Changelog
What's new in aish.
Unreleased
Fixed
- Container worker image self-builds the aish binary —
Dockerfile.workeris now a multi-stage build that compilesaishinside arust:1-bookwormstage and copies it into a matchingdebian:bookworm-slimruntime, killing theGLIBC_2.39 not foundclass of opaque container-worker failures. Adds a.dockerignoreand a pre-launch--versionprobe that degrades to the host subprocess with a real diagnostic. - Shift-Tab with no coordinators is a silent no-op — no more screen clear/redraw when there's nothing to cycle to.
Changed
- Dev release reuses same-commit CI/CD builds instead of recompiling —
release-dev.ymlre-publishes byte-identical binaries when a matchingci-<run>-<sha>release exists, saving ~4–6 min of runner time per unchanged commit. - Collapsed tool-output activity stream + symmetric Ctrl-O toggle — the activity stream shows a
… N lines — Ctrl-O to expandsummary; Ctrl-O toggles the full last-turn output in and out.
Added
:goallong-horizon goals — durable, cross-session goal → milestones → tasks + blockers hierarchy in~/.aish/aish.db, injected into every turn.- Audible finish-bell when a background worker/batch/coordinator completes (
AISH_WORKER_BELL=0to mute,AISH_WORKER_BELL_CMDfor a custom sound). - Plugin system — skill-registry expansion — aish discovers plugins under
~/.aish/plugins/<id>/and merges each enabled plugin's skills into the catalog.
0.21.1
- Interactive system-prompt refresh to the current LightHeart persona.
- NEVER FABRICATE, ALWAYS VERIFY guardrail added to system + worker prompts: agents must confirm claims with real tool output, not assert unverified results.
0.21.0
- Lifecycle hooks —
PreToolUseblocking gate: hooks can block a tool call before it runs, not just observe it. :stopcoordinator stand-down channel — harder than:tell.- Parent session wakes when fanned-out coordinators complete.
read_filegains 1-based line-range slicing.- Ctrl-C interrupts an attached worker's current turn without killing the run.
- Local backend auto-downloads the detected GGUF from Hugging Face on first use.
0.20.0
- Release binaries now ship with the local backend built in (
--features local) — the publishedaishincludes the llama.cpp / GGUF backend out of the box, no from-source rebuild to run offline.
0.19.3
- Hardware-aware local model selection — the local backend inspects the host and picks an appropriate GGUF model/parameters for the detected hardware.
- Final, clean release of the llama.cpp local-backend line (recovering the burned v0.19.0 tag onto
main).