Council Review: Sprint 35 LLM-First Delta Audit — Seat Synthesis

Date: 2026-06-23
Scope: MemorySmith.Agent head commit 073175b03387e05592780016db345f7ae48217c0
Self-simulated seats (no subagents: user permission not requested)

Decision

Proceed with P0/P1 urgent fixes now (five concrete items); defer strategic architecture decomposition and data model modernization to Sprint 36+ with explicit evidence gates.

Evidence Reviewed

Seat Findings Summary

Synthesis: What Changes NOW vs What is DEFERRED

🔴 CHANGE NOW — P0/P1 Urgent Fixes

These are runtime bugs or data-loss risks that are actively degrading agent behavior today. They should be implemented before any strategic work.

🟡 CHANGE NOW — P1 Code Health Items

🔵 DEFERRED — Sprint 36+ Strategic Improvements

These are important architectural improvements that should NOT block the P0/P1 fixes above. Each has an evidence gate.

Where Seats Agree vs Disagree

High-Confidence Consensus (≥3 seats, ≥0.80 confidence)

Disagreements (need resolution)

Evidence Gates for Deferred Items

Every deferred item must pass its gate before implementation begins:

Risks & Assumptions

Risks

1. P0 fix #1 (terrain skip) may require adapter-level changes. If the blockPlaced event cannot distinguish "block placed in world" from "block metadata updated at occupied position," the fix may need a Mineflayer adapter patch — which increases scope and testing burden. - Mitigation: Start with C#-side filtering (compare pre-place world state vs post-event world state). Only escalate to adapter changes if filtering proves insufficient.
2. P0 fix #3 (wire SearchMemory) has unknown scope. TSK-0004 is tracked but not scoped. Wiring results into planning may require changes to HtnTaskLibrary, ActionData flow, and decision logic. - Mitigation: Start with the simplest possible consumer (log results, use bestPageId as a hint in the next tool dispatch). Iterate from there.
3. P0 fix #2 (smelt→craft) may reveal furnace tool is missing entirely. If no furnace-specific tool exists, the fix requires a new tool, decomposer, and adapter action. - Mitigation: Audit existing tools first. If absent, implement a SmeltItemTool that wraps the furnace adapter action (likely already supported in Mineflayer).

Assumptions

• The 5 P0/P1 items are independent and can be implemented in parallel by separate branches.
• No P0/P1 fix requires a schema migration or breaking API change.
• The IKnowledgeResolver interface (Agent.Memory/IKnowledgeResolver.cs) is stable enough to wire after TSK-0004, even if its implementation evolves.
• The 608 unit tests will catch regressions in P0/P1 changes for non-checkpoint code paths. Checkpoint/occupancy tests must be written alongside fix #1.

Open Questions

1. Does the adapter distinguish blockPlaced (real placement) from blockUpdate (metadata change at occupied position)? This determines the fix strategy for TSK-0074/TSK-0075. Need to check MineflayerAdapter/index.js event emissions.
2. Is there already a SmeltItemTool or furnace-related action in the adapter that's simply not wired through the chat interpreter? The code audit found no evidence, but a quick adapter grep would confirm.
3. Should SearchMemory results be consumed at plan-construction time (in HtnTaskLibrary) or at dispatch time (in AgentBackgroundService)? TSK-0004 needs a design decision.
4. How many of the ~15 SearchMemory calls per cycle are genuinely useful (i.e., would produce actionable results) vs. noise? A brief instrumentation pass before TSK-0004 implementation would inform the deduplication strategy.

Directional Assessment

Is the project on the right track? — Yes, with a caveat.

The LLM-first architecture direction (Sprint 35) is sound. The IntentDraft → GoalRequest → IntentManager pipeline is a meaningful improvement over the old parsers-create-goals model. The code is materially ahead of where it was 5 sprints ago.

However: The gap between architectural intent and runtime behavior has widened in three areas:

1. Silent failures — The TSK-0074 terrain skip is a regression that introduces less observable failure modes than the stalls it replaced. This is a net negative for debuggability.
2. Orphaned infrastructure — SearchMemory is called ~15 times per cycle and produces zero behavioral effect. This is wasted latency, API cost, and code complexity. It suggests a pattern of "build the tool, forget to wire the consumer."
3. Testing gap — The most critical new behavior (checkpoint/occupancy) has zero tests. This is a predictable outcome when unit-test coverage (608 tests) masks the absence of integration and E2E tests.

Bottom line: Fix the five P0/P1 items immediately. They are the difference between "architecturally improving but regressing in observable behavior" and "architecturally improving AND running correctly." After those are done, the strategic decomposition and data model work in Sprint 36+ will rest on a solid foundation.

Acceptance Criteria

This council review is complete when:

• [x] Decision statement is explicit: Proceed with P0/P1 fixes now; defer strategic work to Sprint 36+
• [x] Evidence links are source-grounded (code paths confirmed)
• [x] All 5 seats include confidence values and blocking concerns
• [x] Dissent is visible (GoalFactory rename priority, decomposition order)
• [x] Acceptance criteria are testable (evidence gates table)
• [ ] Report is saved to Data/Pages/council/sprint35-llm-first-seat-synthesis-20260623.md
• [ ] Council report is linked from the active task tracker