haozheou/dsh-discipline-guard

Your agent burned through a weekly token quota in 3 minutes — while politely telling you it was "almost done". This plugin stops it at the harness boundary, where text-based reminders can't reach.

Prompt-level "please don't repeat yourself" reminders are advisory: a determined (or confused) model walks straight through them. dsh-discipline-guard is not a prompt. It hooks the harness waterfall (tools/pre-execute, agent/pre-step, session/event) and refuses dispatches, blocks turns, and pops native approval cards — the model cannot argue with the tool pipeline never running.

The four gates

#	Gate	Default	What it does
①	Loop breaker	on	Counts identical tool+arguments calls per agent. 2nd repetition → visible warning injected; 4th → the call is denied before dispatch (2 consecutive failures trips it even faster). Any new user message releases it.
②	Cost fuse	on, fail-closed	Meters every step from real usage data (input tokens net of cache, cache-hit rate, per-turn cumulative). Typical token-plan pattern — huge full-price resend with ~0% cache hits — gets a warning first; a repeat violation in the same turn hard-stops the turn. Reply 继续/放行/别停/照常 (or continue) to release one turn.
③	Route watcher	on	When the session switches model endpoints with a long history, injects a one-time warning: the new endpoint may re-bill your whole context at full price.
④	Plan gate	off	Opt-in: the first write/execute tool call of each user turn routes through the native approval card. One approval covers the turn.

Every gate is individually switchable with its own thresholds, and every listener is fail-open internally: a bug in the guard can never brick the harness it protects. Money-class gates are fail-closed by design: they stop and wait for you.

Control surfaces

• Main UI: a 🛡 chip in the session header — one click arms/disarms everything (the escape hatch must never be buried).
• Settings → Discipline Guard（纪律卫士）: per-gate switches, thresholds, and three presets:
  • off — everything disabled
  • daily (default) — tuned for cached routes (deepseek-style: 1–5K net input/step will never trip it) while capping an uncached runaway at ~500K tokens/turn instead of millions
  • wild — warns but never interrupts; for experimental heavy work
• Settings namespace discipline-guard in ~/.dsh/settings.yaml, hot-reloaded — no restart needed after the first install.

Typical release flow (the guard never decides for you)

gate trips  →  you look  →  worth it?  →  reply 「继续」  →  the rest of the turn runs free
                                         not worth it? →  you stopped a fire in one turn

One nuance on purpose: replying anything resets the loop breaker; the release words only waive the current turn. The next turn starts a fresh ledger — the guard never accumulates grudges, and never quietly forgets.

Install

Requires DSH with the web profile (client UI is web-only for now; the host gates work everywhere).

A. Via plugin market (once listed)

Open DSH → Plugins market → search "discipline guard" → install.

B. Manual install

Add the plugin to your profile (~/.dsh/profiles/<profile>/package.json):

{
  "dependencies": {
    "dsh-discipline-guard": "github:haozheou/dsh-discipline-guard"
  },
  "dsh": {
    "profile": {
      "bundles": [
        // ... your existing bundles ...
        "dsh-discipline-guard"   // ← add this line
      ]
    }
  }
}

Then pnpm install in the profile directory and restart DSH. The bundle's own cordis.patch.yml mounts the plugin row automatically.

Thresholds: how to tune

Measured baselines (per step, net-billable input):

• cached route (deepseek-style): 1–5K tokens, ~97% cache hits → daily defaults never fire
• uncached "token plan" route: the entire context re-billed every step → that is what ② exists to cap

If your workload…	adjust
gets stopped while doing legitimate long work	raise costTurnBudget (500K → 800K+)
runs on a route where caching is broken	that is exactly the pattern ② exists to cap — check the context size, not the thresholds
polls tools on purpose (job waiting etc.)	add them to loopExclude

The incident replay test: 16 steps × 360K full-price resend = 5.74M tokens in 3 minutes → with daily defaults, execution stops at ~550K (≈90% saved), mid-run, with a visible reason.

One switch: "cached" vs "no-cache" route

The cost fuse normally infers an uncached full-price resend from the measured cache-hit rate (auto). If you know which endpoint you're on, you can tell the guard directly and it evaluates more precisely, without guessing from numbers:

• cached (has cache) — e.g. deepseek official with prompt caching. The fuse no longer flags a single low cache-hit step (cold-start / brand-new long topic can also hit low), keeping only the per-turn budget hard cap. This is the "never fires" daily baseline.
• nocache (no cache) — e.g. token-plan / "套餐" endpoints that re-bill the whole context each step. The fuse gets strict: any step with net input ≥ costRequestInputMax and context ≥ costContextMin counts as a full-price resend, with no cache-hit-rate precondition.
• auto (default) — keep the original inference based on cache-hit rate.

Find it under Settings → Discipline Guard → ② Cost fuse → 当前路由缓存判定 (three-way toggle). Hot-reloaded, no restart.

Caution: labeling a truly nocache route as cached removes the full-price-resend guard (only the turn budget remains as a backstop). Mark routes truthfully.

Known limitations

• Client UI (chip + settings page) targets the web profile; desktop profile gets host gates only (config via settings.yaml directly).
• Cost metering is per-turn/per-session; parallel sub-agents' usage is metered per agent, session aggregate in v2.
• Coexists with the built-in repeat-tool-reminder (advisory at 3/5/8 repetitions). If the double notices annoy you, disable the built-in one.

Credits

• Loop-chain canonicalization pattern borrowed (in spirit) from DSH's built-in repeat-tool-reminder.
• Built on Cordis / DeepSeek Harness plugin APIs: tools/pre-execute, tools/post-execute, agent/pre-step, session/event, approval, dsh-settings.

MIT © haozheou

---

中文文档见 README.zh.md。