dsh-fail-guard

dsh-fail-guard

English | 中文

Stops the agent when it keeps failing on the same tool: after N consecutive failures it pushes a four-part stop instruction (what it tried / the full error text / the first guess at why / what you need to do) so it reports instead of trying a fourth way to spell the same command.

Why it has to come from outside

Repeated short lines in reasoning ("Do it. Do it.") are usually not model degradation — they are the residue of consecutive failure: the same thing fails, the parameters change, it fails again, nothing new comes in, so the same "time to act" decision gets re-made over and over.

Prompt-level rules ("stop after two consecutive failures") do not work in that state — the model is not reading the prompt then, it is reading its own thinking. The interruption has to come from outside.

Measured on a 74-turn session: turns 1–67 kept the share of short single-line fragments between 7% and 48%; the moment a task hit a dozen consecutive failed steps it jumped to 81.6%, then 85.9% / 89.1% / 91.4%.

Install

dsh plugin --profile desktop add dsh-fail-guard

Restart once. cordis.patch.yml is applied through the dsh.bundle manifest — no need to edit your profile config.

Configuration

Edit it in Plugins → dsh-fail-guard → details. Fields are marked .volatile(): saving takes effect immediately, no restart.

Key	Default	Meaning
threshold	2	Consecutive failures before stopping
sameToolOnly	true	Count per tool; turn off to share one counter across all tools
stopText	built in	The instruction pushed into the conversation
stopSummary	built in	The one-line summary shown in the panel
debug	false	Write events, counts and trips to a log file
logFile	<plugin dir>/fail-guard.log	Where debug writes

Or set it directly in cordis.patch.yml:

- id: dsh-fail-guard
  config:
    threshold: 3

What counts as a failure

A tool/result counts if either holds: message.isError === true, or meta.opsShell.exitCode exists and is non-zero. Both are needed — a shell command exiting 1 is not marked isError, and some tool-level errors carry no exit code. Counters are per session and per tool, and any success resets them.

Limits

• It does not abort the call that just failed — that already happened
• It does not judge whether a failure is "real"; any non-zero exit counts
• It does not reset the conversation; the stop instruction is just one more user message
• It is not a rate limiter: after a trip, further failures start counting again and will trip again at the threshold

License

MIT