# AGENTS.md
Guidance for AI coding agents (and humans) working in this repository.
## What this project is
`ffsync` is a file synchronization tool written in Rust that syncs data at the
chunk level to minimize transfer and I/O, similar in spirit to rsync. As of
v0.3, single-file sync **and directory-tree sync** between two peers over TCP
work end-to-end: parallel chunking/hashing pipeline, file
reconstruction, manifest/plan exchange, zstd-compressed transfer,
chunk-root verification, mtime/size fast path (sub-second mtimes),
resume after mid-transfer disconnects, the persistent chunk index for warm
starts (D15), TLS transport with mutual TOFU fingerprint pinning (D16), and
tree sync (D17: protocol v3 — TREE_BEGIN/TREE_END envelope, sequential
per-file conversations, mtime preservation on commit, deletes deferred to
session end, `--no-delete` opt-out) and tree-mode excludes (§16.7:
`--exclude`/`--exclude-from`, sender's patterns rule the session,
receiver's patterns guard the delete pass), and parallel per-file
conversations (§16.8, D13 step 1: `--parallel K` over K connections,
measured 3.4×–6.1× at 30 ms RTT), and the QUIC transport (`--quic`, §17,
D13 step 2: one QUIC connection per session, tree conversations ride
streams, same identity/pinning). Not implemented yet:
permissions/xattrs/symlink policy (D9).
Design documents (read before making architectural changes):
- `concept/Positioning.md` — why this project exists, the niche, the decision
framework (link speed decides what matters), and the rejected
"build around rsync" alternatives
- `concept/Design Decisions.md` — recorded decisions with reasoning and
alternatives; add new decisions there, do not silently change old ones
- `concept/Protocol v0.1.md` — the wire protocol (named for its first
release; it tracks the protocol across versions — currently v2: HELLO
gained `file_mtime_nanos`): framing, handshake, manifest/plan/data flow,
verification, resume. The implementation follows it; deviations are
protocol bugs unless the spec is amended first
- `concept/Deep Research Digest.md` — what the two deep-research runs
(architecture + P2P connectivity, saved alongside as markdown) got right,
and what we rejected
- `concept/File Synchronisation Workflow.md` — the four-phase sync pipeline
(shallow → medium → deep diff, then pipelined transfer and reconstruction)
- `concept/File Reconstruction.md` — out-of-place reconstruction with a
pre-allocated temp file and atomic rename
- `evaluation/chunking.md` — measured performance of chunking approaches and
the reasoning behind the current pipeline design
- `evaluation/memory.md` — measured memory footprint of the sync paths
(heap is O(chunks) ≈ 0.05% of file size; trees stay flat with file count)
## Commands
| `cargo build` | Build the binary |
| `cargo check --tests` | Fast compile check including tests |
| `cargo test` | Run unit and integration tests (`tests/integration.rs`) |
| `cargo run` | Run the CLI (`ffsync send/receive`, see `--help`) |
| `cargo run --bin bench` | Run the benchmark harness (fixture variants, delta simulation, reconstruction, link-speed projections; `--quick`, `--size-mib N`, `--seed N`, `--rsync`, `--disconnect`, `--warm`, `--tree`, `--latency` (+ `--rtt-ms`, `--parallel K,...`), `--json`) |
| `cargo fmt` | Format (default rustfmt style; no `rustfmt.toml`) |
| `cargo clippy` | Lint (no `clippy.toml`) |
| `cargo audit` | Check `Cargo.lock` against RustSec advisories (see D14 in `concept/Design Decisions.md`) |
CI (`.github/workflows/`) runs on push to `main` and on PRs, on
Windows **and** Linux: `cargo fmt --check`, `cargo clippy --all-targets
-- -D warnings`, `cargo test`, `cargo audit` — keep all four green
locally before pushing. A weekly scheduled workflow re-runs `cargo audit`
so advisories published between pushes are caught (D14).
Toolchain: current stable Rust (1.99.x as of 2026-10), edition 2024. The crate
is a lib + two bins (`ffsync`, `bench`); tests import via `ffsync::chunking` /
`ffsync::reconstruction`. Run `cargo audit` whenever dependencies change; on
fastcdc upgrades additionally fingerprint chunk boundaries before/after with
`cargo run --release --bin bench -- --quick --seed 42` (identical chunk counts
and missing-byte counts required — boundaries are wire-protocol-relevant).
The `--rsync` bench flag compares against rsync when available; on Windows it
falls back to WSL (`wsl.exe` + `wslpath` translation). Paths are passed as
separate args (never through a shell) so spaces are safe.
