Skip to content

00 — Executive Summary

Scope: all 8 crates (~23k lines of Rust), every document in docs/ and ref-docs/ including the 58-page AHTL PDF, CI and supply chain, licensing, and the agent integrations verified against current upstream APIs. Subject: v1.0.0 at c176d03.

Result: 51 findings.

Critical High Medium-High Medium Low Info
1 8 1 22 18 1

Bottom line

The engineering hygiene is good:

  • zero unsafe in first-party code;
  • clean clippy-pedantic;
  • real 95% coverage;
  • no dependency advisories;
  • a pinned and verified secret scanner;
  • careful bounded parsing in most places.

The product claims are not supported by the shipped system:

  • Latency. The hook path is about 1,900× slower than documented at 332 stored tool outputs (~940 ms vs <500 µs), and the slowdown grows linearly with every further output (ICX-AUD-001).
  • Cache hits. The headline ≥94% cache-hit rate is produced by benchmark padding (ICX-AUD-002).
  • Provider caching. The proxy's caching strategy probably works against the provider caches (ICX-AUD-003).
  • Integration. Several adapters send context to fields the agents ignore or show only to the user (ICX-AUD-004 to ICX-AUD-006).
  • "Lossless". The store is not lossless (ICX-AUD-021).
  • Recall. Associative recall degrades to near chance at a thousand documents (ICX-AUD-049).
  • Security. One security finding is High: a cloned repository can plant persistent prompt injections in the agent's first turn (ICX-AUD-010).

Top 10 findings

# ID Sev. Finding Evidence
1 ICX-AUD-001 Critical Each hook reloads and re-encodes the entire HDC history. Real icx hook latency: 36 ms at ~14 artifacts, 296 ms at ~104, ~940 ms at 332; perf puts 87.5% in HdcBundler::add_delta. The <500 µs SLO measures an in-memory path. measured, perf
2 ICX-AUD-010 High A repo-shipped .icx/ directory is loaded without trust checks, and its text is injected into SessionStart/UserPromptSubmit context. Full bodies are served by deref with no untrusted marker. PoC (private)
3 ICX-AUD-002 High The ≥94% hit rate is a byte-prefix ratio inflated by a 22.8 KB static head. Without the padding it is 69.9%. It is not a provider measurement. PoC
4 ICX-AUD-004/005/006 High Claude Code systemMessage outputs go to the user, not the model. Codex updatedMCPToolOutput is unsupported and additionalContext adds tokens. The documented Hermes/Codex configurations cannot load. upstream docs, local sources
5 ICX-AUD-003 High The Anthropic proxy places cache_control on message objects, strips client breakpoints, and rewrites already-cached history every turn. docs + code
6 ICX-AUD-049 High HDC recall of a unique token: top-1 28% at 100 docs, 5.1% at 1,000 docs. The Stage-1 pre-screen uses 64 of 10,240 bits. PoC
7 ICX-AUD-021 High "Lossless" storage collapses CRLF/LF, trailing newlines, and ⟪ vs << into the same artifact. PoC
8 ICX-AUD-011 Med-High Content that starts with icx's own markers (⟦, ⟪ICX:STUB│, …) bypasses Tier-0 escaping and redaction in the proxy and the Hermes adapter. code path
9 ICX-AUD-012/013/014 Medium Unauthenticated proxy routes serve CAS data; .icx/ is not git-ignored; unbounded stdin (300 MiB → 1.12 GB RSS, 4.3 s). PoC, measured
10 ICX-AUD-015/016 Medium The redactor misses URL credentials, dotted tokens and multi-line secrets. A one-line PEM deletes the rest of the input. The redactor's mutation score is 68%. PoC, mutants

Claim-vs-reality scorecard

Claim Verdict
Zero unsafe across 8 crates True (first-party; dependencies use unsafe as expected)
≥ 95% coverage True (95.75%; 95.78% production-only)
100% lossless CAS / DAG False (ICX-AUD-021, -022)
Sub-50 µs associative index Latency true; recall quality false (ICX-AUD-049)
< 500 µs synchronous hook False for the shipped binary (ICX-AUD-001)
≥ 94% prefix-cache hit rate Unsupported: a fixture artifact (ICX-AUD-002, -003)
Lock-free /dev/shm seqlock telepathy bus Inaccurate: file I/O with a writer lock; silent loss (ICX-AUD-046, -054)
Bi-temporal graph Synthetic timestamps (ICX-AUD-048)
Tier-0 escaping prevents packet forgery Bypassable (ICX-AUD-011, -010)
Loopback/UID isolation, UDS + SO_PEERCRED Not implemented (ICX-AUD-012, -017)
Claude Code / Codex / agy / Hermes integration Partially working; several outputs never reach the model (ICX-AUD-004 to -008)
AHTL v1.0 specification volume Stale Chapter 08 (ICX-AUD-030)
Benchmarked SLO numbers Inconsistent across documents, up to 100× (ICX-AUD-028, -031)
Research citations Mostly accurate; several misattributed or unsourced figures (ICX-AUD-033)

Prioritized remediation roadmap

Ticket IDs follow the project's T-PS-NNN convention. They are proposals, and nothing has been added to to-dos/ yet.

  1. Now (security and correctness):
  2. Move workspaces out of the repository and refuse untrusted .icx/ directories (010, 013).
  3. Remove the prefix-based trust checks (011).
  4. Add a token to the proxy and refuse non-loopback binds (012).
  5. Bound stdin, MCP line length and fold depth (014).
  6. Fix the PEM swallow and the URL-credential leak (015, 016).
  7. Remove the checked-to-unchecked CAS fallback (022).
  8. Next (make the core promise true):
  9. Persist the HDC index and run a resident daemon, so the hook p99 stays under 50 ms at 10k artifacts (001, 047, 050).
  10. Replace the Stage-1 pre-screen and add a token index (049).
  11. Move virtualization to the hook path and stop rewriting history in the proxy (003).
  12. Fix the adapter outputs against contract fixtures (004 to 008).
  13. Then (honesty and governance):
  14. Regenerate all published numbers from one benchmark JSON, drop the padded cache metric, and measure provider cache tokens (002, 028, 031).
  15. Rebuild or unpublish the stale PDF (030).
  16. Correct the citations (033).
  17. Align the licenses and have counsel review PATENTS/CLA (034 to 036).
  18. Add a security contact and make the CLA check required (037).
  19. Continuously (test strength):
  20. proptest round-trips, cargo fuzz on every parser, and a mutation gate of at least 85% on the codec (039).
  21. Agent contract fixtures pinned to each agent version (006 to 008).

What is solid

  • Integrity on read: BLAKE3 verification on every CAS read, and fail-closed ambiguous handles.
  • Bounded parsing: panic-free parsing on almost every path; recursion depth (026) is the one exception.
  • Seqlock memory ordering: the in-process seqlock ordering is correct.
  • Atomic file writes: temp file, 0600, rename.
  • Dependencies: a small dependency set with no advisories, and a well-built gitleaks gate.
  • CLA workflow: it avoids the classic pull_request_target checkout pitfall.
  • Real-input compression: inline footprint reduction on real inputs (34–55×) is within the documented range.