Posthorse starts a fresh context window instead of summarizing the old one

Posthorse is an extension for the fitchmultz fork of the Pi coding agent that changes what happens when a context window fills up. Instead of summarizing the conversation to make room, it starts a fresh window and keeps the complete transcript recoverable. The repository was created on 31 August and had 177 stars when checked on 5 September, under an MIT licence.

The problem it names is familiar to anyone running long agent sessions. When a model hits its token limit, the usual fix is compaction: the agent writes a summary of what happened and continues from that. The summary is lossy by design, and what it drops is chosen by a model that does not know which detail will matter in twenty minutes. The failure is quiet, because the agent carries on confidently with less information than it had before.

Rollover instead of compression

Posthorse's answer is to treat the full window as the point to start a new window rather than as text to shrink. A new_context call requests an atomic rollover after the current batch of tool calls succeeds, and it can carry an optional handoff that becomes the starting state of the new window. The old transcript stays on disk, unsummarized, and stays searchable.

Two supporting tools do the recovery work. A notes store at .pi/notes/ holds plain text that survives the package being removed and is shared across linked worktrees, with list, read, write, append and search operations. A history tool searches the stored transcript and returns entries with window identifiers and character offsets, so an agent that needs an earlier detail can go and read it rather than hoping the summary preserved it.

The thresholds are stated rather than implied. Posthorse needs at least 10,000 usable tokens after its reserve to operate. It issues one best-effort checkpoint reminder in the last 10 percent of usable context, capped at 32,000 tokens. A handoff is capped at 20,000 characters, or half the fresh window's operational capacity, whichever is smaller.

The adoption cost is high and stated plainly

Here the project is honest, and that also limits how many people can use it. Posthorse does not run on the official Pi agent. It requires the fitchmultz fork pinned to one named commit, because it depends on native context window entries and specific compaction hooks that the upstream project does not expose. Node 22.19.0 or later is required. Anyone not already on that fork is looking at an idea rather than a tool they can adopt this week.

The idea is the part other teams can use. Most agent harnesses today treat a full context window as a compression problem, and the design consequence is that the agent's memory quality degrades every time it fills up. Posthorse treats it as a paging problem instead, the way an operating system moves memory to disk and back when needed: keep everything, pass forward only what is needed to continue, and let the agent fetch the rest on demand. That distinction is available to any team building its own harness, whatever agent sits underneath it.

The 177 stars in five days say other people recognise the problem. Whether the rollover model beats summarization in practice is not something the repository measures, and no benchmark is published alongside it.

Source

fitchmultz/pi-posthorse, GitHub, created 31 August 2026.