@clear-capabilities/agentic-security-scanner 0.195.9 → 0.195.11

package/src/language/haskell-ir.js

@@ -43,6 +43,26 @@ const IO_NAMES = new Set(['putStr', 'putStrLn', 'print', 'getLine', 'getContents
 
 // name -> positions of the function, the collection and (folds) the seed.
 const BODY_SETTER = /^Network\.HTTP\.(?:Simple|Conduit|Client)\.setRequestBody(?:JSON|LBS|BS|URLEncoded|File)$/;
+// A mutable cell carries what was written to it: `writeIORef ref v` assigns the cell's variable, `readIORef ref` reads it, `modifyIORef ref f`
+// is `ref = f ref`. The same for MVar, TVar and STRef. Only a cell held in a LOCAL variable of the function that reads it is modelled
+// (created there, or handed in as a parameter): a cell written in one function and read in another is not followed (see the guide's limits).
+const CELL_OPS = Object.freeze({
+  'Data.IORef.writeIORef': 'write', 'Data.IORef.readIORef': 'read', 'Data.IORef.modifyIORef': 'modify', "Data.IORef.modifyIORef'": 'modify', 'Data.IORef.atomicWriteIORef': 'write',
+  'Control.Concurrent.MVar.putMVar': 'write', 'Control.Concurrent.MVar.takeMVar': 'read', 'Control.Concurrent.MVar.readMVar': 'read', 'Control.Concurrent.MVar.modifyMVar_': 'modify', 'Control.Concurrent.MVar.swapMVar': 'write',
+  'Control.Concurrent.putMVar': 'write', 'Control.Concurrent.takeMVar': 'read', 'Control.Concurrent.readMVar': 'read', 'Control.Concurrent.modifyMVar_': 'modify',
+  'Control.Concurrent.STM.writeTVar': 'write', 'Control.Concurrent.STM.readTVar': 'read', 'Control.Concurrent.STM.readTVarIO': 'read', 'Control.Concurrent.STM.modifyTVar': 'modify', "Control.Concurrent.STM.modifyTVar'": 'modify',
+  'Control.Concurrent.STM.TVar.writeTVar': 'write', 'Control.Concurrent.STM.TVar.readTVar': 'read', 'Control.Concurrent.STM.TVar.readTVarIO': 'read', 'Control.Concurrent.STM.TVar.modifyTVar': 'modify', "Control.Concurrent.STM.TVar.modifyTVar'": 'modify',
+  'Data.STRef.writeSTRef': 'write', 'Data.STRef.readSTRef': 'read', 'Data.STRef.modifySTRef': 'modify', "Data.STRef.modifySTRef'": 'modify',
+});
+// `instance Runner Shell`, `instance Show a => Runner (Box a)`: the type constructor the instance is for (the first constructor after the class name).
+function instanceTypeOf(inst) {
+  const head = String((inst && inst.headText) || '').replace(/^.*?=>\s*/, '');
+  const cls = inst && inst.name ? String(inst.name).replace(/[.*+?^${}()|[\]\\]/g, '\\$&') : null;
+  if (!cls) return null;
+  const after = head.replace(new RegExp(`^\\s*(?:[A-Z][\\w']*\\.)*${cls}(?![\\w'])`), '');
+  const m = /^[\s([]*(?:[A-Z][\w']*\.)*([A-Z][\w']*)/.exec(after);
+  return m ? m[1] : null;
+}
 const HOF = {
   map: { fn: 0, coll: 1, kind: 'map' }, fmap: { fn: 0, coll: 1, kind: 'map' }, concatMap: { fn: 0, coll: 1, kind: 'map' },
   traverse: { fn: 0, coll: 1, kind: 'traverse' }, mapM: { fn: 0, coll: 1, kind: 'traverse' },
@@ -232,6 +252,11 @@ function exportedLookup(proj, target, name, depth, seen) {
   if (depth > HS_IR_LIMITS.maxResolveDepth) return { kind: 'unknown', reason: 'resolution-depth-cap' };
   const next = new Set(seen); next.add(key);
   if (!target.exports) return target.defs.has(name) ? defResult(target, name) : null;
+  // A wildcard re-export of a module the project does not contain (`module CorePrelude` over `import CorePrelude hiding (..)`) can only say the
+  // name MIGHT come from there. It is kept as the fallback while the rest of the export list is searched: a definition the project holds, or a
+  // name an import lists, is the answer when one exists. IHP's preludes re-export `CorePrelude` first and `IHP.Controller.Param` later, and
+  // returning the possibility first made every `param` an unknown external name instead of the in-tree request read.
