Context for coding agents (Codex, Claude, and others) working in this repository.
This file is a pointer. The authoritative, detailed agent context lives in
CLAUDE.md — read it before making changes. It carries the current
sprint state, full architecture map, build/test commands, and the
non-negotiable design constraints.
- Project: Context Synapse — a local-first Bayesian prompt orchestration engine. Experimental / research-grade, not a consumer product.
- Language / platform: Swift 5.8+, macOS 13+. No external dependencies.
- Targets:
SynapseCore(library),contextsynapse(CLI),ContextSynapseApp(SwiftUI app). Tests inTests/. - Build:
swift build -c release - Test:
swift build && swift test --parallel— build first; some tests exec the CLI binary. - State: plain JSON under
~/Library/Application Support/ContextSynapse/. - Core types:
SynapseCore,Weights/Priors/Prior,Region,SynapseWeightState,SynapticCircuit. (There is noContextRegiontype — named embedding vectors areRegion.)
- Local-first: no required network calls; the AI clients are opt-in.
- Interpretability: all weights, priors, and similarity scores are human-readable JSON — nothing hidden.
- Fragility is intentional: fault injection is a first-class feature, not a bug to fix.
- No operational/affective-state modelling (ADR-002, referee track) — a permanent design boundary, not a roadmap gap.
See ROADMAP.md for the version plan, CHANGELOG.md
for history, and docs/adr/ for architecture decisions.