CodeRabbit review caught two issues:
1. Empty feedback array → division by zero → NaN avgScore
2. allFailing=true branch logged but never renegotiated (try block
   was inside else-if only)

Fix: add feedback.length > 0 guard, restructure to outer condition
gates renegotiation with inner if/else for accurate log messages.

…ation logger

New mixed-harness/ — cross-model adversarial dev inspired by GAN architecture:
- Generator (Claude Opus 4.6) builds code against sprint contracts
- Evaluator (Codex GPT-5.4) rips apart the work in fresh context
- Zero sycophancy: evaluator has no emotional investment in the code

Includes all 3 hardening fixes from parent harnesses:
- Iterative contract negotiation (3 rounds)
- Fail-closed contract parsing (throws on garbage)
- Mid-sprint renegotiation when all criteria fail

ConversationLogger (shared/conversation-logger.ts):
- Captures every agent prompt, response, tool call, score, and error
- Saves as Obsidian-friendly markdown (.md) + machine-readable JSONL
- Default output: agent-brain-vault/Projects/brane-code/debates/
- Collapsible tool calls, score badges, duration tracking

Tests (29 passing):
- parseContract: 9 tests (fail-closed, code blocks, garbage rejection)
- Renegotiation trigger: 7 tests (thresholds, division-by-zero guard)
- parseEvalResult: 3 tests (threshold recalculation, extraction)
- Negotiation rounds: 3 tests (early approval, max rounds)
- ConversationLogger: 7 tests (entries, markdown, JSONL, disk save)

New gemini-harness/ — cross-company adversarial dev:
- Generator: Claude Opus 4.6 (Anthropic) builds code via Agent SDK
- Evaluator: Gemini 3.1 Pro Preview (Google) rips it apart via @google/genai
- Gemini evaluator has tool calling: readFile, runCommand, listFiles
- Multi-turn chat loop handles tool calls until Gemini is done evaluating
- 1M context on BOTH sides — true heavyweight matchup

Same 3 hardening fixes as other harnesses:
- Iterative contract negotiation (3 rounds)
- Fail-closed contract parsing
- Mid-sprint renegotiation on bad criteria

Includes ConversationLogger integration for full transcript logging.

SDK: @google/genai@1.48.0
Model: gemini-3.1-pro-preview (1M input, 65K output)

- Gemini evaluator: sandbox tool handlers with path confinement,
  command allowlisting (execFileSync instead of execSync), and
  fs.readdir instead of shell-interpolated find
- Fix fragile "APPROVED" exact-match in all 4 harness negotiation
  loops — now case-insensitive startsWith
- Remove personal vault path from logDir defaults (use ./logs)
- Remove brane-streaming-fix.md (task doc, not project code)
- Remove unused imports in mixed/gemini harnesses
- Fix copy-paste comment ("Codex" -> "Gemini" in gemini harness)

- Gemini sandbox: use realpath() instead of resolve() to prevent
  symlink-based path traversal escape
- Gemini sandbox: reject runCommand args with absolute paths outside
  workspace (prevents `cat /etc/passwd`, `grep -r secret /etc`)
- All evaluators: guard against empty feedback array where
  [].every() returns true — prevents silent false pass
- ConversationLogger: empty JSONL returns "" not bare newline

Security (Gemini evaluator sandbox):
- Remove node/npm/npx/bun/bunx from runCommand allowlist — prevents
  arbitrary code execution via `node -e "..."`
- Restrict git to read-only subcommands (log, status, diff, show,
  ls-files, rev-parse) — prevents data exfiltration via git push
- Block find -exec/-execdir/-delete flags — prevents subprocess spawn
- Keep path containment (realpath + absolute path rejection) from
  previous commit

Polish:
- Fix wrong evaluator name in gemini generator prompt ("Codex" -> generic)
- Remove unused readContract import from claude-harness
- Fix README: wrong default model (sonnet -> opus), add mixed/gemini
  harness sections to Quick Start and Project Structure

RESULTS.md — full scoreboard, key findings, and analysis from 4 harness
runs (Claude 5/5, Codex 0/1, Mixed 11/13 on S4, Gemini 5/5).

examples/gemini-run-excerpt.md — first 150 lines of actual Gemini
evaluator conversation log showing the ConversationLogger output format.