Posts claim Codex Desktop system prompt leaked with GPT-5.6 Sol tool list

• Posts claimed a full Codex Desktop prompt and tool list for GPT-5.6 Sol leaked: the first post from elder_plinius said the system prompt alone was over 42,000 words, and the follow-up pointed readers to full files.
• The dump remains unverified: OpenAI’s public Sol docs describe the API model surface, while the leak post is social evidence, not an OpenAI confirmation.
• The useful engineering detail is the harness surface: mattrickard recognized desktop-injected thread tools, and mattlam_ showed inline review comments rendered through response metadata.
• Sol’s agent behavior is already being debugged in public: 0xblacklight moved back toward low thinking, while pvncher said subagents depend on user requests, repo instructions, or skills.
• The Codex versus Claude Code boundary got blurrier: AI Builder Club claimed OpenAI was encouraging GPT-5.6 Sol inside Claude Code via CLIProxyAPI.

OpenAI’s model page gives Sol a 1,050,000-token context window, 128,000 max output tokens, and tools ranging from hosted shell to MCP. The prompting guide tells developers to adapt prompts, tool descriptions, agent instructions, and prompt stacks for GPT-5.6 Sol; the Codex app-server docs explain the rich-client thread interface that one commenter spotted in the alleged desktop tool list. OpenAI’s system card says GPT-5.6 shows a greater tendency than GPT-5.5 to go beyond user intent in agentic coding tasks, which makes the alleged desktop harness dump more interesting than prompt gawking.

The claim was blunt: a full “System Prompt + Tools” dump for GPT-5.6 Sol in Codex Desktop, with only a fraction fitting in the post because the system prompt allegedly exceeded 42,000 words. The same thread used a follow-up post to point to the full files.

A second post from synthwavedd boosted another link to the material, with the follow-up crediting whylifeis4.

The leak should be treated as alleged until OpenAI confirms it. The public evidence still matters because the excerpt is not just personality text, it describes a complete desktop coding-agent harness: communication channels, workspace sharing, sandbox permissions, escalation policy, file editing rules, inline review metadata, automations, and thread coordination.

OpenAI’s GPT-5.6 Sol model page confirms the model slug, gpt-5.6-sol, and says the gpt-5.6 alias routes to Sol. The same page lists these Responses API tools:

• Web search
• File search
• Image generation
• Code interpreter
• Hosted shell
• Apply patch
• Skills
• Computer use
• MCP
• Tool search

The alleged prompt posted by elder_plinius goes below that API catalog into desktop-client policy. OpenAI’s public docs establish the model and tool families, but not the leaked desktop instructions.

OpenAI’s GPT-5.6 launch post also frames Programmatic Tool Calling as a way for the model to write lightweight programs that coordinate tools, process intermediate results, monitor progress, and choose the next action with fewer round trips.

mattrickard recognized a separate Codex Desktop surface: helper tools for thread management such as create_thread, fork_thread, send_message_to_thread, and set_thread_title. OpenAI’s Codex app-server docs describe the app server as the interface that powers rich clients with authentication, conversation history, approvals, and streamed agent events.

The open-source app-server README describes primitives for thread/start, thread/resume, thread/fork, and turn notifications. In mattrickard’s workflow, a manager thread spawns new threads, gets heartbeat reports, sends messages, and cancels active work through the shared app server.

The same desktop-specific layer showed up in review UI. mattlam_ demonstrated inline code-review annotations rendered from response metadata, including file, line range, and priority fields.

Subagents became the first real argument after the prompt circulated. 0xblacklight said GPT-5.6 felt alien until they treated low thinking as the default, then escalated only when needed.

pvncher pushed back on the claim that Codex always defaults to subagents. In one reply, pvncher said subagents are not default in Codex, but skills can ask for them; another reply said the system prompt tells the model to use them only when the user asks, or when repo instructions or skills reference them.

OpenAI’s preview post gives the public version of the same capability: max gives Sol more time to reason, while ultra uses subagents to accelerate complex work. A community thread on OpenAI’s forum reported that Ultra “spawns a bazillion subagents,” and a separate forum post said Sol parent sessions appeared to drain quota when Luna subagents allegedly inherited the parent model.

The leaked-prompt cycle immediately turned into workflow packaging. kevinkern described a reusable skill built around GPT-5.6 prompting guidance and best practices.

That reaction matched the community mood. kevinkern’s reply compared the prompt-hunting to a social algorithm change or game balance patch: people were trying to understand the new meta, share what works, and update it as behavior changes.

One screenshot from dexhorthy shows why skills matter in practice: a library-researcher pulled guidance from React’s official docs that contradicted another model’s outline. That is the agentic coding version of prompt engineering moving out of the single prompt and into callable specialist routines.

The leak landed during a strange cross-harness moment. AI Builder Club claimed OpenAI was openly encouraging developers to run GPT-5.6 inside Claude Code via CLIProxyAPI: set up the proxy, point Claude Code at gpt-5.6-sol, and run.

Third-party guides made the same route explicit. A Valletta Software guide described a local proxy that rewrites Claude Code requests to a model provider, while a Nerdschalk guide described Sol acting either as a reviewer or as the model driving the session.

The social read was less diplomatic. dejavucoder framed the Claude Code route as OpenAI rubbing it in Anthropic’s face, after AI Builder Club reposted GPT-5.6 running inside Claude Code.

A separate Codex privacy fight was unfolding at the same time. dbreunig called it wild to upload a whole local codebase unless a user opts out, while Ben Hylak called “slurping up arbitrary files” malware.

Gergely Orosz objected that sending a whole codebase upstream was the wrong design, then acknowledged a correction in the thread. David Gomes said the change had been reverted and might require a window reload.