Gemini CLI headless failure persists after v2.0.0 update + sync still missing schemas/benchmarks

[EdgeCaser]

Follow-up to #9

Two issues from #9 remain after pulling the latest Shipwright update (post-v2.0.0, commit e6c5bd4):

1. Gemini CLI headless failure — NOT FIXED

Retested with the updated harness. Tried a new invocation pattern:

gemini --output-format text --approval-mode plan -p "$(cat case-packet.md)"

Result: Identical failure. Process spawns (201MB memory, actively running), produces 0 bytes on stdout, no stderr output, hangs indefinitely until killed. This was with --output-format text (which should force text-only output) AND --approval-mode plan (read-only, no tool use).

All patterns now tested and failed:

Attempt	Flags	Result
1 (prior session)	cat | gemini -p "..."	0 bytes
2 (prior session)	gemini -p "$(cat ...)"	0 bytes
3 (prior session)	gemini -p "..." -y	141 bytes preamble, then hung
4 (this session)	gemini --output-format text --approval-mode plan -p "$(cat ...)"	0 bytes

Claude and Codex continue to work perfectly with the same ~4KB prompt. This blocks Gemini from participating as a conflict harness side or as a triple-panel tiebreaker judge.

2. Sync script still missing schemas/ and benchmarks/ mappings

The updated shipwright-sync.sh (from the latest pull) added output-styles and ROOT_FILES mappings but still does not include:

"schemas:schemas"
"benchmarks/scenarios:benchmarks/scenarios"

The conflict harness scripts (build-case-packet.mjs) import from schemas/conflict-case.schema.json at runtime. Without the schemas directory synced, any destination repo that runs the harness will fail with missing file errors.

Current MAPPINGS array (from latest sync):

MAPPINGS=(
  "skills:.claude/skills"
  "skills:.codex/skills"
  "agents:.claude/agents"
  "commands:.claude/commands"
  "output-styles:.claude/output-styles"
  "docs:.claude/docs"
  "evals:.claude/evals"
  "scripts:.claude/scripts"
  "scripts:.codex/scripts"
)

Workaround: Manual cp -r of schemas/ and benchmarks/scenarios/ after each sync.

Impact

The Gemini failure prevented a triple-panel tiebreaker on a real strategic decision (gaming use case prioritization for SaventData). The analysis completed with 2/3 panels (Claude + Codex) but the orchestrator's judge escalation policy calls for Gemini as the third panel when two judges disagree — which they did (UC04 vs UC10 for the third slot). Had to resolve with orchestrator judgment instead.

[EdgeCaser]

Thanks for the follow-up. Two things to separate:

Sync fix: Both the schemas and benchmarks mappings ARE in the fix — they're in commit c3cd1f6, which landed after the e6c5bd4 you pulled. A git pull will bring them in. Verified MAPPINGS now include schemas:schemas, benchmarks/scenarios:benchmarks/scenarios, benchmarks/fixtures:benchmarks/fixtures.

Gemini headless on macOS: Our fix in #9 addressed the alias failure path (missing shipwright-gemini-* config → silent failure). Your attempt 4 shows the issue runs deeper — even with --output-format text and a real model, macOS Gemini CLI produces 0 bytes when invoked as a headless subprocess. That's confirmed as a separate macOS-specific Gemini CLI bug.

Four attempts tested, all fail on macOS. Claude and Codex work fine on the same prompt. This is not fixable in the harness without replacing the CLI invocation entirely.

Workaround for now: Use Claude + Codex as both sides and judge (with different role prompts), or use Gemini only from a Linux/Windows environment. The 2-provider path (Claude + Codex) still gets you a double panel — you just lose the third-family judge.

Real fix path: Gemini API adapter instead of CLI subprocess. That would bypass the macOS headless issue entirely. Filing as a separate issue — the CLI approach is not reliable on macOS regardless of flags.