frxd/AGENTS.md

AGENTS.md

Repo shape

• rfc.txt (FRX — Federated Retrieval Exchange, Draft 0.5) is the normative spec; src/ is the Phase 1 frxd implementation (single crate, two binaries).
• DESIGN.md is the partner-facing architecture and decision log (with rationale for the spec cuts); keep it in sync when architecture decisions change.
• frxd is the member node (init/add/index/serve/relay/query/status); frx is the thin client (search/query/status). Relay and node roles are separate subcommands.
• DEPLOY.md documents the TLS/deployment story: members need no TLS (outbound HTTPS), relays terminate TLS with Caddy or a tunnel, ca_cert adds private CAs, allow_insecure opts into plain http on private networks, and non-loopback http:// is refused by default.
• Commands: cargo build, cargo test (104 tests: unit in src/; e2e tests/phase1.rs; conformance tests/conformance.rs; aggregates + member directory tests/aggregates.rs; registry tests/registry.rs; federation/isolation/admission tests/federation.rs; SSE tests/streaming.rs; encrypted unicast tests/encryption.rs; concurrency/restart tests/concurrency.rs; real subprocess CLI tests/cli.rs; 1000-doc tests/scale.rs; purge-log absence tests/purges.rs; shared fixtures tests/common/mod.rs). No CI/lint config. unit in src/; e2e tests/phase1.rs; conformance tests/conformance.rs; aggregates + member directory tests/aggregates.rs; registry tests/registry.rs; federation/isolation/admission tests/federation.rs; SSE tests/streaming.rs; encrypted unicast tests/encryption.rs; concurrency/restart tests/concurrency.rs; real subprocess CLI tests/cli.rs; 1000-doc tests/scale.rs; purge-log absence tests/purges.rs; shared fixtures tests/common/mod.rs). No CI/lint config.
• E2E pattern: relay + nodes in-process on ephemeral ports with tempdir corpora; use tests/common/mod.rs helpers (spawn_relay*, query_envelope, poll_messages, register) for new coverage. Raw relay polls return envelopes (payload under body), not response bodies.

Editing the spec

• Read all of rfc.txt before editing; it is the sole source of truth and is deliberately terse.
• Keep the plain-text single-file format. Don't restructure into Markdown files unless asked.
• Invariants I1–I9 (§2) are normative; proposals contradicting them (scores in responses, topic taxonomy, announce stream, dispute messages, replayable broadcast, in-protocol pricing/settlement) are out of scope by design.
• Appendix B (Purge Log) is normative: a rejected mechanism may only be re-proposed if the written rationale is addressed.
• §10 Open Issues are known gaps, not oversights (e.g., signature canonicalization blocks Phase-1 interop). Check it before "fixing" something.
• Use the spec's vocabulary — member/querier/responder, aggregates, source/enrichment members — not client/server or search-engine terms.
• A new MUST is only legitimate if it is observable at the boundary, deterministically verifiable by a peer, beneficial to the counterparty, and not derivable from local policy. Ranking/ordering/presentation fails this test and stays local (§5); scores never travel (I6).
• I2 is not blanket anti-centralization: shared coordination (identity, admission, contract) is centralized in the MA because common state is cheaper held once; decisions that consume local information (matching, relevance, sharing, retention) stay local. Off-wire conduct (link handling, retention, gating) is contract, not conformance.

