Repository navigation
Reported as: at export, with cleanup ticked, only the API can be chosen — on a machine where claude is installed and working.
Four separate faults, and the first two were hiding each other.
The CLI could not be found. A double-clicked app inherits /usr/bin:/bin:/usr/sbin:/sbin and reads no shell profile, and claude installs to ~/.local/bin by default — a directory no GUI app has ever had on its PATH. The lookup now checks PATH and then where these CLIs actually install, including every nvm node version, and runs the resolved absolute path.
The backend did not exist. The provider factory had branches for openai and litellm and none for cli, so it returned nothing every time and the export said "no language model available" — untrue: the CLI was installed and working. It had simply never been reachable from the app, and nothing caught it because the first fault greyed the option out before anyone could select it.
codex could not be read. It prints a session banner ahead of the answer, and cleanup needs exactly as many lines back as it sent, so every batch came back the wrong length and was discarded. It answers into a file now.
The timeout was too tight. Measured over a 229-line meeting: 16–90s per batch against a 90s limit, mean 65. Two batches hit the ceiling and lost their whole batch of work. At 300s the same transcript finished with zero failures.
Also
The dialog offered one hardcoded CLI, so an installed Codex was unreachable — it lists whatever you have installed now, and picking one actually runs it.
The speaker count starts at 1. It began at 2, in both the live control and the export merge, for no reason anyone could name. A one-to-one call is the commonest shape there is and the one where over-splitting shows most — a single support agent coming back as S1 and S2.
The README's "20–50s per batch" came from short batches and under-promised by about half; the export estimate used the same figure, so an 11-minute job advertised itself as four. Both say 20–90s now.
uv tool install --force --refresh echoai-helper
453 tests. Verified end to end on a real 229-line meeting: 10 batches, 0 failed, 85 corrections with every original preserved.
Not verified: codex end to end (broken on the test machine for an unrelated reason — its configured model needs a newer Codex), and gemini.
Full changelog: v1.6.0...v1.6.1