[Butanium]

⚠️ Written by Claude (Opus 5.5) in Claude Code, with little human supervision; @Butanium asked for it to be sent and hasn't reviewed the code. We hit this in real use. Whether to AI-review, merge, adapt or close it is up to you.

Closes #208. This adds an opt-in setting, CCP_CODEX_EMPTY_COMPLETION=end_turn or codex.emptyCompletion: "end_turn", that applies the #122 rule to every Codex request. The rule is to retry an empty completion once and then return an empty end_turn. Without the setting, behavior is unchanged: the rule covers only resolved tool results, and other empty completions get the full retry loop and a 503.

Why. With Claude Code, some requests that don't end in tool results come back empty by design, for example a turn answering a Monitor event that needs no reply. Each one cost 11 full-context calls and 3–4 minutes before failing. Details and timings are in #208.

Change

• config.rs: codex_empty_completion_end_turn() reads the env var, then codex.emptyCompletion. It accepts end_turn / end-turn or retry; anything else falls back to retry. The setting is also listed in the config summary.
• providers/codex/mod.rs: both places that call tail_is_successful_tool_result(&body) now call empty_completion_may_end_turn(&body), which is that check OR the setting. accept_empty_completion is untouched, so the one-retry discriminator from codex: treat an empty successful completion after a tool_result tail as a valid end_turn #122 still applies.
• Docs: a row in the configuration table and an "Empty completions" subsection in the Codex page.

Result. We ran the same change on top of v0.1.44 with #122's earlier version. Offline, against a stub upstream that always completes empty, a request costs 2 upstream calls and ~1.5 s instead of 11 calls and a 503, on both the buffered and the streaming path. Claude Code 2.1.293 accepts the empty end_turn (-p returns success, one turn, no client retries). Live with gpt-6-astra, the next 270 requests over three weeks produced no empty-completion 503.

Details for reviewers and agents: tests, scope

Tests

• New codex_empty_completion_mode_reads_env_and_config: default off, config on, env retry overrides config, env end_turn.
• cargo fmt --all -- --check: clean.
• cargo clippy --all-targets --locked -- -D warnings: clean.
• cargo test --all --locked: 1175 passed, 0 failed (on main at fe10718).

Scope

• The setting is read per request, like the other Codex config accessors.
• It only widens which requests may take the existing one-retry path. It doesn't change what counts as an empty completion or the retry count.

Unrelated, noticed while testing: running cargo test appends log lines to the real ~/.local/state/claude-code-proxy/proxy.log, e.g. gpt-5.5 503s at ~4 ms and ~11 s. Some test doesn't isolate the state dir. #164 has a similar leak: one of its tests wrote codex-compaction/session.json into the real state dir.

---

Written by Claude Opus 5.5 in Claude Code on behalf of @Butanium, with little human supervision; not reviewed by a human.

https://claude.ai/code/session_01Xmx9RE7u4ZLBzSUjpjqyBM