Agent Workflows

Published from the source document. Propose an edit through a pull request; direct wiki edits are replaced by the next publication.

Chat-led scheduling

The central server uses scenarios.create|revise, runs.start|get and results.get. Facts are canonical engine input; production-template expansion and live source adapters are not exposed as central tools yet. A supporting chat host displays the MCP App returned with a result. Approval/publication sets the active plan inside APEX and does not write back to MES/ERP. See control platform.

The compatibility executable's web UI is a plan viewer: schedule, KPIs, order completion, commitments and operation details. It does not require the planner to understand scenario forks, patches, Q policies or search budgets. The agent performs those operations through tools and returns a saved-plan link. ?mode=workbench exposes the development controls when needed.

The Discuss in your chat button copies a bounded context with scenario, revision, schedule, baseline and selected operation IDs. It does not send a message or embed another model. A tool-capable chat needs an actual MCP/HTTP connection; a URL alone does not grant tool access. The UI is independent of the chat provider.

Agent workflows

Three portable repository skills live in skills/:

Skill	Responsibility	Handoff
apex-data-intake	Read authorized sources, convert units and relationships, validate and import	Scenario/revision, source mapping, counts, unresolved facts
apex-planning	Discuss domain rules, explain plans, run scenarios and sensitivity cases	Validated results, KPI trade-offs, viewer links
apex-extension	Implement new rules, objectives, hooks and measured Q proxies	Tested code/schema, synthetic examples, documented limits

Customer extensions follow the customization boundary and approval workflow. The agent uses existing package interfaces first and obtains informed, scoped approval before changing shipped base application files, including registration/build wiring. The skill records upgrade implications; it is agent guidance, not a runtime write barrier.

A single agent can switch skills. Separate agents are optional when distinct credentials, domains or independently testable implementation work justify them. They must exchange artifact references and explicit requirements, not duplicate full production exports in their context. These are ordinary Markdown skills; hosts without skill discovery can load the relevant file as workflow guidance.

External writeback is deferred. A later writeback capability needs an explicit target, source revision, agreed change set, conflict handling and retry semantics. The current roles do not publish plans into MES/ERP.

flowchart LR
  Sources[Excel / MES / ERP / files] --> Intake[Data intake skill or optional connector]
  Chat[Planner + agent chat] <--> Planning[Planning skill]
  Intake --> Input[Validated input artifact]
  Input --> Material[Optional existing-supply pegging]
  Material --> Tools[APEX MCP / HTTP tools]
  Input --> Tools
  Planning <--> Tools
  Planning --> Extension[Extension skill]
  Knowledge[Customization knowledge + provenance] --> Planning
  Knowledge --> Extension
  Extension --> Tested[Typed rules / Rust hooks / tests / proxy evidence]
  Tested --> Core[Rust scheduling core]
  Tools --> Core
  Core --> Validation[Independent validation + exact KPIs]
  Validation --> Results[Saved schedules and comparisons]
  Results --> Viewer[Lean plan viewer]
  Results --> Planning

Loading

The following fork/freeze/patch sequence describes the compatibility tools. On the central server, preserve a baseline scenario, submit canonical facts to a new scenario or revision, and compare retained result metrics explicitly.

sequenceDiagram
  actor Planner
  participant Agent
  participant APEX
  participant Viewer
  Planner->>Agent: Compare an outage; keep today's assignments fixed
  Agent->>APEX: Inspect saved baseline + current revision
  Agent->>APEX: Fork scenario; freeze requested dimensions; patch outage
  Agent->>APEX: Quick plan / Improve plan with an explicit total budget
  APEX-->>Agent: Saved validated plan + search report
  Agent->>APEX: Compare baseline and candidate
  Agent-->>Planner: Trade-offs, assumptions and view link
  Planner->>Viewer: Inspect schedule and fixed commitments

Loading

Domain knowledge and implementation

Keep human-readable rules in customization Markdown with: stable rule ID, statement, source, owner, units, scope, hard/soft classification, exceptions, status and links to implementation/tests. Proposed rules and estimates remain distinct from confirmed facts. This record helps the agent discuss requirements; it does not execute constraints.

A new metric needs its exact schedule evaluation and a useful candidate-ranking estimate. Standard Q mappings cover common goals. A coding agent can add a custom Q, but no general method guarantees that automatically generated proxies correlate well with every objective. Measure correlation/regret on full schedules and held-out cases, keep adverse cases visible, and retain exact fitness and independent validation.

XH optimizes policies; it is not the correctness checker. See search semantics, material preparation and transport setup.