+  let fallback = null;
   for (const ex of target.exports) {
     if (ex.kind === 'var' && ex.name === name) return resolveGlobal(proj, target, name, ex.qual || null, depth + 1, next);
     if (ex.kind === 'type' && ex.all) {
@@ -245,11 +270,13 @@ function exportedLookup(proj, target, name, depth, seen) {
         if (imp.items && (imp.hiding ? imp.items.includes(name) : !imp.items.includes(name))) continue;
         const t2 = proj.modules.get(imp.module);
         const r = t2 ? exportedLookup(proj, t2, name, depth + 1, next) : { kind: 'external', module: imp.module, name, certain: !!(imp.items && !imp.hiding && imp.items.includes(name)) };
-        if (r) return r;
+        if (!r) continue;
+        if (r.kind === 'external' && !r.certain) { if (!fallback) fallback = r; continue; }
+        return r;
       }
     }
   }
-  return null;
+  return fallback;
 }
 
 // A top-level resolution (depth 0, nothing seen yet) is a pure function of the module table, which is complete before any function
@@ -285,7 +312,9 @@ function resolveGlobalUncached(proj, mod, name, qual, depth, seen) {
         const owner = r.kind === 'def' ? (r.def.kind === 'field' ? (proj.modules.get(r.mod)?.cons.get(r.def.con)?.type) : r.def.kind === 'method' ? r.def.cls : null) : null;
         if (!owner || !(imp.typeAll || []).includes(owner)) continue;
       }
-      if (r.kind === 'unknown') { unknownReason = r.reason; if (r.candidates) unknownCandidates = r.candidates; } else hits.push(r);
+      if (r.kind === 'unknown') { unknownReason = r.reason; if (r.candidates) unknownCandidates = r.candidates; }
+      else if (r.kind === 'external' && !r.certain) wildcard.push(r.module || imp.module);   // a re-export chain's "might come from there" is a possibility, not a hit that makes a definition ambiguous
+      else hits.push(r);
     } else if (qual || (imp.items && !imp.hiding && imp.items.includes(name))) hits.push({ kind: 'external', module: imp.module, name, certain: true });
     else wildcard.push(imp.module);   // includes `import M (T(..))` of a module we cannot read: possible, never certain
   }
@@ -590,6 +619,11 @@ class FnLowerer {
     const r = this.resolve(e.name, e.qual);
     if (r.kind === 'alias') return r.expr;
     if (r.kind === 'def' && r.def.kind === 'field') return { kind: 'ident', name: e.name, hs: { fieldSelector: r.name } };
+    if (r.kind === 'def' && r.def.kind === 'fun' && isKnownApi(r.mod, r.name)) {
+      // a modelled name the project itself defines (see applyVar): the registry's meaning, the definition still referenced
+      this.noteRef(r.qid, 'modelled-definition', ln);
+      return { kind: 'ident', name: `${r.mod}.${r.name}`, hs: { resolution: 'external', certain: true, modelledDefinition: r.qid } };
+    }
     if (r.kind === 'def' || r.kind === 'localfn') {
       this.noteRef(r.qid, 'reference', ln);
       return { kind: 'ident', name: this.canonical(r), hs: { functionRef: r.qid } };
@@ -683,6 +717,16 @@ class FnLowerer {
       if (f.name === 'uncurry' && argAsts.length === 2 && tup(argAsts[1])) return this.lowerApp(argAsts[0], tup(argAsts[1]), ln, via);
     }
     if (r.kind === 'def' && r.def.kind === 'field' && argAsts.length >= 1) {
+      // `exec ops x` where `ops` is a top-level constant record whose `exec` field holds a function: the application is of THAT function. The
+      // held function is found in the constant's record syntax (`Ops { exec = \x -> .. }` or `Ops { exec = pingHost }`); a positional
+      // construction, a field set by an update, or a record built at run time stay an indirect call.
+      const held = argAsts.length > 1 ? this.heldFieldFunction(r, argAsts[0], ln) : null;
+      if (held) {
+        const lowered = argAsts.slice(1).map((a) => ({ t: 'lowered', expr: this.lowerExpr(a, ln) }));
+        const savedMod = this.mod; const savedOverride = this.lineOverride;
+        this.lineOverride = this.lineOverride || ln; this.inlineDepth = (this.inlineDepth || 0) + 1; this.mod = held.mod;
+        try { return this.lowerApp(held.expr, lowered, ln, via); } finally { this.mod = savedMod; this.lineOverride = savedOverride; this.inlineDepth -= 1; }
+      }
       const obj = this.lowerExpr(argAsts[0], ln);
       const mem = { kind: 'member', object: obj, prop: r.name };
       return argAsts.length === 1 ? mem : { kind: 'call', callee: '<indirect>', args: [mem, ...argAsts.slice(1).map((a) => this.lowerExpr(a, ln))], hs: { status: 'unknown', reason: 'field-applied' } };
@@ -700,7 +744,18 @@ class FnLowerer {
     if (selector) hs.selector = selector;
     let callee = f.qual ? `${f.qual}.${f.name}` : f.name;
     const arity = (r.kind === 'def' && r.def.kind === 'fun') ? r.def.arity : (r.kind === 'localfn' ? r.arity : -1);
-    if ((r.kind === 'def' && r.def.kind === 'fun') || r.kind === 'localfn') {
+    if (r.kind === 'def' && r.def.kind === 'fun' && isKnownApi(r.mod, r.name)) {
+      // The project defines a module the model registry models: the project IS the framework (a framework repository scanning itself) or
+      // carries a copy of it. The registry says what the name means to its caller (a request read, a sink, a sanitizer), which the
+      // definition's body, written in terms the registry does not model (an implicit request parameter, a parser over it), does not show.
+      // The call is lowered as the library call; the definition is still referenced, so it is never an uncalled entry point, and its own
+      // body is still analysed as a function of the project. Pinned both ways in test/haskell/haskell-taint-shapes.test.js.
