Skip to content
MyDSHMy DSH Field Notes
中文
Independently maintained; not affiliated with DeepSeek. No association with mydsh.dev or other MyDSH sites.

Home/Modes & context/Working with the modes

The four built-in presets: standard, PTC, minimal, cordis

Picking a preset without knowing the difference, then blaming the choice when something goes wrong.

VerifiedVersion and environmentChecked 2026-11-14
Source
presets/*/preset.yml in the repo
Host version
0.1.2-alpha.5
Checked
2026-11-14

The four presets are defined as four directories in the repo, one preset.yml each. Quoted verbatim, not summarised:

orderidNameOfficial description
1standardStandardA feature-complete coding agent supporting file editing, shell, file and web search, Skills, plans, goals, subagents, and workflows.
2ptcPTCA feature-complete coding agent, but without workflow tools by default; the other tools are presented through the PTC SDK so the model composes multi-step operations in one TypeScript program.
3minimalMinimalA two-tool coding agent offering only a persistent bash and str_replace_editor.
4cordisCreatorFor authoring custom agent presets: every capability of standard, plus runtime inspection, plugin experiments, and preset authoring guidance.

Choosing

  • Standard by default: the widest tool set and the easiest place to find others who hit the same wall.
  • PTC suits work where multi-step operations can be written as one program; it collapses five round trips into one (next page). It requires a code runtime.
  • Minimal is for 'one shell and one editor'. The smallest surface and the most control — but no Skills, no plans, no subagents, and no context compaction at all.
  • Creator is for people writing custom presets, with runtime inspection and authoring guidance. Daily work does not need it.
The trap: assuming minimal just means fewer features. It also turns off context compaction (agent.cordis.yml says Context compaction is absent), so a long session hits the ceiling. Fewer features is survivable; missing compaction is not.

Same task, two presets

Run it yourself before deciding. One task, once in standard and once in PTC, recording only two numbers: tool calls and model round trips. Skip wall-clock time; the network dominates it.

# standard
In standard mode: read a.md, b.md, c.md and merge them into summary.md
# PTC
Do the same task in PTC mode

If PTC shows markedly fewer tool calls, the task suits programmatic composition. If it is the same, the extra layer buys nothing.

What counts as working

Check every line. All of them pass or the job is not done.

  • You can state the tool surface of each preset
  • You know minimal has no context compaction
  • You ran the same task under two presets once

Where it fails

Symptom, cause, what to do — written in the order it actually happens.

SymptomCauseWhat to do
On minimal, a long session suddenly stops answeringMinimal has no compaction; the context hit its ceilingSwitch to standard, or shorten the session
PTC fails to startPTC needs a code runtime with a registered SDK renderer (TS or Python)Confirm the codeRuntime, or use standard

Sources

Verified means measured here. Compiled means every line carries a source.