dsh-plugin-tool-health
Advanced tools
Remembers which DeepSeek Harness tools have been failing, across sessions, and tells the next session — so an agent stops rediscovering a dead endpoint one failed call at a time.
English | 中文
Remembers which tools have been failing — across sessions — and tells the next session before it starts work.
An agent that hits a dead endpoint learns it the expensive way: one failed call at a time, in the middle of a task, often several sessions in a row. This plugin turns that into something the model is told up front.
dsh plugin --profile web add dsh-plugin-tool-health
No configuration needed. It registers no tools and is invisible to the model except for one prompt paragraph, which is empty while everything works.
Listens on tools/result — an emit-mode observation point, so it cannot
change what a tool returns — and keeps one record per tool: total calls,
failures, the current consecutive-failure streak, the last error, and when the
tool last succeeded. Records persist through the storage domain, so they
outlive the session.
When a tool has failed repeatedly and recently, the next prompt carries:
Recent tool failures (observed in this and earlier sessions, newest first):
- astock_market_quotes: 4 consecutive failures, latest 2 分钟前 (last succeeded 3 小时前)
last error: UND_ERR_SOCKET: fetch failed
This is evidence from past calls, not a prohibition: the cause may have cleared.
Prefer a working alternative when one exists, and if you do call one of these and it
fails again, tell the user the source is unavailable — do not substitute or invent data.
Consecutive failures, not a failure rate. A tool that fails one call in fifty is healthy; one that has failed its last three is not. A ratio cannot tell those apart, and it only gets worse the longer a since-recovered outage stays in the denominator.
Failures expire. An outage from last week says nothing about this session,
and warning about it would teach the model to avoid a tool that works. The
default memory is 24 hours (forgetAfterHours).
Silence when healthy. The report is empty text, which contributes nothing to the prompt. A standing "all tools healthy" banner would spend tokens on every request to say nothing.
| Key | Default | Meaning |
|---|---|---|
unhealthyAfter | 2 | Consecutive failures before a tool is reported |
forgetAfterHours | 24 | How old a failure may be and still be mentioned |
maxTools | 200 | Bound on the store; least recently touched are dropped |
maxListed | 8 | How many tools one report may name |
Persistence is optional: with no storage backend composed, the plugin still works within a single session rather than refusing to load.
MIT
FAQs
Remembers which DeepSeek Harness tools have been failing, across sessions, and tells the next session — so an agent stops rediscovering a dead endpoint one failed call at a time.
We found that dsh-plugin-tool-health demonstrated a healthy version release cadence and project activity because the last version was released less than a year ago. It has 1 open source maintainer collaborating on the project.
Research
/Security News
Ten malicious OpenAPI React Query Codegen versions were published to npm in the Mini Shai-Hulud attack, all with valid provenance.
Security News
Socket joins more than 100 technology, cybersecurity, and financial organizations calling for a global surge in cyber defense.
Product
Enterprise security teams can now detect malware, credential theft, suspicious network activity, and risky updates across Microsoft Edge extensions.