Summary

The end-of-turn tool-error warning (⚠️ 🛠️ \ (agent)` failed) fires for **normal/benign non-zero exits from the Codex app-server's exec_commandtool**, even though the equivalent nativeexec/bash` tool is intentionally silenced in the same situation.

Root cause: resolveToolErrorWarningPolicy suppresses the warning for exec-like tools via

if (isExecLikeToolName(toolName) && !includeDetails) return { showWarning: false, includeDetails };

but isExecLikeToolName checks against a hardcoded set that does not include the Codex app-server tool name:

// src/agents/tool-error-summary.ts
const EXEC_LIKE_TOOL_NAMES = new Set(["exec", "bash"]);

The Codex app-server surfaces its shell tool as exec_command (confirmed in rollout transcripts: function_call name="exec_command"), so it falls through the exec-silencer to the default branch and shows a "failed" warning for any non-zero exit — including a grep/test/diff no-match used as a verification step, which is normal shell usage, not a tool failure.

Notably, the codebase already knows exec_command is the Codex equivalent of exec:

// src/agents/harness/native-hook-relay.ts
const NATIVE_HOOK_TOOL_NAME_ALIASES = { exec_command: "exec" };

…but that alias is only applied in the native-hook-relay path, not before the warning gate. If the alias were honored in resolveToolErrorWarningPolicy (or exec_command added to EXEC_LIKE_TOOL_NAMES), the warning would be correctly suppressed for benign exits, matching native exec/bash behavior.

This is the same class of bug as #89837 (tools.exec.pathPrepend not applied to the Codex app-server's exec_command): an exec-targeted feature that isn't applied to Codex's exec_command.

Affected versions

Confirmed present in 2026.5.20 and in latest stable 2026.6.6 (the relevant code is byte-identical between them).

Repro

1. Run an agent on the openai-codex backend (Codex app-server).
2. In a single turn, run a shell command that exits non-zero as normal usage — e.g. a verification step like grep -E 'SOME LABEL' /tmp/out.txt that finds no match (exit 1), or script > /tmp/out.txt && grep ... /tmp/out.txt where the grep target is absent.
3. The turn completes normally, but the channel shows ⚠️ 🛠️ \ (agent)` failed`.
4. The same workload using the native exec/bash tool (non-Codex backend) does not produce the warning, because isExecLikeToolName matches and showWarning is false.

Expected

Benign non-zero exits from exec_command should be treated like exec/bash — no ⚠️ … failed warning on a normal turn (includeDetails false). Real tool failures (e.g. command-not-found 126/127, timeouts) should still warn, exactly as they do for exec/bash.

Suggested fix

Either:

• add "exec_command" to EXEC_LIKE_TOOL_NAMES in src/agents/tool-error-summary.ts, or
• normalize the tool name through NATIVE_HOOK_TOOL_NAME_ALIASES before resolveToolErrorWarningPolicy (src/agents/pi-embedded-runner/run/payloads.ts) so the existing alias is honored consistently across paths.

Pointers

• src/agents/tool-error-summary.ts — EXEC_LIKE_TOOL_NAMES, isExecLikeToolName
• src/agents/pi-embedded-runner/run/payloads.ts — resolveToolErrorWarningPolicy (the isExecLikeToolName(...) && !includeDetails branch)
• src/agents/harness/native-hook-relay.ts — NATIVE_HOOK_TOOL_NAME_ALIASES = { exec_command: "exec" }