@deepseek-ai/dsh-loop-doctor-tool

中文 | English

The self-referential model-facing tool of the loop-doctor family: loop_doctor_diagnose.

Calling it without a sessionId diagnoses the calling agent's own session — the harness diagnosing its own harness. It runs the deterministic engine over the session's SessionEvent log, derives suggestions, verifies each by replay, and returns a compact JSON report:

{
  "sessionId": "session-7",
  "summary": "3 detour(s), 3 suggestion(s); apply only after human review.",
  "signals": [
    { "kind": "retry-storm", "sub": "repeat:grep", "turn": 1, "metric": 4,
      "metricUnit": "calls", "detail": "4 consecutive identical grep calls …",
      "evidenceCount": 4, "evidenceSeqs": [2, 3, 4, 5] }
  ],
  "suggestions": [
    { "id": "retry-storm:1:repeat:grep", "signalKind": "retry-storm",
      "title": "Stop repeating identical calls", "target": "prompt",
      "action": "Add a system-prompt rule …", "confidence": 0.8,
      "fixPatternId": "retry-storm-repeat",
      "replay": { "before": 3, "after": 2, "avoided": 1, "verified": true,
                  "simulated": false, "unit": "calls", "detail": "…" } }
  ]
}

Evidence is projected as seqs only — locate the events in session.events to audit a finding; the full evidence never rides into the model surface.
exec.signal is honored, unknown sessions fail loud, and no change is ever applied — the tool returns a report for review.
Config forwards detector thresholds to the on-demand analysis.

See the family README for the demo and safety baseline.

Model Experience

The loop_doctor_diagnose report

What the model sees

A compact multi-line report in the tool result: one line per detour signal with its evidence count, and one line per suggestion with target, confidence, action, and an honest badge. Evidence rides as seqs only — full events never enter the model surface.

Report as the model receives it

Loop-doctor diagnosis for session main-session-…:
2 detour(s), 2 suggestion(s) (1 replay-verified); apply only after human review.
- [retry-storm/repeat:grep] turn 1: 4 consecutive identical grep calls (4 evidence events)
- suggestion [retry-storm-repeat] (prompt, 80%): Stop repeating identical calls — … [replay-verified]

Token effect

The report scales with detour count (a few lines per signal/suggestion); a clean session returns a one-line summary.

KV Cache effect

The tool result appends after the reusable request prefix; nothing before it changes, so existing KV-cache entries stay valid.

Known Limitations and Deferred Work

Read-only. The tool returns a report; applying a fix is the apply plugin's job behind human review. A review-approval tool is deferred.
Whole-session analysis. No turn-range or time-window scoping yet; long sessions analyze in one pass (bounded by the detector's evidence caps).
No fix application from the tool. The model cannot self-apply; this is the safety boundary, kept deliberately.

@hubert-hwk/dsh-loop-doctor-tool

Install

README