Severity: High
Category: Shell Compatibility / Error Handling
Platform: See per-site analysis
Confidence: Confirmed
Description
Four shell scripts use the pattern <cmd> | head -1 | ... under set -euo pipefail. When the upstream command produces multiple lines, head -1 terminates after reading line 1 and sends SIGPIPE (exit 141) to the upstream process. With pipefail, the pipeline's exit becomes 141, propagating up and killing the script — even though the intent was "take the first line and continue". This is a known project pattern (memory note: sed -n '1p' is the SIGPIPE-safe replacement). All four sites should be fixed together.
Affected File(s)
dream-server/scripts/pre-download.sh:112 — nvidia-smi --query-gpu=memory.total ... | head -1 | awk ...dream-server/scripts/dream-preflight.sh:87 — docker exec ... nvidia-smi --query-gpu=memory.free ... | head -1 | tr -d ' 'dream-server/scripts/check-offline-models.sh:27 — ls -1 data/models/*.gguf | head -1dream-server/installers/macos/dream-macos.sh:158 — grep -E "^${key}=" "$env_file" | head -n 1 | cut -d'=' -f2- | tr -d '\r' || true
Note: dream-server/scripts/dream-test-functional.sh:66,84,230 — already addressed by open PR Light-Heart-Labs#1003 via if ! VAR=$(...) wrapper; NOT in scope of this issue.
Note: dream-server/dream-cli — line 254 already addressed by open PR Light-Heart-Labs#1008. Remaining dream-cli sites use the local VAR=$(...) form which masks the pipeline exit status and is already safe.
Root Cause
set -euo pipefail (enforced at the top of each of the 4 scripts) means the first non-zero pipeline exit kills the script. head -1 reading 1 line and closing its stdin causes SIGPIPE on the upstream producer (nvidia-smi, ls, grep) → exit 141. Under pipefail, the whole pipeline exits 141. The surrounding code was written when set -e alone was active (no pipefail), so it assumed a multi-line input was safe.
Evidence
# pre-download.sh:112
nvidia-smi --query-gpu=memory.total --format=csv,noheader,nounits 2>/dev/null | head -1 | awk '{print int($1/1024)}'
# dream-preflight.sh:87
GPU_MEM=$(docker exec "$LLM_CONTAINER" nvidia-smi --query-gpu=memory.free --format=csv,noheader,nounits 2>/dev/null | head -1 | tr -d ' ')
# check-offline-models.sh:27
MODEL_FILE=$(ls -1 data/models/*.gguf | head -1)
# dream-macos.sh:158
grep -E "^${key}=" "$env_file" 2>/dev/null | head -n 1 | cut -d'=' -f2- | tr -d '\r' || truePlatform Analysis
- pre-download.sh / dream-preflight.sh: Linux NVIDIA multi-GPU only (nvidia-smi emits N lines on N-GPU hosts). Single-GPU unaffected. macOS / Windows-WSL2 do not run these paths.
- check-offline-models.sh: Linux and macOS (any user with 2+
.gguf files). Windows-WSL2 not affected (script not invoked there).
- dream-macos.sh: macOS only. SIGPIPE fires if
.env contains a duplicate key (e.g. from a buggy upgrade). Trailing || true catches script abort but VAR still holds truncated/empty output, silently returning wrong value for GPU_BACKEND, LLAMA_CPU_LIMIT, LLAMA_CPU_RESERVATION.
Reproduction
- pre-download.sh: Any dual-GPU NVIDIA Linux host. Run
scripts/pre-download.sh — exits with SIGPIPE-related error before model selection.
- dream-preflight.sh: Same hardware. Run
scripts/dream-preflight.sh — aborts without completing service checks.
- check-offline-models.sh: Place two
.gguf files under data/models/; run scripts/check-offline-models.sh — pipeline abort.
- dream-macos.sh: Manually duplicate a line in
.env (e.g. two GPU_BACKEND= lines); call a dream subcommand that uses read_env_value GPU_BACKEND — returned value silently wrong.
