RuTracker — Rust CLI + MCP

Rust workspace with two binaries sharing one parser/HTTP/cookies core:

Tools

Prerequisites

• macOS (osx-arm64). Cookie extraction is macOS-Keychain-bound; Linux/Windows support is future work.

• Rust stable (≥ 1.75). Auto-installed by rustup via rust-toolchain.toml.

• Brave browser with an active rutracker.org session in a profile named "Peter" (or set RUTRACKER_PROFILE=...).

Install

cargo build --release
cargo install --path crates/cli --locked
cargo install --path crates/mcp --locked

rutracker --help
rutracker-mcp --help

CLI usage

# Search
rutracker search "2026" --category 252

# JSON to file, pipe to jq
rutracker search "2026" --out /tmp/search.json
jq '.results | length' /tmp/search.json

# Topic with comments
rutracker topic 6843582 --comments --format text

# Browse a category without a query
rutracker browse 252 --sort-by seeders

# List all categories
rutracker categories --format text

# Download .torrent (sandboxed to $HOME/CWD by default)
rutracker download 6843582 --out-dir $HOME/tmp-rutracker
rutracker download 6843582 --out-dir /Volumes/ext --allow-path

Global flags: --format {json,text} (default json), --out FILE, --base-url URL, --profile NAME.

Claude Code MCP usage

The included .mcp.json points at the rutracker-mcp binary. After cargo install runs, reopen Claude Code in this directory and the 5 tools appear automatically.

Cookies

First run prompts the macOS Keychain for "Brave Safe Storage" access. Cookies are decrypted via AES-128-CBC (PBKDF2-SHA1 key derivation, matches Chromium defaults) and cached locally (gitignored).

Required cookies:

• bb_session, bb_guid, bb_ssl, bb_t — authenticated browsing.

• bb_dl_key — .torrent download via dl.php. Absence triggers an explicit error with a refresh hint.

Testing

# Fixture-driven tests (fast, offline — default)
cargo test --workspace

# Lint + format gates
cargo clippy --workspace --all-targets -- -D warnings
cargo fmt --all -- --check

# Live smoke (manual, opt-in — requires Brave session + Keychain prompt)
cargo test --workspace -- --ignored

Coverage

Run cargo llvm-cov --workspace --lcov --output-path lcov.info for a full workspace report. Open cargo llvm-cov --workspace --html --open for a line-by-line HTML view.

Threshold: ≥95% line coverage, enforced by CI. Exclusions are documented in .cargo/llvm-cov.toml (or the --ignore-filename-regex flag in .github/workflows/ci.yml).

Local invocation that matches CI (same exclusion regex, same threshold):

IGNORE='crates/cookies-macos/src/keychain\.rs|crates/mcp/src/main\.rs|crates/cli/src/main\.rs|crates/cookies-macos/src/lib\.rs'
cargo llvm-cov --workspace --locked --summary-only --ignore-filename-regex "$IGNORE"

Excluded paths and rationale:

Local mirror

rutracker-mirror keeps an incremental on-disk copy of any forums you watch. The JSON-per-topic layer under forums/<id>/topics/<topic_id>.json is the source of truth; state.db (SQLite, WAL) is a derived index rebuildable from the JSONs.

Layout (default root $HOME/.rutracker/mirror/, override via RUTRACKER_MIRROR_ROOT or --root):

$HOME/.rutracker/mirror/
├── structure.json   — forum tree snapshot
├── watchlist.json   — forums to sync
├── state.db         — derived SQLite index
└── forums/<id>/topics/<topic_id>.json

Bootstrap + first sync:

rutracker mirror init
rutracker mirror structure
rutracker mirror watch add 252
rutracker mirror watch add 251
rutracker mirror sync --max-topics 20

Re-running rutracker mirror sync is idempotent: steady-state passes stop after 5 consecutive unchanged rows and rewrite only topics with a new last_post_id. On HTTP 429/503 the forum is parked for one hour (forum_state.cooldown_until); rutracker mirror status surfaces active cooldowns.

Sync automation flags

rutracker mirror sync now auto-resumes through cooldowns instead of forcing a manual rerun after every 429/503 cycle.

• --max-attempts-per-forum 24 caps retry loops per forum before the run marks that forum as gave_up and exits non-zero.

