Copilot CLI should retry on transient CAPIError 400 Bad Request during agentic workflow execution

[pholleran]

Summary

When running the Copilot CLI inside a gh-aw agentic workflow, the CLI fails with a non-retried CAPIError: 400 Bad Request from the Copilot inference API (CAPI) mid-session. The CLI (or the gh-aw execution wrapper) should implement retry logic for transient 400 errors that occur after the session has already completed successful turns.

Reproduction

• Workflow: repo-assist.lock.yml (Repo Assist)
• Repository: octodemo/octocat_supply-glorious-meme
• Failed run: https://github.com/octodemo/octocat_supply-glorious-meme/actions/runs/24136873717/job/70427441392
• Request ID: C818:3ED713:19D401B:1C446B7:69D653CA
• Trigger: workflow_dispatch on main branch
• CLI invocation: /usr/local/bin/copilot --add-dir ... --prompt "$(cat /tmp/gh-aw/aw-prompts/prompt.txt)"
• Model: claude-sonnet-4.6 (444.6k input, 3.2k output, 373.6k cached)
• Date: 2026-04-08

Error

Execution failed: CAPIError: 400 400 Bad Request
 (Request ID: C818:3ED713:19D401B:1C446B7:69D653CA)

The CLI had been running normally for ~3.5 minutes — completing 8+ tool calls (reading files, listing issues, exploring the codebase) — when CAPI returned this 400 error. The CLI exited with code 1, failing the workflow.

Expected Behavior

A 400 that occurs mid-conversation (after multiple successful inference turns) is likely transient rather than a genuinely malformed request. The CLI should:

1. Detect CAPIError: 400 during an active session (after at least one successful turn)
2. Retry the failed inference request with exponential backoff (e.g., 3-5 attempts)
3. Only fail the session if retries are exhausted

Context

• The conclusion job confirmed Inference access error: false — this was not an auth issue
• No code changes were produced (+0 -0), so the entire premium request was wasted
• The workflow auto-created issue octodemo/octocat_supply-glorious-meme#33 to track the failure
• Related: Agent fails with "Failed to get response from the AI model; retried 5 times" after noop safe-output #23265 (agent retries on model response failures — similar pattern, different error class)

[pelikhan]

@copilot create a JavaScript driver to run copilot with an outer that handles this kind of error and retries with the resume flag.
The drivers passes all the arguments to the copilot sub process, transparently standard io streams. Recovers from transient errors.

[pelikhan]

The agentic engine interface should have a function that provides an optional JavaScript driver.