z3/.github/workflows/ostrich-benchmark.md

description	on	permissions	network	tools	safe-outputs	timeout-minutes	steps
Run Z3 string solver benchmarks (seq vs nseq) and ZIPT on all Ostrich benchmarks from tests/ostrich.zip on the c3 branch and post results as a GitHub discussion	schedule workflow_dispatch cron 0 6 * * *	read-all	defaults	bash github true toolsets default	create-discussion missing-tool noop title-prefix category close-older-discussions [Ostrich Benchmark] Agentic Workflows true create-issue true report-as-issue false	180	name uses with Checkout c3 branch actions/checkout@v5 ref fetch-depth persist-credentials c3 1 false

Ostrich Benchmark: Z3 c3 branch vs ZIPT

You are an AI agent that benchmarks Z3 string solvers (seq and nseq) and the standalone ZIPT solver on all SMT-LIB2 benchmarks from the tests/ostrich.zip archive on the c3 branch, and publishes a summary report as a GitHub discussion.

Context

• Repository: ${{ github.repository }}
• Workspace: ${{ github.workspace }}
• Branch: c3 (already checked out by the workflow setup step)

Phase 1: Build Z3

Build Z3 from the checked-out c3 branch using CMake + Ninja, including the .NET bindings required by ZIPT.

cd ${{ github.workspace }}

# Install build dependencies if missing
sudo apt-get install -y ninja-build cmake python3 zstd dotnet-sdk-8.0 unzip 2>/dev/null || true

# Configure the build in Release mode for better performance and lower memory usage
# (Release mode is sufficient for benchmarking; the workflow does not use -tr: trace flags)
mkdir -p build
cd build
cmake .. -G Ninja -DCMAKE_BUILD_TYPE=Release -DZ3_BUILD_DOTNET_BINDINGS=ON 2>&1 | tail -20

# Build z3 binary and .NET bindings SYNCHRONOUSLY (do NOT add & to background these commands).
# Running ninja in the background while the LLM agent is also active causes OOM and kills the
# agent process. Wait for each build command to finish before continuing.
# -j1 limits parallelism to reduce peak memory usage alongside the LLM agent process.
ninja z3 2>&1 | tail -30
ninja build_z3_dotnet_bindings 2>&1 | tail -20

# Verify the build succeeded
./z3 --version

# Locate the Microsoft.Z3.dll produced by the build
Z3_DOTNET_DLL=$(find . -name "Microsoft.Z3.dll" -not -path "*/obj/*" | head -1)
if [ -z "$Z3_DOTNET_DLL" ]; then
    echo "ERROR: Microsoft.Z3.dll not found after build"
    exit 1
fi
echo "Found Microsoft.Z3.dll at: $Z3_DOTNET_DLL"

If the build fails, report the error clearly and exit without proceeding.

Phase 2a: Clone and Build ZIPT

Clone the ZIPT solver from the parikh branch and compile it against the Z3 .NET bindings built in Phase 1.

cd ${{ github.workspace }}

# Re-locate the Microsoft.Z3.dll if needed
Z3_DOTNET_DLL=$(find build -name "Microsoft.Z3.dll" -not -path "*/obj/*" | head -1)
Z3_LIB_DIR=${{ github.workspace }}/build

# Clone ZIPT (parikh branch)
git clone --depth=1 --branch parikh https://github.com/CEisenhofer/ZIPT.git /tmp/zipt

# Patch ZIPT.csproj to point at the freshly built Microsoft.Z3.dll
# (the repo has a Windows-relative hardcoded path that won't exist here)
sed -i "s|<HintPath>.*</HintPath>|<HintPath>$Z3_DOTNET_DLL</HintPath>|" /tmp/zipt/ZIPT/ZIPT.csproj

# Build ZIPT in Release mode
cd /tmp/zipt/ZIPT
dotnet build --configuration Release 2>&1 | tail -20

# Locate the built ZIPT.dll
ZIPT_DLL=$(find /tmp/zipt/ZIPT/bin/Release -name "ZIPT.dll" | head -1)
if [ -z "$ZIPT_DLL" ]; then
    echo "ERROR: ZIPT.dll not found after build"
    exit 1
fi
echo "ZIPT binary: $ZIPT_DLL"

# Make libz3.so visible to the .NET runtime at ZIPT startup
ZIPT_OUT_DIR=$(dirname "$ZIPT_DLL")
if cp "$Z3_LIB_DIR/libz3.so" "$ZIPT_OUT_DIR/" 2>/dev/null; then
    echo "Copied libz3.so to $ZIPT_OUT_DIR"
else
    echo "WARNING: could not copy libz3.so to $ZIPT_OUT_DIR — setting LD_LIBRARY_PATH fallback"
fi
export LD_LIBRARY_PATH="$Z3_LIB_DIR${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"
echo "ZIPT build complete."