+      this.noteRef(r.qid, 'modelled-definition', ln);
+      callee = `${r.mod}.${r.name}`;
+      hs.status = 'external'; hs.certain = true; hs.modelledDefinition = r.qid;
+      hs.effect = r.def.sig ? (r.def.sig.returnsIO ? 'io' : 'pure') : (IO_NAMES.has(f.name) ? 'io' : 'unknown');
+      if (hs.effect === 'io') this.sawIO = true;
+    } else if ((r.kind === 'def' && r.def.kind === 'fun') || r.kind === 'localfn') {
       callee = this.canonical(r);
       hs.status = 'resolved'; hs.target = r.qid;
       if (r.kind === 'def' && arity === args.length && arity > 0 && (this.inlineDepth || 0) < INLINE_MAX_DEPTH) {
@@ -722,6 +777,10 @@ class FnLowerer {
       hs.status = 'typeclass'; hs.cls = r.def.cls; hs.method = r.name; hs.target = r.qid;
       hs.candidates = this.proj.instancesOf(r.def.cls, r.name);
       hs.effect = 'unknown';
+      // `run Shell x`: an argument that is a constructor of a type with exactly one visible instance of the class names the instance, so the
+      // call is a call of that instance's method (a resolved edge, not a set of candidates). An argument of unknown type keeps the candidates.
+      const picked = this.dispatchInstance(r.def.cls, r.name, args);
+      if (picked) { callee = picked.name; hs.status = 'resolved'; hs.target = picked.qid; hs.dispatchedOn = picked.type; delete hs.candidates; }
     } else if (r.kind === 'param') {
       hs.status = 'param'; hs.paramIndex = r.index; hs.effect = 'unknown';
     } else if (r.kind === 'local') {
@@ -747,6 +806,10 @@ class FnLowerer {
       if (ctxTargets.some((t) => t.length)) hs.argTargets = ctxTargets;
     }
     const call = { kind: 'call', callee, args, line: ln, hs };
+    if (hs.status === 'external' || hs.status === 'builtin') {
+      const cell = this.lowerCellOp(callee, args, argAsts, ln, call);
+      if (cell) return cell;
+    }
     // `setRequestBodyJSON body req` returns the REQUEST with a body: the body's taint is a property of the body, and
     // must not make the request's URL tainted. The call still gets its own node so a sink on the body argument fires.
     if ((hs.status === 'external' || hs.status === 'builtin') && BODY_SETTER.test(callee) && args.length >= 2) {
@@ -757,6 +820,64 @@ class FnLowerer {
     return (hs.status === 'external' || hs.status === 'builtin') && isSourceApi(callee) ? this.sourceValue(call, ln) : call;
   }
 
+  // The instance a class-method call dispatches to, when an argument is a constructor (nullary, positional or record) of a type that has
+  // exactly one visible instance of the class. Several candidate types, or no constructor argument, is null (the candidates stay).
+  dispatchInstance(cls, method, args) {
+    const cands = this.proj.instancesOf(cls, method);
+    if (!cands.length) return null;
+    const types = new Set();
+    for (const a of args) {
+      const con = a && a.hs && a.hs.con;
+      if (!con) continue;
+      const c = this.mod.cons.get(con) || this.proj.findCon(con);
+      if (c && c.type) types.add(c.type);
+    }
+    if (types.size !== 1) return null;
+    const [type] = types;
+    const hits = cands.filter((q) => { const m = this.proj.instanceMeta.get(q); return m && m.type === type; });
+    if (hits.length !== 1) return null;
+    return { qid: hits[0], name: this.proj.instanceMeta.get(hits[0]).name, type };
+  }
+
+  // The function a record field of a top-level constant holds: `ops = Ops { exec = \x -> .. }` for `exec ops`. Returns the field's expression
+  // and the module it was written in, or null (the constant is not a single record literal, or the field is not set there).
+  heldFieldFunction(r, objAst, ln) {
+    if ((this.inlineDepth || 0) >= INLINE_MAX_DEPTH) return null;
+    const u = unparen(objAst);
+    if (!u || u.t !== 'var') return null;
+    const r0 = this.resolve(u.name, u.qual);
+    if (r0.kind !== 'def' || r0.def.kind !== 'fun' || r0.def.arity !== 0) return null;
+    const mod = this.proj.modules.get(r0.mod);
+    const fnAst = mod && mod.groups.funs.find((f) => f.name === r0.name);
+    if (!fnAst || fnAst.clauses.length !== 1) return null;
+    const c = fnAst.clauses[0];
+    if (c.pats.length || !c.rhs || (c.rhs.guards && c.rhs.guards.length) || (c.rhs.where && c.rhs.where.length) || !c.rhs.body) return null;
+    const body = unparen(c.rhs.body);
+    if (!body || body.t !== 'rec' || !body.con || body.con.name !== r.def.con) return null;
+    const fld = (body.fields || []).find((f) => !f.wildcard && f.name === r.name && f.e);
+    if (!fld) return null;
+    this.noteRef(r0.qid, 'reference', ln);
+    return { mod, expr: fld.e };
+  }
+
+  // A mutable cell held in a local variable (see CELL_OPS): a write is an assignment of the cell's variable, a read is that variable, a modify is
+  // the function applied to it. Anything else (a cell that is not a local variable, a missing argument) is the plain call.
