@inobit/pi-retry

English | 中文

Manual transparent retry for Pi coding agent: /retry + alt+r re-issues the last failed turn as-is, injecting no prompt into the LLM context.

• Zero prompt injection: the retry trigger is a hidden sentinel message that a persistent filter strips from every LLM request — the model never sees it
• Keep everything as-is: the failed partial response stays as the last message of the context; the model continues seamlessly (prefill-style), no dropping, no regeneration
• Failure-agnostic: connection errors, stream drops, hung requests you aborted — one key re-sends them all

⚠️ Known limitation (read before use)

This extension must stay installed. The retry trigger permanently writes a sentinel entry into the session file (a custom message containing a single "."):

• With the extension installed: the sentinel is stripped by the context filter and never reaches the model; it is not rendered in the transcript either.
• If you uninstall/disable the extension and resume an old session containing sentinels: the filter is gone, and pi's built-in conversion turns each sentinel into a user message sent to the model — i.e. one "." noise message appears in the conversation. Harmless, but unavoidable.
• Sentinel entries cannot be deleted from the session file (pi exposes no public API for entry deletion); this is inherent to the design.

How it works

A hard constraint of pi's public API: every entry point that can trigger an LLM turn writes a message, and anything that writes nothing does not trigger one. A "transparent re-send" therefore has to be split in two halves:

① Trigger: sendMessage({ customType:"pi-retry", triggerTurn:true })
      └─ starts a new turn (the only no-user-content trigger in the public API); sentinel persisted
② Filter: persistent pi.on("context") handler
      └─ before every LLM call, right before messages are converted to provider format,
         strips exactly role==="custom" && customType==="pi-retry"

The interception happens at the AgentMessage[] layer (pi registers the extension context event as the global transformContext, which runs before convertToLlm and the HTTP request), so it is provider-independent. After filtering, the model receives the context ending exactly at the break point — the failed partial assistant is the last message, which is standard prefill shape; the model simply continues:

| | Failed partial assistant | Model behavior | |---|---|---| | pi built-in auto-retry | removed from context | regenerates from scratch | | typing "continue" | kept + followed by a user instruction | continuation (with instruction ambiguity) | | this extension | kept as last message, nothing added after it | continuation (zero instructions) |

Installation

pi install npm:@inobit/pi-retry

Restart Pi or run /reload.

Local dev (isolated, --no-extensions excludes installed old version):

pi -ne -e ./packages/pi-retry
# after a failure/abort: /retry or alt+r

Usage

• /retry — transparently re-issue the last turn; on success it notifies Retry submitted.
• alt+r — same as /retry.
• Failure guard: triggers only when the last assistant message ended with error/aborted; otherwise notifies Nothing to retry — last turn ended normally (or Nothing to retry yet on a fresh session). This prevents accidental presses from starting a free-form continuation and permanently writing a sentinel.
• Only available when idle; while streaming it notifies Agent is busy — retry is only available when idle.

Configuration

Shortcut is configurable via ~/.pi/agent/extensions/pi-retry/config.json (requires /reload):

{
  "shortcut": "alt+r"
}

Default alt+r. When trusted, project config at .pi/extensions/pi-retry/config.json overrides global.

Compatibility & Limitations

• The extension must stay installed (see "Known limitation" above): resuming an old session without it leaks one "." to the model. Same family of paths: compaction / branch-summary inputs bypass the filter, so a sentinel inside the summarized range may surface as "." inside the summary text.
• Interaction with built-in auto-retry: if the re-issued turn fails again with a whitelisted retryable error, pi core takes over and deletes the partial response, regenerating from scratch (the prefill continuation guarantee only holds for the first re-issue).
• A trailing assistant message is natively supported by Anthropic (prefill); some OpenAI-compatible endpoints may behave differently — test with your providers.
• If the stream happened to end exactly at a complete sentence, the model may start a new paragraph instead of continuing — inherent to continuation semantics, same as typing "continue".

Development

pnpm --filter @inobit/pi-retry check   # tsc --noEmit
pnpm --filter @inobit/pi-retry test    # vitest
pnpm --filter @inobit/pi-retry pack:check