Skip to content

Releases

aishell v1.0.4 — one-shot enforcement in aishell's own source

Choose a tag to compare

@nrupala nrupala released this 06 Nov 03:22

aishell v1.0.4 — one-shot behaviour now lives in aishell's own source (supersedes v1.0.3).

v1.0.3's --single-turn worked everywhere we checked, but it was the same kind of dependency that broke v1.0.1 (-no-cnv, removed upstream) and v1.0.2 (--no-conversation, removed upstream). So v1.0.4 removes the dependency class entirely:

  • aishell passes llama-cli only the stable core flags (-m, -c, -n, --temp, --log-disable, -p) — unchanged in meaning since llama.cpp's earliest builds. No conversation flags, no display flags.
  • aishell itself terminates llama-cli the moment a complete response block has been captured, whatever the build's conversation behaviour, and bounds every inference with a 10-minute timeout (the failure mode behind the 2026-10-30 hang).
  • The test suite (16 tests) forbids the whole conversation/display flag class and drives a fake conversation-mode llama-cli that would wait forever, proving termination on completion and on timeout.

Proofs: unit suite 16/16; live run against the current ggml-org llama.cpp (b12140) with no conversation flags passed — proposed and answered cleanly. Platform re-proofs (Aetheris/old llama-cli, Windows on the T15, macOS runner) run against these published binaries; results are recorded in the gate pack.

Binaries (in dist/, SHA-256 in dist/SHA256SUMS.txt):

  • ai-shell.exe — Windows x86-64
  • ai-shell-linux — Linux x86-64
  • ai-shell-linux-arm64 — Linux arm64
  • ai-shell-mac — macOS Apple Silicon (arm64)

The model (~1.9 GB) is fetched by setup.bat / setup.sh; it is not bundled. Licence: Apache-2.0. This is a prerelease (release candidate in soak).

aishell v1.0.3 — the one-shot flag that survives

Choose a tag to compare

@nrupala nrupala released this 06 Nov 02:38

SUPERSEDED by v1.0.4 (one-shot enforcement moved into aishell's own source; no conversation flags passed at all).

The one-shot fix that survives upstream drift. During the v1.0.1 certification round, the macOS and Windows proofs showed the newest llama.cpp builds removed the conversation toggle entirely: --no-conversation (v1.0.2's fix) is now rejected exactly as -no-cnv (v1.0.1) was — every inference failed with error: invalid argument on a current llama-cli.

v1.0.3 passes only --single-turn --no-display-prompt as one-shot controls. --single-turn exists with generate-one-turn-then-exit semantics in every build checked: the Aetheris proof build bb4caa7 (verified by direct invocation), ggml-org b8417, Homebrew's current build, and the MSVC dev build — the last three verified by their own --help output and by the re-run certification proofs. A unit test now forbids both conversation toggles so this cannot regress silently.

Download

Binaries are in dist/ in this repository:

  • ai-shell.exe — Windows (amd64)
  • ai-shell-linux — Linux (amd64)
  • ai-shell-linux-arm64 — Linux (arm64)
  • ai-shell-mac — macOS (Apple Silicon)

SHA-256 sums (dist/SHA256SUMS.txt):

27dc037be66db812a8b8d7b9ada7ef467bf7f429c9e13158c852c1c87bf5e5e3  ai-shell.exe
b8c8ef18dd53eafa6d548c78ef9f9e8aa3bf8699a9b0ad421e5234106c6abe0c  ai-shell-linux
0196e5e1583195d866f48d3c24bc3ec8cabb500342d466e9eb66937779455f75  ai-shell-linux-arm64
402c04a62f54d3891f49b30ea7067482b350821bd7246defc483217209b37255  ai-shell-mac

You also need llama-cli (llama.cpp) in your PATH; setup.sh / setup.bat fetch the Qwen2.5-Coder-3B model for you.

Release candidate (prerelease): certification proofs (Windows execution, macOS CI, three-lane multi-tester review) are recorded against this build; see CHANGELOG 1.0.2/1.0.3 for the full flag-drift story.

aishell v1.0.2 — cross-platform llama-cli flag fix

Choose a tag to compare

@nrupala nrupala released this 06 Nov 02:25

SUPERSEDED by v1.0.3 (same day). The newest llama.cpp builds also reject --no-conversation; use v1.0.3.


