feat: auto-continue after server overload

[AVas112]

What changed

• The TUI now automatically submits Continue after a ServerOverloaded error.
• Auto-continuation runs for every overload event and is not blocked by the quota pause state.
• Recovery state for ServerOverloaded is kept separate from UsageLimitExceeded, preserving the existing usage-limit behavior.
• Duplicate Continue entries are prevented by deduplicating the locally rendered prompt and its app-server echo using both turn_id and message content.
• The same echo deduplication behavior is preserved for safety-buffering retries.
• Added a configuration toggle for enabling or disabling overload auto-continuation.

Why

When the selected model is overloaded, the current turn ends with Selected model is at capacity. Previously, the user had to submit a continuation manually. Codext now starts the next turn automatically so work can continue without manual intervention.

Configuration

Auto-continuation is enabled by default. To disable it, add the following to config.toml:

[tui]
server_overloaded_resume_enabled = false

Validation

• cargo build -p codex-cli
• cargo build --release -p codex-cli
• git diff --check
• Project linters
• Architecture review completed with no findings
• Configuration schema regenerated

Tests and snapshot files were not changed, in accordance with the temporary rust-v0.145.0 reapply guardrails.

[Loongphy]

Thanks for your pr, I'll check this later.

[Loongphy]

The staged Continue is submitted immediately after the failing turn ends. If the model stays overloaded, this becomes an infinite immediate-retry loop that hammers the API and floods the transcript with error / Continue pairs.

   submit → immediate overload error → auto-Continue → immediate overload error → …

Suggested fix (fine as a follow-up): add a delay and a retry budget — e.g. wait 30s before the first retry and increase the delay on each subsequent attempt (30s, 60s, 120s, …), then stop after ~5 attempts and leave the error on screen for the user. Reset the counter once a turn completes successfully.

[AVas112]

@Loongphy
Fixed in the latest commit. Server-overload auto-Continue no longer submits immediately. It now waits 15 seconds before the first retry, followed by 30, 60, 120, and 240 seconds, with a budget of up to five automatic retries. After the retry budget is exhausted, the overload error remains visible and no further Continue messages are submitted. The retry counter is reset after a successful turn.

I chose 15 seconds for the first retry because transient overloads often clear quickly, and an immediate retry frequently succeeds. Waiting the full 30 seconds before the first attempt would add unnecessary latency in those cases.