yvgude--lean-ctx
wehub-resource-sync
26382a7ac6
CI / Clippy (push) Failing after 15m13s
CI / Test (ubuntu-latest) (push) Failing after 16m1s
CI / Test (macos-latest) (push) Has been cancelled
CI / Test (windows-latest) (push) Has been cancelled
CI / Build (no embeddings / no ORT) (push) Has been cancelled
CI / Format (push) Has been cancelled
CI / Cookbook (Node) (push) Has been cancelled
CI / Pi Extension (Node) (push) Has been cancelled
CI / Rust SDK (lean-ctx-client) (push) Has been cancelled
CI / Embed SDK (lean-ctx-sdk) (push) Has been cancelled
CI / Python SDK (leanctx) (push) Has been cancelled
CI / Hermes Plugin (Python) (push) Has been cancelled
CI / SDK Conformance Matrix (push) Has been cancelled
CI / Coverage (push) Has been cancelled
CI / cargo-deny (push) Has been cancelled
CI / Adversarial Safety (push) Has been cancelled
CI / Benchmarks (push) Has been cancelled
CI / Output-Quality Gate (eval A/B) (push) Has been cancelled
CI / Documentation (push) Has been cancelled
CI / CI Green (push) Has been cancelled
JetBrains Plugin / Actionlint (push) Has been cancelled
CodeQL / Analyze (actions) (push) Has been cancelled
CodeQL / Analyze (javascript-typescript) (push) Has been cancelled
CodeQL / Analyze (rust) (push) Has been cancelled
JetBrains Plugin / Validation (push) Has been cancelled
JetBrains Plugin / Build (push) Has been cancelled
JetBrains Plugin / Test (push) Has been cancelled
Security Check / Security Scan (push) Has been cancelled
84 行
2.9 KiB
Markdown
84 行
2.9 KiB
Markdown
# Cognition Interface (v1)
|
|
|
|
LeanCTX cannot modify proprietary model weights. Instead, it ships a **Cognition Interface**: a deterministic control surface that shapes the model’s *effective* reasoning by controlling **what it sees**, **how it is budgeted**, **what is remembered**, and **what must be verified**.
|
|
|
|
This is production-realistic for API LLMs and becomes even stronger when paired with open-weights models in an optional Cognition Lab track.
|
|
|
|
## What “Cognition Interface” means in practice
|
|
|
|
### 1) Context I/O (signal in)
|
|
|
|
- Deterministic reads/search/shell output with explicit tool calls.
|
|
- Bounded outputs (size caps, truncation markers).
|
|
- Sandboxed file access (PathJail / allowed roots).
|
|
|
|
Evidence:
|
|
- `rust/src/core/pathjail.rs`
|
|
- `rust/src/core/output_verification.rs`
|
|
- `rust/src/core/cache.rs`
|
|
|
|
### 2) Orchestration (routing + budgets)
|
|
|
|
- Profile-driven pipelines (“Context as Code”): read modes, budgets, verification, autonomy.
|
|
- Intent/mode prediction and adaptive thresholds (bandits) to keep cost/quality stable.
|
|
- Client constraints compilation: the same policy must compile into client-safe instruction blocks.
|
|
|
|
Evidence:
|
|
- `rust/src/core/profiles.rs`
|
|
- `rust/src/core/intent_engine.rs`
|
|
- `rust/src/core/mode_predictor.rs`
|
|
- `rust/src/core/adaptive_thresholds.rs`
|
|
- `rust/src/core/instruction_compiler.rs`
|
|
- `docs/integrations/client-constraints-matrix-v1.md`
|
|
|
|
### 3) Memory (what persists)
|
|
|
|
- Session continuity (CCP), structured knowledge, contradictions/relations.
|
|
- Exportable handoffs and auditability across agents.
|
|
|
|
Evidence:
|
|
- `rust/src/core/session.rs`
|
|
- `rust/src/core/knowledge.rs`
|
|
- `rust/src/core/gotcha_tracker/*`
|
|
- `rust/src/core/a2a/*`
|
|
|
|
### 4) Verification (what must hold)
|
|
|
|
- Deterministic checks on compressed outputs (paths, identifiers, structure, line numbers).
|
|
- Proof artifacts and CI gates to prevent “it worked yesterday” drift.
|
|
|
|
Evidence:
|
|
- `rust/src/core/output_verification.rs`
|
|
- `CONTRACTS.md`
|
|
- `rust/tests/*_up_to_date.rs`
|
|
|
|
### 5) Delivery (everywhere)
|
|
|
|
- MCP + HTTP MCP + Team Server let the same primitives run locally, in CI, and enterprise setups.
|
|
- SDK + cookbook runs against a real server instance (no mock mode).
|
|
|
|
Evidence:
|
|
- `rust/src/http_server/mod.rs`
|
|
- `rust/src/http_server/team.rs`
|
|
- `cookbook/sdk/src/client.e2e.test.ts`
|
|
|
|
## Contract: deterministic steering, not “prompt magic”
|
|
|
|
The Cognition Interface is only useful if it is:
|
|
|
|
- **Deterministic**: same inputs/policies → same compiled output.
|
|
- **Bounded**: caps enforced per client/model constraints.
|
|
- **Auditable**: evidence artifacts and CI gates catch drift.
|
|
- **Local-first**: no telemetry unless explicitly enabled.
|
|
|
|
## Optional: Cognition Lab track
|
|
|
|
For open-weights experimentation (or internal models), the same interface becomes a research harness:
|
|
|
|
- learned attention/layout drivers
|
|
- calibration/evaluation suites
|
|
- ONNX models with versioning + rollout policies
|
|
|
|
See: `docs/cognition-lab/plan-v1.md` (tracked in GitLab `#2344`).
|
|
|