[bug] Auto-retry backoff and max attempts are hardcoded, causing premature agent stops on shared gateways

What happens

When using jcode with a shared LLM gateway (for example, a proxy that many people use at the same time), the agent stops after only 3 retries. The backoff is 2s, 4s, 6s. This is too fast for a shared service. The gateway often returns 429 Too Many Requests without a retry-after header. The agent runs out of retries before the gateway has time to free capacity, so the turn dies.

Root cause: hardcoded retry constants

The retry parameters are hardcoded in crates/jcode-tui/src/tui/app.rs:

const AUTO_RETRY_BASE_DELAY_SECS: u64 = 2;
const AUTO_RETRY_MAX_ATTEMPTS: u8 = 3;

This is a bug. These values are fixed in the binary and you cannot change them. Other retry settings like stream_idle_timeout_secs and auto_poke are configurable in config.toml or with env vars, but the retry budget is not. There is no way for a user to adjust the backoff or the number of retries for their provider.

When the gateway returns 429 without a parseable retry-after or reset time, jcode uses these hardcoded values. Three attempts at 2s/4s/6s is not enough for a shared service under load. The agent gives up on a temporary error that would recover if it waited longer.

Impact

Any user on a shared LLM gateway will see the agent stop in the middle of a task when the gateway is briefly busy. Since auto_poke keeps the agent working on its own, a premature retry failure kills the autonomous loop silently. The user does not notice until they check the session.

What this is NOT

This is not about auto_poke itself. Auto-poke is already configurable with [features] auto_poke = true/false. The bug is that the retry budget behind auto-poke is hardcoded and cannot be tuned.

The fix

Branch: fix/auto-retry-config on my fork.

Makes both values configurable. You can set a global default in [provider] and optionally override it per provider in [providers.<name>]. Env vars also work.

# Global defaults
[provider]
auto_retry_base_delay_secs = 2
auto_retry_max_attempts = 3

# Per-provider override (used when that provider is active)
[providers.my-shared-gateway]
auto_retry_base_delay_secs = 10
auto_retry_max_attempts = 10

# Env var overrides (highest priority)
export JCODE_AUTO_RETRY_BASE_DELAY_SECS=10
export JCODE_AUTO_RETRY_MAX_ATTEMPTS=10

Priority: env var > per-provider > global default

The backoff is linear: base_delay * attempt_number. The per-provider value is checked at retry time from the active session provider key. If the provider does not set it, the global value is used.

If the gateway provides a retry-after time in the error message, jcode uses that instead of the linear backoff.

Documentation

The fix also adds a new section to the README (Auto-retry backoff) explaining the configuration, environment variables, and precedence. The default config template in crates/jcode-base/src/config/default_file.rs also documents both keys with comments.

Files changed (7 files, +135/-6)

• crates/jcode-config-types/src/lib.rs — new fields on ProviderConfig (global) and NamedProviderConfig (per-provider override)
• crates/jcode-base/src/config/env_overrides.rs — env var parsing
• crates/jcode-base/src/config/default_file.rs — documentation in default config template
• crates/jcode-tui/src/tui/app.rs — runtime fields on App (replaces hardcoded consts)
• crates/jcode-tui/src/tui/app/tui_lifecycle.rs — effective_auto_retry_*() methods that check the active provider override at retry time, falling back to the global value
• crates/jcode-tui/src/tui/app/tests/remote_events_reload_04.rs — update test references
• README.md — new Auto-retry backoff section with configuration, env vars, and precedence docs

Start with the hardcoded retry constants in crates/jcode-tui/src/tui/app.rs, then trace configuration through crates/jcode-config-types/src/lib.rs and crates/jcode-base/src/config/env_overrides.rs. Review the lifecycle methods and remote-events test named in the issue, along with the default template and README. Done means global, per-provider, and environment settings work with the stated precedence and documentation matches.