+  lowerCellOp(callee, args, argAsts, ln, call) {
+    const op = CELL_OPS[callee];
+    if (!op) return null;
+    const ref = args[0];
+    if (!ref || ref.kind !== 'ident' || !ref.name || ref.hs) return null;
+    this.sawIO = true;
+    if (op === 'read') return { kind: 'ident', name: ref.name };
+    if (op === 'write' && args.length >= 2) { this.cfg.add('assign', ln, { target: ref.name, source: args[1], hs: { demand: 'executed', cellWrite: callee } }); return call; }
+    if (op === 'modify' && argAsts.length >= 2) {
+      const applied = this.lowerApp(argAsts[1], [{ t: 'lowered', expr: { kind: 'ident', name: ref.name } }], ln, callee);
+      this.cfg.add('assign', ln, { target: ref.name, source: applied, hs: { demand: 'executed', cellWrite: callee } });
+      return call;
+    }
+    return null;
+  }
+
   // A modelled source read gets its own assignment node, so the evidence chain names the exact line the
   // untrusted data entered, even when the read sits inside a larger expression (`f =<< getLine`).
   sourceValue(call, ln) {
@@ -813,7 +934,8 @@ class FnLowerer {
     try {
       inl.params.forEach((name, i) => this.declare(name, { kind: 'alias', expr: loweredArgs[i] }));
       this.mod = calleeMod;
-      (this.inlinedCalls ||= []).push({ site: null, callee: this.canonical(r), args: [], line: ln, hs: { status: 'resolved', target: r.qid, effect: 'unknown', inlined: true, ...(via ? { via } : {}) } });
+      // the lowered arguments are kept on the edge: a rule asking what a helper's parameters receive reads them (haskell-security-rules.js requestTextParams)
+      (this.inlinedCalls ||= []).push({ site: null, callee: this.canonical(r), args: loweredArgs, line: ln, hs: { status: 'resolved', target: r.qid, effect: 'unknown', inlined: true, ...(via ? { via } : {}) } });
       const out = this.lowerExpr(inl.body, ln);
       return out && typeof out === 'object' ? { ...out, hs: { ...(out.hs || {}), inlinedFrom: r.qid } } : out;
     } finally {
@@ -1297,6 +1419,7 @@ class Project {
     this.fns = [];
     this.byFile = new Map();
     this.instanceIndex = new Map(); // `${cls}.${method}` -> qids
+    this.instanceMeta = new Map();  // instance-method qid -> { type, name }: the type the instance is for and the function's call-graph name
     this.diagnostics = [];
   }
 
@@ -1467,7 +1590,11 @@ function buildHaskellIRUncached(fileContents) {
         const key = `${inst.name}.${f.name}`;
         if (!proj.instanceIndex.has(key)) proj.instanceIndex.set(key, []);
         proj.instanceIndex.get(key).push(qid);
-        instQids.push({ mod, inst, f, qid });
+        // the call-graph name carries the instance's type when it is known, so two instances of one class resolve to different functions by name
+        const type = instanceTypeOf(inst);
+        const name = `${mod.name}.$inst.${inst.name}${type ? `.${type}` : ''}.${f.name}`;
+        proj.instanceMeta.set(qid, { type, name });
+        instQids.push({ mod, inst, f, qid, name });
       }
     }
   }
@@ -1481,8 +1608,8 @@ function buildHaskellIRUncached(fileContents) {
       try { proj.addFn(lowerPatbind(proj, mod, pb)); } catch (err) { proj.diagnostics.push({ file: mod.file, line: pb.line, kind: 'ir-lowering-failure', detail: String(err && err.message) }); }
     }
   }
-  for (const { mod, inst, f, qid } of instQids) {
-    const info = { qid, name: `${mod.name}.$inst.${inst.name}.${f.name}`, line: f.line, parentName: `${mod.name}.$inst.${inst.name}.${f.name}` };
+  for (const { mod, inst, f, qid, name } of instQids) {
+    const info = { qid, name, line: f.line, parentName: name };
     try {
       const L = new FnLowerer(proj, mod, info, []);
       L.lowerClauses(f.clauses, f.line);

package/src/language/haskell-models.js

@@ -398,6 +398,13 @@ export const HS_RULE_APIS = Object.freeze([
   // mwc-random and splitmix are statistical generators, not CSPRNGs (createSystemRandom only seeds from the OS)
   ...E('System.Random.MWC', ['uniformR', 'uniform', 'uniformRM', 'uniformM', 'createSystemRandom', 'create', 'initialize', 'withSystemRandom', 'withSystemRandomST']),
   ...E('System.Random.SplitMix', ['nextWord64', 'nextInt', 'nextDouble', 'mkSMGen', 'newSMGen', 'initSMGen']),
