Codex loader file ordering is nondeterministic, causing Rust test failure #1105

What happened?

The Rust test suite fails in tests::loads_codex_token_count_events_in_parallel because Codex session files are loaded in a nondeterministic filesystem order.

Observed with:

cargo test --manifest-path rust/Cargo.toml --workspace

Failure:

---- tests::loads_codex_token_count_events_in_parallel stdout ----
thread 'tests::loads_codex_token_count_events_in_parallel' panicked at crates/ccusage/src/main.rs:649:9:
assertion `left == right` failed
  left: "session-b"
 right: "session-a"

This is distinct from the existing Codex overcounting issues (#950, #988). This report is about deterministic file/event ordering and test reliability, not token totals or cost inflation.

Steps to reproduce

1. Check out the current repository.

2. Run the targeted test:

   cargo test --manifest-path rust/Cargo.toml -p ccusage --bin ccusage tests::loads_codex_token_count_events_in_parallel -- --exact

3. Or run the full Rust workspace suite:

   cargo test --manifest-path rust/Cargo.toml --workspace

Expected behavior

Codex event loading should be deterministic for a fixed set of session files, or tests should not assume an ordering that the loader does not guarantee.

At minimum:

• load_codex_events_from_directory(..., true) and (..., false) should remain equivalent.
• The test should not depend on filesystem read_dir ordering.
• Any fix should preserve Codex dedupe and token/cost aggregation semantics.

Evidence-gated analysis

The failing test creates two files:

• session-a.jsonl
• session-b.jsonl

It then compares single-threaded and parallel Codex loading. The equality assertion passes before the failing order assertion:

assert_eq!(parallel_events, single_thread_events);
assert_eq!(parallel_events[0].session_id, "session-a");
assert_eq!(parallel_events[1].session_id, "session-b");

So the evidence does not support “parallel loading corrupts the order relative to single-thread loading.” Both modes agree with each other. The failure is that the shared baseline order is session-b, then session-a, while the test expects session-a, then session-b.

Source context:

• rust/crates/ccusage/src/codex_loader.rs collects files with collect_usage_files(sessions_dir, &mut files) and does not sort them before loading.
• The Claude loader sorts discovered usage files with files.sort_by_cached_key(|path| path.to_string_lossy().into_owned()).

Local probes run before a minimal patch:

• Targeted exact test: 3/3 failed with session-b before session-a.
• Targeted filtered test: 10/10 failed with the same assertion.
• Full Rust suite: failed only this test in the ccusage binary suite.

A local one-line patch that sorts Codex files after collection made the targeted test and full Rust suite pass:

collect_usage_files(sessions_dir, &mut files);
files.sort_by_cached_key(|path| path.to_string_lossy().into_owned());

Post-patch verification run locally:

cargo test --manifest-path rust/Cargo.toml -p ccusage --bin ccusage tests::loads_codex_token_count_events_in_parallel -- --exact
cargo test --manifest-path rust/Cargo.toml -p ccusage --bin ccusage tests::builds_codex_daily_json_report -- --exact
cargo test --manifest-path rust/Cargo.toml --workspace

All passed after the sort patch.

Suggested fix

Sort collected Codex usage files before single-thread or parallel loading, mirroring Claude loader behavior:

collect_usage_files(sessions_dir, &mut files);
files.sort_by_cached_key(|path| path.to_string_lossy().into_owned());

Likely patch target:

rust/crates/ccusage/src/codex_loader.rs

Non-goals / caveats

• I am not claiming this changes user-visible token or cost totals.
• I am not claiming this fixes Bug: Massive token overcounting for Codex subagent sessions (91x inflation) #950 or @ccusage/codex double-counts tokens for branched Codex Desktop conversations #988.
• The confirmed issue is narrower: Codex loader file ordering is not deterministic enough for the current Rust test expectation.