dsh-research-first

dsh-research-first

A research-first assistant plugin for DeepSeek Harness (DSH).

Gently nudges the agent to do a cheap, deterministic investigation before editing code — without hard-blocking (so "can't reach docs" never stalls the work).

简体中文

agent about to write/edit/bash without investigating
        │
        ▼
  soft nudge: allow + inject a reminder (model-visible, logged)
        │
  investigation blocked (web_search failed / GitHub unreachable)
        │
        ▼
  remind the user why, offer to continue anyway — fully logged, replayable

---

Why

Three real failure modes of LLM coding, which compound each other:

Failure	Symptom
Guessing before acting	Writes code from memory instead of checking requirements / API / docs
Trial-and-error loops	Repeated failures without looking up known community workarounds
Stale knowledge	Acts on outdated versions / endpoints that have since changed

The core insight: one cheap lookup prevents one expensive rework.

---

Design philosophy

Make "verify first" a preferred attempt + frictionless nudge + blocked-notice + full trace, not a hard prerequisite.

Why not hard-block? Investigation can be genuinely impossible (air-gapped environments, network restrictions, no official docs). If "can't verify" halts development, the cure is worse than the disease. So the default is a soft nudge — remind without blocking, leaving the choice to the agent and the user.

Behavior	Description
Soft nudge (default remind)	Editing without investigating → not blocked, but a reminder is injected (model-visible + logged)
Blocked-research detection	Investigation tools (web_search, etc.) failing → reminder: "research blocked, ask the user whether to continue"
Research-norm injection	Adds "golden research habits" to systemPrompt: search with versions, prefer official tools, explain when blocked
Failure reminder	After repeated failures, remind to check community feedback (GitHub issues) first
Full trace	Every reminder is a plugin-sourced message in the session log — replayable, auditable

Set intensity: warn (hold for approval) or intensity: block (deny) when you want hard enforcement.

---

Comparison with similar plugins

Plugin	Focus	Difference from this one
dsh-doublecheck	Engineering-discipline gates (grill requirements, red/green tests, adversarial review)	Defaults to hard enforcement; this plugin defaults to soft
dsh-pain-point-check	Veto-based stop after failed experiments	It denies non-investigative tools; this plugin reminds without blocking
dsh-research-first	Lightweight research-first assistant	Frictionless, non-blocking, hands decisions back to the user when blocked

They compose — use doublecheck for strict discipline, or this plugin for a lightweight, unobtrusive default.

---

Install

Option 1: official (recommended, needs pnpm)

dsh plugin --profile web add github:outnever/dsh-research-first

Restart:

dsh web

Option 2: manual (no pnpm)

# 1. symlink into the profile's node_modules
ln -sfn /absolute/path/to/dsh-research-first ~/.dsh/profiles/web/node_modules/dsh-research-first

# 2. add to ~/.dsh/profiles/web/cordis.patch.yml:
# - insert:
#     - id: research-first
#       name: 'dsh-research-first'
#       config: { intensity: remind }

# 3. restart dsh web

Verify:

dsh --profile web --dump-config | grep -A2 research-first
# expect: id: research-first / name: dsh-research-first / intensity: remind

---

Configuration

Field	Default	Meaning
intensity	remind	remind (nudge) / warn (approval) / block (deny)
failureThreshold	2	Consecutive failures before the community-check reminder
injectNorm	true	Whether to inject the research norm into systemPrompt
detectBlocked	true	Whether to detect failing investigation tools and remind
investigationTools	read, grep, glob, web_search, read_image, ask_user_question, skill	Tool names counted as "investigation"
mutationTools	write, edit, bash, pwsh, str_replace_editor	Tool names counted as "editing"

---

How it works

Built on DSH's official extension points only — no private machinery:

• tools/pre-execute (waterfall): mark investigation / decide allow · remind · deny
• tools/post-execute (waterfall): inject additionalContexts reminders, detect blocked research, count failures
• agent/pre-step (waterfall): reset the investigation flag on step===1
• systemPrompt.section: inject the research norm

Key implementation detail: reminders are hand-written UserMessage shapes ({ id, role:'user', content, source:{kind:'plugin',...} }) with id from Node's built-in crypto.randomUUID(), keeping zero dsh package imports — so the plugin loads via symlink without pnpm resolving dependencies.

---

Testing

node test.mjs
# 19 mock unit tests: soft nudge, blocked-research, failure reminder, block/warn modes, norm injection, turn reset

Runs with no dependencies — lib/index.js only imports node:crypto.

---

Known limitations

• The remind reminder is a soft hint in the session — not enforced; the model may ignore it. Use block to enforce.
• "Search with versions / prefer official tools" is soft-guided via systemPrompt — rules can't hard-guarantee it (it's ultimately the model's willingness).

---

Contributing

Issues and PRs welcome. When changing:

1. Keep zero dsh package imports (prerequisite for symlink loading without pnpm);
2. Update test.mjs and keep node test.mjs green;
3. Update the README's behavior description.

---

License

MIT