# ADR 0002: Input policy and resource limits - Status: accepted - Date: 2026-09-29 ## Context The detector accepts individual files, directory trees, and ZIP-based containers. Unbounded traversal or decompression can exhaust memory, CPU, file descriptors, or disk. Extension-only detection can also misclassify binary files as text. The input layer needs deterministic defaults before it is redesigned as a streaming API. ## Decision Limits apply per invocation and can be lowered by configuration or CLI flags. Raising them requires an explicit option; there is no unlimited sentinel. | Resource | Default limit | | --- | ---: | | Plain file or archive member | 8 MiB | | Input archive file | 64 MiB | | Total expanded archive bytes | 256 MiB | | Archive entries inspected | 10,000 | | Expanded/compressed ratio per member | 100:1 | | Directory traversal depth | 64 levels | | Nested archive traversal | disabled (depth 0) | An exact file passed by the user is considered regardless of extension. During directory traversal, configurable extension allowlists select likely prose, markup, configuration, and source-code files. Content still has to decode as strict UTF-8 and must not contain a NUL byte. Unsupported/binary content produces a skipped-input result rather than a clean analysis or a lossy conversion. ZIP-based containers are recognized from both a supported extension (`zip`, `docx`, `xlsx`, or `pptx`) and ZIP magic bytes. An extension/magic mismatch is a typed input error. Members are streamed and counted before allocation. Directory members and encrypted members are skipped with a reason. Absolute paths, platform prefixes, and any path containing a parent (`..`) component are rejected. Archive member names are identities only; members are never extracted to the filesystem by the analysis path. Directory walking does not follow symbolic links. Hidden paths are skipped by default. The default ignore set includes `.git`, `.hg`, `.svn`, `target`, `node_modules`, `vendor`, common generated directories, source maps, and minified assets. Repository ignore files are respected when the eventual walker supports them. An explicitly named file can bypass path/extension ignores, but cannot bypass content or resource-safety limits. One unreadable, invalid, or oversized path yields a per-document input error and does not discard successful documents from the same directory or archive. Results remain ordered by normalized source identity even if reading is later parallelized. ## Consequences - The input API must stream per-document results and track cumulative archive counters. - `Document` needs distinct outer-path and optional archive-member identities. - Size checks must use checked arithmetic and occur before allocating from metadata supplied by an archive. - Tests need boundary cases immediately below, at, and above each configurable limit, plus malicious archive names and misleading extensions. - The README and CLI help must distinguish skipped inputs, operational errors, and clean analyzed documents.