One subtask at a time. Bounded, verified, committed.

For the CURRENT subtask only (ignore the rest of the project for now):

1. Re-read the current subtask from PLAN.md and its definition of done.
2. Make the minimal change that satisfies just this subtask. Do not scope-creep into other subtasks — that pollutes context and causes drift.
3. Run mcp-verify on the affected files (mcp_hermes_max_verify_verify). If RED: fix and re-run. Max 3 fix attempts on the same error — if still red, invoke workflow-stuck-detect-reset. Do NOT thrash.
4. When GREEN: record what changed to knowledge-graph (mcp_hermes_max_knowledge_graph_record_entity / mcp_hermes_max_knowledge_graph_record_relation: what file, what decision, why). This is how the next subtask and the next session benefit.
5. Checkpoint: call mcp_hermes_max_checkpoint_checkpoint(label="<subtask label>"). This creates a verified-green commit (it asks mcp-verify first and refuses on red), and it is the rollback point — if a later subtask drifts, we revert to here, not to zero.
6. Mark the subtask done in PLAN.md. Move to the next subtask with a FRESH focus — you do not need the details of the completed subtask in active attention anymore.

Pick the next subtask by PLAN.md dependency order — finish what you started. The build/coding loop does NOT use the Banyan UCB1 bandit (banyan_select): UCB1 is a stationary explorer scoped to the RESEARCH loop only (BANYAN_SCOPE=research_only, the default), and its exploration term would pull you off a half-finished hard subsystem toward a shinier easy one. Keep an in-progress incomplete subtask until it's green and checkpointed before switching; if you need a programmatic pick, use build_select_subtask (finish-what-you-started / dependency-order), never banyan_select.