Retry Count Controller

Adjust your LLM provider's retry limit from the chat box.

English | 中文

A slash command that lets you change how many times your provider retries a failed request — or switch to unlimited retries. No restart, no config file hunting — just type /retry-count and you're done.

In the screenshot above, the retry message now reads "已重试模型请求(1/20)" — retry 1 of 20, instead of the default 2. You can see the current retry count and the remaining time before the next attempt.

Why you might want this

• You keep seeing "retry model request (2/2)" and want more attempts before giving up.
• You're watching your token usage and want to reduce unnecessary retries.
• You're testing provider behavior and want to crank retries up to the max.
• A flaky provider keeps dropping your request — set it to inf and it will keep retrying until it succeeds (or you interrupt it).
• You want to fail fast when something is broken — set it to 0 and move on.

Usage

Type the command in your DSH Web GUI chat input, just like any other slash command:

/retry-count <provider-name> <number|inf>

Examples

Command	What it does
/retry-count my-provider 5	Set max retries to 5
/retry-count my-provider 0	Disable retries entirely
/retry-count my-provider 20	Set max retries to 20 (maximum allowed)
/retry-count my-provider inf	Unlimited retries (always mode) — keep retrying until success or interruption (-1, infinite, infinity and ∞ work too)
/retry-count	Show usage help (no arguments)

After you run it

The plugin will respond with one of these messages:

Success (limited):

Provider "my-provider" retry policy updated to maxRetries=5. Changes take effect immediately.

Success (unlimited):

Provider "my-provider" retry policy updated to unlimited (always mode — retries until success or interruption). Changes take effect immediately.

No change needed (same value already set):

Provider "my-provider" already has maxRetries=5.
Provider "my-provider" already has unlimited (always mode — retries until success or interruption).

Provider not found:

Provider "whoami" is not registered. Available: my-provider, another-provider

Invalid number:

maxRetries must be an integer from 0 to 20, or "inf" for unlimited retries

What happens next

The next time your provider fails a model request, the retry count will use your new setting. If you set it to 5, you'll see:

已重试模型请求(1/5) · 1s
已重试模型请求(2/5) · 2s
...

With inf, retries never give up on their own — the counter keeps climbing (1/∞, 2/∞, ...) until the request succeeds or you stop it. The change is saved to settings.yaml and persists across DSH restarts.

With unlimited retries enabled, the retry counter shows "1/无限" — attempt 1 of infinity. The counter keeps climbing and each retry waits with exponential backoff (capped at 60s) until the request succeeds or you manually stop it.

⚠️ Unlimited retries can hammer a provider that is down. Every retry waits with exponential backoff (capped), and stopping the request cancels any in-flight retry — but a persistently failing provider will keep being retried indefinitely. Use inf deliberately and switch back to a finite number when you're done.

How it works

1. You type /retry-count my-provider 5 (or inf).
2. The plugin checks that the provider exists and the value is valid (0-20, or inf).
3. It writes the new setting to settings.yaml — inf stores an always-mode retry policy.
4. The pi-ai adapter picks up the change instantly and re-registers the provider route with the new retry policy.
5. Done. No restart, no reload.

Installation

# In your cordis.patch.yml
- insert:
    - id: retry-count
      name: '@deepseek-ai/dsh-command-retry-count'

The base layer already has everything else this plugin needs.

A note for the model

• The model never sees this command. It's between you and the settings file.
• It costs zero tokens.
• It doesn't affect the conversation cache or history.

Known limitations

• pi-ai only — this command works with providers configured through dsh-llm-pi-ai. If you're using the native DeepSeek adapter (dsh-llm-deepseek), the plugin won't work for you yet.
• Count or unlimited only — you can set a retry count (0-20) or unlimited (inf). Custom retryable error-code lists and fine-grained backoff tuning aren't exposed by this command.
• Needs a command adapter — surfaces without ctx.commands can't use it. The Web GUI has one built in.

License

MIT