dsh-anchored-standard-profile

A community profile for DeepSeek Harness that measurably raises the output quality and efficiency of deepseek-v4-pro at MAX reasoning effort on Windows — by combining:

1. the Anchored Standard preset (two-phase context bootstrap),
2. a battle-tested AGENTS.md knowledge router (v4.2.30, with 25 frozen backups),
3. a modernized shell stack (PowerShell 7.6.5 instead of 5.1, busybox sh instead of bash),
4. a disciplined tooling/policy stack, and
5. a real end-to-end case study: GARGANTUA, a Schwarzschild black-hole ray tracer, with its full session trajectory included for comparative testing.

• English · 中文说明

Community project. Not affiliated with or endorsed by DeepSeek. All measurements come from one machine (see AGENTS.md for the exact hardware/toolchain facts); reproduce on your own machine before drawing conclusions.

---

TL;DR — why this works

deepseek-v4-pro conditions strongly on the API tool catalog and the first-request output budget. This profile exploits exactly that:

• Request #1 is deliberately minimal: only one shell tool + read, a 1024-token output cap, and a ~1.9 KB preamble (the expression-law + system-instruction header of AGENTS.md). Skills and heavyweight plugin context are suppressed.
• After the first promotion signal (first tool call or first assistant message — promoteOn: either), the session unlocks the full Standard tool catalog, the complete AGENTS.md digest and the normal output budget.
• Result: the first thinking pass stays cheap, fast and unpolluted; MAX reasoning is spent on the actual task, not on absorbing a large injected context.
• Once promoted, AGENTS.md acts as an O(1) lookup router — facts, failure-mode countermeasures, PowerShell 5.1→7.6 migration diffs and task→tool routes — so the model retrieves instead of re-deriving, cutting retry loops and wasted reasoning tokens.

This is exactly the setup under which the Gargantua case study was produced: a complete WebGL Schwarzschild geodesic ray tracer (13 JS modules, headless-CDP acceptance, 0 console errors, physics signatures verified by vision) delivered in 2 turns / 41 steps at deepseek-v4-pro + reasoningEffort: max.

---

Repository layout

dsh-anchored-standard-profile/
├─ README.md / README.zh-CN.md        ← this document
├─ LICENSE / NOTICE                    ← MIT + attribution
├─ AGENTS.md                           ← current profile v4.2.30 (the file injected into DSH sessions)
├─ COMPUTER_PROFILE.md                 ← workspace mirror of the same profile
├─ AGENTS.full.md / COMPUTER_PROFILE.full.md
│                                      ← full archive: original text + every revision entry
├─ SHELL_USAGE_SPEC.md                 ← busybox/pwsh shell discipline referenced by AGENTS.md
├─ backups/                            ← 25 frozen snapshots (AGENTS.md.bak-v3 … v4229b,
│                                        COMPUTER_PROFILE.md.bak-v4212 … v4229b)
├─ config/
│  ├─ settings.yaml                    ← model = deepseek-v4-pro, reasoningEffort = max, 14 parallel tool calls
│  ├─ cordis.patch.yml                 ← root patch layer
│  └─ profile-web/
│     ├─ package.json                  ← DSH web profile bundles (local paths redacted as <DSH_WORKSPACE>)
│     ├─ cordis.patch.yml
│     └─ pnpm-workspace.yaml
├─ preset/                             ← the anchored-standard preset this profile runs on
│  ├─ agent.cordis.yml                 ← agent-plane composition (persona, bootstrap, compaction, delegation…)
│  ├─ preset.yml
│  ├─ tool-bootstrap.mjs               ← bootstrap filter (full variants)
│  ├─ tool-bootstrap-trim4.mjs
│  └─ tool-bootstrap-trim5.mjs         ← the one wired into agent.cordis.yml (preamble-only request #1)
└─ case-studies/gargantua/
   ├─ README.md                        ← reproduction + A/B comparison guide
   ├─ project/                         ← the full Gargantua web app (renderer, shaders, tests, screenshots)
   └─ session-trajectory/
      └─ dsh-session-session-350b5553-….zip   ← 7,727 events of the original DSH session

---

Mechanism, layer by layer

1. Anchored Standard preset — two-phase bootstrap

From preset/agent.cordis.yml (the comments in that file document every design decision):

- id: tool-bootstrap
  name: ./tool-bootstrap-trim5.mjs
  config:
    shellTools: [bash, pwsh]
    commonTools: [read]
    promoteOn: either
    bootstrapMaxTokens: 1024
    suppressedContextSources: [skill-catalog, plugin:hindsight]

What happens across requests:

Phase	Request #1	First promoted request onwards
Tool catalog	one shell + read	full Standard catalog (fs, jobs, skills, goals, plan mode, subagents, workflows, web…)
Output cap	1024 tokens	normal budget
AGENTS.md	preamble only (~1.9 KB: expression law + system instructions)	full digest (~11 KB)
skill-catalog / hindsight	suppressed	re-injected

Why it matters for V4 Pro (from the preset's own design notes): "V4 Pro conditions strongly on the API tool catalog AND the first request output budget." A minimal first request means the model starts acting immediately instead of first absorbing a large system context, and the small cap keeps the first turn cheap. Nothing is lost — the full context returns exactly when the session proves it is doing real work.

tool-bootstrap-trim5.mjs implements the preamble-only substitution and the full-digest re-injection on promotion (the trim4/trim5 line fixes a 1:1 replacement edge case documented in the backups; see backups/AGENTS.md.bak-v4229b and the failure-mode table).

2. AGENTS.md as a knowledge router, not prose

AGENTS.md (v4.2.30, 11,114 bytes) is the profile that gets injected into every session. Its structure is deliberately machine-lookup-shaped:

• Facts (measured) — exact OS/hardware/toolchain versions, so the model never wastes reasoning on uname-style probing.
• Sandbox (measured) — what each sandbox mode allows.
• Failure-mode lookup table — symptom → countermeasure for every previously-observed failure (WMI denied, Git Bash crash, node pipe EPERM, stale PATH cache, ESM reload…). Troubleshooting becomes a table lookup, which saves MAX reasoning budget and eliminates retry loops.
• PS7 migration diff table — every PowerShell 5.1→7.6 breaking change and its fix (UTF-8 NoBOM, -AsByteStream, removed WMI/EventLog cmdlets, Split('pq'), web cmdlets, $LASTEXITCODE…). Compatibility guesswork disappears.
• Routing table — task type → tool (busybox sh for bash semantics, pwsh for launcher/fallback, read/glob/grep/edit for files, dev_* for plugins…).
• Discipline — no dead-end grinding, permission-aware escalation rhythm, cost rules (subagents on flash, MAX only for real work), cleanup discipline.
• Expression law (multi-expert routing persona) — a complete: true persona in the preset system prompt: first-person-plural "we/us" thinking from the very first token, a full-chain ban on the exact sequence "The user", counter-instruction precedence. Plural self-reference keeps the whole expert-routing cluster engaged instead of a single-route mindset. The full archive records the regression evidence (e.g. expression-law compliance Fisher p = 3.05e-10; counter-instruction arms).

The evolution history is shipped in backups/ — from v3 (2.5 KB) to v4.2.30 (11.1 KB) — including the EvoX convergence runs, blind-review scores, double-blind experiments and per-revision rationale documented in AGENTS.full.md. Browsers can watch the profile converge revision by revision.

3. PowerShell 7.6.5 replaces Windows PowerShell 5.1

• DSH resolves pwsh once at daemon start; pin the path in settings.yaml → shell.pwshPath.
• PS7 gives UTF-8 NoBOM defaults, -Parallel, cd -, ??, Get-Uptime, Test-Json, ConvertFrom-Markdown.
• The 5.1→7.6 diff table in AGENTS.md removes every migration trap the model would otherwise discover by trial and error.
• Telemetry off: $env:POWERSHELL_TELEMETRY_OPTOUT='true'.

4. busybox sh replaces bash

• Git Bash crashes inside DSH's named-pipe sandbox (measured; see the failure-mode table).
• busybox-w32 sh gives zero-approval bash semantics: sh, sed, grep, wc, sort, awk -f.
• SHELL_USAGE_SPEC.md codifies the working patterns (script files instead of inline sh -c, file redirection instead of raw stdout capture, awk -f scripts) — so shell work succeeds on the first attempt.

5. Policy stack (config/settings.yaml)

agent-default-model:
  provider: deepseek-modlens
  model: deepseek-v4-pro
  reasoningEffort: max
agent-presets:
  default: anchored-standard
agent-loop:
  maxParallelToolCalls: 14
shell:
  maxOutputBytes: 384000

• Main agent: deepseek-v4-pro + MAX, with modlens as the provider (vision verification of screenshots is part of the Gargantua acceptance loop).
• Subagents: pinned to deepseek-v4-flash in the preset (tool-subagent / tool-subagent-fork agentOptions) — parallelism and fan-out stay cheap; MAX reasoning is reserved for the main trajectory.
• permission.defaultPreset: danger-full-access — this profile assumes a trusted local machine; adjust for your own threat model.