Impact
- Multi-GPU NVIDIA Linux users cannot complete preflight or model pre-download.
- Users with multiple downloaded models see offline check crash instead of a clean report.
- macOS users with an accidentally duplicated
.env line see wrong values used silently (no error, no warning).
Suggested Approach
Replace | head -1 | (and | head -n 1 |) with | sed -n '1p' |. sed reads all input and terminates after printing line 1 without generating SIGPIPE to the upstream process. Works identically on BSD (macOS) and GNU (Linux) sed. For check-offline-models.sh, find data/models/ -maxdepth 1 -name "*.gguf" | sed -n '1p' also normalizes ordering (locale-free).
Labels
bug, shell-compatibility, error-handling, pipefail, multi-platform
Severity: High
Category: Shell Compatibility / Error Handling
Platform: See per-site analysis
Confidence: Confirmed
Description
Four shell scripts use the pattern
<cmd> | head -1 | ...underset -euo pipefail. When the upstream command produces multiple lines,head -1terminates after reading line 1 and sends SIGPIPE (exit 141) to the upstream process. Withpipefail, the pipeline's exit becomes 141, propagating up and killing the script — even though the intent was "take the first line and continue". This is a known project pattern (memory note:sed -n '1p'is the SIGPIPE-safe replacement). All four sites should be fixed together.Affected File(s)
dream-server/scripts/pre-download.sh:112—nvidia-smi --query-gpu=memory.total ... | head -1 | awk ...dream-server/scripts/dream-preflight.sh:87—docker exec ... nvidia-smi --query-gpu=memory.free ... | head -1 | tr -d ' 'dream-server/scripts/check-offline-models.sh:27—ls -1 data/models/*.gguf | head -1dream-server/installers/macos/dream-macos.sh:158—grep -E "^${key}=" "$env_file" | head -n 1 | cut -d'=' -f2- | tr -d '\r' || trueNote:
dream-server/scripts/dream-test-functional.sh:66,84,230— already addressed by open PR Light-Heart-Labs#1003 viaif ! VAR=$(...)wrapper; NOT in scope of this issue.Note:
dream-server/dream-cli— line 254 already addressed by open PR Light-Heart-Labs#1008. Remaining dream-cli sites use thelocal VAR=$(...)form which masks the pipeline exit status and is already safe.Root Cause
set -euo pipefail(enforced at the top of each of the 4 scripts) means the first non-zero pipeline exit kills the script.head -1reading 1 line and closing its stdin causes SIGPIPE on the upstream producer (nvidia-smi, ls, grep) → exit 141. Under pipefail, the whole pipeline exits 141. The surrounding code was written whenset -ealone was active (no pipefail), so it assumed a multi-line input was safe.Evidence
Platform Analysis
.gguffiles). Windows-WSL2 not affected (script not invoked there)..envcontains a duplicate key (e.g. from a buggy upgrade). Trailing|| truecatches script abort but VAR still holds truncated/empty output, silently returning wrong value forGPU_BACKEND,LLAMA_CPU_LIMIT,LLAMA_CPU_RESERVATION.Reproduction
scripts/pre-download.sh— exits with SIGPIPE-related error before model selection.scripts/dream-preflight.sh— aborts without completing service checks..gguffiles underdata/models/; runscripts/check-offline-models.sh— pipeline abort..env(e.g. twoGPU_BACKEND=lines); call adreamsubcommand that usesread_env_value GPU_BACKEND— returned value silently wrong.Impact
.envline see wrong values used silently (no error, no warning).Suggested Approach
Replace
| head -1 |(and| head -n 1 |) with| sed -n '1p' |.sedreads all input and terminates after printing line 1 without generating SIGPIPE to the upstream process. Works identically on BSD (macOS) and GNU (Linux)sed. Forcheck-offline-models.sh,find data/models/ -maxdepth 1 -name "*.gguf" | sed -n '1p'also normalizes ordering (locale-free).Labels
bug,shell-compatibility,error-handling,pipefail,multi-platform