+  // mutable cells (haskell-ir.js CELL_OPS): named here so the IR qualifies them whichever wildcard imports a module has
+  ...E('Data.IORef', ['newIORef', 'readIORef', 'writeIORef', 'modifyIORef', "modifyIORef'", 'atomicWriteIORef']),
+  ...E('Control.Concurrent.MVar', ['newMVar', 'newEmptyMVar', 'putMVar', 'takeMVar', 'readMVar', 'modifyMVar_', 'swapMVar']),
+  ...E('Control.Concurrent', ['newMVar', 'newEmptyMVar', 'putMVar', 'takeMVar', 'readMVar', 'modifyMVar_']),
+  ...E('Control.Concurrent.STM', ['atomically', 'newTVar', 'newTVarIO', 'readTVar', 'readTVarIO', 'writeTVar', 'modifyTVar', "modifyTVar'"]),
+  ...E('Control.Concurrent.STM.TVar', ['newTVar', 'newTVarIO', 'readTVar', 'readTVarIO', 'writeTVar', 'modifyTVar', "modifyTVar'"]),
+  ...E('Data.STRef', ['newSTRef', 'readSTRef', 'writeSTRef', 'modifySTRef', "modifySTRef'"]),
   ...E('Crypto.Random', ['getRandomBytes']), ...E('System.Entropy', ['getEntropy']),
   ...E('Network.Wai', ['strictRequestBody', 'lazyRequestBody']),
   ...E('Data.Time.Clock.POSIX', ['getPOSIXTime']), ...E('Data.Time.Clock', ['getCurrentTime']), ...E('System.CPUTime', ['getCPUTime']),
@@ -441,8 +448,8 @@ export const HS_WEB_FRAMEWORKS = Object.freeze({
   yesod: { modules: ['Yesod', 'Yesod.Core', 'Yesod.Auth'], package: 'yesod', tested: ['1.6'] },
 });
 export const HS_WEB_APIS = Object.freeze([
-  ...E('Web.Scotty', ['get', 'post', 'put', 'delete', 'patch', 'options', 'addroute', 'matchAny', 'middleware', 'scotty', 'scottyApp', 'status', 'finish', 'raise', 'raiseStatus', 'redirect', 'text', 'json', 'html', 'file', 'setHeader', 'header', 'headers', 'param', 'params', 'captureParam', 'queryParam', 'formParam', 'body', 'jsonData', 'files', 'rescue', 'liftIO', 'notFound']),
-  ...E('Web.Scotty.Trans', ['get', 'post', 'put', 'delete', 'patch', 'middleware', 'scottyT', 'status', 'finish', 'raise', 'raiseStatus', 'redirect', 'text', 'json', 'html', 'header', 'param', 'captureParam', 'queryParam', 'formParam', 'body', 'jsonData']),
+  ...E('Web.Scotty', ['get', 'post', 'put', 'delete', 'patch', 'options', 'addroute', 'matchAny', 'middleware', 'scotty', 'scottyApp', 'status', 'finish', 'raise', 'raiseStatus', 'redirect', 'text', 'json', 'html', 'file', 'setHeader', 'header', 'headers', 'param', 'params', 'captureParam', 'queryParam', 'formParam', 'body', 'jsonData', 'files', 'rescue', 'liftIO', 'notFound', 'setMaxRequestBodySize']),
+  ...E('Web.Scotty.Trans', ['get', 'post', 'put', 'delete', 'patch', 'middleware', 'scottyT', 'status', 'finish', 'raise', 'raiseStatus', 'redirect', 'text', 'json', 'html', 'header', 'param', 'captureParam', 'queryParam', 'formParam', 'body', 'jsonData', 'setMaxRequestBodySize']),
   ...E('Network.Wai', ['pathInfo', 'requestMethod', 'requestHeaders', 'responseLBS', 'queryString', 'rawPathInfo']),
   ...E('Network.Wai.Handler.Warp', ['run', 'runSettings']),
   ...E('Network.HTTP.Types', ['status200', 'status201', 'status400', 'status401', 'status403', 'status404', 'unauthorized401', 'forbidden403', 'methodGet', 'methodPost', 'methodPut', 'methodDelete', 'hAuthorization', 'hCookie']),

package/src/language/haskell-security-rules.js

@@ -87,6 +87,54 @@ const BOUNDING = /(?:^|\.)(?:take|takeE|isolate|hGetSome|hGet|splitAt|limit|limi
 const REQUEST_PROVENANCE = new Set(['url-param', 'path-param', 'http-body', 'header', 'cookie', 'upload']);
 // wai-extra's `requestSizeCheck (Just n) req` wraps the request so the body read after it fails past n bytes: a bound on every body read of the function.
 const WAI_SIZE_CHECK = /(?:^|\.)requestSizeCheck$/;
+// scotty's global body limit: `setMaxRequestBodySize kb` in the ScottyM block answers every body read of the application it builds with a 413.
+const SCOTTY_BODY_LIMIT = /^Web\.Scotty(?:\.Trans)?\.setMaxRequestBodySize$/;
+
+// Request text handed to a project helper: in `deleteEnum tableId list`, where the caller's `tableId` was read from the request, the helper's
+// first parameter carries caller text for the partial-function, allocation and bounding rules (which do not follow calls themselves). Decided
+// from the IR's RESOLVED call targets and the caller's own request reads, one hop per round, bounded rounds; a parameter the IR could not name
+// (`$arg0`, a pattern) is skipped. Returns qid -> Set of parameter names.
