kami
968cbfa973
Found while live-QAing freestyle_planning on a 12B local model: - list_dir tool: recursive, .gitignore-aware listing so weak models stop flooding context with `ls -R` over node_modules/build/dist. Wired into the fileRead toggle + advertised to the planner (architect_freestyle). - ContextClassifier: assistantToolCall turns are STRUCTURED, so the token pruner never shreds the model's own tool-call history — that was causing amnesia loops (re-issuing calls it had already made). - Retire instruction-doc LLMLingua pruning (DOC_SOURCE_TYPES emptied): it fused load-bearing procedural text into unparseable soup. The static block stays small by dropping CLAUDE.md at the loader instead. - AgentInstructionsLoader: load only AGENTS.md, not CLAUDE.md — the latter targets the outer assistant and polluted the agent's stage context. - DefaultSessionReducer: WorkflowFailed flips session status to FAILED (was stuck ACTIVE forever, so clients/approval loops never saw a terminal). - ShellTool: run shell command lines (cd/&&/pipes) via `sh -c`; unrunnable program is recoverable instead of an uncaught IOException killing the stage; malformed argv (non-string/collapsed-array) rejected with guidance. - llmlingua sidecar: cap force_tokens to max_force_token (big docs blew the assert and 500'd, so doc pruning silently failed open). Tests added/updated across all of the above.
LLMLingua-2 token-pruning sidecar
Prunes low-perplexity tokens from freeform prose before it hits the local LLM, so more usable
context fits a bounded window. Implements pipeline stage 3 (TOKEN_PRUNE, level 3+) — see
docs/plans/correx-compression-pipeline.md §4.
Python-only because LLMLingua-2 is a torch/BERT classifier with no JVM equivalent. correx calls
it over localhost HTTP via HttpTokenPruner, which fails open: if this sidecar is down, the
kernel passes context through uncompressed. Nothing breaks; you just don't get token pruning.
Run
cd sidecars/llmlingua
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
uvicorn server:app --host 127.0.0.1 --port 8199
First /prune call downloads the model (~1-2 GB) and loads torch; /health responds immediately.
Wire into correx
Set compression level ≥ 3 for the workflow and point the kernel at the sidecar:
[compression]
level = 4
token_pruner_url = "http://127.0.0.1:8199"
API
GET /health→{"status":"ok"}POST /prune{"text": str, "protected": [str], "rate": 0.55}→{"compressed": str}rate= fraction of tokens to keep (0.55 ≈ 45% compression)protectedsubstrings (IDs, numbers, paths, code) are kept verbatim