@mehdisadeghi/pi-retry0.12.0

---

---

Pi 1.0 compatibility (0.10.4)

Tested with Pi 1.0.0. Host-provided Pi packages and TypeBox are peers (*), not bundled runtime dependencies; development uses exact Pi 1.0.0 pins and host-compatible TypeBox where needed.

Verified the actual AgentSession.bindExtensions and _prepareRetry lifecycle, including child retry ownership, cancellation and native nested tools. No replacement retry dispatcher is introduced.

Run bun run test:host for the offline real-host load, native codemode/nested-call, module-identity and reload checks. Set PI1_HOST_PACKAGE to an installed Pi package directory to test that host explicitly; add PI1_HOST_ENTRY=bundle to check the bundled CLI runtime's constructors.

Overview

This extension automatically detects and retries all errors, forever, until you abort. There is no error blacklist and no failure cap: invalid API keys, missing models, quota/budget exhaustion and context overflow are retried like a socket hiccup.

Error Type	Retry Behavior	Use Case
Any error (catch-all)	Uncapped exponential retry	Everything, including unknown errors
HTTP 400/413	Uncapped with exponential backoff	Transient context overflow that might resolve
Credit / payment errors	Uncapped with exponential backoff	"Not Enough Credits", insufficient balance, 402 — top up and the retry loop auto-resumes
Quota / session-limit / budget exhaustion	Retried — labelled in /retry status	"You've hit your limit", insufficient_quota, "out of budget", suspended accounts
Invalid API key / unknown model	Retried — no exception	The key or model may be fixed mid-loop
Context overflow	Retried alongside SDK compaction	Oversized payload; either path can recover the turn
Connection errors	Uncapped with exponential backoff	Network hiccups, connection drops, socket errors, stream exhaustion
Max tokens (stopReason: "length")	Auto-continue indefinitely with hidden continuation turns	Model hits output token limit mid-generation
Empty / think-only stop (stopReason: "stop" with no text or tool calls)	Nudge once with a hidden continuation, then give up	Model ends its turn with no usable output (Anthropic empty responses with end_turn, thinking-only turns)

---

The Problem

By default, pi has built-in retry for some errors (rate limits, 5xx, overloaded), but:

1. 400/413 errors are treated as context overflow → triggers compaction but NO retry
2. Connection errors sometimes get only limited retries before giving up
3. Credit errors ("Not Enough Credits") are never retried
4. Stream exhaustion ("Max outbound streams") and other provider-specific errors are never retried
5. Any unknown error from a new provider is silently ignored

The Solution

This extension provides automatic retry for all errors with configurable exponential backoff (2s → 4s → 8s → ... → 60s by default) and keeps retrying at the maximum delay forever.

Philosophy: retry EVERYTHING, forever. Nothing is excluded — not an invalid API key, not a missing model, not an exhausted quota. Only a user abort (ESC), fresh input, or a session change stops the loop.

Features:

• Catch-all retry — Any stopReason: "error" is retried, regardless of error message
• Automatic detection of 400/413, connection, credit, and stream exhaustion errors
• Auto-continuation when the model hits its max output tokens (stopReason: "length") — indefinite, no cap, hidden from the TUI
• No cutoff — Keeps retrying at the maximum backoff until success, abort, or session change, with a notification explaining why
• Exponential backoff with configurable base delay, cap, and multiplier
• Hidden triggers — provider-valid custom messages use display: false, so retries do not add TUI clutter
• Manual controls via unified /retry command

---

Installation

Option 1: Install via pi package (Recommended)

Install directly from GitHub as a pi package:

pi install https://github.com/monotykamary/pi-retry

Or add to your settings.json:

{
  "packages": [
    "https://github.com/monotykamary/pi-retry"
  ]
}

Option 2: Global Installation

Copy the extension to pi's global extensions directory:

cp retry.ts ~/.pi/agent/extensions/

Option 3: Project-Local Installation

Copy to your project's .pi/extensions/ directory:

mkdir -p .pi/extensions
cp retry.ts .pi/extensions/

Option 4: Quick Test

pi -e ./retry.ts

---

Usage

Once loaded, the extension automatically detects and retries all errors.

Manual Controls

Command	Description
/retry	Manually trigger immediate retry (auto-detects: 400/413, credit, connection, max_tokens, or any other error)
/retry status	Show current retry diagnostics for all error types + continuation state
/retry reset	Reset all retry counters and state

---

Configuration

The extension reads a piRetry object from Pi's settings files:

• ~/.pi/agent/settings.json applies globally.
• .pi/settings.json overrides matching global values for the current project.