If the ZIPT build fails, note the error in the report but continue with the Z3-only benchmark columns.

Phase 2b: Extract Benchmark Files

Extract all SMT-LIB2 files from the tests/ostrich.zip archive.

cd ${{ github.workspace }}

# Extract the zip archive
mkdir -p /tmp/ostrich_benchmarks
unzip -q tests/ostrich.zip -d /tmp/ostrich_benchmarks

# List all .smt2 files
find /tmp/ostrich_benchmarks -name "*.smt2" -type f | sort > /tmp/all_ostrich_files.txt
TOTAL_FILES=$(wc -l < /tmp/all_ostrich_files.txt)
echo "Total Ostrich .smt2 files: $TOTAL_FILES"

if [ "$TOTAL_FILES" -eq 0 ]; then
    echo "ERROR: No .smt2 files found in tests/ostrich.zip"
    exit 1
fi

Phase 3: Run Benchmarks

Run every file from /tmp/all_ostrich_files.txt with both Z3 string solvers and ZIPT. Use a 5-second timeout per run.

For each file, run:

1. z3 smt.string_solver=seq -T:5 <file> — seq solver
2. z3 smt.string_solver=nseq -T:5 <file> — nseq (ZIPT) solver
3. dotnet <ZIPT.dll> -t:5000 <file> — standalone ZIPT solver (milliseconds)

Capture:

• Verdict: sat, unsat, unknown, timeout (if exit code indicates timeout or process is killed), or bug (if a solver crashes / produces a non-standard result)
• Time (seconds): wall-clock time for the run
• A row is flagged SOUNDNESS_DISAGREEMENT when any two solvers that both produced a definitive answer (sat/unsat) disagree

Use a bash script to automate this:

#!/usr/bin/env bash
set -euo pipefail

Z3=${{ github.workspace }}/build/z3
ZIPT_DLL=$(find /tmp/zipt/ZIPT/bin/Release -name "ZIPT.dll" 2>/dev/null | head -1)
ZIPT_AVAILABLE=false
[ -n "$ZIPT_DLL" ] && ZIPT_AVAILABLE=true

# Ensure libz3.so is on the dynamic-linker path for the .NET runtime
export LD_LIBRARY_PATH=${{ github.workspace }}/build${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}

RESULTS=/tmp/benchmark_results.tsv
mkdir -p /tmp/ostrich_run

echo -e "file\tseq_verdict\tseq_time\tnseq_verdict\tnseq_time\tzipt_verdict\tzipt_time\tnotes" > "$RESULTS"

run_z3_seq() {
    local file="$1"
    local start end elapsed verdict output exit_code

    start=$(date +%s%3N)
    output=$(timeout 7 "$Z3" "smt.string_solver=seq" -T:5 "$file" 2>&1)
    exit_code=$?
    end=$(date +%s%3N)
    elapsed=$(echo "scale=3; ($end - $start) / 1000" | bc)

    if echo "$output" | grep -q "^unsat"; then
        verdict="unsat"
    elif echo "$output" | grep -q "^sat"; then
        verdict="sat"
    elif echo "$output" | grep -q "^unknown"; then
        verdict="unknown"
    elif [ "$exit_code" -eq 124 ]; then
        verdict="timeout"
    elif echo "$output" | grep -qi "error\|assertion\|segfault\|SIGABRT\|exception"; then
        verdict="bug"
    else
        verdict="unknown"
    fi

    echo "$verdict $elapsed"
}

run_z3_nseq() {
    local file="$1"
    local start end elapsed verdict output exit_code

    start=$(date +%s%3N)
    output=$(timeout 7 "$Z3" "smt.string_solver=nseq" -T:5 "$file" 2>&1)
    exit_code=$?
    end=$(date +%s%3N)
    elapsed=$(echo "scale=3; ($end - $start) / 1000" | bc)