The crate may be mid-refactor and fail to compile. Before starting work, run
`cargo check --tests` to establish a baseline so you know which errors are
pre-existing.
## Architecture
- `src/lib.rs` — lib target; exposes `chunking`, `index`, `protocol`,
`reconstruction`, and `transport` (tests and the bench binary import
through it).
- `src/main.rs` — the `ffsync` CLI (`send`/`receive`/`trust`/`untrust`/
`fingerprints`): argument parsing, progress display, TCP setup, TLS
wiring, and error reporting. Thin by design; the sync logic lives in
`protocol::sync`.
- `src/bin/bench.rs` — the benchmark harness: generates fixture variants
(identical / modified / appended / fresh), scans both sides through the real
pipeline, simulates the delta against a hash set, reconstructs the target
through the real `Reconstructor` with hash verification, and projects
transfer times across link speeds and zstd ratios. Optional `--rsync`
comparison (native or via WSL), `--disconnect` resume simulation, `--warm`
warm-start index measurement. Zero dependencies beyond the crate's own.
- `src/chunking/` — the core data pipeline. The public API is re-exported from
`mod.rs`.
- `pipeline.rs` — `process_file()` is the entry point: it memory-maps the
file (`memmap2`), spawns one chunker thread per core (each striding over
256 MiB regions), and pushes `(offset, length)` work items through bounded
`crossbeam` channels to hash-worker threads, which slice the mmap directly.
Returns a `Receiver<ChunkEvent>`.
- `chunker.rs` — `Chunker` trait with two implementations: `FixedChunker`
(baseline) and `FastCDCChunker` (content-defined, via `fastcdc::v2020`).
- `hasher.rs` — `Hasher` trait; `Blake3Hasher` (in-memory and streaming).
- `types.rs` — `Chunk` (offset, length, BLAKE3 hash) and `ChunkEvent`.
- `src/index.rs` — the persistent per-file chunk index (D15, warm starts):
hand-rolled binary codec (magic `FFSYIX01`, 68-byte header, 44-byte chunk
entries), load-time self-verification (root re-derived from the chunk
list), size + nanosecond-mtime validation, atomic temp+rename writes, and
the `index_path_for` naming convention (suffix **appended**, never
`with_extension` — the orphan cleanup recovers target names by stripping
the literal suffix).
- `src/reconstruction/` — writes a new file out-of-place from chunk sources.
- `reconstructor.rs` — `Reconstructor`: pre-allocates a temp file, writes
each chunk at its target offset (local copy from the old file, or remote
payload), then verifies the final BLAKE3 hash and atomically renames.
- `io_ext.rs` — `FileExt` trait for positional I/O (`read_at`/`write_at`)
with separate `#[cfg(unix)]` and `#[cfg(windows)]` implementations.
- `src/protocol/` — the wire protocol. `framing.rs` (length-prefixed frames),
`message.rs` (HELLO/manifest/plan/DATA/VERIFY/… codecs), `sync.rs` (the
sender/receiver state machines: manifest exchange, exact-hash diffing (D12),
zstd delta transfer, locals-first write order, mtime/size fast path,
resume (D10), warm-start integration (D15)), and `tree_sync.rs` (the v3
tree session orchestrators: TREE_BEGIN/TREE_END envelope, deterministic
triage on both sides, sequential per-file conversations, deferred
deletions; plus the v4 parallel sessions of §16.8 — `--parallel K`
spreads conversations over K connections, one native thread per
connection, membership-validated via TREE_FILE). State machines are
generic over `S: Read + Write` — transport-agnostic by design.
- `src/tree.rs` — the tree layer (D17): metadata-only directory scan
(sorted, ffsync sidecar files excluded, optional exclude patterns with
subtree pruning — §16.7), relative-path validation (no traversal, wire
paths always `/`-separated), the hand-rolled glob matcher
(`ExcludeSet`/`ExcludePattern`), the `triage` pure function
(skip/sync/transfer/create-dir/delete/kind-conflict), and the tree
manifest type. Pure computation — no protocol frames here.
- `src/transport/` — the transport layer. `mod.rs` is transport security
(D16): rustls TLS 1.3 as a thin wrapper. `Identity` (self-signed cert +
PKCS#8 key, raw-DER codec in `identity.bin`, created on first run),
SHA-256 fingerprint codec (`AA:BB:…`), `PinPolicy` (`Pinned(set)` vs
`FirstUse`), and `client_session`/`server_session` — each completes the
TLS handshake eagerly and returns the presented fingerprint; the CLI
applies the TOFU decision (prompt or `--trust`) *after* the handshake,
before any sync data. The verifier never makes trust decisions by
itself. Pins live in `<state-dir>/pins.conf` (one fingerprint per line,
atomic rewrite). `--plaintext` bypasses TLS entirely (TCP only).