{
  "piRetry": {
    "baseDelayMs": 10000,
    "maxDelayMs": 3600000,
    "multiplier": 2
  }
}

The example above waits 10 seconds before the first retry, doubles each delay, caps the delay at one hour, and keeps retrying at that cap until you abort. Supported values are:

Setting	Default	Description
baseDelayMs	2000	Delay before the first retry, in milliseconds
maxDelayMs	60000	Maximum delay between retries, in milliseconds
multiplier	2	Exponential backoff multiplier; must be at least 1
messageFreeRetries	false	Retry errors without a hidden retry message (see below)

piRetry is separate from Pi's built-in retry object so the two retry policies do not share ambiguous settings. The extension disables Pi's native retry scheduler while it is loaded, while preserving Pi's compaction handling, so only one retry loop owns the backoff schedule.

Settings are read when the extension starts. Restart pi or use /reload after editing them.

Message-free error retries (opt-in)

By default, each error retry queues a hidden Retry the previous request. message, which stays in the model context and the session file. With "messageFreeRetries": true, an automatic error retry instead waits out the backoff inside the original prompt, then uses Pi's agent_before_settle boundary to omit the failed response with a context_edit entry and continue the unchanged context. The provider receives the same request again, and the omission persists when the session is resumed. The failed response stays in the raw session history.

• Max-token continuations, empty-stop nudges, and manual /retry still use hidden messages, because they need a new user turn to continue.
• Escape, new input, the backoff schedule, and the retry cap work as before. The prompt stays busy during the backoff, so input typed then is steered into the run and cancels the pending retry.
• Requires Pi with the agent_before_settle boundary (0.87.0 or later). On older Pi the extension warns once and keeps hidden retry messages.

---

How It Works

1. Listen to agent_end event — Fires after each agent turn completes
2. Check for any error — Examine the last assistant message for stopReason === "error"
3. Categorize for messaging — Classify into 400/413, credit, connection, quota, or other for /retry status diagnostics (labelling only; it never gates retrying)
4. Retry or continue with hidden turns — Wait with exponential backoff, then trigger a provider-valid custom user turn via pi.sendMessage() with display: false and triggerTurn: true (with messageFreeRetries, error retries omit the failed response and continue instead)
5. Valid provider context — Hidden retry and continuation messages remain in context so providers never receive a trailing assistant message
6. Indefinite continuation — Max_tokens auto-continues are uncapped; repeated length stops keep producing continuation turns until the model terminates normally
7. Empty-stop recovery — A stop turn with no usable output (only thinking, or nothing at all) gets exactly one hidden "nudge" continuation, then gives up. Matches Anthropic's documented "empty responses with end_turn" remedy (continuation prompt in a new user message) — retrying an empty response in place doesn't help because the model has already decided it's done.
8. Lifecycle exposure — Emits pi-retry:started, pi-retry:completed, and pi-retry:cancelled on Pi's shared extension event bus with a matching retryId, allowing status integrations to suppress intermediate completion signals

The pi's built-in transform-messages already strips aborted/errored assistant messages from the LLM context, so the model never sees the failed attempts.

---

Detected Error Patterns

Catch-All (Any Error)

• Any assistant message with stopReason === "error" is retried — nothing is excluded
• Unknown provider errors, stream errors, unexpected failures — all handled automatically

Retried Like Everything Else

These used to be treated as permanent. They are now retried too, because the cause can disappear mid-loop (a key gets refreshed, a quota window rolls over, the context gets compacted):

• Invalid API key / invalid authentication / revoked or missing key
• Model not found / unknown model / unsupported model
• Quota, session-limit and budget exhaustion — "You've hit your limit · resets …" (Claude Code, Codex, ChatGPT plans), insufficient_quota (OpenAI), "You have exhausted your capacity on this model" (Gemini Code Assist), free-models-per-day (OpenRouter), "out of budget" (LiteLLM-style proxies), suspended accounts (Kimi)
• Context overflow — "prompt is too long", "exceeds the context window", and friends

These messages are still labelled quota in /retry status, so an endless retry loop tells you whether you are waiting on billing or on the network.

Max Tokens (stopReason: "length")

• The model hit its max_tokens / output token limit
• The model's response was truncated mid-generation
• Auto-continuation sends a provider-valid custom message hidden from the TUI

400/413 Errors

• HTTP 400 Bad Request
• HTTP 413 Payload Too Large
• "bad request" messages
• "payload too large" messages

Credit / Payment Errors

• "Not Enough Credits"
• "insufficient credits"
• "insufficient balance"
• "out of credits"
• "Payment Required"
• HTTP 402 status code

