S1CAP

System-1 Selective Context and Adaptive Planning

S1CAP: Selective Context and Adaptive Planning via System-1 Models for Efficient LLM Agents · 中文说明

Authors: Yuanming Chen · LI Changzhe

S1CAP puts a cheap System-1 decision model (Jev / Laya / Kev class, speaking the /v1/systemone protocol) in charge of an LLM agent harness's context lifecycle — instead of the expensive System-2 LLM. The S1CAP control layer intervenes at exactly two points:

Selective Context (context lifecycle) — every session segment (user turn, assistant message, reasoning trace, tool call/result) is a node in a growing association graph scored by the System-1 model. Each turn, bounded BFS + relevance threshold + token budget decide what the LLM sees, assembled in Trace-as-State order: [pinned prefix | state proxy T | recalled blocks | recent tail | current input].
Adaptive Planning (decision priority) — when the LLM proposes candidate plans, the same System-1 model scores them as a choice question with probabilities; plans execute in probability order under a verification oracle, and unexecuted alternatives are discarded on first success.

Everything is measured, not assumed: solve rate, token cost split by prompt-cache hit/miss (the dominant cost lever — cache-hit tokens are ~50× cheaper than misses on DeepSeek), and wall time excluding approval waits.

Why now (September 2026)

enabler	fact
Trace as State (arXiv:2609.02702)	placing the reasoning-trace state proxy before the long context beats trace-append in 26/27 model×task×metric combos — training-free, inference-time only
Decision models arrive	Jev (TypeSafe AI): $0.042/M input, output free, parallel question batches 12× cheaper than serial · open Laya (Apache-2.0) · EdgeJev local runtime: 322M INT8, 324 MB, 15.6 ms/decision on 4 vCPU
Cache economics	DeepSeek `deepseek-flash`: $0.006/M cache-hit vs $0.30/M cache-miss — context assembly is a cost decision, not just a quality decision

Architecture

                      ┌──────────────────────────────────────────────────┐
                      │           SYSTEM-1 CONTROL PLANE (S1CAP)         │
                      │   SEGMENTER → two-tier RECALL → RG STORE (SQLite)│
                      │   ASSEMBLER: BFS-τ-d recall, token budget,       │
                      │   Trace-as-State layout, cache-aware prefix      │
                      │   PLAN GATE: choice-question plan ranking        │
                      │   TELEMETRY: versioned JSONL (cache hit/miss,    │
                      │   cost, latency, S1 calls)                       │
                      └──────────┬───────────────────────┬───────────────┘
                 surface rewrite │ (DSH surfaceOp /      │ /v1/systemone
                 (model view only│  proxy message edit)  │ (Jev | Laya | EdgeJev | Kev)
                                 ▼                       ▼
                      ┌──────────────────────────────────────────────────┐
                      │   SYSTEM-2 COMPUTE PLANE (unmodified host LLM)  │
                      │   deepseek-flash / GLM-5.3, called by DSH /     │
                      │   opencode / Claude Code / pi                   │
                      └──────────────────────────────────────────────────┘

The user-facing transcript stays strictly chronological; only the model view is reassembled (native in DSH's session/surface split, replicated by the portable proxy elsewhere).

Evaluation design (pre-registered)

2×2 within-task paired factorial — factor A: Trace-as-State ordering; factor B: S1 governance (selection + plan gate):

Cell	A	B
C1 baseline	off	off (native compaction only)
C2	on	off
C3	off	on
C4 full	on	on

Benchmarks (all automated scoring, no GUI, no LLM judges): SWE-bench Verified (100/cell) · Terminal-Bench 4.0 (66/cell) · τ²-bench (full base split). Model: deepseek-flash (DeepSeek-V4.1-Flash), temperature 0.

Success rule: solve-rate non-inferiority vs C1 (paired McNemar, one-sided α=0.05, margin −2 pp) AND ≥10% improvement in cost/task or time/task (paired bootstrap 95% CI excluding 0, Holm-corrected). Winning cost while losing >2 pp solve rate is not a win.

