cgh-summarize 0.1.0

File summaries for cgh behind a confidentiality egress gate: agent CLIs, Ollama or any OpenAI-compatible endpoint, plus corpus-level insights

cgh-summarize

Prose summaries of indexed files for cgh, produced by whatever model you already have, behind a confidentiality egress gate.

pip install cgh-summarize
cgh summarize status      # detected backends, gate posture, coverage
cgh summarize run         # summarize what the gate clears
cgh findings --key summary
cgh insights              # cross-file patterns from the summaries

Backends

Backend	Runs	Egress
cli:claude	claude -p, light model	cloud
cli:gemini	gemini -p, flash tier	cloud
cli:codex	codex exec	cloud
ollama	local Ollama daemon, model is a config line	none
openai	any OpenAI-compatible endpoint (vLLM, LM Studio, watsonx, ...)	cloud
structural	cgh's own outline, no model at all	none

backend = "auto" (default) picks the first available in that order. Third-party plugins add backends through the summarize.backend extension namespace without touching this plugin.

The egress gate

Before any cloud backend sees a file, its findings are checked: a confidential flag or any block-severity finding (private keys, cloud credentials) stops it, PII findings stop it unless allow_pii = true. With mode = "secure" in the cgh config, the gate switches to allowlist: only files explicitly labeled non-confidential go out. Local backends (ollama, structural) bypass the gate since nothing leaves the machine. Every cloud call is logged to .codegraph/activity.log.

Configuration

[plugin.summarize]
# backend = "auto"          # or cli:claude, ollama, openai, structural...
# min_kb = 4                # skip files smaller than this
# allow_pii = false
# language = "en"
# claude_model = "haiku"
# gemini_model = "gemini-2.5-flash"
# ollama_model = "qwen2.5:1.5b"
# ollama_url = "http://127.0.0.1:11434"
# openai_base_url = ""      # e.g. http://localhost:8000/v1
# openai_model = ""
# openai_api_key_env = "OPENAI_API_KEY"

Re-summarize policy: unchanged content is never re-scanned (blob SHA); when content changes, the old summary is carried forward while the drift stays under 30% of lines and fewer than 5 changes accumulated, then a fresh summary is produced.