Implementation notes

• Wire format matches Draft 0.5: envelope {type, from, key, ts, nonce, body, sig} with from = identifier, key = pubkey; signatures cover the JCS canonical form of the unsigned envelope under prefix FRX/0.5 (src/crypto.rs). Our canonical_json is JCS-compatible only for the restricted schema (ASCII keys, integers, no floats) — golden vectors in tests/conformance.rs pin the bytes and signature; revisit before claiming interop with non-Rust stacks.
• Relay addresses transport mailboxes by key (unicast to = recipient pubkey; queues keyed by pubkey); the identifier is protocol identity only. Registry binding checks map[key].id == from.
• Built: envelope/query/response, Tantivy index, aggregates, member directory. Not built: dashboard UI, directory watching, TLS, lineage/delegation. Responses travel relay-mediated unicast; transport is HTTP long-poll, not SSE.
• Economics is out of protocol scope (I4: aggregates advise, contracts govern): no receipt, citation, pricing, or settlement fields or message types exist or may be added.
• Relay verifies signatures and ±300s timestamp skew, carries only query broadcasts, holds no history (queues drained on poll), and requires challenge–response proof of key possession for mailbox polls. Nodes prefer SSE streams (/v1/stream, authenticated like polls) and fall back to long-poll on 404/405. Per-member queues are isolated: a lagging member gets 429 + Retry-After with a missed count; publishers and other members are never stalled. Relays MAY flood to configured peers (/v1/federation, --peer+--url, hop-bounded, seen-set dedup without suppressing identical direct publishes) and MAY gate senders against a registry (--registry+--ma-key).
• Responder searches only collections marked shared (I9), stays silent when nothing matches, and emits results with honest truncated/more_available and no scores (I6). BM25 order is a local implementation detail, not protocol surface.
• Index layout: Tantivy at <data_dir>/index, collections manifest at <data_dir>/collections.toml; exposure (metadata|full) gates whether content is returned.
• Egress checks live in the responder path (src/node.rs respond), not the relay — keep private collections unreachable there.
• Appendix B is executable policy: tests/purges.rs has one absence test per rejected mechanism (12 rows). Add a negative test there before ever re-proposing one, and only if the rationale is addressed.
• add/reindex reset a collection (delete by manifest name) before re-adding, so deleted files don't linger; collection identity is its name, and same-named collections replace each other.
• Relay backpressure is global: any member's full queue 429s every publisher until drained (visible per §3, but one lagging member can stall the firehose — revisit before scale).
• Member authority (Draft 0.5 §6): the MA-signed registry snapshot is authoritative when configured ([node] registry = file path or URL, ma_key pinned; monotonic version — rollback and forgery close the node; file path is mtime-reloaded, URL is fetched at start + every 60s and cached to <data_dir>/registry-cache.json, so outage fails static). Keys carry optional validity windows (not_before/not_after); rotation = registry add-key then revoke-key.
• <data_dir>/members.toml (name, pubkey, class, previous keys, mtime-reloaded) is a dev/local fallback used only when no registry is configured; empty directory without a registry is open bootstrap only when dev_bootstrap = true (RFC §6: explicit dev flag). Receivers drop content-bearing responses from enrichment-class senders (metadata-only, §6).
• MA tooling: frxd registry init|add|add-key|revoke-key|remove|list|applications|approve|invite|set-relays|show|serve (signed registry.json + ma-key.hex in --dir); frxd init --id/--registry/--ma-key; frxd key show|rotate; member add --previous <old> for the fallback path. A node with no [node] relays discovers them from the registry snapshot (doc.relays).
• Aggregate semantics are our implementation choices from a terse spec: requests are aggregate envelopes carrying only period; replies carry sent (broadcasts that month) / passed (responses consumed from that member); granularity floor is enforced as YYYY or YYYY-MM only (finer rejected), yearly rolls up months. Revisit with §10 sufficiency review.

Known gaps (Phase 2/3, intentional — don't fake them)

• No dashboard UI, no directory watching (new files need reindex), no TLS, no user-supplied URL ingestion, no node-side (bilateral) rate limiting.
• Unicast confidentiality is an implementation profile, not yet normative (RFC §10): response/aggregate bodies are encrypted to the recipient's registry-listed X25519 enc_key (x25519-hkdf-sha256-chacha20poly1305, src/crypto.rs); members without an enc_key get plaintext; relays verify signatures over ciphertext but cannot read payloads.
• Receipts/settlement are out of protocol scope, not unimplemented (I4; Appendix B row 2). Lineage and delegation remain §10 open issues — unspecified without a supply stream, so not buildable as written; don't invent them silently.
• Node query dedup is by qid only; replay inside the ±300s skew window remains possible (no nonce cache at nodes), relays have no directory/admission, and there is no end-to-end encryption — relays see everything in clear.

