Morrowind: Navigate Vivec Foreign Quarter with spatial awareness #101
Reference in New Issue
Block a user
Delete Branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Task
Prove Timmy can navigate a complex interior/exterior cell:
Acceptance Criteria
Parent: EPIC #99
🔨 Artisan Review — Issue #101: Navigate Vivec Foreign Quarter #bezalel-artisan
Verdict: KEEP OPEN — This is the spatial reasoning proof. Critical for the thesis.
Analysis
Vivec Foreign Quarter is an excellent test environment — a complex multi-level structure with interior/exterior transitions, multiple named NPCs, and distinct cell boundaries. If Timmy can navigate here, the architecture works.
Key Challenges
Door detection ambiguity — Perception reports doors by name, but Morrowind door names are often generic ("Door" or "Wooden Door"). The agent needs to correlate door positions with spatial context, not just names.
Cell transition verification — When you activate a door, the cell name in perception should change. This is the ground truth for "did I actually move?" But there's a lag — perception parsing may need a wait-and-retry after door activation.
3D navigation in 2.5D thinking — The agent reasons in text (no spatial model), but Vivec has multiple vertical levels connected by ramps. Moving "forward" toward a door on a different level requires pathfinding the agent can't do.
Describing without hallucinating — Acceptance criterion #3 is crucial. The agent must report what perception says, not what it "knows" about Morrowind from training data. This is the integrity test.
Suggested Approach
{door_name: relative_position}from perceptionDependency
Blocked by #100 — all tools must be verified first.
Deep triage pass: keep this open. This is a good evaluation task because it tests perception-grounded navigation in a space where language priors are actively dangerous. Vivec is exactly the kind of environment where an agent can sound competent while actually hallucinating.
The strongest acceptance criterion here is the requirement to describe the environment from perception output rather than prior game knowledge. I would tighten execution around a simple proof bundle:
The main blocker remains #100/tool verification. Once that is cleared, this issue should produce a compact navigation transcript proving that the agent can bind text observations to spatial action without making up the world.
Rerouting this issue out of the Gemini code loop.
Reason: it does not look like code-fit implementation work for the active Gemini coding lane. Leaving it unassigned keeps the queue truthful and prevents crash-loop churn on non-code/frontier issues.
Restoring this issue to the Gemini code lane.
Reason: this is concrete implementation work, not a non-code epic/KT/report item.