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; install pinned runtime dependencies with npm ci.

Join the community at clank.rodecomp.online/discord, which redirects to the Discord invite. The header's Discord link includes an inline SVG icon from Simple Icons, with no external image request.

contributors + their agents                       coordinator (clank.rodecomp.online)
┌──────────────────────────┐   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: data/decomp.json, data/toolchain.lock.json; the project's own root AGENTS.md is the only agent guide
deploy/	install script, systemd unit, Caddyfile, nftables firewall
deploy/forgejo/	Forgejo git dashboard at /git/: installer (checksum + signature verified), config, unit

Maintainer bootstrap

Run npm ci in the coordinator checkout before starting the server or running tests.

# 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 data/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://clank.rodecomp.online/install.ps1 | iex # Windows

curl -fsSL https://clank.rodecomp.online/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.

Batch visibility

#/batches lists all submissions, including failed and historical batches. The In review tab requires an actual in_review or peer_review attempt; an open batch alone is not proof that anything is reviewable. Generated-code claims in skip batches are included while their attempts await review. Awaiting merge separately shows approved or merging attempts. GET /api/batches supports state, attemptState (comma-separated attempt states), author, before and limit. Filters apply before pagination.

Skip-claim reviews show the claim category and reason beside the original instructions/pseudocode. They intentionally contain no submitted C++ or byte comparison. Attempt details expose the original symbol identity independently of compiled results so these reviews can load evidence without inventing a target.

The merge worker settles abandoned open code batches when they contain failed attempts and no attempt remains live. It records the reason and an audit event, preserving attempt feedback, task states, review votes and scores. This does not approve historical code: missing bindings and stale payloads need current-source verification and a new submission. Maintenance and actively merging batches are not settled by this reconciliation.

Resubmissions with identical compiled code retain actual peer verdicts or an explicit prior review waiver. Waiver carry-over is audited and does not create peer votes or carry a score-regression override; changed code still needs review.

Header-includer discovery reads committed source blobs in one git cat-file --batch operation and reuses that snapshot until HEAD or source roots change. Submitted overrides apply only to the current query. Reverse include traversal handles cycles without losing transitive dependents; uncommitted working-tree text is never used as the baseline.

The original-string gate uses each source unit's configured MSVC execution charset (/execution-charset:.949, for example) when checking readable Unicode literals. Numeric byte escapes remain literal bytes, and wide strings remain UTF-16LE. Unrepresentable characters are rejected rather than accepted as replacement ? bytes. Configure source UTF-8 and the target execution code page separately; this does not permit new strings absent from the original executable.

Repository synchronization before verification

When server.json sets pushRemote, the single worker fetches that remote's branch matching its checked-out branch at startup and before admitting each code/maintenance verification cohort. Existing verification, maintenance, merge and remote-check jobs drain first; new jobs do not start during synchronization. No remote is guessed when pushRemote is unset.

Only a fast-forward is automatic. Local-ahead commits are retained. Divergence reports both full commit IDs and requires a maintainer to reconcile the histories; there is no automatic reset, rebase or merge commit. Dirty tracked, staged and untracked files block synchronization, and ignored files cannot be overwritten by incoming tracked files. Preserve local edits before retrying.

Fetch/authentication failures and unsafe states leave batches queued with an explicit result.repositorySync.error; merge and remote-check admission pauses. The worker retries at its next pump and removes the diagnostic after a successful sync, so operator repairs do not require resubmission or a restart. Fetch is noninteractive and bounded to two minutes.

After synchronization the worker reloads data/decomp.json, its selected compile-command file and data/toolchain.lock.json from one committed snapshot, updates the shared project object and invalidates checks cached under old settings. The separate web process adopts only the published snapshot, never partially updated working-tree configuration. Direct uncommitted configuration changes are not loaded this way; commit them, or restart web and worker when no remote is used.

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.