Connection Errors

• Connection / network errors
• Fetch failures
• Socket hang up / socket errors
• ECONNRESET, ECONNREFUSED, ETIMEDOUT, ENOTFOUND
• DNS lookup failures
• "Request ended without sending any chunks"
• Upstream connect errors
• TLS handshake errors
• Timeouts awaiting response
• Stream exhaustion ("Max outbound streams is 100, 100 open")
• Stream limit errors

---

Development

Running Tests

# Run all tests
npm test

# Run tests in watch mode
npm run test:watch

# Run tests with coverage
npm run test:coverage

# Type check
npm run typecheck

# Dead code detection
npm run lint:dead

Project Structure

.
├── retry.ts                   # Main unified extension
├── src/                       # Shared utilities (testable, DRY)
│   ├── config.ts             # Settings-backed retry configuration
│   ├── error-patterns.ts      # Error pattern matching, custom types, hasMaxTokensStop
│   ├── retry-logic.ts         # Retry utilities (calculateDelay, RetryState, ContinuationState, etc.)
│   └── index.ts               # Barrel exports
├── __tests__/                 # Unit tests
│   └── unit/
│       ├── config.test.ts
│       ├── error-patterns.test.ts
│       └── retry-logic.test.ts
├── vitest.config.ts           # Test configuration
└── knip.json                  # Dead code detection config

Code Quality

# Run all quality checks
npm test              # 224 tests
npm run typecheck     # TypeScript type checking
npm run lint:dead     # Dead code detection with knip

---

Troubleshooting

Extension not working?

Check that it's loaded in the startup header:

Loaded extensions: retry.ts

Retry not triggering?

Use the status command to diagnose:

/retry status

Want to see what's happening?

The extensions send notifications on retry attempts. Look at the footer status line for retry status updates, and /retry status for the per-category attempt counters and the last error text.

Too many retries?

Use /retry reset to clear the counters, or press Ctrl+C to abort the session.

---

Comparison with @georgebashi/pi-retry

The npm package @georgebashi/pi-retry handles "aborted" streaming errors but explicitly excludes "connection error" (assuming pi's built-in retry handles it). This extension:

1. Handles ALL errors via a catch-all — no more playing whack-a-mole with new error patterns
2. Handles connection errors that pi might not retry sufficiently
3. Handles 400/413 errors without compaction
4. Handles credit errors and stream exhaustion

They can work together for maximum coverage:

pi install npm:@georgebashi/pi-retry
# Plus install this extension

---

Limitations

• Pi's native retry scheduler is disabled while this extension is loaded so native and extension retries cannot interleave; Pi's compaction check still runs normally
• Error messages remain in the session history (but are invisible to the LLM)
• Every error class is retried, so a permanent-looking failure (invalid API key, exhausted quota, genuinely oversized context) keeps firing at the maximum backoff until the cause is fixed
• Escape aborts the loop; fresh input or /new also stops it
• Warning: Retrying 400/413 without reducing context may fail repeatedly if the payload is genuinely too large — run /compact

---

Related

• Pi Coding Agent Extensions Docs
• @georgebashi/pi-retry — Handles "aborted" streaming errors
• Issue #252: Connection error with no retry

License

MIT

Subagent retry policies

Sessions that should retry with a different policy can be selected by matching their effective system prompt. This is user configuration: the extension ships no built-in marker or identity check for any specific subagent tool.

{
  "piRetry": {
    "subagents": {
      "enabled": true,
      "match": {
        "systemPromptRegex": [
          {
            "pattern": "^<active_agent name=\"[^\"\\r\\n]+\"/>$",
            "flags": "m"
          }
        ]
      },
      "baseDelayMs": 1000,
      "maxDelayMs": 10000
    }
  }
}

Semantics:

• Sessions whose effective system prompt matches any listed rule use the inline child policy; rules combine with OR.
• Omitted child fields inherit from the effective top-level policy, field by field. Project matcher lists replace global ones.
• A matching rule with enabled: false leaves pi's native retry scheduler untouched; a missing, empty, malformed, or nonmatching configuration keeps the ordinary pi-retry takeover.
• Unmatched sessions in print or JSON mode (pi -p, pi --mode json, and SDK hosts that bind extensions without a mode, such as pi-subagents children) retry inline with the top-level policy, inside the original prompt. These hosts read the result as soon as prompt() settles, so a detached retry would arrive too late. A matcher is only needed for a different child policy or for enabled: false.
• Patterns are compiled when settings resolve. Invalid syntax, unsupported or duplicate flags, and malformed groups warn and become no-match instead of matching everything.
• The extension must also be loaded in the target session (for pi-subagents that means listing it in the child's subagentOnlyExtensions), not only in the parent.

The example pattern above matches the <active_agent name="…"/> line that pi-subagents prefixes to native child system prompts; substitute your own marker if you select child sessions differently. Keep the m flag: some children have a <sub_agent_context> block before that line.