dsh-anchored-flash

English | 中文

An experimental DeepSeek Harness agent preset: Minimal-aligned anchoring with low-injection post-promotion, an indirect AGENTS.md injection pattern, and subagent anchoring — the "anchored intelligence without the IQ drop" experiment series.

Fork of xiaobright/dsh-anchored-standard with local modifications; the experiment report is in upstream issue #49.

Maintenance status (2026-08-17): upstream xiaobright/dsh-anchored-standard has entered maintenance mode (FAREWELL — the official API price change made the experiment series cost-prohibitive). This repository, an experiment branch of that series, is frozen: it no longer tracks upstream, and the experiment conclusions live in upstream issue #49. Related work has moved to the zhu1090093659/dsh-web-ui ecosystem (the non-assertive AGENTS.md hint, instructionHint, was merged there as #388). Local installs keep working; no further releases are planned here.

What makes this different from upstream

Experiment summary (measured 2026-08-16, DeepSeek V4 Pro, reasoningEffort=max)

Fingerprint = completed-reasoning counts of we / let's / let me (case-insensitive, whole words). Sessions are real DSH sessions on Windows + rc.6.

Key conclusions:

1. The promoted system-prompt mutation (46→6620 chars) is itself a perturbation. Keep the system prompt at the single Minimal sentence in every phase.
2. Imperative wording in a user-role injected message correlates with the style switch (we report correlation, not mechanism).
3. Non-imperative reference wording does not flip the trajectory, and read willingness is driven by task need, not hint wording.
4. Anchored intelligence and AGENTS.md availability are not mutually exclusive: tell the model (non-imperatively) that reference documents exist, and let it load them on demand.

Install

$target = Join-Path $env:USERPROFILE '.dsh\.agent-presets\anchored-flash'
if (Test-Path -LiteralPath $target) { throw "Preset already exists: $target" }
New-Item -ItemType Directory -Force -Path (Split-Path -Parent $target) | Out-Null
Copy-Item -Recurse -LiteralPath '.\preset' -Destination $target

Restart DeepSeek Harness, create a fresh session, pick 锚定标准·满血子代理 (anchored-flash). Do not switch presets mid-session.

Recommended AGENTS.md layout (the indirect-injection pattern)

Split the user-global ~/.dsh/AGENTS.md into:

• ~/.dsh/AGENTS.md — ~1 KB topic index (3 iron rules + pointer table)
• ~/.dsh/AGENTS-<topic>.md — topic files (environment / network / dsh / workflow / browser …)
• ~/.dsh/skills/env-* — the same content as skills, discoverable via skill_search and loadable via skill_load

Verify

node verify/verify-anchored-flash.mjs      # 18 logic checks, no DSH runtime needed
node verify/trace-session.mjs <session-dir>        # header + reasoning fingerprint trace
node verify/count-pronouns.mjs <session-dir>       # we / let's / let me counters
node verify/dive-session.mjs <session-dir>         # full system + message sources
node verify/dump-session.mjs <session-dir>         # header + first-line dump

License

MIT. The preset is based on the DeepSeek Harness Standard preset and on xiaobright/dsh-anchored-standard; see NOTICE and LICENSE. Part of the experiment design and this documentation were assisted by DeepSeek (an AI assistant); all measurements are from real DSH sessions.