Certification-round fix release. Found by the macOS CI proof (a real Apple Silicon runner, Homebrew llama.cpp) during the v1.0.1 multi-tester certification: newer llama.cpp builds removed the short -no-cnv alias, so every inference failed on a current llama-cli with error: invalid argument: -no-cnv (graceful [Inference Error], clean exit — banner, OS detection and session flow unaffected). The older Aetheris proof build (bb4caa7) still accepted the short form, which is why the Linux proof could not see it.

Download

Binaries are in dist/ in this repository:

  • ai-shell.exe — Windows (amd64)
  • ai-shell-linux — Linux (amd64)
  • ai-shell-linux-arm64 — Linux (arm64)
  • ai-shell-mac — macOS (Apple Silicon)

SHA-256 sums (dist/SHA256SUMS.txt):

6089056406f295f241f1a124aa6e7c3e7b5cf2c2fdb33ffd451341287758e235  ai-shell.exe
12d2980e8add8181dfd42aa766b6b330f5fb486b01ef603356461f31f43700f7  ai-shell-linux
285705ba7bff530aff4632e96c4d9dc90a23e833cc5300e23b98e425eee989b2  ai-shell-linux-arm64
05f7a6200e35f1b499d728e2395da99308d15957861fd6a80ed682f9a95e24dd  ai-shell-mac

You also need llama-cli (llama.cpp) in your PATH; setup.sh / setup.bat fetch the Qwen2.5-Coder-3B model for you.

Changes in v1.0.2

  • Inference uses the long flag forms only (--no-conversation --single-turn --no-display-prompt), accepted by both the older proof build and the newest builds; a unit test pins the argument list (suite: 14/14).
  • README Quick start now leads with the prebuilt-binary route (independent cold-user walk finding); stale test count removed.

Release candidate (prerelease): the macOS CI proof and Windows execution proof are being re-run against this build to complete the v1.0.1 certification round; see CHANGELOG for detail.

aishell v1.0.1 — release candidate (Aetheris-proven)

Choose a tag to compare

@nrupala nrupala released this 30 Oct 23:38

First public release candidate of aishell — the AI-native shell: natural language in, local llama.cpp inference, clean OS commands out, executed only after your manual confirmation.

Download

Binaries are in dist/ in this repository:

  • ai-shell.exe — Windows (amd64)
  • ai-shell-linux — Linux (amd64)
  • ai-shell-linux-arm64 — Linux (arm64)
  • ai-shell-mac — macOS (Apple Silicon)

SHA-256 sums (dist/SHA256SUMS.txt):

391107e0f8268ec129e1f44f96fec7a98cb8e23bac3f23c66f7387305a981bd5  ai-shell.exe
23eaf419c175cfad43dd791cbede9ff2c6c72c5357254e160294af74f86b6138  ai-shell-linux
f023a55b1b8545b1a0f12b6dbd39a97b0367b528d7decb1885ef3759d3d345c3  ai-shell-linux-arm64
9fbd52af9aebfc0032a57d64ae045ba728cbd67eb6da82e3ec781ed1dede21da  ai-shell-mac

You also need llama-cli (llama.cpp) in your PATH; setup.sh / setup.bat fetch the Qwen2.5-Coder-3B model for you.

End-to-end proof (2026-10-30, Aetheris box, real model)

The linux-arm64 binary above is byte-identical to the one exercised live with the real Qwen2.5-Coder-3B model:

  • Benign + approve → ls executed in a systemd-run scope, real output shown
  • Benign + decline → nothing ran
  • Destructive intent → model refused; no dangerous command produced
  • Write intent → intercepted, rerouted to the editor; declined; no file created
  • help / aliases / alias interception exercised; history archives only on explicit yes
  • setup.sh downloaded the 1.8 GB model to the exact filename the binary loads

That proof found and fixed three real defects (llama-cli conversation-mode hang, prompt-echo parsing poison, Linux sandbox invocation that could never run) — see CHANGELOG 1.0.1.

Gate verdict

PASS WITH FINDINGS — release candidate. Machine scan PASS (0 MUST failures, 0 secrets); go vet clean; tests 13/13.

Known limits, stated plainly: no OS sandbox on Windows; on Linux the confirmed command runs in a systemd-run scope (no extra OS sandbox) — on both, the screening plus your confirmation is the control. Screening is finite string matching by design. Windows/macOS binaries are built and unit-tested but were not executed on those OSes in this proof; multi-tester certification is still to come.

Licence: Apache-2.0 — Copyright © 2026 2351485 ALBERTA INC.