[ docs · concepts ]

SoL-Pi

SoL-Pi is Grist's per-run efficiency layer. It sits in the harness, not in a cheaper model: the same model does the thinking, but the run stops feeding it noise. Same quality, fewer tokens.

Where the tokens actually go

Most of an agent run's context isn't thinking — it's plumbing. A test command dumps ten thousand tokens of log to deliver one failing line. An edit takes one call to make and a second call to check. A specialist restarts cold and re-reads files it read an hour ago. SoL-Pi attacks exactly those three leaks.

The three mechanisms

• ObservationPack. Huge tool output is packed into a handle plus an excerpt. The model sees the shape of the result and the part that matters; the full text stays on disk and is fetched only if the run asks for it.
• Fused edit-and-verify. An edit runs its check in the same move — one call, not a round-trip of edit, wait, check, read. The verification result comes back with the edit itself.
• Turn compression. Session context is carried compactly from step to step, and repeat specialist calls resume warm instead of restarting cold and re-reading the world.

What it is not

SoL-Pi is not a smaller model, a summarised-away codebase, or a quality trade. The rung still decides how smart the run is; SoL-Pi decides how much of the run's budget is spent on judgment instead of log spam.