auto_build

ai.easyterritory/ezt-mcp · Verify this server

[Tier 1 — Balanced Territory Builder] When: partition account points into N balanced territories over a part layer (for example, ten territories; Mode A count, Mode B workload target, Scoped Split). Execution gate: after sizing, balance, dwell, and scope are known, call THIS tool immediately — searching the catalog, get_guidance, or polling Tasks never starts a build. A successful response with a new task_id is the only proof of submit; do not claim started/restarting until then. Then follow do_this_next only with that new task_id. When workflow_advisor returns next_tool=auto_build, call auto_build next. Canonical example — 3 TX ZIP territories, workload-only, 30-min dwell: build_mode={mode: fixed_territory_count, territory_count: 3}, objective={workload_bias: 100}, dwell_time={type: scalar, value: 30, unit: minutes}, part_scope=explicit, part_filter={state_abbr: TX}, map_session_id=<open MC>. Omit ts and ts_handle when the session id argument is already set. That session is the TS. Do not repost inline ts. build_mode.mode must be exactly one of: fixed_territory_count, fixed_workload_target, scoped_split. For workload targets (e.g. 40-hour territories), use build_mode={mode: fixed_workload_target, target_workload: 40} (territories only approximate the target). ALWAYS ask for dwell/onsite time before calling — Auto Build always computes territory workload hours (drive + dwell), even when balancing on a metric or account count. Pass dwell_time only after they confirm a column ({type: field, field, unit}) or scalar ({type: scalar, value, unit}). Never invent default dwell such as 1 hour or 30 minutes per visit. Server rejects auto_build without resolved dwell_time (CLARIFICATION_REQUIRED). VISIT FREQUENCY: when the point layer has a visit-frequency column, ask whether to aggregate workload across a schedule period (pass visit_frequency_field) or ignore it (pass ignore_visit_frequency=true). Do not inherit the column silently. If no visit-frequency column exists, omit both. Modern clients that negotiate protocol >= 2026-07-28 and declare form elicitation may answer missing sizing/dwell (and optional balance) in-band via Resolve/Elicit; legacy/Cursor hosts without form elicitation keep CLARIFICATION_REQUIRED / ask_user (HITL-025 — not a server failure). build_mode may be omitted when elicitation can fill it. Workload is drive time plus dwell — not Revenue or any metric column. When the user says 'balanced on workload' or 'workload-balanced', use workload_bias=100 with no objective.metric; do not ask which metric column workload means — ask dwell (Q5) only. Confirm balance dimension (Q2) and bias (Q3) with the user when unclear — a workload-hour target is sizing, not permission to silently set workload_bias=100 unless workload is already named as the balance dimension. Declare metric_fields at ingest but ask which column (if any) to balance on before auto_build when the user did not already choose workload balance. After build completes: analyze with map_session_id and analysis_panel=single (or load_analysis_panel) so stats appear in the MC dock — not chat-only. Part scope defaults to bbox_intersect (ZIPs intersecting the account-point bbox) — geographic proximity, not state/attribute scope. When the user names a state or region ('TX ZIPs only', 'Texas only'), pass part_scope=explicit and part_filter={state_abbr: TX} immediately; do not default to bbox_intersect (can pull tens of thousands of national ZIPs) and do not query_parts first for an obvious state filter. Scope fields are top-level part_filter/part_ids — never under objective. A genuinely nationwide ZIP request remains supported: when bbox_intersect resolves at national scale, the job continues at full ZIP-level quality and returns a NATIONAL_PART_SCOPE warning rather than silently switching geography. Long partitions publish named 35–42% subphases (travel cache, power solve, border refinement, compactness, seam polish) and renew their worker lease independently; keep polling via next_action/sleep_ms and do not infer a stall from one long quality pass. Prerequisites: ingest_accounts done, part_layer chosen, viewer connected (human-in-loop). Use direct_build for known assignments; account_build to group by attribute. Appends a TAL; does not replace existing alignments. Dissolved leaves report min_solidity beside mean_polsby_popper and min_polsby_popper. Gate shape from those figures rather than recomputing them from a download. Next: analyze with analysis_panel=single. Full atom: ezt://guidance/workflows/build-from-accounts. Scenarios: S003, AB-001..017, MC-011.