# AGENTS.md — tines Instructions for AI agents working on **tines**: opportunistic shared memory for mostly-offline robots. Read this before touching code. ## What tines is Small robots collect facts about the world while mostly offline. When two of them happen to meet in range, they swap what they know so each one's picture of the world improves. No robot depends on receiving anything, but there is no reason not to share. `tines` is the library that does this: a store of timestamped-by-age observations, a belief layer derived from them, and a delay-tolerant exchange protocol over any datagram link. Read the full intent, goals, anti-goals, and architecture in the **spec first**: - **Core design (authoritative):** `docs/superpowers/specs/2026-10-16-tines-design.md` - **Addressed delivery (authoritative):** `docs/superpowers/specs/2026-10-18-tines-addressed-delivery.md` - **Fleet deployment v1 (authoritative):** `docs/superpowers/specs/2026-10-18-tines-fleet-deployment-design.md` — supersedes the addressed-delivery spec's epidemic forwarding (custody chain with handoff receipts). - **Fleet fixes (done):** `docs/superpowers/specs/2026-10-19-tines-fleet-fixes.md` — security-assessment findings F1–F5 and F17. - **tines-core 1.0 (next, core):** `docs/superpowers/specs/2026-10-18-tines-1-0-design.md` — wire version, public API, signature slot, core hardening (§10). - **Host hardening (next, hosts):** `docs/superpowers/specs/2026-10-19-tines-host-hardening.md` — findings F10–F12. - **Implementation plans:** `docs/superpowers/plans/` The plans argue from the specs. Where a plan is silent, the **spec rules**. Do not edit the specs or the plans on your own initiative — they are design history, not source. Change one only to record something Kyle has discussed and agreed, and record it as a dated decision in that document. Anything else that needs changing in a spec goes to Kyle first. `README.md` is the user-facing overview and should track the code. `THREAT_MODEL.md` is the living security assessment: when a change fixes or changes a finding, update its status there (§1 and §10) with the commit. `CHANGELOG.md` records each release. ## Architecture (three layers) ``` memory — "my priors": beliefs per topic, derived from storage under a policy storage — append-only log of raw observations; local trimming policy comms — contact-driven exchange of observations; transport trait + adapters ``` Only observations cross the wire. Memory is a pure function of storage plus a policy and is never transmitted. See the spec's "Architecture" section for the `Observation`, `Store`, `BeliefPolicy`, `Transport`, and exchange protocol details. ## Workflow Changes are TDD: write the failing test → see it fail → implement → see it pass → commit. 1. Keep the public API small. New public names in `tines-core` (types, trait methods, functions) need a reason; call them out when you add them. 2. Before committing, run the full gate, which is what CI runs: `sh scripts/ci.sh` (fmt, clippy and rustdoc with warnings as errors, tests with and without `alloc`, and the `thumbv7em-none-eabihf` build). 3. Use conventional commit messages (`fix(core): …`, `ci: …`, `docs: …`). The `Node` facade is core, not application glue: an application on any target should only need to construct a `Node` with its `NodeId`, store, and transport, call `poll()` in its loop, and `observe()` when it sees something. ## Hard constraints (non-negotiable in `tines-core`) - `tines-core` is `#![no_std]`; `alloc` is a cargo feature (default on). With `--no-default-features` it must build for `thumbv7em-none-eabihf`. Only `MemStore` may require `alloc`. - No async, no threads, no I/O, and no panic on untrusted input in core. Every frame from the wire is untrusted: malformed bytes are dropped, never unwrapped. - Core never maps a transport `Addr` to a `NodeId`; every frame carries the sender's `NodeId`. - Identity hash covers `(observer, seq, topic, payload, to)` **only**. `age_ms`, `observed_at` and `lifetime_ms` are excluded. - Addressed observations (`to = Some`) never appear in a summary Bloom filter, and an addressed observation past its `lifetime_ms` is never stored or forwarded. - `first_seq` passed to `Node::new` must not repeat across reboots (`boot_count << 32`, with a persisted boot counter; `LogStore::first_seq` provides it). A repeated seq after reboot collides in the ID hash and must dedupe silently (`insert` returns `Ok(false)`). - Receipts: a receive the store rejects (`Err`, e.g. `Full`) is never acked, addressed or broadcast; an addressed duplicate whose local copy is released is never acked; an `Ack` is accepted only from a peer the id was sent to this contact. - `NodeId = [u8; 16]`, `ObservationId = [u8; 16]`, `Tick = u64` (local monotonic ms, supplied on every call). - 1.0 trait extension rule: any method added to `Store`, `Transport` or `BeliefPolicy` after 1.0 must have a default whose behaviour is safe — it may lose performance or telemetry but never data or custody. A method that can't have such a default is a 2.0 change (the rule `release` violated in 0.3.0). ### Review Focus (each covered by a test; keep it that way) 1. Reused `seq` after reboot → store dedupes, `insert` returns `Ok(false)`. 2. Oversized `topic`/`payload` at `observe()` → `Err(TooLarge)`, never truncate/panic. 3. Garbage frames → dropped, no panic, no store change. 4. Frame from an address with no prior `Contact` → treated as implicit contact. 5. Tick going backwards → age saturates at stored value, never underflow. 6. Every emitted frame fits the transport's `mtu()`, including summaries. 7. Store-full receive → no `Ack`; custody stays with the sender (unit test and the full-store sim property). 8. Released addressed duplicate → no `Ack`. ### Compile-time constants (`tines-core/src/consts.rs`) Split into **wire** (frozen for wire v1 — changing one changes how a v1 peer decodes or interprets a frame) and **local** (never crosses the wire; may change in a minor release): - **Wire:** `WIRE_VERSION=4`, `MAX_TOPIC=64`, `MAX_PAYLOAD=1024`, `MAX_FRAME=1500`, `SUMMARY_BYTES=512`, `ACK_IDS_CAP=90`, `MAX_SIG=64`, `BLOOM_K=3`, plus the private `MAX_FRAGMENTS=64`. - **Local:** `MAX_PEERS=8`, `SUMMARY_INTERVAL_MS=2000`, `REASSEMBLY_TIMEOUT_MS=5000`, `SEND_QUEUE_CAP=256`, `MIN_MTU`, `EVENT_QUEUE_CAP=64`, `DROP_TABLE_CAP=32`, `BCAST_SENT_CAP=256`, `ACK_BUFFER_CAP=256`. Match the exact values above. ## Crate layout ``` tines/ tines-core/ no_std core: Observation, Store/Transport/BeliefPolicy traits, exchange state machine, Node facade, in-memory store, reference policies tines-sim/ std: simulator, contact traces, convergence, relay and custody property tests tines-store/ std: LogStore, the persistent append-only backend tines-bt/ std: BLE transport over BlueZ (bluer) tines/ std: standalone service, Unix-socket JSON IPC tines-ros2/ std: ROS2 node; rclrs behind the `ros2` feature ``` Only `tines-core` is headed for 1.0; the host crates stay 0.x (1.0 spec, D4). They're std-only and may use threads and I/O, but they're still the trust boundary for radio, IPC and disk input (see `THREAT_MODEL.md`). `tines-ros2` with `--features ros2` is outside the default gate: check it with the ROS2 Docker job (`Dockerfile.ros2`, `scripts/check-ros2.sh`). ## Not started `tines-udp`, `tines-sqlite`, `tines-cli`, and `tines-serial` (from the core spec's crate layout) belong to a future plan. Don't start them unless asked.