dsh-loop-recover

中文文档

A DeepSeek Harness plugin that rescues agents stuck in repetition loops:

detect → prune → compact → resume — in-place, no session forking.

Why

Long-running local models (Ornith, Qwen, Sage…) on long contexts degrade into repetition: sssssss char spam, abc×N literal loops, or semantic spins where reasoning repeats the same sentences in different words. Most loop plugins only interrupt — the task dies.

dsh-loop-recover goes further:

• the loop tail never lands in the session log (the block-end payload is rewritten mid-stream)
• the poisoned context is physically compacted — a checkpoint summary replaces all earlier turns via surfaceOp: replace
• the agent gets a corrective follow-up and continues the same task — verified end-to-end on real sessions (turn/end → completed)

Detection layers

Layer	Catches
periodic	literal abc123×N loops
ngram	64-char blocks recurring 3+× (reworded spins)
anchor	large paragraph rewrite
charrun	sssssss collapse — substantive char run ≥128 (---- won't trigger)
pool	short sentence repeated across the stream (real-session verified)
intent	reworded same-topic web_search queries → steer correction

Modes

mode	Behavior
rescue (default)	cut + prune + compact + auto-resume
cut	cut + prune, task stops
observe	log detections only — run first to check false positives

maxConsecutiveRescues (default 3) bounds auto-rescue; normal output resets the counter. Optional retryTemperature/retryReasoningEffort overrides apply to resumed requests.

Install

# desktop app runs the `desktop` profile (NOT `web`)
dsh plugin --profile desktop add github:terry2010/dsh-loop-recover

Then restart the app. The plugin appears in the installed list as loop-recover.

Mutually exclusive with dsh-agent-loop-guard / @mrweicodes/dsh-loop-guard / dsh-client-auto-continue — they wrap llm/stream or auto-resume and would double-intervene.

Config (cordis.patch.yml)

- id: loop-recover
  config:
    mode: rescue              # rescue | cut | observe
    watch: both               # both | reasoning | text
    maxConsecutiveRescues: 3
    retryTemperature: 1.0     # sampling override on resumed requests
    retryReasoningEffort: low
    retryRequests: 1
    compactOnRescue: true     # shadow old context before resuming
    detector:
      runMinChars: 128        # single-char spam threshold
      ngramSize: 64
      anchorSize: 512
      poolMinRepeats: 4       # scattered-sentence repeat threshold
    intentEnabled: true

Development

npm install          # deps: @deepseek-ai/schemastery, dsh-llm, dsh-compaction
npm test             # 46 node:test unit + integration tests
npm run check        # plugin load smoke test

Real-session regression fixtures live in test/fixtures-real-loop.txt (an actual reasoning-loop capture from an Ornith long-context run).

License

MIT. Vendored/adapted from: dsh-loop-detector (viego-qiin) · dsh-loop-rescue (citydirector) · rice-patrol (LeeLizuoLiu)