DSH Error Improvement

English | 中文

A DeepSeek Harness plugin that turns user-confirmed experience into a durable agent memory: mistakes become prevention rules, proven fixes become recipes, and battle-tested entries graduate into standalone skills.

It is intentionally independent from EverOS, memory services, databases, browsers, and network APIs. All persistent state lives in plain files under $DSH_HOME/error-improvement/.

Targets DSH 2.0.15+ (Node 22.19+ or 24+). For 0.1.x hosts use release 0.2.x.

Safety contract

• A memory entry is injected only when confirmed: true. Every auto-captured lesson lands in drafts/ first and requires explicit user approval — the plugin never writes model output into memory silently. (Deliberate exception: the agent-invoked improve_record_recipe tool, which records an already-verified solution.)
• It never executes tools or changes permissions, Desktop mode, browser access, network exposure, sandbox presets, or approval policy.
• Injected text uses plugin provenance (form: instructions), never direct-user provenance; subagents are told not to re-inject or re-record (<SUBAGENT-STOP> recursion guard).
• Stored text is control-character cleaned, role-tag encoded (all < and > are escaped), count-limited, and character-limited; files are written atomically (tmp + rename).
• Lessons are rendered atomically: an entry is included only if the entire block fits the character budget; otherwise it is omitted entirely (no mid-rule truncation).
• Every hook and every file/LLM operation fails open: a plugin error never blocks the tool pipeline or the agent loop.

How it works

Memory layer (0.3.0)

Plain Markdown (with structured fields) is the single source of truth:

$DSH_HOME/error-improvement/memory/
  lessons.md       # confirmed mistakes → prevention rules
  recipes.md       # confirmed proven solutions
  decisions.md     # user decisions (e.g. rejected drafts, with reason)
  candidates.jsonl # zero-LLM capture queue
  drafts/          # pending drafts awaiting user confirmation
$DSH_HOME/error-improvement/state.json   # error statistics + enforcement rules (v2)
$DSH_HOME/skills/<slug>/SKILL.md         # graduated skills

1. Capture (zero LLM). Turn outcomes and tool errors from direct sessions (delegated sessions are skipped) are queued into candidates.jsonl — capped per session and per text size.
2. Distillation (LLM, gated). When the queue reaches capture.minQueueSize and capture.minIntervalMs has passed, a two-stage Curator→Writer run proposes draft entries into drafts/. Empty results return a sentinel (FLUSH_OK) and retry later; nothing is written to memory.
3. Confirmation gate. Drafts appear in Settings → 错误改进 → 待确认草稿 (list, edit, approve, reject). The agent can also list and act on them via the improve_review_drafts tool. Approval writes the entry (confirmed) into the memory file; rejection logs a decision entry. Approving a draft whose supersedes points at an older entry disables the predecessor (never deletes).
4. Injection. On every model step of each turn the agent/pre-step waterfall selects relevant confirmed lessons (assist mode scores by scope/keywords with CJK bigram matching; strict injects all eligible entries; off disables injection) and renders them plus recipes inside one bounded, budget-measured block.
5. Turn signals & maturity. Entries presented during a turn that ended successfully (no new enforcement promotions, failures ≤ successes) gain a hit. Maturity evolves draft → validated (≥2 hits) → core (≥5 hits).
6. Skill graduation. Validated/core entries reaching graduation.minHits (default 3) are proposed as skills — again through a confirmation draft — and on approval written as $DSH_HOME/skills/<slug>/SKILL.md (agentskills.io-style frontmatter; refuses to overwrite existing files; the source entry is marked core and gains a skill: evidence note).

Repeated errors graduate into rules (0.2.0)

1. The Host listens to tools/post-execute and counts identical failures (same tool + same normalized error signature, long hex/digit runs collapsed). Counts persist across restarts in state.json.
2. Once a failure crosses enforcement.threshold (default 3), it is promoted into a rule — linked to a matching confirmed lesson when one exists, otherwise synthesized from the error sample as an auto-rule.
3. Matching future calls (same tool + sufficiently overlapping arguments) are intercepted on tools/pre-execute:
   • warn (default): intercepts once per cooldown window (warnCooldownMs, default 1h), showing the model the prevention reason; an immediate retry is allowed — a reminder, not a blockade;
   • deny: always blocks, for calls that must never happen again.
4. Every hook fails open and never weakens an existing downstream deny/ask into an allow. Rules are capped by maxRules (default 20); only the oldest auto-rules are evicted beyond the cap, lesson-linked rules are kept.

Success recipes

• After solving a non-trivial problem, the agent can call the improve_record_recipe tool to persist the verified solution (title/problem/solution/scope/keywords) into recipes.md as a confirmed entry.
• Recipes share the lessons' relevance engine (including CJK bigrams) and render in a <success_recipes> block next to the lessons, so a recurring problem is answered with the proven fix instead of being re-derived from scratch.

Configuration

Configuration is declared as a Schemastery schema and lives in the profile's cordis.patch.yml (~/.dsh/profiles/<profile>/cordis.patch.yml). Fields marked volatile are live-editable from the settings page without a plugin remount.