Status & roadmap

Pre-alpha — M0 scaffolding landed: monorepo, @s1cap/core (segmenter · association graph · assembler · plan gate · telemetry v1), @s1cap/s1-client, dsh-s1cap skeleton, 2×2 cell presets; node --test 20/20 offline. Remaining M0: live-backend smoke against Jev / laya-serve. Full spec: docs/AGENT_BRIEF.md §10.

M	Scope
M0	monorepo scaffold, core + `s1-client`, telemetry v1, DSH plugin skeleton — scaffold done, live-backend smoke pending
M1	assembler/recall replay-correctness tests, proxy MVP, DSH hook wiring (`agent/pre-step`, surface ops)
M2	plan gate wiring, degradation paths, settings UI, Terminal-Bench 10-task cost pilot
M3	full 2×2 on SWE-bench Verified + τ²-bench (+ TB), optional Laya fine-tune
M4	Terminal-Bench cells, opencode transfer check, GLM model-swap check
M5	paper: Pareto + cache-waterfall figures, case studies, LaTeX draft

Development

node --test --experimental-strip-types "packages/*/test/*.test.ts"   # 20 tests, zero deps, offline

Type-checking needs TypeScript ≥ 5.8 (erasableSyntaxOnly): npm i -D typescript@^5.8 && npm run typecheck. pnpm is the intended workspace manager; on Windows PowerShell call pnpm.cmd (the .ps1 shim is blocked by the default execution policy).

Repository layout

s1cap/
  packages/core        # segmenter, association graph, assembler, plan gate, telemetry v1
  packages/s1-client   # /v1/systemone client + provider matrix
  packages/proxy       # OpenAI-compatible middleware (M1)
  packages/dsh-plugin  # dsh-s1cap: first-class DeepSeek Harness plugin (skeleton)
  bench/               # cells/ presets; runners + stats land with M2
  docs/                # proposal, implementation brief, related-work dossier, formulas
  paper/               # LaTeX (M5)

Documentation

doc	audience
docs/PROPOSAL.zh.md	research proposal — supervisor / cooperator (Chinese)
docs/AGENT_BRIEF.md	implementation brief for coding agents — verified facts base, interfaces, algorithms, milestones
docs/RELATED_WORK.md	verified related-work dossier + novelty audit

Environment

Verified on the dev machine (2026-09-28): Node v22.23.1 · npm 12.0.2 · git 2.45.2 · Python 3.13.13 (miniforge) · i7-12700H (AVX2 — EdgeJev-compatible) · DSH runtime bundles Node 24.18.1.

Requirements: Node ≥ 22.19 (DSH plugin engines contract) · pnpm for the monorepo (on Windows PowerShell, invoke pnpm.cmd or relax the execution policy) · Python ≥ 3.10 for local S1 runtimes (pip install "laya[serve]", pip install edgejev) · a System-1 backend: cloud Jev key, or local EdgeJev/Laya for weak CPUs (324 MB, offline).

Name

S1CAP = System-1 Selective Context and Adaptive Planning. Note: "Selective Context" is also the name of an unrelated EMNLP 2023 token-level compression work (Xiao et al.); S1CAP's mechanism is segment-level association-graph recall for agent context lifecycle — the paper cites and explicitly distinguishes the two.

Acknowledgments

TypeSafe AI (Jev) · Convai Innovations (Laya) · yzfly (EdgeJev) · jaredpalmer (Kev) · Benchmark Heaven (JevBench) · Xu Zou & Jie Tang (Trace as State, arXiv:2609.02702) · snailium (DSH plugin engineering template) · DeepSeek Harness

Citation

Paper in preparation. Authors: Yuanming Chen, LI Changzhe. Title (locked): S1CAP: Selective Context and Adaptive Planning via System-1 Models for Efficient LLM Agents.

@misc{s1cap2026,
  title  = {S1CAP: Selective Context and Adaptive Planning via System-1 Models for Efficient LLM Agents},
  author = {Chen, Yuanming and LI, Changzhe},
  year   = {2026},
  url    = {https://github.com/Jaffe2718/s1cap},
  note   = {Paper in preparation}
}

License

TBD (MIT proposed).

dsh-s1cap

安装

说明文档