+function requestTextParams(ir) {
+  const byQid = new Map();
+  for (const fir of Object.values(ir.perFile)) for (const fn of fir.functions) byQid.set(fn.qid, fn);
+  const out = new Map();
+  const requestRead = (c) => { const s = sourceInfo(c.callee); return !!s && REQUEST_PROVENANCE.has(s.provenance); };
+  const callsIn = (e) => { const r = []; for (const x of walk(e)) if (x.kind === 'call' && typeof x.callee === 'string') r.push(x); return r; };
+  for (let round = 0; round < 4; round++) {
+    let grew = false;
+    for (const fn of byQid.values()) {
+      const from = new Set(out.get(fn.qid) || []);
+      for (let k = 0; k < 8; k++) {
+        let g = false;
+        for (const n of fnNodes(fn)) {
+          if (n.kind !== 'assign' || !n.target || !n.source) continue;
+          if (!from.has(n.target) && (callsIn(n.source).some(requestRead) || [...names(n.source)].some((x) => from.has(x)))) { from.add(n.target); g = true; }
+        }
+        if (!g) break;
+      }
+      const derived = (e) => [...names(e)].some((x) => from.has(x)) || callsIn(e).some(requestRead);
+      const mark = (x) => {
+        if (!x || !x.hs) return;
+        // a call with every argument, or a partial application (`updateSchema (deleteEnum tableId)`: the IR keeps the supplied arguments as elements)
+        const target = x.kind === 'call' && x.hs.status === 'resolved' ? x.hs.target : (x.kind === 'array' && x.hs.partial ? x.hs.partial.target : null);
+        if (!target) return;
+        const callee = byQid.get(target);
+        if (!callee || callee === fn || !Array.isArray(callee.params)) return;
+        ((x.kind === 'call' ? x.args : x.elements) || []).forEach((a, i) => {
+          const p = callee.params[i];
+          if (!p || typeof p !== 'string' || p.startsWith('$') || !derived(a)) return;
+          if (!out.has(callee.qid)) out.set(callee.qid, new Set());
+          if (!out.get(callee.qid).has(p)) { out.get(callee.qid).add(p); grew = true; }
+        });
+      };
+      for (const n of fnNodes(fn)) for (const ex of [n.value, n.source, n.cond, ...(n.args || [])]) if (ex) for (const x of walk(ex)) mark(x);
+      // a helper the IR inlined at its call site keeps its edge, with the arguments, on `fn.calls`
+      for (const c of fn.calls || []) if (c && c.hs && c.hs.inlined) mark(c);
+    }
+    if (!grew) break;
+  }
+  return out;
+}
 // `replicateM n getLine` reads n lines: the count is an allocation (and an input loop) sized by whoever supplied n.
 const ALLOCATORS = new Set(['replicate', 'Prelude.replicate', 'Data.List.replicate', 'Data.ByteString.replicate', 'Data.ByteString.Char8.replicate', 'Data.ByteString.Lazy.replicate', 'Data.Text.replicate', 'Data.Vector.replicate', 'Control.Monad.replicateM', 'Control.Monad.replicateM_', 'Data.Vector.replicateM', 'Data.List.genericReplicate']);
 const CLAMPS = /^(?:(?:Prelude|Data\.Ord)\.)?(?:min|clamp)$/;
@@ -331,6 +379,44 @@ function weakDigestUse(fn, node, hashCall, hashArgs, fnBase) {
   return reasons;
 }
 
+/**
+ * The migration shape: a fast password hash kept ONLY to verify hashes stored before a password KDF was introduced. True when every
+ * use of the function, in every file, is an operand of an equality (nothing new is stored with it) and a function that compares with
+ * it also computes a password KDF, directly or through one project function (`verify stored salt pw = if legacy stored then stored ==
+ * oldHash salt pw else stored == kdfHash salt pw`). The weak hash is still computed, so the finding is kept at `info` with this reason;
+ * a function that is also used to STORE a hash, or whose comparisons sit beside no KDF, is the full finding. Decided from the IR only,
+ * never from a name (`hashPasswordLegacy` is not evidence; the comparison and the KDF are).
+ */
+function legacyVerifierOnly(fn, ir, memo) {
+  if (memo.has(fn.qid)) return memo.get(fn.qid);
+  const allFns = Object.values(ir.perFile).flatMap((x) => x.functions || []);
+  const byQid = new Map(allFns.map((x) => [x.qid, x]));
+  const usesF = (e) => !!e.hs && ((e.kind === 'call' && e.hs.target === fn.qid) || e.hs.inlinedFrom === fn.qid);
+  const kdfCall = (c) => typeof c === 'string' && (STRONG_KDF.test(c) || PBKDF2_CALLS.has(c));
+  const hasKdf = (g, depth) => (g.calls || []).some((c) => kdfCall(c.callee) || (depth < 1 && c.hs && c.hs.target && c.hs.target !== g.qid && byQid.has(c.hs.target) && hasKdf(byQid.get(c.hs.target), depth + 1)));
+  let uses = 0; let compared = 0; let besideKdf = false;
+  const visit = (e, parent, g, depth) => {
+    if (!e || typeof e !== 'object' || depth > 60) return;
+    if (usesF(e)) {
+      uses++;
+      if (parent && parent.kind === 'binary' && EQUALITY_OPS.has(parent.op)) { compared++; if (hasKdf(g, 0)) besideKdf = true; }
+      return;     // under an inlined root the nodes are the helper's own body, not further uses of it
+    }
+    for (const k of Object.keys(e)) {
+      if (k === 'hs') continue;
+      const v = e[k];
+      if (Array.isArray(v)) { for (const x of v) visit(x, e, g, depth + 1); } else if (v && typeof v === 'object') visit(v, e, g, depth + 1);
+    }
+  };
+  for (const g of allFns) {
+    if (g.qid === fn.qid) continue;
+    for (const n of Object.values(g.cfg.nodes)) for (const root of [n.source, n.value, n.cond, ...(n.args || [])]) visit(root, null, g, 0);
+  }
+  const r = uses > 0 && compared === uses && besideKdf;
+  memo.set(fn.qid, r);
+  return r;
+}
+
 /**
  * @param {Record<string,string>} files path -> source (only Haskell files are read)
  * @param {object} [opts]
@@ -353,7 +439,22 @@ export function analyzeHaskellRules(files, opts = {}) {
     if (/import\s+(?:qualified\s+)?Network\.Wai\.Middleware\.RequestSizeLimit\b/.test(text) || /\brequestSizeLimitMiddleware\b/.test(text)) { requestLimit = { file, line: (text.split('\n').findIndex((l) => /RequestSizeLimit|requestSizeLimitMiddleware/.test(l)) + 1) || 1 }; break; }
   }
 
+  const requestParams = requestTextParams(ir);
+  // A ScottyM block that calls setMaxRequestBodySize bounds every body read of the application it builds: certain for a body read in the SAME
+  // function (an inline route) and for a handler the block references; a block elsewhere in the project is a control whose reach is not verified
+  // (the finding stays, at low, as for the wai middleware).
