setup-baseline

In: contract.md, as init settled it. Out: a workspace whose base row is recorded in records/candidates.jsonl and benchmark.csv, and whose base commit carries a trace and a diagnosis.

1. Create a separate worktree and make every change there. Leave the original checkout alone.
2. Use <worktree>/workspace-who-ate-my-flops/ as <ws>. Add /workspace-who-ate-my-flops/ to the local exclude file returned by git rev-parse --git-path info/exclude from the worktree, preserving existing entries. Create the workspace with python <plugin root>/tools/workspace.py init <ws> <contract>. The plugin root is the directory two levels above this SKILL.md, and Claude Code also exposes it as ${CLAUDE_PLUGIN_ROOT}. Every tool this plugin ships resolves from there, because you are standing in the target's worktree and never in the plugin. Read workspace.py first: it defines the file formats everything downstream uses, and the workspace is the common ground the skills coordinate through. Keep process records in records/, including any planning notes. Leave contract.md, latest-report.md, and benchmark.csv at the workspace root. Put helper scripts in tools/ and run artifacts in runs/. Avoid adding ad hoc files to the workspace root.
3. Use the extract-computation-graph skill to build the job's computation graph and execution schedule inside the workspace: records/computation_graph.json for what is computed, and records/execution_schedule.json for when and where it is computed and what moves between devices. Both get read against the profile results later, to locate the performance problems and find fixes for them.

4. Record the original code's speed, memory and correctness using the scripts defined in run of contract.md, and diagnose it. The baseline covers all four, and all four get written down:

   • Diagnosis: run the torch-profile skill on the unmodified code, in full. Its traces go under commits/<base sha>/trace/ and its report to commits/<base sha>/report.md. This is the one diagnosis of the user's own code: every later profile only says whether a change removed something from it, and the handover points back to it. finish refuses a campaign without it.
   • Speed: by default measure the warm-up cost, which is paid once, and the steady step time, which is paid many times over.
   • Memory: measure CPU and GPU memory, and watch how each moves across the warm-up phase and the steady-step phase.
   • Correctness: a. Read the ## Correctness Instrumentation Plan section of contract.md for where the checkpoints go and which recorder it names. b. Use the recorder the repo already has. In imp_genai that is research/core/probe, the same code vendored under its own name. Install one only when the repo has none, and then install parity (https://github.com/OpenPerfAgent/parity): import parity, recorded when PARITY=1 with PARITY_OUT=<file>, compared with parity compare. c. Read its README first, then put the planned calls in. d. Smoke run it 3 times, and work out the tolerance its numbers actually hold to. parity derive does this, and it needs three runs of unmodified code.

   Record the result with workspace.py add --verdict base, and carry the tolerance forward: every later comparison is read against it.

The questions ended with init, but setup can still fail in a way worth handing back: a runner that does not run, a node that is busy, a tolerance that will not hold. Say what failed and stop, rather than working around it.

Arm nothing here. What keeps a campaign turning for hours belongs to the caller, and a diagnosis has no loop to keep turning.