# GPT-6 Astra · DeepSeek Harness Agent Preset

[中文说明](README.md) | **English**

A **DeepSeek Harness (dsh)** agent preset whose persona layer is the **GPT-6 / Codex
“Astra”** prompt set. Install it and pick “GPT-6 Astra” in dsh to get an agent with the
full dsh coding toolchain but the identity, tone, judgement, and autonomy of the Astra
text.

> In one line: **dsh's tools, Astra's persona.**

**Published: <https://github.com/Zioove/gpt-6-astra-dsh-preset>** — method A installs it in
one command.

---

## Features

- **The persona is selected block by block, not copied wholesale** — the source is a
  library of fragments for several Codex surfaces; this preset inlines only the two
  contiguous blocks that are actually a coding agent's instructions (see “How the persona
  was selected”).
- **The tool surface stays complete** — the preset derives from dsh's shipped `standard`
  preset as a whole directory, so file read/write, search, pwsh, background jobs, Skills,
  plan mode, goals, subagents, workflows, and the web tools are all kept.
- **One run-environment adaptation section** — `prompt/harness-bridge.md` maps the names
  the source uses (`functions.exec`, `apply_patch`, `update_plan`, `request_user_input`,
  `skills.list`, …) onto tools that exist here, and states which Codex product surfaces
  (Apps/Connectors, Plugins, `request_permissions`, `end_conversation`, …) do not.
- **Spawned subagents do not carry the persona** — the `tool-subagent` row (provider
  `spawn`) installs a 444-character child persona. The `subagent_fork` row deliberately
  keeps the parent persona: inheriting history and reusing the parent's KV cache is its
  whole point.
- **The excluded text travels with the preset** —
  `skills/gpt-6-astra-reference/` (mounted through `customSkillDirs` + `baseUrl`) carries
  the ~4,600 excluded lines, the selection map, and the rebuild recipe as an on-demand
  skill: zero resident tokens.
- **The model name comes from a variable** — the persona refers to `{{model}}` instead of
  hardcoding a name.
- **Rebuildable** — `scripts/build.mjs` regenerates the composition from `base/` +
  `prompt/`.
- **Mount-verified** — the composition passes a real mount through
  `agentPresets.standingKeyFor()`: all 18 rows activate, with no process-global service
  leakage.

---

## Repository layout

```
.
├── package.json              # also a dsh bundle plugin (dsh.bundle.patch)
├── cordis.patch.yml          # bundle patch: inserts one host row
├── lib/
│   └── index.js              # host row: syncs preset/ (files + skills/) into $DSH_HOME/.agent-presets/
├── preset/
│   ├── agent.cordis.yml      # ready-to-install composition (inlines the persona, ~60 KB)
│   ├── preset.yml            # display name and description
│   └── skills/
│       └── gpt-6-astra-reference/
│           ├── SKILL.md      # preset-local skill: selection map, exclusions, how to change them
│           └── reference/GPT-6-Astra_Prompts.md   # the source document, verbatim (339 KB)
├── base/
│   └── standard.agent.cordis.yml   # upstream baseline (dsh 0.1.5-rc.3 `standard`)
├── prompt/
│   ├── GPT-6-Astra_Prompts.md      # the source document (verbatim, unmodified)
│   └── harness-bridge.md           # appended run-environment adaptation
├── scripts/
│   ├── build.mjs             # selection + row patches → preset/agent.cordis.yml
│   ├── check.mjs             # artifact self-check (rows, persona schema, patches, skill files)
│   ├── check-plugin.mjs      # bundle self-check (DSH_HOME, files + skills/, idempotency)
│   ├── ci-check.sh           # local CI: reproducible build + variable guard + both self-checks
│   ├── install.ps1           # Windows installer (method B)
│   └── install.sh            # macOS / Linux installer (method B)
├── docs/
│   └── notes.md              # how it works, why this selection, schema gotchas, upgrades
├── NOTICE.md                 # provenance and ownership of the prompt text
├── README.md                 # 中文说明
└── LICENSE                   # MIT (scripts and configuration only)
```

---

## Install

A preset's directory name is its id; dsh only reads local presets from
`${DSH_HOME:-~/.dsh}/.agent-presets/<id>/`.

