Error: unknown tool "read": only run_code is callable directly — call read from inside a run_code program instead

[zhao999]

@deepseek-ai/dsh@0.1.1-rc.2 PTC, The tool invoked is incorrect. The read is a subtool. Every time start to read the code, this error is reported first, and then the run_code executes correctly. The read is successfully executed within the subtool

[weijiafu14]

This is PTC/Code-Mode's tool surface working as designed, colliding with the model's habits — and there are two validated ways out depending on what you want:

What's happening: in PTC mode only run_code is directly callable; read/edit/bash are sub-tools to be called from inside a run_code program. Models that grew up on native function calling reflexively try read directly first, eat this error, then fall back correctly — that first failed call is pure waste (same pattern reported in #1543: first call goes standard, only after failing does it use run_code).

Options: ① if you don't specifically want code-mode, switch the tool presentation to native function calling (DSH_TOOLS_MODE=native in the launching environment) — direct read/edit calls then just work; ② if you want PTC, the per-turn waste is a prompt-adherence issue upstream — the mode's instructions exist but models skip them on the first call; worth reporting with your model name so the mode's system prompt can be hardened.

Related family datapoint while you're here: if you ever see tool calls failing with empty names/arguments on third-party gateways, that's a different (streaming-assembly) family — #3090 / our measurements cover it.

[muyuanjin]

Try dsh-ptc-plus — it auto-rewrites top-level tool calls like read into run_code, cutting that unknown tool noise before it reaches you.