---

Plugins & ecosystem (so customers can find everything)

Component	Where
DeepSeek Harness (the host)	https://github.com/deepseek-ai/deepseek-harness
dsh-anchored-standard preset plugin (canonical upstream of preset/)	https://github.com/xiaobright/dsh-anchored-standard
modlens (vision bridge used by the provider)	npm @liustack/modlens
dsh-evox (EvoX evolution harness used to converge AGENTS.md)	local build in DSH plugins/dsh-evox; loaded as bundle @dsh-external/dsh-evox
dsh-super-injector (runtime plugin injection/healing)	bundle @dsh-external/dsh-super-injector
dshmarket (plugin market)	npm dshmarket
dsh-better-sidebar	npm dsh-better-sidebar
dsh-web-ui-all (panels: task board, git graph, ssh, aionui…)	npm @linxin666/dsh-web-ui-all
dsh-deep-whale (skins/maid-atelier)	https://github.com/Small-tailqwq/dsh-deep-whale

---

Case study — GARGANTUA (Schwarzschild black-hole ray tracer)

See case-studies/gargantua/README.md and case-studies/gargantua/project/README.md for the full story. In short:

• Every pixel integrates the Schwarzschild null geodesic in a fragment shader — real physics, no fake black sphere: event horizon, photon sphere, photon ring, lensed disk, Doppler beaming + gravitational redshift, volumetric accretion cloud, ACES/Bloom/FXAA post chain, 21 runtime parameters, debug heatmaps.
• Built end-to-end by deepseek-v4-pro + MAX under exactly this profile: 2 turns / 41 steps, 7,727 session events, delivered with a zero-dependency server, a headless-CDP acceptance suite and vision verification.
• Automated acceptance (tests/results.txt): 13/13 JS modules syntax-OK; CDP READY, 0 console errors, WebGL2 + half-float render target; pixel statistics (e.g. Cinematic render meanLum 0.279 / maxLum 1.0 / 26% bright / 42% deep shadow); modlens semantic checks (black circular shadow, photon ring, left-bright/right-dim Doppler asymmetry, lensed striations, face-on volumetric disk, redshift heatmap).
• The original session trajectory (session-trajectory/*.zip, 948 KB, unzips to session.jsonl with 7,727 lines) lets anyone inspect what the model actually did step by step, including the per-request headers showing the preset and model.

---

Install & reproduce

1. Install DeepSeek Harness.
2. Install PowerShell 7.6+ and add C:\Program Files\PowerShell\7\pwsh.exe; pin it in DSH settings (shell.pwshPath).
3. Put busybox.exe (busybox-w32) somewhere stable and add its bin/ directory to PATH.
4. Copy this repo's files:
   • AGENTS.md → ~/.dsh/AGENTS.md (and COMPUTER_PROFILE.md → your DSH workspace, if you use the mirror convention)
   • preset/ → ~/.dsh/.agent-presets/anchored-standard/
   • config/settings.yaml → merge into ~/.dsh/settings.yaml (mind shell.pwshPath and your permission preset)
   • SHELL_USAGE_SPEC.md → your DSH workspace
5. Restart DSH and create a new session with the anchored-standard preset.
6. Run the Gargantua prompt from case-studies/gargantua/README.md and verify with the project's test suite.

---

Comparative (A/B) testing — the invitation

The whole point of this repository is falsifiable comparison. The session export API makes it easy:

1. Create session A with the DSH default preset and session B with anchored-standard, same prompt (the Gargantua build prompt).
2. For each session record: total tokens, reasoning tokens, tool calls, number of steps/turns, wall-clock time, and failure/retry count (browser console + session telemetry).
3. Export both sessions (session.export → zip → session.jsonl) and diff the trajectories side by side.
4. Post your results as an issue or PR into docs/comparisons/. Include machine spec, model, effort tier and DSH version.

Suggested metric table:

Metric	Default preset	anchored-standard
Turns / steps
Total tokens
Reasoning tokens
Tool calls
Failed/retried tool calls
Wall time

We — the maintainers of this profile — welcome independent negative results too; they are the only way the claims above get stronger or get corrected.

---

Privacy notes

• .credentials.yaml and any API keys are never part of this repository.
• Machine-specific absolute paths in config/profile-web/package.json were redacted to <DSH_WORKSPACE> placeholders.
• AGENTS.md keeps its original local paths (e.g. C:\deepseek_harness\DSH, C:\Users\<you>\.dsh) — adapt them to your machine when installing.

License & attribution

• MIT — see LICENSE.
• The profile derives from the DeepSeek Harness Standard preset (MIT, DeepSeek) and the dsh-anchored-standard community plugin (MIT, xiaobright) — see NOTICE.
• Gargantua ships vendored three.js r160 (MIT) — see case-studies/gargantua/project/vendor/three.