Skip to content

Releases

v1.6.1 — the CLI cleanup backend works now

Choose a tag to compare

@colakang colakang released this 11 Sep 04:14
27d0e58

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

v1.6.0 — a menu bar, and four things it uncovered

Choose a tag to compare

@colakang colakang released this 01 Sep 03:33
c75688a

The button layout changes in this release. Nine 26px buttons were wedged around three dropdowns, plus Export and Clear in the button column — in the corner of a window whose whole job is to be glanceable while somebody is talking. Import and delete get used a handful of times ever; export happens once, at the end.

File   Import ▸ System Role… / Case Detail… / Knowledge Base…
       Manage Templates ▸ <category> ▸ <template> ▸ Edit… / Delete…
       ─────
       Export Conversation…                                  ⌘E

Edit   Copy                                                  ⌘C
       Select All                                            ⌘A
       ─────
       Clear Transcript…

If you are looking for the + and − buttons: they are File ▸ Import and File ▸ Manage Templates. The Export and Clear Transcript buttons are in File and Edit.

The panel keeps what gets touched mid-call: Answer Selection, Pop Up, the checkboxes, and the three dropdowns. The dropdowns stay because they are as much a status display as a control — reading which role and model are live should not cost a click. They also take back the space that was reserved for the buttons, so long template names stop being truncated.

Four things this uncovered

Clear Transcript asked nothing before discarding a meeting, from a button sitting between two harmless ones. It now says how much goes — "4 turns will be discarded", because "are you sure?" is not a question anyone can answer — and deliberately has no keyboard shortcut.

Deleting a template could silently swap the prompt in use. It only ever removed whichever template was selected, and moved the selection unconditionally afterwards — so tidying up an unused template changed which one was applied. It works by name now, and only moves the selection when the deleted one was the live one.

Templates can be edited. Shipped ones are copied into your own directory first: the package directory can be read-only, and an upgrade replaces the file, so an in-place edit would survive until the next release and then disappear with nothing to explain it. Your imported templates are edited in place. The file opens in whatever you normally use for it.

Copy and Select All were always there and nothing said so. Tk binds them at the Text class level, so ⌘C on a selected passage has always worked. They are listed now, which is the point — an undiscoverable feature is close to an absent one.

uv tool install --force --refresh echoai-helper

439 tests.

Full changelog: v1.5.3...v1.6.0

v1.5.3 — every answer stays on screen

Choose a tag to compare

@colakang colakang released this 29 Aug 01:19
0815c39

Reported as: the right panel only ever shows one reply, and the earlier ones are gone.

responder.response is a single string that each new answer overwrites, and the panel mirrored it — so the second answer erased the first from the screen, while both were written to the session file. That connects something not previously linked: an export could contain answers you had watched disappear.

The behaviour made sense where the feature started. In an interview only the newest suggestion matters. It stops making sense once you can select turns and ask about them, which is the meeting case built for exactly this: you ask about one passage, read the answer, ask about another, and the first is gone — while the transcript panel immediately beside it accumulates.

Answers are now kept and rendered oldest first. Thinking... and errors still show, since those are states rather than answers. The view follows the newest answer only when you are already at the bottom, matching the transcript — being yanked away from an older answer mid-read is worse than not seeing the new one instantly. Clear Transcript clears them too.

uv tool install --force --refresh echoai-helper

419 tests.

Full changelog: v1.5.2...v1.5.3

v1.5.2 — the driver actually appears

Choose a tag to compare

@colakang colakang released this 29 Aug 00:35
d976b29

Reported as: setup says BlackHole installed, and Sound settings has no BlackHole in it.

Both were true. The driver was installed. It was never published.

A HAL driver is a file in /Library/Audio/Plug-Ins/HAL, and CoreAudio only offers it as a device once coreaudiod re-reads that directory. That restart was in the code as sudo -n killall coreaudiod — non-interactive sudo, which fails rather than ask for a password, so it failed on every machine that asks for one. It failed quietly, both callers discarded the result, and the driver sat on disk unpublished while the install reported success.

Worth stating plainly why this looked fine in development: sudo -n fails on the development machine too. BlackHole worked there because a later reboot published it. The install step never worked for anybody.

The restart now runs inside the installer's own administrator prompt — one password rather than two, since two prompts for one operation is how people grant the first and cancel the second, which produces exactly the state reported. It is braced so a failed install does not restart the machine's audio.

