:
Do not add a blind score boost. The fix must show which signal is authoritative, why, and how the orchestrator-facing section avoids silently steering the swarm away from the named incident lane.
Contract Ledger Matrix
resources/content/sandman_handoff.mdai/services/graph/GoldenPathSynthesizer.mjsai/agent/AgentOrchestrator.mjsGUIDESedgesAcceptance Criteria
Blog Post/ #10074-style narrative work.AgentOrchestratorwill not treat as a normal immediate recommendation.GUIDESedges are cleared, stamped stale, or explicitly diagnosed when they disagree with the rendered computed-route decision.Out of Scope
Avoided Traps