Clank Together

Coordinator, sandboxed verifier and agent client for a collaborative, LLM-assisted matching decompilation of an MSVC 2010 C++ executable. Node.js ≥ 22.13, no npm dependencies.

contributors + their agents                       coordinator (saty.dev)
┌──────────────────────────┐   HTTPS   ┌──────────────────────────────────────────┐
│ td CLI / td mcp (agent)  │◄─────────►│ Caddy (TLS) → node coordinator :8080      │
│ git clone of the repo    │  git+api  │  SQLite: users, symbols, versions, tasks, │
│ .td/target.exe (verified)│           │          attempts, reviews, func_status   │
│ pinned Ghidra (headless) │           │  canonical git repo (read-only clones)    │
│ cl.exe via docker/native │           │  verifier → docker sandbox (cl.exe/Wine)  │
└──────────────────────────┘           │  merge queue → commits to main            │
                                       └──────────────────────────────────────────┘

Git dashboard

The source is browsable at https://<host>/git/ (Forgejo, self-hosted). The coordinator is the only writer: every merge is pushed by the clank-bot account, main is protected, pull requests are off and Forgejo sign-up is disabled. Contributors clone from there; /repo.git on the coordinator is a read-only fallback.

Core ideas

• Stable symbol IDs. Source annotations hold IDs (// FUNCTION: fn_00123), never addresses. The coordinator's database maps each ID to its address per target version, so a new exe version means mapping addresses (td-server version map), not retagging source.
• Same pipeline everywhere. td check (local) and the server verifier run the same code: compile with the pinned MSVC, read the COFF object and /FAcs listing, compare bytes against the original with relocations masked, then check that every relocation points where the original points (calls via REL32, data via DIR32 + the PE's base relocations).
• Learning references. A call to a not-yet-decompiled function is checked against the database's address for that function; on merge the decorated name is learned. References to data without a symbol create new symbols automatically.
• Untrusted input everywhere. Static gate (paths, banned constructs such as inline asm, #pragma code_seg, #line, byte blobs), a second lint on the preprocessed output, compilation in a container with no network, read-only root, dropped capabilities, memory/CPU/pid limits and userns-remap. Notes from other contributors reach agents wrapped as untrusted data.
• Humans decide. Matching work goes to review (1 approval, 2 for header changes, no self-review), then a serial merge queue three-way merges onto current main and re-verifies before committing.

Task flow

open ─claim→ leased ─submit→ verifying ─┬→ rejected (static gate / cheating) → open
  ▲                                      ├→ needs_cleanup (compile error, mismatch, regression) ─claim→ …
  │                                      └→ in_review ─approve→ approved → merging → done
  └──────── lease expired / released / changes requested / merge conflict ─────────┘

After maxCleanupCycles failed attempts a task becomes hard. The author of the last attempt gets first pick on the follow-up for 6 hours.

Layout

Path
src/shared/	annotation parser, PE/COFF/listing readers, compare, reconcile, lint, build runners
src/server/	coordinator: API, tasks, verifier, merge queue, git http-backend, review UI
src/client/	td CLI, MCP server for agents, Ghidra integration
ghidra_scripts/	ExportFunctions (bootstrap), ApplySymbols, DecompileAt, DisassembleAt
toolchain/msvc2010/	Dockerfile for cl.exe under Wine (you supply the compiler files)
template/	skeleton of the decomp repo: decomp.json, toolchain.lock.json, AGENTS.md
deploy/	install script, systemd unit, Caddyfile, nftables firewall
deploy/forgejo/	Forgejo git dashboard at /git/: installer (checksum + signature verified), config, unit

Maintainer bootstrap

# on the server (as the service user):  td-server <cmd> --data /var/lib/togetherdecomp
td-server version add 1.0 game.exe              # register the target
# analyze with the pinned Ghidra and export the function list
analyzeHeadless /tmp/g boot -import game.exe -scriptPath ghidra_scripts -postScript ExportFunctions.java functions.csv
td-server symbols import functions.csv          # one stable ID per function, tasks get created
td-server symbols units units.csv               # optional: start,end,src/file.cpp
td-server user add alice --role reviewer
# build the compiler image (see toolchain/msvc2010/Dockerfile), pin clSha256 in toolchain.lock.json

New target version later: td-server version add 1.1 game11.exe, then td-server version map 1.1 mapping.csv --from 1.0 (old_addr,new_addr from Ghidra Version Tracking or BinDiff), td-server version set-current 1.1, td-server rebuild.

Contributor quick start

Register on the website, wait for approval, create a token under Account, then one line:

$env:TD_TOKEN='td_...'; irm https://saty.dev/install.ps1 | iex          # Windows

curl -fsSL https://saty.dev/install.sh | TD_TOKEN=td_... sh               # Linux

It installs td (a single executable, no Node.js; git is the only other requirement), sets up the workspace and connects the AI (Claude Code automatically, other MCP clients via the printed settings). Compiling and reading the original go through the server; a local compiler (td setup ... --toolchain) and a local Ghidra (td ghidra install + td ghidra download) are optional. td updates itself.

Work comes in batches: td claim --unit <file> takes every open function of a file, td submit sends them together. Byte-identical functions merge automatically; the others are peer-reviewed by three agents, then approved by a human with one click per batch.

Building td

node tools/build-td/build.mjs (Node 22.21) writes dist/td-windows-x64.exe, dist/td-linux-x64 and SHA256SUMS; copy them to <dataDir>/downloads on the server.

Tests

npm test uses synthetic PE/COFF files, so it doesn't need a compiler.