    if echo "$output" | grep -q "^unsat"; then
        verdict="unsat"
    elif echo "$output" | grep -q "^sat"; then
        verdict="sat"
    elif echo "$output" | grep -q "^unknown"; then
        verdict="unknown"
    elif [ "$exit_code" -eq 124 ]; then
        verdict="timeout"
    elif echo "$output" | grep -qi "error\|assertion\|segfault\|SIGABRT\|exception"; then
        verdict="bug"
    else
        verdict="unknown"
    fi

    echo "$verdict $elapsed"
}

run_zipt() {
    local file="$1"
    local start end elapsed verdict output exit_code

    if [ "$ZIPT_AVAILABLE" != "true" ]; then
        echo "n/a 0.000"
        return
    fi

    start=$(date +%s%3N)
    # ZIPT prints the filename on the first line, then SAT/UNSAT/UNKNOWN on subsequent lines
    output=$(timeout 7 dotnet "$ZIPT_DLL" -t:5000 "$file" 2>&1)
    exit_code=$?
    end=$(date +%s%3N)
    elapsed=$(echo "scale=3; ($end - $start) / 1000" | bc)

    if echo "$output" | grep -qi "^UNSAT$"; then
        verdict="unsat"
    elif echo "$output" | grep -qi "^SAT$"; then
        verdict="sat"
    elif echo "$output" | grep -qi "^UNKNOWN$"; then
        verdict="unknown"
    elif [ "$exit_code" -eq 124 ]; then
        verdict="timeout"
    elif echo "$output" | grep -qi "error\|crash\|exception\|Unsupported"; then
        verdict="bug"
    else
        verdict="unknown"
    fi

    echo "$verdict $elapsed"
}

COUNTER=0
while IFS= read -r file; do
    COUNTER=$((COUNTER + 1))
    fname=$(basename "$file")

    seq_result=$(run_z3_seq "$file")
    nseq_result=$(run_z3_nseq "$file")
    zipt_result=$(run_zipt "$file")

    seq_verdict=$(echo "$seq_result" | cut -d' ' -f1)
    seq_time=$(echo "$seq_result" | cut -d' ' -f2)
    nseq_verdict=$(echo "$nseq_result" | cut -d' ' -f1)
    nseq_time=$(echo "$nseq_result" | cut -d' ' -f2)
    zipt_verdict=$(echo "$zipt_result" | cut -d' ' -f1)
    zipt_time=$(echo "$zipt_result" | cut -d' ' -f2)

    # Flag soundness disagreement when any two definitive verdicts disagree
    notes=""
    declare -A definitive_map
    [ "$seq_verdict"  = "sat" ] || [ "$seq_verdict"  = "unsat" ] && definitive_map[seq]="$seq_verdict"
    [ "$nseq_verdict" = "sat" ] || [ "$nseq_verdict" = "unsat" ] && definitive_map[nseq]="$nseq_verdict"
    [ "$zipt_verdict" = "sat" ] || [ "$zipt_verdict" = "unsat" ] && definitive_map[zipt]="$zipt_verdict"
    has_sat=false; has_unsat=false
    for v in "${definitive_map[@]}"; do
        [ "$v" = "sat"   ] && has_sat=true
        [ "$v" = "unsat" ] && has_unsat=true
    done
    if $has_sat && $has_unsat; then
        notes="SOUNDNESS_DISAGREEMENT"
    fi

    echo -e "$fname\t$seq_verdict\t$seq_time\t$nseq_verdict\t$nseq_time\t$zipt_verdict\t$zipt_time\t$notes" >> "$RESULTS"
    echo "[$COUNTER] [$fname] seq=$seq_verdict(${seq_time}s) nseq=$nseq_verdict(${nseq_time}s) zipt=$zipt_verdict(${zipt_time}s) $notes"
done < /tmp/all_ostrich_files.txt

echo "Benchmark run complete. Results saved to $RESULTS"

Save this script to /tmp/run_ostrich_benchmarks.sh, make it executable, and run it. Do not skip any file.

Phase 4: Generate Summary Report

Read /tmp/benchmark_results.tsv and compute statistics. Then generate a Markdown report.

