npx agentspec-cli install ba23294b-bd1d-4524-a020-fce985cacb06 --file-type agents-md→ .agents/
# Grasshopper repository instructions Before substantive work, read `integrations/policy/AGENTS.md`. That file is the one canonical shared memory-use policy; this file governs repository work. If startup has not supplied memory context, call the connected Grasshopper `context` tool once when available. Read nested AGENTS.md before editing files in its scope. ## Product boundary Grasshopper is one authenticated, self-hosted SQLite memory service. Codex Desktop, Cursor, and Claude Code use one stateless client. Clients do not maintain writable memory databases. The remote service exposes only `context`, `store`, `search`, `get`, and `archive`; it cannot index server files. Make setup and everyday use straightforward across machines and agent harnesses; keep the memory model flexible as the product evolves. Remember explicit preferences, accepted decisions, verified lessons, and short handoffs. Scope global/project/device/platform records before retrieval or writes. A missing Git remote or durable project ID is unresolved, never global project context. Keep confirmed preferences stable until an explicit correction with expected revision. Preserve record IDs, provenance, revision history, legacy quarantine, and idempotency. New writes must be searchable immediately. Local code search from the old Rust CLI is retired for this release. Do not restore a second language runtime, separate writable store, transcript mining service, or parallel instruction files. ## Commands ```sh go test ./... go vet ./... go test -race ./internal/gomemory ./internal/gomcp ./internal/goclient go build ./cmd/grasshopper # Stateless client bridge and hooks go build ./cmd/grasshopper-go-server # Fresh or converted database go run ./cmd/grasshopper-go-backup --help go run ./cmd/grasshopper-go-migrate --help go run ./cmd/grasshopper-go-bundle --help ``` The real BGE ONNX tests need `GRASSHOPPER_BGE_TEST_ROOT` and, when outside the Mac research layout, `GRASSHOPPER_ONNX_RUNTIME_LIBRARY`. Do not call skipped model tests a real-model pass. The model and tokenizer digests are pinned in `internal/goembed/bge.go`. ## Layout and safety - `internal/gomemory/` owns scoped SQLite reads, writes, history, search, migration, and backup. `internal/goembed/` owns local ONNX embeddings. - `internal/gomcp/` exposes the authenticated five-tool HTTP service. `internal/goclient/` supplies the one stateless harness bridge and AGENTS.md hook adapter. `cmd/` contains their binaries and the archive builder. - Keep shared behavioral policy only in `integrations/policy/AGENTS.md`; use root or genuinely narrower nested AGENTS.md for project instructions. - Use synthetic or shadow database copies for tests. The migration never mutates its source. Back up and rehearse recovery before any live cutover. - Keep credentials out of Git, output, logs, exports, and repository identity. Bind the server to explicit loopback; private routing requires approval. - Record actual harness, version, OS, and observed behavior in `docs/memory-acceptance.md`; keep `tasks.md` short and current. Do not claim cross-harness or deployed success from unit tests alone. The old Linux service names in historical documentation do not prove a running deployment. Inspect the target host, supervisor, database, ports, credentials, and existing private routes before proposing a cutover. ## Linear tracking Repository: https://github.com/MylesMCook/grasshopper Team: Lab (`aed4ab18-a5db-4a5a-964e-4be027d7e529`). Project: [Grasshopper](https://linear.app/mcook/project/grasshopper-9d092487eee7) (`bff19887-4f6d-4256-956f-80d2521be8fe`). The owner authorizes finding, creating, and maintaining outcome issues and concise progress comments for meaningful work in this repository. Reuse an existing issue when applicable.
Codex Agent Instructions
These instructions apply to the entire repository.
by AgentHub Crawler
Cursor Rules — ai-memory
Use `memory.*` MCP tools as a persistent memory loop when available. **Namespace for all operations in this repo:** { "repo_url": "https://github.com/matt-antone/ai-memory" } Do not pass `agent` or `repo_name` — set automatically. Global items (no `repo_url`) are always included in repo-scoped searches.
by AgentHub Crawler
Substrate
This repo is an AI agent orchestration shell (roles, file-based substrate memory).
by AgentHub Crawler
Repository instructions
Sign in to leave a comment.
These instructions apply to all AI coding agents working on this repository. Everything here is resident: it loads at launch, in every session, before anything has been read. It is therefore limited to what a session could get
by AgentHub Crawler