Summary

Pi interactive session can crash with an uncaught ERR_HTTP2_INVALID_SESSION from undici HTTP/2 transport during a long-running Objective Loop session.

This appears to be a provider/transport-level crash because the stack trace is entirely inside Node/undici/Pi package dependencies, not the Objective Loop extension.

Environment

• Pi package: @earendil-works/pi-coding-agent 0.75.1
• Pi CLI: pi --version => 0.75.1
• Node: v22.16.0
• OS: Ubuntu 24.04 under WSL
• Model/provider in TUI: openai-codex / gpt-5.5 low
• Invocation path: pi-objective-loop wrapper loading Objective Loop extension, but stack has no extension frames

What happened

While pi-objective-loop was running an interactive session, after several approved iterations, Pi exited with an uncaught exception:

pi exiting due to uncaughtException:
Error [ERR_HTTP2_INVALID_SESSION]: The session has been destroyed
    at ClientHttp2Session.request (node:internal/http2/core:1766:13)
    at requestStream (/home/ubuntuadmin/.nvm/versions/node/v22.16.0/lib/node_modules/@earendil-works/pi-coding-agent/node_modules/undici/lib/dispatcher/client-h2.js:737:22)
    at writeH2 (/home/ubuntuadmin/.nvm/versions/node/v22.16.0/lib/node_modules/@earendil-works/pi-coding-agent/node_modules/undici/lib/dispatcher/client-h2.js:925:12)
    at Object.write (/home/ubuntuadmin/.nvm/versions/node/v22.16.0/lib/node_modules/@earendil-works/pi-coding-agent/node_modules/undici/lib/dispatcher/client-h2.js:210:14)
    at _resume (/home/ubuntuadmin/.nvm/versions/node/v22.16.0/lib/node_modules/@earendil-works/pi-coding-agent/node_modules/undici/lib/dispatcher/client.js:658:50)
    at resume (/home/ubuntuadmin/.nvm/versions/node/v22.16.0/lib/node_modules/@earendil-works/pi-coding-agent/node_modules/undici/lib/dispatcher/client.js:576:3)
    at /home/ubuntuadmin/.nvm/versions/node/v22.16.0/lib/node_modules/@earendil-works/pi-coding-agent/node_modules/undici/lib/dispatcher/client.js:350:28
    at node:internal/process/task_queues:151:7
    at AsyncResource.runInAsyncScope (node:async_hooks:214:14)
    at AsyncResource.runMicrotask (node:internal/process/task_queues:148:8) {
  code: 'ERR_HTTP2_INVALID_SESSION'
}

The wrapper then reported:

Objective Loop run ended: 20260518T074903Z-118215 (exit 1)
Analyze: objective-improve-last --run-id 20260518T074903Z-118215

Expected behavior

A destroyed HTTP/2 session should not crash the Pi process as an uncaught exception. Ideally Pi/provider transport should either:

• retry once with a fresh HTTP/2 session when safe, or
• surface a recoverable provider error to the agent/session, or
• gracefully end the current request with a useful error message and preserve session state.

Actual behavior

The whole Pi process exited from an uncaught exception.

Notes

Because this happened inside a long-running autonomous Objective Loop session, the impact is high: a transient provider HTTP/2 session failure terminates the run. Objective Loop can potentially add wrapper-level recovery, but the root crash looks like Pi/undici/provider transport should guard this error.

Please let me know if there is a debug flag or provider transport setting that would help capture more detail if it recurs.