Skip to content

Configuration

RustySqueezer's extraction rules are tunable without recompiling. Mount a JSON file with CONFIG_PATH, or override individual knobs with environment variables (env wins over the file).

Mount a config file

docker run -d -p 8000:8000 \
  -v ./my-config.json:/config.json:ro \
  -e CONFIG_PATH=/config.json \
  ghcr.io/handtrixx/rustysqueezer:1.0.6

The canonical set of keys is in config.example.json.

Keys

Key Default Meaning
model_name rustysqueezer Reported model id on discovery endpoints
max_summary_chars 12000 Hard cap on the generated summary
tail_turns / tail_chars 6 / 600 How many recent turns are kept verbatim; tail_chars is a per-turn floor, and the newest turn gets half of the tail section's budget (v1.2.2)
max_commands 25 Shell commands retained (most recent first)
max_files 60 File paths retained
max_errors 20 Error lines retained
max_key_points 40 Salient assistant sentences retained
instruction_patterns regex list Phrases that identify a compaction instruction (Zed, /compact, "summarize the conversation so far")
instruction_max_chars 4000 Treat the matched instruction as content if longer than this
error_pattern regex What counts as an error line
file_extensions list Extensions considered source files
sections ordered list Which summary sections to emit, in order
section_budgets shares Per-section share of max_summary_chars (tail first)
section_titles strings Headline text for each section
template_headings false Echo the harness template's own ## headings from the compaction instruction (OpenCode validates them verbatim)
assist_provider openai Backend dialect for the key-point condensation: openai (llama.cpp etc.) or google (Gemini REST)
assist_base_url "" OpenAI-compatible server for the assist; empty = assist off
assist_api_key "" API key for provider google; empty = assist off (base URL then optional)
assist_model qwen3.5-2b Assist model id (gemini-3-flash for Google)
assist_timeout_ms / assist_max_chunks / assist_chunk_chars / assist_max_tokens 45000 / 3 / 700 / 140 Assist chunking and per-chunk limits
forward_provider openai Forwarding dialect: openai or google; forwarding is off unless the provider's credential and forward_model are set
forward_base_url "" OpenAI-compatible upstream for forwarding
forward_api_key "" API key for forwarding with provider google
forward_model "" Upstream model to forward to; empty = forwarding off
forward_timeout_ms / forward_max_tokens 90000 / 2048 Upstream timeout and answer cap

Small-model assist

The key-points section may be condensed by a second, small model (map-reduce over bullet chunks). assist_provider selects the dialect:

  • openai - a local llama.cpp server, e.g. assist_base_url: "http://rsq-aux:8080/v1" with assist_model: "qwen3.5-2b".
  • google - the Gemini REST API, e.g. assist_provider: "google", assist_api_key: "AIza...", assist_model: "gemini-3-flash"; the base URL defaults to https://generativelanguage.googleapis.com/v1.

Keys are read from the config file or ASSIST_API_KEY / ASSIST_BASE_URL env and are never logged. A failed or timed-out chunk keeps its original extractive bullets, so the summary is never worse than the pure extractive build.

Forwarding mode

With forward_model plus the provider's credential (forward_base_url or forward_api_key) set, requests are passed through to a real model instead of being answered extractively - useful when a harness wants a narrative compaction (OpenCode/OpenChamber /compact pointing at Gemini). Tool schemas are sanitized to the keyword subset Gemini accepts (exclusiveMinimum/ exclusiveMaximum become inclusive integer bounds, local $refs inlined, draft-2020-12 keywords like $schema, additionalProperties, patternProperties dropped); on the native Google provider tools are dropped outright. Any upstream failure (timeout, HTTP error, unusable body) falls back to the extractive summary, so the endpoint never fails because of the upstream. Forwarding stays off by default: the extractive path remains the local default.

Environment overrides

Variable Overrides
PORT Listen port (default 8000)
MODEL_NAME model_name
MAX_SUMMARY_CHARS max_summary_chars
TAIL_TURNS / TAIL_CHARS tail settings
MAX_COMMANDS / MAX_FILES / MAX_ERRORS / MAX_KEY_POINTS section caps
ASSIST_BASE_URL / ASSIST_PROVIDER / ASSIST_API_KEY / ASSIST_MODEL assist backend
FORWARD_BASE_URL / FORWARD_PROVIDER / FORWARD_API_KEY / FORWARD_MODEL forwarding upstream
TEMPLATE_HEADINGS 1/true enables template heading echo

Determinism is not affected by tuning

Changing budgets changes how much is kept, not which content is chosen at random. The same input with the same config always produces byte-identical output.