[ docs · concepts ]
The concepts
A plain agent CLI does two things: it calls a model, and it forgets. Grist is the harness around the model that fixes both — it prices every task before spending, and it keeps what was proven so the next task starts ahead.
The pieces
| Concept | What it does |
|---|---|
| The Jev gate | Scores each task once and picks the cheapest rung that can handle it, with an effort budget to match. |
| The ladder | Four cost-tiered rungs, each holding the best model in its band. Clients only ever see rung names. |
| Specialists | Named subagents for exploring, planning, reviewing, and verifying — delegated work re-gates under the parent's rung as a ceiling. |
| SoL-Pi | The per-run efficiency layer: observations are packed, edits verify in the same move, context is reused instead of rebuilt. |
| Verified memory | Outcomes that were proven — tests passed, you approved, you corrected — are kept and recalled next session. |
| Dreaming & calibration | The across-session learning loop: memory is curated mechanically, and gate thresholds adjust from measured outcomes. |
| The code map | A graph of the repository — files, symbols, routes — that briefs the agent on structure before it reads a line. |
| The gateway | Accounts, API keys, hard spend caps, and BYOK routing. The only hosted piece; execution stays local. |
Two layers of efficiency
SoL-Pi is the efficiency layer inside a run — it makes a single session cheaper per unit of progress. Memory, the code map, and Dreaming are the experience layer across runs — they make the tenth session better than the first. The gate sits between them, pricing every task with everything the harness has learned so far.
The design rule underneath all of it: nothing enters memory unverified, nothing spends unpriced, and no source leaves your machine.