fsm-engine

---

name: fsm-engine description: CroqTuner harness — robust FSM-driven GPU kernel tuning loop with guard flags, idea dedup, and compaction-safe state. Wraps ai-tune-from-current-best and ai-tune-from-scratch with mechanical enforcement of the tuning protocol. Use when the user asks to "tune", "ai-tune", "croqtune", "tune from best", "tune from scratch", or "start tuning". argument-hint: <mode: from-best|from-scratch> <dtype: f16|e4m3|all> [shape_key]

CroqTuner — Router Skill

This skill manages the tuning loop through a finite state machine. Instead of relying on LLM memory to follow a long protocol, it uses:

• Mode-specific state files — each tuning mode has its own FSM state file so they can run concurrently
• Validation scripts — pre/post checks that block illegal transitions
• Idea dedup log — append-only JSONL preventing repeated ideas
• Compaction summary — structured resume instructions surviving context loss

Dual State Files (Concurrent Mode Support)

Two tuning modes can run in parallel on separate agent sessions because each has its own state file:

Mode	State file	Env var
from_current_best	state/loop-state.json (default)	not needed
from_scratch	state/loop-state_from_scratch.json	CROQTUNER_STATE_FILE=.claude/skills/fsm-engine/state/loop-state_from_scratch.json

All FSM scripts (state-transition.sh, pre-step-check.sh, post-step-check.sh) respect the CROQTUNER_STATE_FILE environment variable. If unset, they default to state/loop-state.json.

For from-scratch mode, ALWAYS set the env var before calling any FSM script:

export CROQTUNER_STATE_FILE=".claude/skills/fsm-engine/state/loop-state_from_scratch.json"

FIRST THING: Read State

On every invocation, your FIRST action is:

# Determine which state file to use based on mode
if [ "<MODE>" = "from_scratch" ]; then
    export CROQTUNER_STATE_FILE=".claude/skills/fsm-engine/state/loop-state_from_scratch.json"
else
    export CROQTUNER_STATE_FILE=".claude/skills/fsm-engine/state/loop-state.json"
fi

if [ -f "$CROQTUNER_STATE_FILE" ]; then
    cat "$CROQTUNER_STATE_FILE"
else
    echo "NO_STATE"
fi

Based on the result:

If NO_STATE → Fresh Start

1. Read protocol/identity.md — your role and inviolable constraints
2. Read protocol/loop-contract.md — FSM definition
3. Read kernels/manifest.json — scenarios, reference kernels, build config
4. Read tuning/state.json — which shapes are already done (if file exists)
5. Determine the shape to tune based on user arguments or sweep schedule
6. Initialize FSM (set CROQTUNER_STATE_FILE first if from-scratch):

   # For from-scratch: export CROQTUNER_STATE_FILE=".claude/skills/fsm-engine/state/loop-state_from_scratch.json"
   bash .claude/skills/fsm-engine/scripts/state-transition.sh INIT \
       shape_key=<KEY> dtype=<DTYPE> mode=<MODE> max_iteration=<N> shape='[M,N,K]'
   

7. Proceed to execute the INIT state actions

If state file exists → Resume

1. Read protocol/identity.md — refresh your constraints
2. Read the mode-appropriate state file — determine exact state
3. Read state/compaction-summary.md — what happened before (if exists)
4. Read protocol/step-checklists.md — ONLY the section matching fsm.current_state
5. If in IDEATE or later: read idea dedup log at the path in paths.idea_log
6. Resume executing from fsm.current_state

Executing a Step

For EVERY FSM step, follow this exact sequence:

1. Pre-Check

bash .claude/skills/fsm-engine/scripts/pre-step-check.sh <CURRENT_STATE>

If it exits non-zero → fix the reported issue before proceeding.

2. Execute Step Actions

Follow the actions listed in protocol/step-checklists.md for the current state.

3. Update Guard Flags

After each mandatory action, update the corresponding guard flag via jq:

jq '.guard_flags.<flag_name> = true' .claude/skills/fsm-engine/state/loop-state.json > /tmp/ls.tmp && mv /tmp/ls.tmp .claude/skills/fsm-engine/state/loop-state.json

For metrics updates:

jq '.metrics.<metric_name> = <value>' .claude/skills/fsm-engine/state/loop-state.json > /tmp/ls.tmp && mv /tmp/ls.tmp .claude/skills/fsm-engine/state/loop-state.json

4. Post-Check

bash .claude/skills/fsm-engine/scripts/post-step-check.sh <CURRENT_STATE>

If it exits non-zero → complete the missing actions.

5. Transition

bash .claude/skills/fsm-engine/scripts/state-transition.sh <NEXT_STATE> [key=value ...]

Then immediately proceed to the next step's pre-check. Do NOT pause between steps.

