Retarget persistent runtime to Codex app-server

[jeremyplichta]

Summary

• retarget the persistent tt agent-loop runtime to codex app-server instead of the earlier Claude-first streaming adapter
• keep persistent mode Codex-first (codex, codex-mini) and fall unsupported CLIs back to the existing one-shot loop
• map Codex JSON-RPC thread/turn lifecycle into Tinytown runtime events, session persistence, urgent mid-turn steering, and interrupt handling
• replace the Claude-specific fake streaming harness with a fake Codex app-server and refresh the focused persistent-runtime integration coverage

Testing

• cargo test --no-run
• target/debug/deps/integration_tests-9ee69f1b1d41bbad test_agent_loop_persistent_runtime_ --nocapture
• cargo test --test integration_tests test_agent_persistent_config_parse -- --nocapture
• cargo test --test integration_tests test_persistent_agent_loop_exits_cleanly_after_idle_timeout -- --nocapture

Closes #83.

---

Note

Medium Risk
Introduces a new persistent subprocess execution path for tt agent-loop (including interruption/restart behavior) and stores new state in Redis, which could affect worker reliability and message handling if edge cases are missed.

Overview
Adds an opt-in persistent streaming runtime for tt agent-loop ([agent].persistent = true) that keeps a long-lived coding CLI session across turns (currently implemented for codex/codex-mini via codex app-server), while preserving the existing one-shot subprocess-per-turn default.

Agent state now persists a runtime_session_id in Redis to resume streaming sessions after restarts, and the persistent loop is driven by structured runtime events (turn started/completed, tool calls, deltas, awaiting input, runtime errors) which are also emitted onto the per-agent Redis event stream. Documentation is updated to describe the runtime model, and integration tests add coverage for persistent reuse, urgent mid-turn injection, stop interruption, idle-timeout behavior, and event emission.

^{Reviewed by Cursor Bugbot for commit ba5dd39. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, have a team admin enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit ba5dd39. Configure here.}

[cursor]

Persistent backlog section missing overflow and guard notices

Medium Severity

The persistent loop's backlog_section is missing two pieces of logic present in the one-shot version: the overflow notice when total_matching > tasks.len() ("...and X more role-matching backlog task(s)") and the guard message when total_backlog > 0 but no tasks match the role ("No backlog tasks currently match your role hint. Do not claim unrelated work by default."). Without the guard message, persistent-mode agents may claim unrelated backlog tasks.

Additional Locations (1)

• src/main.rs#L3947-L3980

 

^{Reviewed by Cursor Bugbot for commit ba5dd39. Configure here.}

[cursor]

Raw string literal contains spurious backslash-escaped quotes

Low Severity

Line 878 uses \"what was done\" inside a r#"..."# raw string literal. In raw strings, backslashes are not escape characters, so the output contains literal \" characters. The one-shot prompt correctly uses unescaped "what was done". The persistent agent sees garbled instructions with stray backslashes.

 

^{Reviewed by Cursor Bugbot for commit ba5dd39. Configure here.}