From b9c655ad46d02e10757038a7bc53e48231be52b6 Mon Sep 17 00:00:00 2001
From: Ahmed Ibrahim <219906144+aibrahim-oai@users.noreply.github.com>
Date: Sat, 14 Mar 2026 03:00:42 +0000
Subject: [PATCH] Stabilize approval matrix write-file command

Replace the approval-matrix shell redirection write path with deterministic Python file I/O and keep scenario-level diagnostics in CI while the flake is being verified.

Co-authored-by: Codex <noreply@openai.com>
---
 .codex/flaky-test-triage.md            |  9 ++++++++-
 codex-rs/core/tests/suite/approvals.rs | 13 ++++++++++++-
 2 files changed, 20 insertions(+), 2 deletions(-)

diff --git a/.codex/flaky-test-triage.md b/.codex/flaky-test-triage.md
index 649f71c0c9..b986a02bfd 100644
--- a/.codex/flaky-test-triage.md
+++ b/.codex/flaky-test-triage.md
@@ -38,6 +38,12 @@ Older failures also appeared on Linux, but the repeated cross-PR signal is stron
 1. Approval-related tests in `codex-rs/core/tests/suite/approvals.rs` still have timing-sensitive behavior, especially in cross-platform CI.
 2. Windows-specific approval UI tests in `codex-rs/tui/src/chatwidget/tests.rs` may depend on partially implicit sandbox state and can fail intermittently on Windows runners.
 
+## Current Fix In Progress
+
+- Replace the approval-matrix write-file command from shell redirection (`printf > file && cat file`) with a deterministic `python3 -c` file write/readback command.
+- Keep targeted scenario diagnostics in the matrix so CI logs include the exact command, exit code, and stdout when a scenario fails again.
+- Rationale: the known `read_only_unless_trusted_requires_approval` flake was previously "fixed" by increasing timeout budget. This change removes shell-redirection timing sensitivity instead of stretching the timeout further.
+
 ## Constraints
 
 - Do not run tests locally.
@@ -49,4 +55,5 @@ Older failures also appeared on Linux, but the repeated cross-PR signal is stron
 
 | Commit | Purpose | PR CI result | Notes |
 | --- | --- | --- | --- |
-| _pending_ | PR bootstrap | _pending_ | Tracking starts once the PR is open. |
+| `60f44b4d7` | PR bootstrap | partial pass | PR opened and non-Rust checks passed after rebasing to current `main`, but `rust-ci` skipped because only the tracking doc changed. This commit does not count toward the five full-suite green commits. |
+| _pending_ | First full-suite flaky-test fix | _pending_ | Approvals matrix write-file command now uses deterministic Python I/O instead of shell redirection. |
diff --git a/codex-rs/core/tests/suite/approvals.rs b/codex-rs/core/tests/suite/approvals.rs
index 7a9dba038e..53978c1766 100644
--- a/codex-rs/core/tests/suite/approvals.rs
+++ b/codex-rs/core/tests/suite/approvals.rs
@@ -122,7 +122,11 @@ impl ActionKind {
             ActionKind::WriteFile { target, content } => {
                 let (path, _) = target.resolve_for_patch(test);
                 let _ = fs::remove_file(&path);
-                let command = format!("printf {content:?} > {path:?} && cat {path:?}");
+                let path_str = path.display().to_string();
+                let script = format!(
+                    "from pathlib import Path; path = Path({path_str:?}); content = {content:?}; path.write_text(content, encoding='utf-8'); print(path.read_text(encoding='utf-8'), end='')",
+                );
+                let command = format!("python3 -c {script:?}");
                 let event = shell_event(call_id, &command, 5_000, sandbox_permissions)?;
                 Ok((event, Some(command)))
             }
@@ -1611,6 +1615,9 @@ async fn run_scenario(scenario: &ScenarioSpec) -> Result<()> {
         .action
         .prepare(&test, &server, call_id, scenario.sandbox_permissions)
         .await?;
+    if let Some(command) = expected_command.as_deref() {
+        eprintln!("approval scenario {} command: {command}", scenario.name);
+    }
 
     let _ = mount_sse_once(
         &server,
@@ -1692,6 +1699,10 @@ async fn run_scenario(scenario: &ScenarioSpec) -> Result<()> {
 
     let output_item = results_mock.single_request().function_call_output(call_id);
     let result = parse_result(&output_item);
+    eprintln!(
+        "approval scenario {} result: exit_code={:?} stdout={:?}",
+        scenario.name, result.exit_code, result.stdout
+    );
     scenario.expectation.verify(&test, &result)?;
 
     Ok(())