fabricatio-context

Branch-based append-only prompt context for prefix-cache hits.

Overview

LLM providers cache prompts by exact token prefix. Composition code that interpolates request-specific content into the middle of a prompt invalidates the cache on every call. This package provides the shared data structure and capability for building prompts as stable heads plus cheap branches:

• [ContextEntry] — one frozen block of composed context (kind/title/body).
• [ContextLog] — append-only sequence of frozen entries held in an immutable tuple: with_entry/with_entries append purely, branch forks history in O(1), clear hands out a fresh log, render joins bodies deterministically.
• [AssembleContext] — capability mixin: seed a stable head once, fork a per-request branch, append request-specific entries, render byte-stable text.

Identical logs always render byte-identical strings, so a rendered head is a cache-stable prefix across every request and every retry round.

fabricatio-novel uses this pattern for running manuscript context; fabricatio-tool uses it to render execution feedback.

Usage

from fabricatio_context.capabilities.context import AssembleContext
from fabricatio_context.models.context import ContextEntry, ContextLog


class MyAgent(AssembleContext):
    def context_head(self) -> ContextLog:
        return ContextLog(entries=(
            ContextEntry(kind="rules", title="Rules", body="You SHALL ..."),
        ))


agent = MyAgent()
branch = agent.branch_context()
branch = branch.with_entry(agent.make_entry("request", "Req", "Count the lines"))
prompt_tail = agent.render_context(branch)

License

MIT — see LICENSE