The advice was wrong as well as absent. It said restarting the app would settle it; the app is not what has to re-read /Library. The dialog now separates a driver on disk that is one command away from one that never installed, and points at echoai-helper setup instead of a script path that someone who installed the wheel does not have.

If you hit this on 1.5.1, sudo killall coreaudiod fixes it without upgrading — though upgrading is better, since setup will otherwise leave the next machine in the same state.

Also

Wear headphones if you record your microphone. Played through speakers, the far end lands in your own track: on a measured session the microphone came on 234 seconds in and 5 of 21 "You" turns after that were the far end repeated back. Meeting apps cancel echo on the stream they send, inside themselves — this app opens the input device separately and gets the raw microphone, which that cancellation never touched and could not. Now documented under Known limitations.

uv tool install --force --refresh echoai-helper

413 tests.

Full changelog: v1.5.1...v1.5.2

v1.5.1 — a first run that finishes

Choose a tag to compare

@colakang colakang released this 28 Aug 22:29
bc02a54

Fixes a first run that looked like the app had hung. It had not — three separate problems stacked up, each hiding the next.

The download host could not serve the files. The default was ModelScope, whose site answers in under a second but whose file downloads timed out: 213 retries across an hour with nothing written to disk. The client has no timeout, so it retries forever. The default is HuggingFace now, which served the same models immediately. Model ids are hub-specific, so those moved too — including the speaker embedder, which had its own hardcoded id and never followed the setting at all.

Nothing was on screen while it downloaded. A 1.5GB fetch happened before the window was created, so macOS offered Force Quit — which is what the first person to try it did. The window comes up first now, with a progress bar and, past twenty seconds, how long it has been waiting. The wait is capped at twenty minutes, after which it names the setting to change.

The failure dialog was unusable. It left an event-loop-less splash window behind it, so it could not be dismissed, and it was handed the raw exception — several thousand characters, which sized the alert past the edge of the screen along with its OK button. It gets one line and a pointer to the log now.

If ModelScope is the faster host where you are, hub: "ms" still works — set it in conf.yaml, with the model ids noted in the comments beside each one.

pip install --upgrade echoai-helper    # or: uv tool install echoai-helper --force

Full changelog: v1.5.0...v1.5.1

v1.5.0 — the API key is optional

Choose a tag to compare

@colakang colakang released this 28 Aug 19:40
57cbd80
uv tool upgrade echoai-helper

Transcription no longer requires an OpenAI account

The app showed "No API key" and quit. It should never have been fatal: transcription runs on your machine — FunASR, locally — and needs no account of any kind. The key is for the reply suggestions and the cleanup pass at export, both optional, and demanding it before the window appeared stopped people recording a meeting over a feature they had not reached.

Verified with every key source removed: the app starts, captures, records a session, transcribes, and exports — making no model calls at all.

Setting it, when you want those features

When How
At the moment it is needed The app asks, saves it, and says to try again
Up front echoai-helper key
During install echoai-helper setup says transcription does not need one, and how to set it

Choosing cleanup at export prompts as well, rather than exporting with a note saying cleanup was skipped — true, and not what was asked for. The Claude CLI backend is exempt: it bills a subscription and needs no key.

Where it goes

The user config directory, 0600. The old alert told you to create a file called .llm without saying where — and the place it looked was PathConfig.get_project_root(), which installed from a wheel is site-packages: not somewhere to keep a secret, and a reinstall would have deleted it.

A value that is obviously not a key is refused rather than written, since it would otherwise surface as an authentication failure much later.

396 tests.

v1.4.5 — the documented upgrade command now exists

Choose a tag to compare

@colakang colakang released this 28 Aug 19:15
3ea6043

Documentation only.

The README told you to run uv tool upgrade --refresh echoai-helper. There is no --refresh on upgrade — it belongs to install — so that command errors out.

uv tool upgrade echoai-helper                    # normally
uv tool install --force --refresh echoai-helper  # straight to the newest

uv caches the package index, so an upgrade run soon after a release can land on the version before the one you came for, reporting Updated … -> v1.4.3 and looking entirely successful. Running it again gets there; the second command skips the wait.