`quic.rs` is the QUIC carrier (D13 step 2, §17): quinn + the same
identity/pins, `QuicStream` adapts quinn streams to sync
`Read + Write + Send` (block_on at the boundary), `QuicSession` owns the
connection (worker conversations = extra streams); tokio lives here and
nowhere else (D3 edge exception).
- `tests/loopback.rs` — two-peer integration tests over in-memory duplex
streams: identical/modified/fresh/fast-path/disconnect-resume plus the
warm-start, stale-index, and orphan-cleanup scenarios.
- `tests/soak.rs` — leak guard (ignored by default): 100 sequential
sessions in one process under a counting global allocator, asserting net
live heap stays flat; run with `cargo test --test soak -- --ignored`.
- `tests/tls.rs` — TLS transport tests over real TCP loopback: pin-match
converges, third-party cert rejected by the client, unpinned client
rejected by the pinned receiver (error names the attacker's
fingerprint), FirstUse handshake + sync, pins-file codec round trip.
- `tests/quic.rs` — QUIC transport tests over real UDP loopback (§17):
single-file sync converges, pin-match + MITM rejected by the client,
unpinned client rejected by the pinned receiver, tree convergence,
parallel sessions over K streams (§16.8 on one connection), deletes
after success.
- `tests/tree.rs` — tree-session tests over real TCP loopback: converge a
mixed tree (nested dirs, empty files/dirs), second session skips
everything (mtime preservation fast path), modified/new-only deltas,
deletes after success, `--no-delete`, `--checksum`, empty-file
convergence, kind conflicts, v2-flow peers rejecting TREE_BEGIN,
wire path traversal refused, vanished sender files fail closed,
parallel sessions (§16.8: convergence, sequential-fallback rejection,
fail-closed on worker error).
- `tests/integration.rs` — end-to-end test: reconstruct a modified file from
an older local version plus simulated "network" chunks.
## Conventions
- Conventional Commits for commit messages.
- Doc comments (`///`) on public traits, structs, and functions.
- Swappable behavior goes behind traits (`Chunker`, `Hasher`, `FileExt`); add
new strategies as implementations instead of hardcoding or adding enums with
match arms everywhere.
- Keep submodules private and re-export the intended public API explicitly in
each `mod.rs`.
- Prefer native threads + `crossbeam` channels over async/tokio for the data
plane. The one registered exception (D3 amendment, 2026-10-23) is the
QUIC transport edge: `src/transport/quic.rs` owns a small tokio runtime
because quinn is an async library; everything it returns is sync
(`QuicStream: Read + Write + Send` via block_on at the boundary), and
`ffsync::protocol` MUST NOT depend on tokio — new transport code keeps
that split (async at the edge, sync at the seam). The async-file-I/O
measurement (`evaluation/chunking.md`, ~5× slower) still rules out any
async in the chunking/reconstruction paths.
## Gotchas (non-obvious invariants)
- Chunk data is never copied or sent through channels — only
`(offset, length)`. Hash workers index into the mmap themselves. Don't
reintroduce data copies.
- `process_file` takes a `chunker_factory` closure because each worker thread
needs its own stateful chunker instance; the `Hasher` is shared via `Arc`
because it is stateless (`Send + Sync`). Preserve this split.
- The mmap is created with `unsafe { Mmap::map }`; the underlying file must not
be modified while the pipeline holds the mapping.
- The reconstructor's temp path is derived with
`Path::with_extension("ffsync-tmp")`, which *replaces* the last extension
(`a.mkv` → `a.ffsync-tmp`). This differs from the `<.file.ext.ffsync-tmp>`
naming in the concept doc (recorded as D6 in `concept/Design Decisions.md`).
The index path does the OPPOSITE (appends `.ffsync-index`) because the
orphan cleanup must recover the target name by stripping the suffix; never
"fix" one to match the other.
- Sub-second mtime is load-bearing in two places: the index validation and
the §4 wire fast path (HELLO `file_mtime_nanos`, protocol v2). Whole-second
comparisons let a same-second same-size rewrite pass as unchanged — a
silent wrong-content sync. Any new metadata-based fast path must compare
at the finest available granularity.
