Tags: kdeps/kdeps

Nightly release 2026-08-06-05:31

Nightly release 2026-08-05-05:26

fix: agent loop startup hang, GGUF crash remediation, system-role fol…

…ding, /prompt command

- Bound the unbounded llmfit exec call at startup with a timeout and add a
  startup status line, fixing a silent multi-minute hang on first run.
- Fix extractTarGzSiblings to preserve symlinked sibling libraries (e.g.
  libllama-common.0.dylib), and extend macOS crash remediation to reinstall
  llama-server for any missing-dylib crash, not just libomp.
- Fold system-role messages into the leading message for local GGUF/llamafile
  backends on both the raw-map (workflow) and langchaingo (streaming/REPL)
  request paths, fixing llama-server rejecting requests outright for chat
  templates that don't support a system role (e.g. Gemma).
- Fix a data-loss bug where CacheControl-wrapped system preamble content
  (memory rules, tool guidance, skills, date) was silently dropped instead of
  folded/displayed, because langchaingo wraps cached parts in CachedContent
  rather than plain TextContent.
- Exclude mmproj (multimodal projector) files from the model harvester so
  they're never catalogued as standalone servable models.
- Soften the single-task goal directive to consider tools for time-sensitive
  queries instead of unconditionally answering from memory.
- Add /prompt and /prompt raw REPL commands showing the exact messages and
  tool definitions sent to the LLM on the last call, with local token
  estimates.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

Nightly release 2026-08-04-05:29

fix: decoy process in a Serve() test now misread as a crash

TestLlamafileManager_Serve_StartsAndBecomesHealthy spawns a "#!/bin/sh
exit 0" placeholder solely to satisfy cmd.Start() -- the actual health
endpoint it polls is a separate httptest server, unlocked 100ms later.
The new process-exit crash detection now races that: the decoy exits
in a few milliseconds, well before the 100ms health-ready delay, so it
was misread as *ServerCrashedError instead of the test ever reaching a
successful health check.

Have the decoy sleep briefly instead of exiting immediately, so it
stays alive across the health-ready window like a real (successful)
llamafile process would.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

chore: update coverage badge [skip ci]

fix: GGUF/llamafile local backends ran CPU-only despite GPU builds

Both local backends were leaving GPU hardware idle:
- GGUF's installed llama-server binary is now GPU-first (Vulkan on
  Windows/Linux, Metal on macOS via the existing generic build) and
  passes --n-gpu-layers, which was never set before -- so even a
  GPU-capable binary was silently running every layer on CPU.
- llamafile never passed -ngl either, so its embedded tinyBLAS GPU
  backend was never engaged.
- Dropped --no-mmap from llama-server's launch args so it can lazily
  page a model in instead of reading the whole file into RAM up front
  on every start.

A GPU launch that fails to become healthy (no compatible driver) now
falls back to CPU-only once and remembers that decision on disk
(~/.kdeps/bin/.gguf_cpu_fallback, .llamafile_cpu_fallback), so later
runs skip straight past a GPU path already known not to work on that
machine.

Also pin the llama.cpp release kdeps installs to a tested tag (b10236)
instead of always fetching GitHub's "latest", since an untested
upstream release can change asset names or behavior without warning.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

Nightly release 2026-08-03-05:58

fix: tasklist flags mangled by Git Bash's POSIX-path auto-conversion

The previous fix used tasklist /FI ... /FO CSV /NH, but Git Bash's
automatic path conversion rewrites a bare "/FI" into a filesystem path
before tasklist ever sees it (confirmed: "/FI" became "C:/Program
Files/Git/FI", producing an "Invalid argument/option" error that the
count logic silently swallowed as 0). taskkill already used the
doubled-slash workaround (//F //IM); apply the same //FI //FO //NH
form here.

Verified end-to-end locally: a real long-named dummy process is
correctly found and counted via the doubled-slash invocation.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

chore: release v2.2.0 — Windows compatibility

Full Windows support: test suite passes on windows-latest CI, path
handling fixed across templates/cmd/tui, PowerShell/batch installers
and build scripts, ANSI console color handling, coverage-badge CI fix.

#minor

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>