adaptive-grok-build-pro: Instructions file for Claude Code

Adaptive Grok Build Pro Engineering Contract

Mandatory startup algorithm: measure, then dispatch

Step zero precedes all other startup work, including backlog/route inspection, dependency planning, agent spawning and CPU-heavy commands. Complete resource discovery below and record its snapshot locally first; attach it to the change package after selecting the route. Only then inspect backlog/routes, plan dependencies, dispatch work, assign isolated writers and apply verification/delivery gates, in that order.

1. Observe host physical-core and online logical-CPU topology (lscpu, nproc --all), current process capacity (nproc) and affinity (taskset -pc <pid>). Physical cores, logical CPUs and allowed CPUs are different quantities.
2. Resolve the process's actual cgroup membership/mounts, effective cpuset and finite CPU quota, including applicable ancestor limits. Do not infer usable capacity from host CPU count alone; if a bound cannot be established, report the uncertainty and use a conservative capacity.
3. When affinity appears narrower than the effective cpuset and policy permits it, run one bounded child-only affinity-widening probe over the candidate allowed CPU IDs. For example, taskset -c 0-27 nproc previously exposed 28 CPUs where default nproc showed 22. Recheck that child's affinity and quotas; a failed probe leaves the existing limits in force. Never assume this example's CPU IDs exist on another host or change the controller/system affinity as a shortcut.
4. Compute verified effective CPU capacity from the child/process allowed online CPUs, effective cpuset and finite quota (conservatively round quota capacity down, with one worker minimum). Record timestamp, commands/results, topology, affinity, cpuset/quota bounds, probe result and chosen capacity. Remeasure at every startup and when the execution environment changes. The September 26 host observation was 14 physical cores / 28 logical CPUs, not a permanent capacity guarantee.
5. After the CPU snapshot is recorded, inspect repository handoff/backlog/routes and build the dependency plan. Separately observe available agent slots: the current platform exposes one controller plus 12 child-agent slots; route max_parallel_analysis=10 remains the routing cap, and test-process worker counts are a third independent limit. Never manufacture route permissions or add unselected agents to fill slots.
6. Dispatch all independent route-permitted analyses, checks and reviews in parallel when their prerequisites are satisfied, scheduling available slots in waves. Spread eligible CPU-heavy child work across the verified effective CPU capacity, using the successfully probed affinity when needed and coordinating worker totals to avoid oversubscription. Record dependencies, isolation or resource limits that require serialization.
7. Keep exactly one write owner per isolated task/route/branch/worktree. Independent writers may run concurrently only on separate isolated task contours and worktrees; no two writers share a mutable candidate. Reviews remain independent and read-only.