dsh-gpt56-ptc

GPT-5.6's raw coding ability is rarely the bottleneck. The real friction shows up over long agentic sessions, where boundaries quietly dissolve: a small edit snowballs into a governance framework, a reviewer comment spawns three rounds of new features, passing tests get re-verified just to be safe, and the actual deliverable gets buried under planning docs and status reports. dsh-gpt56-ptc targets exactly this gap. It does not replace the model's capability; it layers a minimal trajectory-control contract on top of the session so the work stays scoped.

The preset keeps the full official Code preset composition, the in-session fixed SDK, and the pure PTC tool presentation intact, then adds only a lean trajectory contract and a single static persona. The rules are deliberately restrained: treat the user-visible outcome and acceptance evidence as the task boundary; do not auto-expand scope from tangential findings; let plans, tests, and reviews serve as evidence, never as substitutes for the deliverable; verify only the affected paths and refrain from re-testing without new evidence; stop immediately once acceptance criteria are met; run independent read-only calls in parallel while keeping writes and approval-sensitive actions ordered. The contract also asks the model to explain results and causality in plain language by default, reaching for technical terms only when precision demands it and glossing them on first use. There is no auto-routing, no Native/PTC flip-flopping, and no per-turn tool pruning. The session stays stable for its entire lifetime, and you pick the model manually at creation time.

If you are running agentic coding on gpt-5.6-sol, terra, or luna inside DeepSeek Harness and keep hitting the wall of a model so capable it simply cannot stop, this preset is a low-intrusion complement. It does not configure a provider, a sandbox, or an approval policy, and it does not auto-switch models. After installation you get one extra custom entry, GPT-5.6 PTC, in the preset list; select it, manually choose a target model, and you are ready. It is best suited to developers who value a clean delivery cadence and want to keep the model's strength tethered to the current task's boundary.