The managed llama-server default ctx-size was 8192, but the full OpenClaw
agent system prompt alone is ~31K tokens, so the first agent turn overflowed
the context window and forced immediate compaction (observed live on the Mac
app local-model onboarding). Raise the default to 65536 so a fresh local-model
install can run a real agent turn out of the box.
The default-download 16 GiB RAM floor already bounds weaker machines, and
Gemma 4 supports far more than 64K, so this only changes headroom, not the
offer gate. Docs updated to match.
* feat(codex): pass explicit skill selections as structured turn input
Thread OpenClaw-resolved explicit $skill selections through the reply
pipeline to the Codex app-server harness and emit them as structured
UserInput::Skill turn items when the Codex catalog (skills/list) knows the
same SKILL.md path. Codex then injects the skill natively and blocks its
text scanner from double-selecting the same name, while unmatched names
stay scannable for Codex-native-only skills and the prompt instruction
block keeps owning skills Codex cannot see. Catalog misses and RPC errors
fail open to the instruction path; the turn never fails on skill lookup.
Part of #123367
* fix(plugins): defer llama-cpp ssrf-runtime load out of the setup closure
The doctor-contract closure guard forbids the SSRF barrel in the
doctor/legacy-setup closure (cold-loads DNS, proxy state, logging); the
static import landed with the managed llama-server switch (#123105) and
first fails when a PR selects the guard lane. Defer to a dynamic import at
download time per the guard's stated remedy.
Move llama.cpp chat and local embeddings onto a verified externally managed llama-server runtime. Remove the in-process native runtime, forked embedding workers, and node-llama-cpp dependency while preserving guided setup, local GGUF models, tool-capable agent runs, diagnostics, and operator docs.
* feat(llama-cpp): add in-process text inference
* test(llama-cpp): narrow setup provider fixture
* fix(llama-cpp): trim public surface and refresh docs map
* fix(llama-cpp): import Context type in inference test
* feat(memory): apply outputDimensionality truncation to local GGUF embeddings
The outputDimensionality config field was passed through to the local
embedding provider but never applied. Local GGUF models (e.g.
Qwen3-Embedding-0.6B) always returned their full dimension vector.
Apply slice(0, N) after normalization so MRL-capable models can benefit
from dimension truncation — matching the behavior already supported by
Gemini embedding-2 and OpenAI providers.
Fixes#58765
* fix(memory): preserve local embedding dimensions through worker
---------
Co-authored-by: Vincent Koc <25068+vincentkoc@users.noreply.github.com>
* fix(memory): accept local default model path migration
Treat the official local default embedding model's hf URI and downloaded GGUF path identities as equivalent so upgraded local memory indexes do not pause solely on path-format changes.
* fix(memory): satisfy local identity lint
Avoid filtered array tail access in the local model filename helper while preserving the same compatibility behavior.
* fix(memory): preserve local embedding identity aliases
---------
Co-authored-by: Vincent Koc <25068+vincentkoc@users.noreply.github.com>
Peter Steinberger