Compute:

• Total benchmarks: total number of files run
• Per solver (seq, nseq, and ZIPT): count of sat / unsat / unknown / timeout / bug verdicts
• Total time used: sum of all times for each solver
• Average time per benchmark: total_time / total_files
• Soundness disagreements: files where any two solvers that both returned a definitive answer disagree
• Bugs / crashes: files with error/crash verdicts

Format the report as a GitHub Discussion post (GitHub-flavored Markdown):

### Ostrich Benchmark Report — Z3 c3 branch

**Date**: <today's date>
**Branch**: c3
**Benchmark set**: Ostrich (all files from tests/ostrich.zip)
**Timeout**: 5 seconds per benchmark (`-T:5` for Z3; `-t:5000` for ZIPT)

---

### Summary

| Metric | seq solver | nseq solver | ZIPT solver |
|--------|-----------|-------------|-------------|
| sat | X | X | X |
| unsat | X | X | X |
| unknown | X | X | X |
| timeout | X | X | X |
| bug/crash | X | X | X |
| **Total time (s)** | X.XXX | X.XXX | X.XXX |
| **Avg time/benchmark (s)** | X.XXX | X.XXX | X.XXX |

**Soundness disagreements** (any two solvers return conflicting sat/unsat): N

---

### Per-File Results

<details>
<summary>Click to expand full per-file table</summary>

| # | File | seq verdict | seq time (s) | nseq verdict | nseq time (s) | ZIPT verdict | ZIPT time (s) | Notes |
|---|------|-------------|-------------|--------------|--------------|--------------|--------------|-------|
| 1 | benchmark_0001.smt2 | sat | 0.123 | sat | 0.456 | sat | 0.789 | |
| ... | ... | ... | ... | ... | ... | ... | ... | ... |

</details>

---

### Notable Issues

#### Soundness Disagreements (Critical)
<list files where any two solvers disagree on sat/unsat, naming which solvers disagree>

#### Crashes / Bugs
<list files where any solver crashed or produced an error>

#### Slow Benchmarks (> 4s)
<list files that took more than 4 seconds for any solver>

---

*Generated automatically by the Ostrich Benchmark workflow on the c3 branch.*

Phase 5: Post to GitHub Discussion

Post the Markdown report as a new GitHub Discussion using the create-discussion safe output.

• Category: "Agentic Workflows"
• Title: [Ostrich Benchmark] Z3 c3 branch — <date>
• Close older discussions with the same title prefix to avoid clutter.

Guidelines

• Always build from c3 branch: The workspace is already checked out on c3; don't change branches.
• Synchronous builds only: Never run ninja (or any other build command) in the background using &. Running the build concurrently with LLM inference causes the agent process to be killed by the OOM killer (exit 137) because C++ compilation and the LLM together exceed available RAM. Always wait for each build command to finish before proceeding.
• Release build: The build uses CMAKE_BUILD_TYPE=Release for lower memory footprint and faster compilation on the GitHub Actions runner. The benchmark only needs verdict and timing output; no -tr: trace flags are used.
• Run all benchmarks: Unlike the QF_S workflow, run every file in the archive — do not randomly sample.
• 5-second timeout: Pass -T:5 to Z3 (both seq and nseq) and -t:5000 to ZIPT (milliseconds). Use timeout 7 as the outer OS-level guard to allow the solver to exit cleanly before being killed.
• Be precise with timing: Use millisecond-precision timestamps and report times in seconds with 3 decimal places.
• Distinguish timeout from unknown: A timeout is different from (unknown) returned by a solver within its time budget.
• ZIPT output format: ZIPT prints the input filename on the first line, then SAT, UNSAT, or UNKNOWN on subsequent lines. Parse accordingly.
• Report soundness bugs prominently: If any benchmark shows a conflict between any two solvers that both returned a definitive sat/unsat answer, highlight it as a critical finding and name which pair disagrees.
• Handle build failures gracefully: If Z3 fails to build, report the error and create a brief discussion noting the build failure. If ZIPT fails to build, continue with only the seq/nseq columns and note n/a for ZIPT results.
• Large report: Always put the per-file table in a <details> collapsible section since there may be many files.
• Progress logging: Print a line per file as you run it (e.g., [N] [filename] seq=...) so the workflow log shows progress even for large benchmark sets.