### Method A — through dsh (recommended: one command, self-updating)

This repository is also a **dsh bundle plugin**. Once installed it syncs the composition,
its metadata, and the whole `skills/` tree into the preset directory on every boot:

```bash
dsh plugin --profile web add github:Zioove/gpt-6-astra-dsh-preset
# then restart dsh web (bundle layers load on the next start)
```

A local checkout works too: `dsh plugin --profile web add ./gpt-6-astra-dsh-preset`

- upgrade: `dsh plugin --profile web update dsh-preset-gpt-6-astra`
- uninstall: `dsh plugin --profile web remove dsh-preset-gpt-6-astra`
  (the synced preset directory is left behind; delete
  `<DSH_HOME>/.agent-presets/gpt-6-astra/` yourself if you want it gone)
- The plugin has **zero dependencies** (Node built-ins only), is **idempotent**
  (per-file byte comparison), and **never takes the boot down** (failures are logged
  only). It treats this repository as the source of truth and overwrites same-named local
  edits, including inside `skills/` — but it never deletes.

### Method B — run the installer

```powershell
pwsh -File scripts/install.ps1          # Windows (-Force to overwrite)
```

```bash
bash scripts/install.sh                 # macOS / Linux (FORCE=1 to overwrite)
```

### Method C — copy the files by hand

Copy `preset/` into `%DSH_HOME%\.agent-presets\gpt-6-astra\`
(or `~/.dsh/.agent-presets/gpt-6-astra/`):

```
.agent-presets/gpt-6-astra/agent.cordis.yml
.agent-presets/gpt-6-astra/preset.yml
.agent-presets/gpt-6-astra/skills/…          # optional: the reference skill
```

### Zero-copy (advanced)

The `dsh-agent-presets` row accepts a `roots` configuration (earlier roots win), so a
clone can itself be mounted as a preset root: rearrange `preset/` into
`presets/gpt-6-astra/agent.cordis.yml`, then add this to the profile's
`cordis.patch.yml` or the host configuration:

```yaml
- id: agent-presets
  config:
    roots:
      - path: ~/src/gpt-6-astra-dsh-preset/presets
        trust: user
```

`path` supports `~`; relative paths resolve against the process cwd.

---

## Verify and rebuild

```bash
node scripts/build.mjs        # selection + patches → preset/agent.cordis.yml
node scripts/check.mjs        # self-check: rows, persona schema, patches, skill files
node scripts/check-plugin.mjs # bundle self-check: DSH_HOME resolution, install, idempotency
bash scripts/ci-check.sh      # all of the above plus reproducible-build and variable guards
```

`preset/agent.cordis.yml` is a **generated artifact**: after changing the selection, the
bridge, or `base/`, re-run `build.mjs` and commit the result, or CI fails on
`git diff --exit-code`.

---

## Usage

1. Reopen or reload the dsh Web GUI and pick **GPT-6 Astra** in the session's preset
   picker.
2. To make it the default for new sessions:

   ```yaml
   agent-presets:
     default: gpt-6-astra
   ```

---

## How the persona was selected (important)

The source document is titled **“Collected prompts and templates”**: a library of prompt
fragments for **several Codex surfaces** (5,051 lines), not one system prompt. This preset
inlines only two **contiguous** blocks that are actually a coding agent's instructions:

| block | source lines | what it is |
|---|---|---|
| `## Base model instruction template` → just before `# Apps (Connectors)` | 5–157 | the model layer: identity, permission posture, autonomy, personality and writing style, working with the user, rules for getting work done, using skills |
| `## Engine default base instruction template` → end of file | 4,775–5,051 | the engine layer: the assembled Codex CLI base instruction — how you work, AGENTS.md, responsiveness, planning, task execution, validating work, ambition vs. precision, progress updates, final answer structure, tool guidelines |

The other ~4,600 lines are **deliberately excluded** because they are addressed to *other*
models: `# Classification` ends with "Output that single token immediately and nothing
else." (a classifier), plus the Guardian review templates, the approval and
`request_permissions` fragments, the browser confirmation policies, the code-review
rubric, the realtime backend, the desktop/UI fragments, and the orchestration and
compaction templates. None of it is lost — it lives in
`preset/skills/gpt-6-astra-reference/reference/` and loads on demand.

There is also a purely technical reason: line 362 contains
`{{ tenant_policy_config }}`. dsh interpolates `{{…}}` **strictly**, so an unregistered
group fails the whole turn — and `dsh-persona` has no `interpolate: false` (only sections
registered through `ctx.systemPrompt.section({ interpolate: false })` do). Inlining the
whole file would simply not mount.

To change the selection, edit `SELECTION` in `scripts/build.mjs` (start/end heading
markers) and re-run build + check; a missing anchor fails the build.

