Skip to content
Merged
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 6 additions & 3 deletions docs/guides/run-tasks-with-benchflow.md
Original file line number Diff line number Diff line change
Expand Up @@ -141,17 +141,17 @@ Probe the local subscription auth:
CODEX_HOME="$(mktemp -d /tmp/codex-home.XXXXXX)"
mkdir -p "$CODEX_HOME"
cp "$HOME/.codex/auth.json" "$CODEX_HOME/auth.json"
printf 'model = "gpt-5.5"\nmodel_reasoning_effort = "xhigh"\n' > "$CODEX_HOME/config.toml"
printf 'model = "gpt-5.5"\nmodel_reasoning_effort = "xhigh"\n[features]\napps = false\n' > "$CODEX_HOME/config.toml"
env -u OPENAI_API_KEY -u OPENAI_BASE_URL CODEX_HOME="$CODEX_HOME" \
codex exec -m gpt-5.5 "Reply exactly ok"
codex exec --disable apps -m gpt-5.5 "Reply exactly ok"
```

Run the six-task environment coverage matrix:

```bash
unset OPENAI_API_KEY OPENAI_BASE_URL
export CODEX_AUTH_JSON="$(tr -d '\n' < "$HOME/.codex/auth.json")"
export CODEX_CONFIG='{"model":"gpt-5.5","model_reasoning_effort":"xhigh"}'
export CODEX_CONFIG='{"model":"gpt-5.5","model_reasoning_effort":"xhigh","features":{"apps":false}}'

uvx --from "benchflow==${BENCHFLOW_VERSION}" bench eval run \
--tasks-dir tasks \
Expand All @@ -175,6 +175,9 @@ uvx --from "benchflow==${BENCHFLOW_VERSION}" bench eval run \

Do not pass `--reasoning-effort` to `codex-acp` on BenchFlow 0.6.4. Put Codex
reasoning settings in `CODEX_CONFIG` as shown above.
Keep Codex `apps` disabled for these runs. The task runtime should exercise only
the local Docker mock services, not hosted app connectors attached to the
operator's Codex account.

The validated run for this revision started and scored all six environments
with `errors=0` and `idle_timeouts=0`. Model pass rate is not the same as
Expand Down
Loading