🤖 bench: enable PTC + hard-restart by default, fix mux-run.sh fatal-on-exit

[ammar-agent]

Summary

Enable Programmatic Tool Calling (PTC) and exec-subagent-hard-restart experiments by default for all users. Fix token extraction loss in Terminal-Bench runs when agent exits non-zero.

Background

The Terminal-Bench baseline (Opus 4.6 / xhigh) scores 57.3% (51/89 pass). Analysis of the failure modes shows:

• 11 tasks time out — the agent exhausts its 30-min budget doing sequential tool calls that PTC could batch
• Sub-agent delegation is used in only 10% of tasks — the hard-restart experiment prevents sub-agents from permanently failing on context overflow
• Token tracking is lost for all timed-out tasks because mux-run.sh calls fatal() before the token-extraction python block runs

Implementation

1. PTC enabled by default (experiments.ts)

• programmatic-tool-calling: enabledByDefault: true
• The code_execution tool is now available by default in both the desktop app and CLI, letting the model batch 2+ tool calls in a single turn

2. Exec sub-agent hard restart enabled by default (experiments.ts)

• exec-subagent-hard-restart: enabledByDefault: true
• Sub-agents that hit context overflow will hard-restart instead of failing permanently

3. CLI respects enabledByDefault (run.ts)

• buildExperimentsObject() now auto-enables experiments with enabledByDefault: true when no explicit --experiment flags are passed
• If the user passes any --experiment flags, those override defaults entirely (no mixing)

4. Fix mux-run.sh fatal-on-exit (mux-run.sh)

• Replace fatal "mux agent session failed" with a warning + continue
• Token extraction now runs unconditionally (even for timed-out / crashed agent runs)
• Original exit code is propagated at the end of the script

Validation

• make typecheck — pass
• make lint — pass
• make fmt-check — pass
• make test — 3446 pass, 0 fail

---

Generated with mux • Model: anthropic:claude-opus-4-6 • Thinking: xhigh • Cost: $92.04

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 0b9aad504b

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Preserve the failing exit code before ! negates it

Because the pipeline is wrapped in if ! …; then, the ! operator negates the pipeline status, so inside the block $? is always 0 when the command failed. With set -o pipefail, this means mux_exit_code becomes 0 on failures/timeouts, and the script exits 0 even when the agent crashed, so the harness can’t detect failures. Capture the pipeline status before applying ! (or use if ...; then ... else mux_exit_code=$?) to preserve the real exit code.

Useful? React with 👍 / 👎.