Deep Dive Skill

What it does Deep Dive is a portable research skill, not a one-off app: point it at a question and it runs a five-phase pipeline — Scope → Search → Fetch → Verify (Adversarial) → Synthesize — and returns a claim-by-claim report where every accepted claim has already survived independent cross-examination by 2-3 separate verifier calls. Claims that only get a tie vote, incomplete support, or no cited source are dropped instead of being presented as fact. It runs identically across 15 different AI coding/agent platforms (Hermes, Claude Code, generic Python/DeepSeek, Aider, Codex CLI, Cline, Cursor, Gemini, GitHub Copilot, Amazon Q, Windsurf, Kimi, GLM, MiniMax, and Google's new Antigravity CLI) by auto-detecting whichever platform it's running under and adapting to that platform's own tools.

How Codex and GPT-5.6 were used The base pipeline existed before Build Week; for this event, OpenAI Codex (GPT-5.6-Terra) was run directly against the codebase to do two rounds of genuine engineering work, file-by-file, each file read and reasoned about independently rather than blindly patched:

1. A hardening pass across 10 core files (harness.py, the full core/ pipeline, and the main runtime adapters). Codex found and fixed real bugs on its own: an unbounded thread-spawning race condition in the concurrency layer, a fetch extractor that was silently never receiving the actual fetched page text (plus missing guards against fetching local/private-IP targets), greedy JSON parsing across every LLM-facing module that accepted truncated/malformed model output, cross-call configuration leakage between pipeline runs, and unlocked concurrent state mutation in the live status monitor. None of this was prompted with the answer — Codex diagnosed each issue itself after reading the code.
2. Extending the platform surface. Codex added 9 new platform adapters (including a brand-new adapter for Google's just-launched Antigravity CLI, the successor to Gemini CLI) and then gave every CLI-based adapter genuine parallel execution: when a platform's local CLI is present, worker threads now run real concurrent invocations of that CLI itself through a shared bounded executor, instead of silently falling back to a generic API path.

Every change is logged in CODEX_CHANGES.md and the per-module notes in CODEX_CHANGES/ — nothing here is asserted without a file-by-file paper trail.

Challenges we ran into Making the same pipeline behave correctly across 15 unrelated CLI tools without a shared SDK meant every adapter had to independently detect its own CLI, shell out safely, and degrade gracefully — Codex's fix for the thread-safety bugs in the shared concurrency layer was what made adding 9 more adapters on top of it safe to do quickly.

What's next Native (non-interactive) multi-agent delegation for platforms that expose it only through interactive TUI commands (e.g. Antigravity's /agent), once/if those platforms ship a scriptable equivalent.

Built With

• bash
• cli
• concurrent-futures
• deepseek-api
• duckduckgo
• git
• google-custom-search
• gpt-5.6
• multi-agent
• openai-codex
• python
• requests
• searxng
• threading