These download names are mutable: the server sends Cache-Control: no-store on /download/ responses, including the manifest and compressed binaries. Self-updates revalidate the manifest, request the binary with ?sha256=<expected hash> to bypass an older unversioned CDN entry, and still verify its SHA-256 before replacing the installed executable. When deploying this policy behind a CDN, purge existing /download/ cache entries once: new origin headers do not evict already-cached responses used by older clients.

Tests

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

Global storage checks

Set "globalRules": true in the target repository's decomp.json to enable the source-level global gate. It is opt-in; deploying this server does not enable it for other projects. Committed configuration is refreshed by repository synchronization when pushRemote is configured; otherwise restart web and worker after changing it.

The verifier applies the payload over current main before checking it, and the merger checks again before committing. Both ordinary and maintenance batches use the same rules:

• An external data declaration needs an actual external definition in main plus the submitted files. Another extern, a local/static namesake, a different namespace, a pending batch or a symbol-database record does not supply storage.
• Static-storage definitions need a stable data annotation. This includes function pointers, multiline tables, out-of-class static members and local statics. Scalar constant expressions may be folded; arrays and mutable pointers are not exempt merely because their element or pointee type is const.
• Data annotations belong on definitions. Source units include their owner's header instead of adding extern-variable declarations. DLL-import declarations are distinguished from missing project definitions.
• Existing violations are grandfathered for unrelated edits. New/changed declarations and removed annotations are checked, as is deletion of a definition still required by an existing extern.

The scanner handles declaration scopes and balanced initializers; it is not a C++ type checker or a substitute for a link. Compilation still validates types and conditional configurations. These checks establish source ownership, not the correct binary address, object extent, initializer values or constructor order. Those require original-binary evidence and review.

Structural maintenance

Maintainers can use td maintenance submit --files a.cpp,b.h --notes "reason" for structural changes without function tasks. The server runs the submission gates, compiles affected units and compares effective code against a fresh main build. Behavior-changing function work must use the normal task/review path. td maintenance symbol <id> --size N --reason "evidence" corrects metadata; --retire is allowed only after main no longer annotates that global ID.

From a contributor workspace, a maintainer can correct a decorated/linker name without changing a symbol's ID, address, task state, match state or confirmations:

td maintenance symbol fn_00123 --decorated '?CorrectName@@YAXXZ' --reason "original-binary evidence"
td maintenance symbol fn_00123 --swap-decorated-with fn_00456 --reason "call targets show the names were reversed"

The maintained HTTP route is POST /api/admin/symbols/:id. Send either {"decorated":"?CorrectName@@YAXXZ","reason":"original-binary evidence"} or {"swapDecoratedWith":"fn_00456","reason":"call-target evidence"}. MCP exposes the same fields through td_correct_symbol, with an additional id argument. Both clients refresh their symbol snapshots after a successful correction.

decorated must be a string of 1-2048 visible ASCII characters, with no whitespace or control characters; it is never coerced, trimmed or truncated. C-linkage names are supported as well as C++ decorated names. Null and empty values are rejected. The evidence reason must be a nonblank string of at most 500 characters. Unknown request fields are rejected. Existing size, name and retire fields remain available; a swap cannot be combined with other corrections.

A name already mapped to another ID returns HTTP 409 without applying any part of the request. A swap requires two distinct, existing, unambiguous mappings whose symbols have the same kind; it exchanges only their decorated names in one database transaction. Do not clear a mapping or use a temporary name as an intermediate step. Each affected ID receives a symbol_corrected event containing the acting user, old and new values, and the evidence reason (plus the other ID for a swap). Audit records and corrections commit together. Symbol snapshot caches in every web process and target version are invalidated by the committed correction, so an immediate conditional refresh observes the new names rather than a stale 304.

Neither an unchanged-code fingerprint nor a successful unit compile proves that new global storage reproduces the original initializer/destructor behavior. Maintenance authors must preserve those semantics; an incomplete object layout must not be published as a recovered full size.