v1.1.0 remediation status¶
Status of every finding in this audit after the v1.1.0 release work on branch claude-audit.
Commit hashes refer to that branch. "docs" means the fix is in the documentation updated for v1.1.0.
Legend: fixed = remediated and covered by a test or a doc change; partial = the code-side or main part is fixed and the rest is stated; open = not addressed in v1.1.0; deferred = not addressed by owner decision.
| ID | Severity | Finding | Status | Where |
|---|---|---|---|---|
| ICX-AUD-001 | Critical | Real hook latency grows linearly with history; load re-encodes all HDC state | fixed | c03d9ea, 61cb0f2 |
| ICX-AUD-002 | High | >=94% prefix-cache hit rate is an artifact of 22.8 KB padding | fixed | 37db9a2 |
| ICX-AUD-003 | High | Anthropic proxy cache strategy likely reduces cache hits (message-level cache_control; strips client breakpoints; rewrites history) | fixed | 110e116 |
| ICX-AUD-004 | High | Claude Code hook outputs routed to user or discarded; SubagentStop reads wrong field | fixed | 129988d |
| ICX-AUD-005 | High | Codex updatedMCPToolOutput unsupported; additionalContext adds tokens; documented config wrong; zstd rollouts unsupported | partial | 129988d; upstream Codex does not apply updatedMCPToolOutput; rollout tailer zstd not implemented |
| ICX-AUD-006 | High | Hermes shim cannot load; hooks.yaml and events do not exist upstream | partial | docs only; no Hermes shim ships |
| ICX-AUD-007 | Medium | agy config locations contradictory; schema and Stop reasons wrong; PreToolUse allow default | fixed | 129988d, c35f15c; docs |
| ICX-AUD-008 | Medium | MCP protocolVersion hard-coded 2024-11-05; no negotiation or server/discover | fixed | 129988d |
| ICX-AUD-009 | Info | PreToolUse echo of output-only fields is dead code | fixed | 684a4ab, c35f15c |
| ICX-AUD-010 | High | Repo-shipped .icx workspace auto-loaded and injected into agent context | fixed | c03d9ea, 61cb0f2, 7dd8e00 (trust adopts only after load), 1fa72ca (atomic install key) |
| ICX-AUD-011 | Medium-High | In-band trust markers bypass Tier-0 escaping (proxy and Hermes) | fixed | 110e116, 684a4ab, e0f225b |
| ICX-AUD-012 | Medium | Proxy has no client auth; /health and /sse-expand serve CAS data; any bind accepted | fixed | 110e116 |
| ICX-AUD-013 | Medium | icx init writes no .gitignore; workspace dir 0755 | fixed | c03d9ea |
| ICX-AUD-014 | Medium | Unbounded hook stdin (300 MiB -> 1.12 GB RSS / 4.32 s) | fixed | 110e116 |
| ICX-AUD-015 | Medium | Secret redactor recall gaps (URL userinfo; dotted tokens; multi-line) | fixed | 684a4ab |
| ICX-AUD-016 | Medium | Single-line PEM block deletes rest of input | fixed | 684a4ab |
| ICX-AUD-017 | Medium | Threat model documents absent controls (XDG/UDS/SO_PEERCRED/loopback/bounded streaming) | fixed | docs (threat model rewrite) |
| ICX-AUD-018 | Low | agy hook hangs on FIFO transcriptPath | fixed | 684a4ab |
| ICX-AUD-019 | Low | Bus file open lacks O_NOFOLLOW and owner check | fixed | 684a4ab |
| ICX-AUD-020 | Low | curl header temp file with API keys survives crash | fixed | 110e116 |
| ICX-AUD-021 | High | Store not lossless: CRLF/trailing newline/delimiter escaping collapse distinct inputs | fixed | 684a4ab |
| ICX-AUD-022 | Medium | put_bytes truncates silently; MMU falls back from checked to unchecked on any error | fixed | 684a4ab |
| ICX-AUD-023 | Low | 64-bit ArtifactId as primary key; DAG insert overwrites on collision | fixed | e0f225b |
| ICX-AUD-024 | Low | Epoch header written decimal and parsed hex | fixed | e0f225b |
| ICX-AUD-025 | Low | Unescaped slice_hint enables stream-level packet forgery | fixed | e0f225b |
| ICX-AUD-026 | Low | Unbounded clause recursion aborts process | fixed | e0f225b |
| ICX-AUD-027 | Low | CDC benchmark input is a single chunk; low-bit Gear masks | fixed | 046b461, 0bfeb18 |
| ICX-AUD-028 | Medium | Benchmark numbers inconsistent; AVX-512 timings on non-AVX-512 host; native-CPU builds | fixed | 37db9a2; docs |
