Retry errored agent run response

[shramee]

Summary

Interactive sessions for unattended tasks stop when a task crashes.

Expected Behavior

A failed or crashed turn in interactive mode should auto-retry on transient API errors.

Actual Behavior

The task crashes and work stops.

Steps to reproduce the issue

Run and wait until API errors, the task stops.

Command Code Version

Updated 1.32.1 → 1.32.2
1.32.2

Operating System

macOS

Terminal/IDE

VSCode

Shell

zsh

Session file (optional)

command-code-session-85c91527.txt

Fix prompt (optional)

No response

Additional context

OS: macOS 27.0

[ahmadawais]

PR-OPENED — fix up at CommandCodeAI/commandcode#4070 (trust score 95/100).

Your session file had the answer in it, thank you — both failures were 500 Provider returned error arriving after the turn had already streamed thinking. That routes into a separate "mid-stream restart" retry budget, and those restarts were spaced off the flat floor of the shared backoff schedule, so the whole budget burned in 3.0 seconds. The byte-identical error arriving before the first byte gets 35.2 seconds. Measured at the current release:

Failure arrives	Attempts	Window
after thinking deltas (your case)	4	3.0s
before first byte	10	35.2s

Since a reasoning model always emits thinking before it dies, the 3-second window was the one nearly every real failure landed in — so any provider blip lasting longer than that ended the turn and left you typing "continue". Restarts now span ~16s instead, without increasing the number of retried (and paid-for) prompts.

One thing your report asks for that this PR does not do: a provider outage lasting minutes will still end the turn. Auto-resuming a fully-failed turn needs a policy decision (retry cap, what the UI shows, how it interacts with /goal), so I left it out rather than widening the fix silently — worth its own issue if that's the behavior you're after.

---

Generated by Claude Code