Choose a preset¶

A preset is a named, ordered pipeline. Set it once and every explicit field you add overrides it.

context-guru-proxy --preset codesmart      # or PRESET=codesmart, or preset: in YAML

Which preset?¶

What each one runs¶

Order is deliberate: lossless repack first, then the cheap structural offloaders, then anything that costs a model call, cache directives last — except in house and housellm, whose order is the operator's on purpose: dedup and cmdfilter run ahead of the lossless pair and toolfilter sits after cachesplit. That costs per-component attribution in /stats, never content; the reasons are recorded in config/config.go and the exemption is noted in the preset reference.

The last three rows are not options in the chooser above. house and housellm are the service configs — what a hosted account runs unless it asks otherwise — and agentdiet reproduces a published baseline for A/B comparison, not a recommendation. They are in the table so it lists every preset that exists; pick from the table above this one.

Notes on the ones people pick¶

cache — start here¶

cachesplit and nothing else. Pick it when what you want is to find out whether this thing helps you, with the smallest possible claim to check:

• Nothing is dropped, summarised, or replaced. No <<cg:HASH>> markers, no context_guru_expand tool added to your requests, no model calls. It splits one oversized system block into two adjacent text blocks whose concatenation is byte-identical, so the model sees exactly the prompt your agent sent — and moves the cache breakpoint onto the half that does not churn.

• What it is worth is regime-dependent, and the funnel's regime is the weak one. The headline −34.1% cost / 0% → 96.7% hit comes from a benchmark harness running tasks back-to-back inside the provider's 5-minute cache TTL (cacheinject), which is precisely the regime where the split pays — and is one task measured three times, not a fleet average. On this project's own interactive traffic the figure is $0.0298 across 1,127 sessions / 11,361 requests (dashboard): Claude Code captures the environment snapshot once per session, and 1,105 of 1,127 session starts read zero tokens from cache because the previous prefix had already expired. It is also exactly zero outside a git repository, on a system prompt under the 1,024-token minSplitTokens floor, and on any implicit prefix-cache backend (vLLM, llm-d). Neither figure is wrong; they differ by three orders of magnitude because the mechanism needs a second session inside five minutes.

• Anthropic-family only. cachesplit is a no-op against implicit prefix-cache backends (vLLM, llm-d), which match to the divergence on their own — so on those it costs nothing and buys nothing.

Move to codesmart once you want the offloaders too. safe is the next step up and is still lossless in meaning, but it does rewrite JSON, so cache is the one whose promise you can confirm by reading a single line of config/config.go.

codesmart is the shipped default and the cheapest arm in the benchmarks at the highest reward. It is the one preset that ships tuned per-component settings rather than a bare name-list, which is why most turns make no model call at all.

codesafe is codesmart with the LLM pass swapped for a blind collapse. Zero model calls by policy — the choice for a shared box, an air-gapped deployment, or a benchmark arm that must be reproducible byte for byte. collapse keeps a head/tail window and stashes the middle, so nothing is unrecoverable, but it has no idea what mattered.

safe is two lossless components. Nothing is ever dropped, so there is nothing to expand and no reversibility surface at all.

agent and general are the long-session choices, and mask — age-based GC of tool outputs older than keep_recent — is the biggest lever in both. In the SWE-bench sweep it delivered ~27% mean content-token savings (up to 93.5% on a long session) with no reward loss. balanced omits mask and is not the choice for a long agentic session: it delivered 6% against general's 31% on the Terminal-Bench replay.

coding is deterministic only: the components measured to actually act on real Claude Code traffic. Until August 2026 it named skeleton instead, which is behind a build tag and so is absent from a normal binary — meaning the preset could not start at all. skeleton still exists in a cg_skeleton build and is still not recommended; its page has the numbers and the reason.

summarize must run alone. It restructures the whole transcript, so no other component's in-place edits can share the request with it.

Presets are only defaults: an explicit pipeline: or components: block always wins, so start from the closest preset and tune. Full expansions live in config/config.go; per-component behaviour is in Components and the preset reference.