How the wrong command shipped

It was "verified" as:

uv tool upgrade --refresh echoai-helper 2>&1 | grep -E "Updated|echoai-helper" | tail -2

The grep swallowed the error, and the version printed afterwards came from a previous, different command that had already succeeded. Filtering the output of a check is how a check comes to agree with whatever you expected.

A test now asks uv itself whether the documented flags exist, and inspects only commands inside fenced code blocks.

387 tests.

v1.4.4 — the icon starts the app, and nothing fails invisibly

Choose a tag to compare

@colakang colakang released this 28 Aug 19:10
01e223d
uv tool upgrade --refresh echoai-helper
echoai-helper install-launcher      # installs the icon and opens the app

Two reports from the same first install on an M1, and a fix for the thing that most likely caused it.

The icon did nothing, with no message

It had not crashed. It printed a complaint and returned — and from the Launchpad icon that text goes to a log nobody has been told about while no window ever appears.

Any startup problem now shows an alert when there is no terminal to print to. One path was worse than invisible: a missing API key called input("Press Enter to exit..."), and a GUI launch has no stdin.

If something still stops before the window appears:

tail -20 "$HOME/Library/Logs/EchoAI Helper.log"

ffmpeg is no longer required

The app refused to start without it — for a program that never calls it. Capture is sounddevice → numpy → FunASR, with no file decoding anywhere; verified by transcribing with ffmpeg removed from PATH entirely.

It was undocumented and normally comes from Homebrew, so v1.4.2 — which made Homebrew unnecessary — turned a hidden requirement into a silent failure for exactly the people it was meant to help.

Only scripts/calibrate_vad.py and check-audio --record decode files, and each checks for itself now.

install-launcher opens the app

Installing an icon is not the goal; having the app running is. It is the last of the three commands in the README, and it used to end by telling you where to find something you then had to go and click. --no-open for scripted installs.

386 tests.

v1.4.3 — say how to upgrade

Choose a tag to compare

@colakang colakang released this 28 Aug 18:44
8cc2ebe

Documentation only. No code changes.

uv tool upgrade echoai-helper

Not uv tool install again. On something already installed that does nothing and reports

`echoai-helper` is already installed

which reads as "you are up to date" when you are not. The README only ever described installing, so anyone on an older version had no way to move off it and no reason to think they needed to.

Add --refresh if a release landed in the last few minutes: uv caches the package index, and without it an upgrade can quietly fetch the version before the one you came for.

Re-run echoai-helper install-launcher after upgrading if you use the Launchpad icon. The bundle holds a generated script rather than a copy of the program, so it keeps working across upgrades — the script and icon are only rewritten when you run that again.

v1.4.2 — no Homebrew needed, and the prerequisites written down

Choose a tag to compare

@colakang colakang released this 28 Aug 18:40
dd6d631

Both of these came from someone installing it on their own Mac for the first time.

setup no longer needs Homebrew

It stopped at "Homebrew is not installed", and told you to fetch the driver by hand from a page that asks for an email address first. Requiring a package manager to install one audio driver is a poor trade.

Homebrew is still used where it exists — the driver is tracked and brew uninstall --cask blackhole-2ch removes it. Where it does not, the package is downloaded from the vendor and installed directly, with its checksum verified before anything runs. The URL and checksum come from Homebrew's cask metadata, so the version tracks what is current without Homebrew being present.

Two things found by actually running it:

  • The vendor's CDN answers urllib's default User-Agent with 406 Not Acceptable, so without a header this failed for every user on the first request. It worked under curl while investigating, which is exactly how it would have shipped.
  • The installer path was interpolated with Python's repr() into an AppleScript literal. repr switches to double quotes when the value contains a single one, and the literal it sits inside is double-quoted — a home directory belonging to anyone called O'Brien would end the string early, and it would surface at the password prompt. quoted form of handles the shell layer now, explicit escaping the AppleScript one.

The prerequisites are written down

Command Line Developer Tools. macOS ships stubs for git and the compilers; touching one pops a dialog demanding the tools, and it can appear part-way through an install, which reads as a failure.

xcode-select --install

When they are already present this says so and exits non-zero — which also reads as a failure, so the README says that too.

uv. The first install command assumed it and never said where to get it. That matters more than convenience: this needs Python 3.12 or 3.13, and the Python on a Mac is 3.9 (Apple's) or 3.14 (Homebrew's).

curl -LsSf https://astral.sh/uv/install.sh | sh

Homebrew is documented as optional, because it now is.

379 tests.