# Agent Start Here Use this file to avoid loading large context up front. ## 90-Second Startup 1. Read `AGENTS.md` for repository policy. 2. Read `documentation/agents/ci_lane_contract.md` to select lane: - upstream GitHub lane (no hardware) - local LXD lane (hardware/HIL) 3. If the request requires remote-role work (`infra`/`runner`), read: - `documentation/agents/remote_delegation_contract.md` (handshake + `thread_ref` correlation) 4. If the request references a short-term goal, read: - `documentation/short-term/coordination/workstream_board.md` - the referenced file under `documentation/short-term/active/` - `documentation/short-term/coordination/handoff_log.md` - owner-governance rules in the board (`Owner Governance` section) - role activation tracker/state in the board 5. Read one task route doc only: - hardware/integration execution: `documentation/agents/integration_test_worklist.md` - documentation linting/gardening: `documentation/agents/document_gardening.md` - frontend payload/route/API changes: `documentation/agents/frontend_requirements_context.md` ## Load-On-Demand Rule - Start with the smallest possible doc set. - Pull in more docs only when the current task hits a dependency. - Keep updates local to the active goal/workstream. ## Session Close-Out 1. Update active goal/workstream status if changed. 2. Add evidence pointers (logs, run IDs, artifacts, commands). 3. Add a handoff line in `documentation/short-term/coordination/handoff_log.md` when work remains. 4. If you are the owner (especially `arch`) and work is not `done`, ensure the workstream is either: - `blocked` with concrete blockers + owners + operator gate + next action, or - explicitly delegated with a `thread_ref` and a scheduled follow-up check (do not leave work as "implicitly pending") - if stopping due to a required pivot or a stalled loop, log `action_type=replan` (pivot) or `blocked` (stall) with the new decision/unknowns captured 5. If closing a goal: verify all deliverables checklist items in the goal file are checked. If any are unchecked, keep goal status active and do not archive. 6. For remote-role work: - do not set `in_progress` when owner is `tbd` - include `context`, `action_type`, and `operator_required` fields in handoff entries - update role activation tracker (`unassigned`/`assigned`/ `first_heartbeat`/`active`) when state changes 7. For user/operator course-correction signals (for example: \"we missed...\", \"easier way...\", \"we should prioritize...\"): - treat as formal `action_type=replan` - update goal workstreams + board dependencies before resuming execution 8. For ad-hoc user requests: - proactively ask if user wants ticket tracking (`yes`/`no`) - if `yes`, create/update `documentation/short-term/coordination/ad_hoc_ticket_queue.md` - if `no`, proceed and log `ticket_tracking=declined` in handoff