Field	Default	Meaning
enabled	true	Master switch for injection, enforcement, capture and graduation.
mode	assist	assist = inject relevant lessons as advice; strict = inject all confirmed lessons as rules; off = capture only, no injection.
maxLessons	5	Maximum lessons injected per turn (1–50).
maxChars	6000	Hard character budget for the lessons + recipes injection block (500–50000).
maxRecipes	3	Maximum success recipes injected per turn (0–20).
enforcement.enabled	true	Intercept tool calls that repeat a known failure signature.
enforcement.threshold	3	Identical tool errors before a runtime rule is promoted (2–10).
enforcement.defaultMode	warn	warn = intercept once per cooldown; deny = block repeats.
enforcement.warnCooldownMs	3600000	Minimum interval between two warnings of the same rule.
enforcement.maxRules	20	Cap of runtime auto-rules; oldest unlinked rules are evicted.
capture.enabled	true	Observe turns and queue candidate lessons/recipes (no LLM).
capture.distillEnabled	true	Periodically distill the queue into user-confirmation drafts (LLM).
capture.minQueueSize	3	Minimum queued candidates before a distillation run.
capture.minIntervalMs	1800000	Minimum interval between two distillation runs.
capture.provider / capture.model	empty	LLM route for distillation; empty = the turn's routed provider/model.
capture.maxCandidateChars	16000	Character budget of candidates fed into one distillation run.
graduation.enabled	true	Propose turning validated experience into reusable SKILL.md files.
graduation.minHits	3	Hits before a validated entry is proposed for skill graduation.

Migrating from 0.2.x: legacy ~/.dsh/settings.yaml values are not auto-migrated into cordis.patch.yml — defaults apply until you reconfigure. Only data (lessons, recipes, error statistics, rules) is imported automatically on first run.

Context compaction was removed in 0.3.0. Use the official compaction-basic plugin, configured in the same cordis.patch.yml.

A memory entry looks like:

## lesson-abc123
- title: Verify the active profile
- mistake: Edited a different profile from the one used by Desktop
- prevention: Confirm the active profile before editing any profile files
- applies_when: DSH Desktop profile changes
- keywords: desktop profile package.json
- confirmed: true
- maturity: validated
- hits: 3

Install into DSH Desktop

Do not edit DSH Desktop security settings (mode, openBrowser, networkExposure, sandbox presets). Install through the normal DSH profile plugin mechanism, then restart DSH Desktop once.

From GitHub (recommended)

Add this to your DSH Desktop profile's package.json (the file at C:\Users\<you>\.dsh\profiles\desktop\package.json):

{
  "dependencies": {
    "dsh-error-improvement": "github:wbushihenshuai-design/dsh-error-improvement"
  }
}

Also add "dsh-error-improvement" to the dsh.profile.bundles array in the same file (and remove it from dsh.desktopDeselectedBundles if present).

Then run in the profile directory:

pnpm install

Restart DSH Desktop. The settings section appears under Settings → 错误改进, with a 待确认草稿 subsection for pending drafts.

From a local checkout (development)

{
  "dependencies": {
    "dsh-error-improvement": "link:D:/work/DS/dsh-error-improvement"
  }
}

Preserve all existing bundles and dependencies when merging. The plugin's cordis.patch.yml mounts exactly one Host row; its dsh.client metadata loads the settings-section UI.

Before installing into a live Desktop profile

• Run npm run ci in this repository.
• Make a backup of the profile's package.json, lockfile, and cordis.patch.yml.
• This repository does not include an auto-installer because silently rewriting a live profile is unsafe.

Build and verify

Requires Node.js 22.19+ or 24+.

npm install
npm run ci
npm run pack:check

The CI command performs formatting/lint checks, TypeScript checks, 32 Host/store/matching/enforcement/drafts/graduation tests, production build, safety preflight scan (all published JS files), and a clean-consumer pack smoke test that installs the tarball into a fresh node_modules and imports the real artifact.

Development layout

• src/index.ts — Host registration, injection waterfall, tool/RPC mounting.
• src/memory.ts — memory entry schema, Markdown parse/render, atomic writes.
• src/lessons.ts / src/recipes.ts — selection, sanitization, budget-measured rendering.
• src/capture.ts / src/distill.ts — zero-LLM capture and Curator→Writer distillation.
• src/drafts.ts — draft lifecycle (write/review/approve/reject).
• src/signals.ts / src/graduate.ts — turn signals, maturity, skill graduation.
• src/enforcement.ts — error statistics and pre-execution interception.
• src/store.ts — v2 state file (error signatures + promotion rules).
• src/client.js — immediately loaded settings-section UI (pending drafts + stats).
• test/ — 32 node:test cases.
• scripts/preflight.mjs — scans all published JS for forbidden coupling.
• scripts/pack-smoke.mjs — clean-consumer tarball install + import verification.

Privacy

No telemetry and no network requests beyond the LLM calls used for distillation (opt-out via capture.distillEnabled: false). Memory stays in local files under $DSH_HOME and is sent only as part of the local agent context when selected.

License

MIT © 2026 wbushihenshuai-design