Anthony Dawson (solo)

Contextus for Muse Code is a local-first context layer for coding agents. An agent forgets architecture decisions, project conventions, and past gotchas between sessions, so a developer ends up re-teaching it the same things. Contextus keeps that context in plain markdown you can read and inspect, with LanceDB as the retrieval index, not the source of truth, and injects it automatically at the prompt. Capture happens as you work, through post-tool-use and stop hooks, and through transcript mining, scraping past sessions for decisions and gotchas and turning them into corpus, not as a separate documentation phase. Retrieval is proactive, not reactive: relevant context is injected at the prompt, the model works from it without a separate query or search for that data, and a context-used footer at the end of the turn shows what was actually used. This entry ports Contextus from Claude Code to Muse Code, with a host-first layout, adapter/muse-code for Muse-specific hooks and manifests, and adapter/shared for host-neutral emission. Meta is a first-class citizen in this port, not a second-class wrapper around a Claude system: the same capture, retrieval, injection, and evaluation paths run on Meta models as run on Claude, with the same touchpoints and the same standing. Muse support is in PR #125 on the muse-support branch, pending review, not merged to main. Contextus is pre-existing work, authorised for hackathon use by Equip, with Equip credit. Demo: one diagnostic session for vector-inspector issue #42, order-dependent test failures, showing injection at prompt, no query or search during the turn, and the footer at the end. No fix is claimed, and no full suite result is claimed. Measured, same simplified diagnostic prompt: with-plugin 2m50s, without-plugin 6m44s.