• --cooldown-wait=true keeps the process alive until the stored cooldown expires; --cooldown-wait=false preserves the old "stop on first rate-limit" behavior for CI or tight smoke tests.

• --log-file PATH writes NDJSON progress events. Omit it to use <mirror-root>/logs/sync-<YYYYMMDD-HHMMSS>.log, pass - for stderr, or pass an empty string to disable file logging entirely.

rutracker mirror sync --forum 252 --max-attempts-per-forum 24
rutracker mirror sync --forum 252 --cooldown-wait=false
rutracker mirror sync --forum 252 --log-file -
rutracker mirror show 252/6843582 --format text
rutracker mirror status
rutracker mirror rebuild-index   # reconstruct state.db from the JSON layer

Caveats: mirror root is APFS/ext4 only (NFS/SMB not tested). Watchlist soft-cap is ~100 forums. Only the latest revision of edited posts is retained — prior revisions are intentionally not kept (plan §5.3). Add .rutracker/ to a repo-level .gitignore if the root lives inside a checkout.

Increase log verbosity with RUST_LOG=rutracker_mirror=debug.

Ranking films

rutracker-ranker layers an objective community-consensus quality score on top of the mirror. It first groups every release topic by film identity, then aggregates comment-sentiment across all rips of the same film, and finally ranks rips within each film by tech quality, format, health, and recency.

The NLP step (Russian-language comment analysis) runs through a Claude Code subagent defined in .claude/agents/rutracker-film-scanner.md — no API key needed. Rust owns the deterministic pipeline stages; only the per-topic Haiku scan hops through the Claude Code harness via a thin /rank-scan-run skill.

Three-step user workflow:

# Stage A — parse titles, populate film_index + film_topic (idempotent).
rutracker rank match --forum "Фильмы 2026"

# Stage B.1 — emit the scan-queue.jsonl manifest (Rust, fast, offline).
rutracker rank scan-prepare --forum "Фильмы 2026"

# Stage B.2 — execute the queue inside a Claude Code session (Haiku scans).
# (in Claude Code)  /rank-scan-run --forum "Фильмы 2026"

# Stage C — aggregate scan outputs into film_score + rank rips.
rutracker rank aggregate --forum "Фильмы 2026"

# Query the results.
rutracker rank list --top 20
rutracker rank show "Альфа" --format text
rutracker rank parse-failures

Forum names are resolved via structure.json (populated by rutracker mirror structure). Numeric ids still work for scripting (--forum 252). Quote multi-word names to prevent shell word-splitting.

rutracker rank aggregate prints a warning when topics in the target forum have no .scan.json yet — that is the cue to run rutracker rank scan-prepare and then /rank-scan-run in Claude Code. Re-running the whole pipeline after mirror sync fetches new topics is incremental: cached scans with matching agent_sha + last_post_id are skipped automatically.

Calibration (release gate)

Before shipping new prompt/agent changes, validate scanner output against a hand-labelled holdout:

1. Create crates/ranker/tests/fixtures/ranker/labels.jsonl with at least 20 entries: {"topic_id":"<tid>","human_score":<0-10>,"note":"..."}.

2. Run the three-step scan workflow for the labelled topics.

3. scripts/calibrate-scanner.sh — computes Spearman ρ vs. labels. Release-blocking: ρ ≥ 0.6 is required.

Manual release gate

bash scripts/soak.sh          # parser soak: 20 random topics, asserts title+desc non-empty
bash scripts/soak-mirror.sh   # mirror soak: 2-pass (initial >= 6 files; second pass 0 files)

Each script logs to soak-*-<date>.log; commit the log with the release.

Architecture

• crates/parser — pure HTML parsers + dataclasses + shared text_format. No I/O.

• crates/http — async reqwest client, cp1251 decoding, login-redirect recovery.

• crates/cookies-macos — Brave cookie AES-CBC decrypt + Keychain lookup + SQLite reader.

• crates/cli — rutracker binary: clap subcommands, JSON/text output, path sandbox.

• crates/mcp — rutracker-mcp binary: hand-rolled JSON-RPC stdio server.

• crates/mirror — local-mirror engine: SQLite index, atomic JSON writes, delta-aware sync.

Full design: .omc/plans/full-mcp.md.