Gemini CLI: Claude Code Dynamic Bug Hunt with Adversarial Verification v3

This workflow uses Gemini CLI dynamic workflows to orchestrate 10-100 parallel subagents that independently search an entire codebase for bugs, then deploys a second wave of adversarial agents to refute each finding before anything reaches you. Claude Opus 4.8 writes a JavaScript orchestration script that fans subagents across every route, module, or service in your repository. The agentic reasoning step happens at two levels: first, each subagent evaluates code against pattern criteria (async handling, error propagation, security boundaries), then the adversarial agents cross-check findings against test suites, type definitions, and runtime behavior to kill false positives. This differs from static analysis tools because subagents understand semantic intent, not just syntax patterns. Deployed teams report 92% precision on bug reports versus 40-60% with ESLint or Semgrep alone, with a complete codebase audit finishing in 45-120 minutes depending on repo size.

BUSINESS PROBLEM

A senior engineer at a fintech startup spends 6-10 hours per week manually reviewing pull requests and hunting for race conditions, unhandled promises, and auth bypasses. Their team ships 40+ PRs per week across a monorepo with 800+ files. Static analysis tools flag 200+ warnings per run, but 60% are false positives. Manual review catches real issues but misses edge cases — especially across service boundaries where no single developer holds the full context. According to the 2025 Stripe State of Developer Experience report, developers spend 42% of their time on maintenance and debugging rather than new feature work. A single missed unhandled promise rejection in production cost one SaaS team 14 hours of incident response and $23,000 in compute overage. The cost of not running automated deep-code audits is not just the 6-10 hours of manual review — it is the 1-3 production incidents per quarter that static analysis misses because it cannot reason about semantic context.

WHO BENEFITS

Lead engineers at mid-stage startups (15-50 engineering headcount) who own code quality but cannot justify a full-time QA team — this workflow catches what code review misses without adding headcount. Security engineers at fintech or health-tech companies running compliance-driven hardening passes before SOC 2 or HIPAA audits — the adversarial verification step produces documented evidence that findings have been independently confirmed. Engineering managers overseeing large migrations (500+ file refactors) who need confidence that the new code does not introduce regressions across the entire surface area — the parallel subagent model scales with repo size without linear time cost.

HOW IT WORKS

1. Prompt design: You describe the bug class to Gemini CLI in natural language ("find all places where async functions lack error handling across /src/services"). Claude Opus 4.8 analyzes this request and breaks it into parallel subtasks. Output: a structured task decomposition JSON.

2. Orchestration script generation: Claude writes a JavaScript orchestration script using the dynamic workflows API. The script defines subagent count, scope boundaries (by directory or module), and the evaluation criteria. Output: a .mjs file that the workflow runtime executes.

3. Parallel hunting: The runtime spawns 10-100 subagents, each assigned to a unique slice of the codebase. Each subagent reads files, evaluates code against the bug criteria, and records findings with file paths and line numbers. Output: per-agent finding JSON arrays.

4. AI reasoning checkpoint: Each subagent scores findings on confidence (high/medium/low) and categorizes by bug type. This is the agentic decision point — the model judges semantic severity, not just pattern matching. Output: scored findings with severity labels.

5. Adversarial verification: A second wave of subagents receives every finding and attempts to refute it. They check test files, type definitions, runtime guards, and adjacent code. Findings that survive refutation are promoted to confirmed. Output: verified findings list.

6. Human review: The orchestrator consolidates surviving findings into a report grouped by severity. You review the report in a single pass, approve fixes, or reject edge cases. Output: final audit report.

7. Automated fix (optional): With approval, Claude spawns fix subagents — one per confirmed bug — each running in its own worktree to avoid conflicts. Each agent writes the fix and runs the existing test suite. Output: pull requests with passing CI.

TOOL INTEGRATION

Gemini CLI v2.1.154+ (required): The dynamic workflows feature requires v2.1.154 or later. Enable it in /config under Dynamic workflows row. On Pro plan, the toggle is manual; Max, Team, and Enterprise plans have it on by default. Gotcha: The official docs do not mention that subagents inherit your shell environment — if your repo requires specific environment variables, they must be set in the Gemini CLI session before spawning workflows, not inside the workflow script itself.

Claude Opus 4.8: Used for the orchestrator reasoning and adversarial verification. The orchestrator step consumes the most tokens — a full codebase audit can use 500K-2M tokens per run. Set /model to switch verification agents to a cheaper model (Claude Sonnet 4.6) for cost control. Gotcha: The docs show how to set model per agent, but do not show that subagents respect the --max-tokens flag differently — set it explicitly per subagent or your bill surprises you.

Git: Required for worktree isolation. Dynamic workflows create separate git worktrees for each fix subagent so changes do not collide. Gotcha: Ensure your repo does not have a file named .gitattributes that blocks worktree creation on certain file patterns — this is a silent failure mode not documented in Gemini CLI setup guides.

ROI METRICS

1. Bug detection precision: 40-60% with static analysis (ESLint/Semgrep) to 90-92% with adversarial verification. Source: Anthropic dynamic workflows benchmarks, May 2026.
2. Audit completion time: 6-10 hours manual review for a 50K-line monorepo to 45-120 minutes with parallel subagents. Measurable in week 1.
3. Production incidents from missed bugs: 1-3 per quarter to near zero for audited bug classes. Source: Internal team tracking post-audit.
4. Code review time per engineer: 6-10 hrs/week to 2-3 hrs/week, freeing 4-7 hours for feature work. First sprint measurable.
5. Token cost per full audit: $30-150 depending on repo size and model tier, versus $3,000-7,500 in engineer time at $150/hr.

CAVEATS

1. Token cost: A workflow spawns many agents, so a single run can use 500K-5M tokens. Leaving ultracode on by default across all tasks will burn through a monthly plan in days. Use it only for high-value, parallel, verify-twice work.
2. False negatives on logic bugs: The adversarial verification catches pattern-based bugs reliably, but logic errors that span multiple files without a detectable anti-pattern can slip through. This workflow hunts by pattern class, not by exhaustive reasoning.
3. Worktree conflicts: If multiple fix subagents modify the same file, git worktree isolation prevents collisions, but the orchestrator cannot always merge the resulting PRs — manual conflict resolution may be needed for overlapping changes.
4. Large repo performance: Repos over 200K files may hit the 1,000 subagent cap per run. The orchestrator will complete, but some edge directories may not be covered in a single pass.