+  const scottyLimitFns = [];
+  for (const fir of Object.values(ir.perFile)) for (const fn of fir.functions) if ((fn.calls || []).some((c) => typeof c.callee === 'string' && SCOTTY_BODY_LIMIT.test(c.callee))) scottyLimitFns.push(fn);
+  const scottyLimitFor = (fn) => {
+    if (!scottyLimitFns.length) return null;
+    if (scottyLimitFns.includes(fn)) return { certain: true, fn };
+    const refs = (g) => new Set([...((g.hs && g.hs.refs) || []).map((r) => r.target), ...(g.calls || []).map((c) => c.hs && c.hs.target).filter(Boolean)]);
+    const g = scottyLimitFns.find((x) => refs(x).has(fn.qid));
+    return g ? { certain: true, fn: g } : { certain: false, fn: scottyLimitFns[0] };
+  };
+
   let currentFn = null;
+  const legacyMemo = new Map();
   const emit = (rule, file, line, extra = {}) => {
     const r = RULES[rule];
     const sev = extra.severity || r.severity;
@@ -407,8 +508,10 @@ export function analyzeHaskellRules(files, opts = {}) {
       // A value READ from the request inside the function (`raw <- param "page"`, `body <- strictRequestBody req`, `lookup h (requestHeaders req)`)
       // is caller-supplied text wherever the function sits: the framework's own read says who supplied it, so no parameter label is needed.
       const requestRead = (c) => { const s = sourceInfo(c.callee); return !!s && REQUEST_PROVENANCE.has(s.provenance); };
-      const derivedText = new Set([...(fnHs.textParams || []).filter((n) => strongParam.has(n) || entryLike), ...(entryLike ? (fnHs.inputParams || []) : [])]);
-      const fromRequest = new Set();
+      // a parameter a caller fills with request-derived text (`requestTextParams`) is caller text here too, whatever this function's own label
+      const helperParams = requestParams.get(fn.qid) || new Set();
+      const derivedText = new Set([...(fnHs.textParams || []).filter((n) => strongParam.has(n) || entryLike), ...(entryLike ? (fnHs.inputParams || []) : []), ...helperParams]);
+      const fromRequest = new Set(helperParams);
       for (let round = 0; round < 8; round++) {
         let grew = false;
         for (const n of fnNodes(fn)) {
@@ -553,9 +656,14 @@ export function analyzeHaskellRules(files, opts = {}) {
               if (e.hs.con === 'VerifyNone' && e.kind === 'literal' && importsFrom(['OpenSSL.Session'])) emit('hs-tls-validation-disabled', file, e.line || node.line, { detail: 'OpenSSL verification mode VerifyNone', evidence: { mode: 'VerifyNone' } });
             }
             // `xs !! n` raises on a short list: a partial function, written as an operator, applied to caller-supplied text
-            if (e.kind === 'binary' && e.op === '!!' && e.left && mentionsText(e.left) && !isTested([e.left])) {
-              emit('hs-partial-function-on-input', file, e.line || node.line, { detail: '!! on a caller-supplied list', evidence: { partial: '!!' } });
-              continue;
+            // (an INDEX from caller text is the same flaw, `list !! tableId`: out of range raises; a relational test on the index in this function covers it)
+            if (e.kind === 'binary' && e.op === '!!' && e.left && e.right) {
+              const listText = mentionsText(e.left) && !isTested([e.left]);
+              const indexText = !listText && mentionsText(e.right) && !isTested([e.right]) && ![...names(e.right)].some((v) => sizeTestedVars.has(v));
+              if (listText || indexText) {
+                emit('hs-partial-function-on-input', file, e.line || node.line, { detail: listText ? '!! on a caller-supplied list' : '!! with a caller-supplied index', evidence: { partial: '!!' }, inlineCtx: e.hs && e.hs.inlineCtx });
+                continue;
+              }
             }
             if (e.kind !== 'call' || typeof e.callee !== 'string') continue;
             const c = e.callee;
@@ -595,7 +703,11 @@ export function analyzeHaskellRules(files, opts = {}) {
             // password hashed with a fast hash
             // (a hash call the IR inlined from a project helper, `digestFor (credPassword creds)` at the comparison, is reported once, in the helper)
             if (ANY_HASH_CALLS.has(c) && !STRONG_KDF.test(c) && !(e.hs && e.hs.inlinedFrom) && args.some((a) => [...names(a)].some((n) => PASSWORD_VAR.test(n.replace(/^\./, ''))))) {
-              emit('hs-fast-password-hash', file, line, { detail: `${c} applied to a password-shaped value`, inlineCtx: e.hs && e.hs.inlineCtx });
+              if (!(e.hs && e.hs.inlineCtx) && legacyVerifierOnly(fn, ir, legacyMemo)) {