Mode-Specific Configuration

from-best (adapt from reference kernel)

• max_iteration: 30
• ncu: required at iter 1, every 5-10 iters, and when consecutive_discards >= 3
• Seed: reference best .cu from manifest.json
• Build: nvcc with NVCC_FLAGS from manifest
• Read protocol/idea-diversity-rules.md for the from_current_best phase schedule

from-scratch (deep tuning from baseline)

• max_iteration: 150
• ncu: MANDATORY before every IDEATE step (no exceptions)
• Seed: baseline .co from manifest.json
• Build: choreo compiler + nvcc
• Read protocol/idea-diversity-rules.md for the from_scratch phase schedule

Build Commands Reference

from-best (nvcc direct)

CHOREO_REPO=/home/albert/workspace/choreo
nvcc -std=c++20 -O3 -arch=sm_90a --use_fast_math \
  -DSPMM_DEFAULT_M=$M -DSPMM_DEFAULT_N=$N -DSPMM_DEFAULT_K=$K \
  -I"$CHOREO_REPO/runtime" -I"$CHOREO_REPO/extern/cutlass/include" -I"$CHOREO_REPO" \
  -L/usr/local/cuda/lib64 -lcuda \
  -o /tmp/${KEY}_iter<NNN> <kernel.cu>

For e4m3, add: -DSPMM_DTYPE_E4M3

from-scratch (choreo + nvcc)

CHOREO_REPO=/home/albert/workspace/choreo
$CHOREO_REPO/choreo -gs -t cute -arch=sm_90a --use-warpspec --use-prepack \
  <kernel.co> -o /tmp/${KEY}_iter<NNN>.cute.result

Timing

CHOREO_TIMING_WARMUP=10 CHOREO_TIMING_REPEAT=500 CHOREO_SKIP_VERIFY=1 \
  /tmp/${KEY}_iter<NNN> --skip-verify

Or for choreo-compiled:

CHOREO_TIMING_WARMUP=10 CHOREO_TIMING_REPEAT=500 \
  bash /tmp/${KEY}_iter<NNN>.cute.result --execute

ncu Profiling

/usr/local/cuda/bin/ncu --set full --target-processes all \
  -o /tmp/${KEY}_ncu_iter<NNN> \
  /tmp/${KEY}_iter<NNN> --skip-verify

GPU Health Check

nvidia-smi --query-gpu=index,name,utilization.gpu,memory.used,memory.total,temperature.gpu --format=csv,noheader

Healthy: utilization ~0%, memory ~14 MiB, temperature <80°C.

Compaction-Safety

Before context fills, proactively update state/compaction-summary.md:

# Template — fill in actual values
cat > .claude/skills/fsm-engine/state/compaction-summary.md << 'COMPACTION_EOF'
# Compaction Summary
## Last Updated
<timestamp>
## IMMEDIATE ACTION
Read `loop-state.json` and resume from state `<STATE>` at iteration <N>.
## Current Task
- Shape: <KEY> (<mode>, <max_iter> iters max)
- Current best: <TFLOPS> TFLOPS at iter <N>
- Baseline: <TFLOPS> TFLOPS
- Last bottleneck: <category>
- Consecutive discards: <N>
## Recent History (last 5 iterations)
<fill in>
## DO NOT
- Do NOT re-run completed iterations
- Do NOT ask the user what to do — resume autonomously
COMPACTION_EOF

Sweep Schedule — NON-STOP (CRITICAL)

The sweep is NON-STOP. After completing a shape, you MUST immediately start the next one. There are ~260 shapes pending. Completing 1-5 shapes is NOT a reason to stop. The sweep runs until ALL shapes are done or the session physically drops.

After completing one shape, immediately proceed to the next:

1. Read tuning/state.json for completed shapes
2. Pick the next pending shape (prefer "near" region, then "far")
3. Reset loop-state.json via state-transition.sh INIT with new shape params
4. Continue immediately — do NOT pause, do NOT summarize-and-wait, do NOT ask the user

ANTI-STOP RULES:

• Do NOT output "Shape X complete, shall I continue?" — just continue.
• Do NOT output a progress table and wait — write it to compaction-summary.md instead.
• Do NOT treat shape boundaries as natural stopping points — they are NOT.
• The user will interrupt you if they want to stop. Until then, keep going.

When Stuck

If you're genuinely stuck (compile error that won't resolve, ncu timeout, GPU unavailable):

1. Log the failure in idea-log.jsonl with appropriate result code
2. If it's a single iteration failure: mark as DISCARD, continue loop
3. If it's a systemic issue (GPU down): report to user with diagnostics
4. NEVER silently stop the loop. Either continue or explicitly report why you can't.
5. A completed shape is NEVER a reason to stop. Immediately proceed to the next shape.

Protocol Files Reference

File	When to Read	Content
protocol/identity.md	Every invocation	Who you are, inviolable constraints
protocol/loop-contract.md	Fresh start or when confused about FSM	State machine definition
protocol/step-checklists.md	Before each step (only relevant section)	Entry/exit conditions per step
protocol/compaction-protocol.md	After context compaction	How to resume
protocol/idea-diversity-rules.md	Before IDEATE	Dedup rules, category definitions
state/loop-state.json	Every invocation (FIRST thing to read)	Machine-readable FSM state
state/compaction-summary.md	After compaction	Human-readable state summary