DSH | Retry Guard | Pause after consecutive identical tool failures

[sjh9714]

Unofficial community plugin, independently maintained and not endorsed by DeepSeek.

Stop repeating the same failed tool call

When DSH keeps sending the same invalid edit, missing argument or failed MCP
request, Retry Guard can pause before the next model step after three
identical consecutive tool errors. Inspect the original failure, fix the cause,
and send a new instruction to continue.

GitHub and installation guide · MIT · DSH Web 0.2.0-rc.2 · Node.js 24

20-second recording.
This is real DSH Web with a scripted model, not a production AI session.

Install from npm

dsh plugin --profile web add dsh-retry-guard@alpha

Or open Plugins → Add plugin, enter dsh-retry-guard@alpha, then click
Install → Enable now. The command selects the current published package tag.
Restart Web after CLI installation. In the plugin's Configure screen, choose
Pause and save. Observe is the default: it shows notices while work continues.

What we checked

We inspected seven public user reports and ran 12 report-inspired scenarios
on the supported DSH version. In four matching patterns—missing description,
identical old/new edit strings, empty tool names, and the reported MCP error—our
bounded tests ran six calls without the guard and three in Pause. A related
invalid permission-escalation test also paused. The original old Bash same-mode
permission failure did not reproduce on this version.

The model is scripted, the MCP server is local, and original user sessions were
not replayed. These are reproduction counts, not a general cost-saving claim.
The full results, original reports and keyless reproduction
include the cases this guard misses.

Only identical consecutive DSH isError results count. Successful polling,
ordinary Bash exit-code failures, changing arguments, interleaved successes and
text-only loops are not stopped. Already-running parallel calls may finish.
Original results are preserved; failed calls are not automatically replayed.
There are other loop guards
if you need broader detection.

Building in public

Try it in your normal workflow and tell me what happened: a useful catch, a missed
failure, a false pause, or installation friction. Reply here or open a
GitHub issue.
Please keep private prompts, credentials and full session logs out of public feedback.

I'll use concrete reports to guide fixes and share the results. There is no
participant quota; actual installation and continued use are still unmeasured.
If this is useful to you, star the repository and follow its updates.