- mtime preservation on commit (§16.6, v3) is what makes tree syncs skip:
the receiver sets the target's mtime to the sender's *before* writing the
D15 index (order: rename → set_modified → stat → write_index). Reversing
the order leaves the index recording the wrong mtime and every subsequent
tree sync re-checks the file.
- Tree manifests must exclude `*.ffsync-index` / `*.ffsync-tmp` — they are
ffsync's own bookkeeping; syncing them would sync internal state and
desync warm starts (an index written at commit changes mtime → re-sync
loop). The scan filter in `src/tree.rs` is the single place this lives.
- Tree triage must be deterministic and identical on both peers (same two
manifests, same flags — the receiver derives its flags from the sender's
TREE_BEGIN). Any nondeterminism desyncs the per-file conversation
sequence; the receiver fails loudly on the first unexpected HELLO.
- Excludes (§16.7) follow the same contract, with a twist: the receiver
must read the sender's TREE_BEGIN *before* scanning its own tree (the
sender's patterns govern both scans). The receiver's own patterns never
touch manifests — they only guard its delete pass. Never "optimize" the
receiver into scanning first; that deadlocks nothing but desyncs triage.
- Tree sessions make the fast-path branch explicit (FILE_SYNC frame);
single-file mode still relies on the drain-after-NOT_MODIFIED dance.
Do not "simplify" one to match the other — they differ because the
session continues after a tree conversation but ends after a single-file
one.
- Parallel sessions (§16.8) replace wire ORDER with membership: every
conversation (control connection included) is announced by a TREE_FILE
frame and validated against the expected set, exactly once. Do not
"optimize" the membership check away or reintroduce order-based
expectations — the sender's worker threads race, and arrival order is
arbitrary on every connection. The receiver's control loop consumes the
sender's TREE_END itself (that is how it stops serving partition 0), so
the sequential §16.5 barrier read is gated on `sender_end_consumed` —
removing that gate deadlocks the sender on its TREE_END reply.
- Worker connections authenticate with the SAME pin policy as the control
connection; the CLI's connect/accept closures handle that, and the
state machines stay transport-agnostic (the closures are injected).
Over QUIC (§17) the same closures open/accept extra streams on the one
session connection — the membership protocol is unchanged.
- QUIC teardown is NOT TCP teardown (§17.6): closing a QUIC connection
discards unACKed *stream* data (RFC 9000 §10.2.2), while TCP `close()`
flushes. A final frame that is ANSWERED (TREE_END report, COMMIT_ACK)
is only safe when the writer holds the connection until the reply/close
is observed — the CLI's `wait_for_peer_close` guards both receive paths
and the tree sender. Dropping a `QuicStream` auto-finishes (FIN), which
is what the receiver's Eof path expects; a dropped `RecvStream` sends
STOP_SENDING, so never rely on reading a stream after its `QuicStream`
was dropped. quinn's `Endpoint::connect`/`Incoming::accept` and
`Endpoint::new` all require a tokio runtime context (they spawn the
driver) — route them through the module's `block_on`, and give
`client_session` its bounded handshake timeout (quinn retransmits
Initials forever; a dead peer must fail loudly, §17.5).
- Windows is the primary development platform; keep platform-specific I/O
behind `FileExt` cfg branches.
- The identity file (`identity.bin`) holds the private key: `0600` on unix,
inherited user-profile ACLs on Windows; corrupt identity files are
refused loudly (never silently regenerated — that would break every
peer's pin). Pins are a set, not per-name entries; any pinned fingerprint
may connect in either role.
- Never commit generated artifacts. `.gitignore` covers `*.bin` and
`/bench_tmp/` (bench fixtures); `.opencode/` holds local-only configuration
and credentials.
## Roadmap (from the concept docs)
Implemented: single-file sync end-to-end (Phases 2–4 of the workflow doc:
root-hash identity, chunk diffing and plan generation, transfer + zstd),
resume, the persistent chunk index (D15), transport security (D16:
rustls TLS 1.3, mutual TOFU fingerprint pinning), directory-tree sync
(D17, protocol v3: Phase 1 metadata triage + per-file conversations,
mtime preservation, deferred deletes), tree-mode excludes (§16.7), and
parallel per-file conversations (§16.8, D13 step 1: `--parallel K` over
K connections — measured 3.4×–6.1× on a 100-file full transfer at
30 ms RTT), and the QUIC transport (§17, D13 step 2: `--quic`, one
connection per session, conversations ride streams). Not yet implemented:
permissions/xattrs/symlinks as first-class metadata (D9). See
`concept/Design Decisions.md` D5 (checksum-based change detection) and
D8 (link-adaptive diff strategy) for the design questions the pipelining
step must answer.