+                emit('hs-fast-password-hash', file, line, { severity: 'info', confidence: 0.5, detail: `${c} applied to a password-shaped value, in a function used only to verify hashes made before a migration (every use of it is an equality, beside a password KDF): re-hash with the KDF at the next successful login and remove it once no stored hash needs it`, inlineCtx: e.hs && e.hs.inlineCtx, evidence: { legacyVerifier: true } });
+              } else {
+                emit('hs-fast-password-hash', file, line, { detail: `${c} applied to a password-shaped value`, inlineCtx: e.hs && e.hs.inlineCtx });
+              }
             }
             if (PBKDF2_CALLS.has(c)) {
               for (const a of args) for (const x of walk(a)) if (x.kind === 'object') for (const p of x.props || []) if (p.key === 'iterCounts' && p.value && p.value.kind === 'literal' && typeof p.value.value === 'number' && p.value.value < 10000) {
@@ -620,12 +732,18 @@ export function analyzeHaskellRules(files, opts = {}) {
               emit('hs-unbounded-allocation', file, line, { detail: `${c} sized by parsed caller text`, evidence: { allocator: c } });
             }
             // a partial function applied to caller text: one malformed value raises an uncaught exception
-            if (PARTIALS.has(c) && args.some((a) => mentionsText(a)) && !isTested(args)) emit('hs-partial-function-on-input', file, line, { detail: `${c} on caller-supplied input`, evidence: { partial: c } });
+            if (PARTIALS.has(c) && args.some((a) => mentionsText(a)) && !isTested(args)) emit('hs-partial-function-on-input', file, line, { detail: `${c} on caller-supplied input`, evidence: { partial: c }, inlineCtx: e.hs && e.hs.inlineCtx });
             // a session/token cookie assembled without HttpOnly or Secure
             if (c === 'Web.Cookie.defaultSetCookie') { /* handled on the record update below */ }
             // unbounded body
             if (BODY_CALLS.has(c) && !sizeChecked && !(c === 'Network.Wai.Conduit.sourceRequestBody' && bounded)) {
-              if (requestLimit) {
+              const scottyLimit = /^Web\.Scotty/.test(c) ? scottyLimitFor(fn) : null;
+              if (scottyLimit && scottyLimit.certain) {
+                controls.push({ kind: 'scotty-body-limit', file: scottyLimit.fn.file, line: scottyLimit.fn.line, appliesTo: `${file}:${line}`, verified: true });
+              } else if (scottyLimit) {
+                controls.push({ kind: 'scotty-body-limit', file: scottyLimit.fn.file, line: scottyLimit.fn.line, appliesTo: `${file}:${line}`, verified: false });
+                emit('hs-unbounded-request-body', file, line, { severity: 'low', confidence: 0.4, detail: c, uncertainty: [{ kind: 'unresolved-import', detail: `a scotty body limit is set (${scottyLimit.fn.file}:${scottyLimit.fn.line}); that this handler is registered in that application is not verified` }] });
+              } else if (requestLimit) {
                 controls.push({ kind: 'request-size-limit', file: requestLimit.file, line: requestLimit.line, appliesTo: `${file}:${line}`, verified: false });
                 emit('hs-unbounded-request-body', file, line, { severity: 'low', confidence: 0.4, detail: c, uncertainty: [{ kind: 'unresolved-import', detail: `a request-size-limit middleware exists (${requestLimit.file}:${requestLimit.line}); that it wraps this handler is not verified` }] });
               } else emit('hs-unbounded-request-body', file, line, { detail: c });

package/src/language/support-registry.js

@@ -243,6 +243,8 @@ export function evaluateSupport(language, input) {
   for (const [cap, req] of Object.entries(REQUIRES_TOOL[language] || {})) {
     const r = out.rows[cap];
     if (r && !tools[req.tool]) out.rows[cap] = { ...r, status: 'blocked', reasons: [`${req.tool} is not available on the measuring host: ${req.why}`, ...r.reasons] };
+    // a GHC other than the recorded toolchain compiles the fixtures against an empty package set: not a verdict on the engine
+    else if (r && req.tool === 'ghc' && tools.ghcMatchesRecorded === false && r.status !== 'supported') out.rows[cap] = { ...r, status: 'blocked', reasons: [`ghc ${tools.ghcVersion} on the measuring host is not the recorded toolchain (${tools.ghcRecorded}); the criterion that needs it could not be judged`, ...r.reasons] };
   }
   const rows = Object.values(out.rows);
   out.summary = { required: rows.length, supported: rows.filter((r) => r.status === 'supported').length, notSupported: rows.filter((r) => r.status !== 'supported').map((r) => `${r.capability}:${r.status}`), dataOk: dataOk.ok };