| ICX-AUD-030 | Medium | AHTL PDF Chapter 08 stale (v0.3.0) under v1.0.0 cover | fixed | 4ad33b2 |
| ICX-AUD-031 | Medium | Cross-document numeric contradictions incl. epoch defaults 4096 vs 16384/8192 | fixed | docs; docs/numbers.md |
| ICX-AUD-032 | Low | Exactness review self-contradictions and wrong 48-bit birthday figure | partial | corrections in ref-docs/corrections-2026-11-18.md; review text unchanged (immutable corpus) |
| ICX-AUD-033 | Medium | Mistitled ACM paper; misattributed 10x/4-8x; unsourced 19.0%; 128-bit claim | fixed | ref-docs/corrections-2026-11-18.md (immutable corpus) |
| ICX-AUD-034 | Medium | Cargo vs REUSE vs LICENSING.md license declarations disagree; GPL s13 misstatement | fixed | 8b1dc28 |
| ICX-AUD-035 | Medium | PATENTS/CLA further restrictions vs GPLv3 s7/s10/s11; patent-pending unsubstantiated | deferred | owner decision: governance text not redrafted |
| ICX-AUD-036 | Medium | 10 arXiv non-exclusive PDFs redistributed; Zep NC-SA; Wikimedia/Unicode labels | partial | 8b1dc28 (arXiv PDFs, Zep label); Wikimedia/Unicode labels open |
| ICX-AUD-037 | Low | SECURITY.md has no contact; CLA not a required check | partial | 8b1dc28 (SECURITY.md); CLA as required check is a repository setting |
| ICX-AUD-038 | Low | Personal paths in public docs; large/duplicate binaries; patents/ only locally excluded | partial | 8b1dc28 (paths, patents/ ignore); large binaries and git history unchanged |
| ICX-AUD-039 | Medium | Mutation score 71.7% (redact.rs 68.1%); no proptest/fuzz/bench; weak asserts | partial | 37db9a2 (proptest, fuzz); redact.rs mutation re-run 2026-11-20 at 7270ae6: 472 caught + 42 timeouts of 688 viable (74.7%; 73.4% on 2026-11-19, 68.1% at v1.0.0), 174 missed remain |
| ICX-AUD-040 | Low | Actions tag-pinned; unpinned llvm-cov install; unpinned toolchain; no deny.toml | fixed | 8b1dc28 |
| ICX-AUD-041 | Low | Unused deps icx-codec+serde (proxy) and serde (cli) | fixed | 110e116 |
| ICX-AUD-042 | Low | Tail slot recalls the same artifact repeatedly | fixed | 046b461 |
| ICX-AUD-046 | Medium | Bus is not /dev/shm lock-free shared memory (file pread/pwrite + writer lock) | fixed | b404ca6; docs |
| ICX-AUD-047 | Medium | Three-file save not atomic; DAG/epochs not verified on load | fixed | c03d9ea |
| ICX-AUD-048 | Medium | Bi-temporal timestamps are a synthetic counter | fixed | 046b461 (opt-in wall clock) |
| ICX-AUD-049 | High | Associative recall collapses with corpus size (top-1 5.1% at 1k docs) | fixed | 046b461 |
| ICX-AUD-050 | Medium | i16 bundler saturation; scalar per-bit accumulate dominates CPU | fixed | c03d9ea |
| ICX-AUD-051 | Low | Titans surprise gate is a hand-tuned heuristic | fixed | docs: Titans-inspired naming |
| ICX-AUD-052 | Low | Fractal suffix depth inferred from distance (false positives) | open | deferred |
| ICX-AUD-053 | Low | hv_digest_u64 linear fold collides trivially | fixed | e0f225b (documented) |
| ICX-AUD-054 | Medium | Bus drops events while returning success; readers skip gaps silently | fixed | b404ca6 |
| ICX-AUD-055 | Low | 4-agent propagation latency not measured cross-process | fixed | b404ca6 (two-handle file-backed test) |
Corrections to the audit itself¶
- ICX-AUD-003: the proxy's folded history turned out to be byte-stable across turns (verified by test in
crates/icx-proxy/tests/cache_placement.rs). The claim that it "rewrites cached history every turn" was overstated; the real defects were client-breakpoint stripping and message-level placement, both fixed. - ICX-AUD-002: including seal turns changed the synthetic figure only slightly (95.38% to 95.46% byte-weighted). The benchmark still uses the same synthetic transcript with its large static prefix, so the number remains a property of that fixture; v1.1.0 labels it as a byte proxy rather than a cache hit rate, and real-transcript measurement is still to do.