---

## Deliberate deviations from `standard`

`build.mjs` carries an explicit `PATCHES` list; each entry names the upstream fact it
overrides and why. If an anchor is missing (or matches more than once) in a newer
`standard`, the build **fails loudly** instead of silently dropping the change:

| patch | why |
|---|---|
| `skill-filesystem` gains `customSkillDirs` (`!!js` + `baseUrl` pointing at the preset's own `skills/`) | lets the excluded text and the selection map travel with the preset and load on demand; the same mechanism the shipped `cordis` preset uses |
| `tool-subagent` (spawn) gains a 444-character `persona` | the persona is resident cost; a freshly spawned child does not need the full version. `subagent_fork` is left alone — it exists to inherit history and reuse the parent's KV cache |

---

## How it works

dsh assembles its system prompt from **ordered sections**; the `dsh-persona` row supplies
two of them:

| section | order | content in this preset |
|---|---|---|
| `deployment:persona-prefix` | 0 | the two Astra blocks + the run-environment adaptation |
| `deployment:persona-suffix` | 10200 | `Your working directory is {{cwd}}.` |

Ordering is **identity-first** (the Claude Code / Codex convention): harness identity
(−100) → persona (0) → each tool row's own guidance (100–199) → the working-directory
suffix (10200). dsh's tool guidance therefore comes *after* the persona and wins on
conflict.

The three `isolate` realm rules (`planning` / `compaction` / `delegation`) are a hard
requirement dsh places on services a preset owns; they are kept exactly as upstream.

---

## Cost note

The persona section is resident: roughly **44 K characters ≈ 11k tokens** on every
request — about one eighth of inlining the whole 339 KB document. On top of that, a
spawned subagent pays ~100 tokens of persona instead of the full text, and the reference
skill costs nothing until it is loaded.

---

## Upgrade and maintenance (important)

dsh changes row config schemas. Gotchas this repository has hit:

- **dsh ≤ 0.1.4**: the `dsh-persona` fields were `text: <string>`.
- **Since dsh 0.1.5-rc.3**: the fields are `prefix` (required) and `suffix`; the old form
  **fails to mount** with `invalid config: $.prefix missing required value`. If that
  preset is also the default, the symptom is that **starting a new conversation fails
  outright and no session directory is created on disk**.

After upgrading dsh, re-sync the baseline and rebuild:

```bash
# 1) replace base/ with the new shipped standard
cp "$DSH_INSTALL/node_modules/@deepseek-ai/dsh-agent-presets/presets/standard/agent.cordis.yml" \
   base/standard.agent.cordis.yml
# 2) rebuild + self-check
node scripts/build.mjs
node scripts/check.mjs
# 3) install/update, then start one session in the GUI to confirm
dsh plugin --profile web update dsh-preset-gpt-6-astra     # method A
#   or: bash scripts/install.sh / pwsh -File scripts/install.ps1 -Force
```

Step 2 is where an upstream `standard` rewrite surfaces: the `PATCHES` anchors are
checked, so a model, skill, or subagent row that upstream restructured fails the build
with the patch label instead of quietly reverting to upstream behaviour.

---

## Known limitations

- The Codex product surfaces the source mentions (Apps / Connectors, Plugins, Artifacts,
  `request_permissions`, `end_conversation`, the Guardian reviewer, the realtime backend)
  do not exist in dsh and are explicitly excluded by the adaptation section; for
  cross-session memory, use this machine's memory service (OpenViking) tools.
- The source is written against OpenAI's product surface. This preset uses it only as
  persona and behavioural guidance; it does not change model routing and does not
  represent any official position of OpenAI or DeepSeek — `{{model}}` states which model
  is actually answering.
- Any edit to the persona text changes the system-prompt prefix and therefore rebuilds
  the KV cache.
- A spawned subagent's persona is shorter than its parent's. Delete the `persona:` line
  from the spawn row in `preset/agent.cordis.yml` (and rebuild) to give children the full
  persona again.

---

## Contributors

- **Author**: [Zioove](https://github.com/Zioove) — the preset composition, the bundle plugin, the selection and run-environment adaptation, the build/check scripts, and the documentation.
- **Contributor**: [T0mor1](https://github.com/T0mr1) — repository collaborator.

Every commit in this repository is authored by Zioove; the repository lives under Zioove, with T0mor1 as a collaborator.

---
## License and attribution

- **Scripts, configuration, and documentation** in this repository: MIT, see `LICENSE`.
- **`prompt/GPT-6-Astra_Prompts.md`** (and the same text inside
  `preset/skills/.../reference/`) is product prompt text for OpenAI's GPT-6 / Codex
  surfaces, included verbatim for technical research and interoperability work only, and
  **not** covered by the MIT license. See `NOTICE.md`.