Technical plans (deliberately not in the RFC)

• Record plans here — not as spec edits — when they are implementation/demo choices rather than protocol surface.
• Demo plan: build a useful end-to-end demo on GDELT and Common Crawl (CC-NEWS; sometimes called "OpenCrawl" in discussion) as enrichment members / backfill seeding. RFC §6 and Appendix A already name both as example derived corpora, so no new mechanisms are required; enrichment members are metadata-only exposure.
• Phase 1 (two-node query/response) is built and tested; the enrichment demo layers on top of it.
• Language: Rust (settled, matches §7). Decided by the engine requirement, not preference: Tantivy gives in-process Lucene-class BM25 + incremental indexing; C/C++ embedded alternatives are worse (Xapian GPL-2+, CLucene unmaintained, SQLite FTS5 thin), plus single static musl binaries for the install story and memory safety on the untrusted network/crypto path. Don't re-litigate.
• frxd modes (one binary, config toggles, no code required of publishers): querier (broadcast/local-first search), responder (match incoming queries against shared collections, sign), local index (watch dirs, extract text, explicit shared marking per I9). Use RFC terms querier/responder, not "subscriber/publisher".
• Roles are not exclusive: a single node may issue queries and answer them concurrently (I5, §3 "any member"). Implement querier/responder as independent enable flags — never an exclusive mode enum or fixed deployment role.
• Matching floor: boundary tokenizer (src/tokenizer.rs — letter/digit splits so 5555 matches DLEX5555, lowercase, ASCII fold, English stopwords+stemmer) → coverage gate ([match] min_coverage, default 0.4; 1–2 term queries require all terms) → title boost 2.0 + phrase boost 3.0 + query-time snippets. Schema changes require a fresh index dir (open_or_create errors on mismatch).
• Engine seam: src/engine.rs SearchEngine trait (search → EngineOutput { hits, total: Option<u64> }, doc_count); respond() in src/node.rs is the conformance wrapper (budget clamp, truncation from engine total — unknown total forces truncated = true). Power users can implement the trait (HTTP adapter or subprocess to an external engine).
• Onboarding: frxd --onboarding runs a wizard consuming a credential block (id=.. token=.. registry=.. ma_key=..) issued by the MA (registry serve; HTML page at /, POST /v1/signup queues a pending application, POST /v1/enroll binds keys and re-signs). Identity registration stays MA-side; the wizard never creates identities, only binds locally generated keys. Applications live in <registry dir>/applications.json (mode 600, MA contract data — never in the signed snapshot); frxd registry approve <id> promotes one (member stub + invite + credential block), and frxd registry invite <id> mints another single-use 24h token — one per node the member runs. Invites live in <registry dir>/invites.json. Prompts accept empty input as the default; scripted stdin works for tests.
• Next matching steps: eval harness with a small golden set (precision@k + false-silence rate), then a dense recall leg (model2vec-rs 0.2.1 exists but needs default-features = false, features = ["fancy-regex", "local-only"] for musl/airgapped; verify crate + model licenses before bundling), then an optional cross-encoder reranker. Embeddings are for recall; reranking is the precision tier.
• Identity/registry (RFC Draft 0.5 §4/§6): MA-hosted FQDN identifiers first (<label>.frx.<ma-domain>, no DNS needed by users), signed versioned registry snapshot with the MA key pinned; envelope from = identifier, key = pubkey; registry outage fails static. Member-hosted identities, MA anchor rollover, and unicast confidentiality are §10 open. Implementation phases: A (signed registry snapshot) and B (identifier + key + JCS on the wire) are built and tested. Prioritize frictionless onboarding (users may be department-level and cannot create DNS).