Skip to content

icx v1.1.0 — audit remediation

Choose a tag to compare

@doublegate doublegate released this 20 Nov 14:27
· 9 commits to main since this release
9d59826

icx v1.1.0 is a remediation release for the internal adversarial audit of v1.0.0 (findings ICX-AUD-001 to ICX-AUD-055, in docs/claude/). Of 51 findings, 42 are fixed, 7 partially fixed, 1 open and 1 deferred by owner decision.

Read the Migration section before upgrading. Three changes are visible to existing users: workspaces move to a private per-project directory under $XDG_STATE_HOME/icx/workspaces/ (a repository-local .icx/ is no longer loaded and must be adopted with icx trust), icx proxy requires a per-install token, and icx deref frames its output as untrusted data (--raw for exact bytes).

Install

Prebuilt Linux x86_64 binaries, built from this tag with cargo build --release --locked on Rust 1.100.0, in two C-library flavours:

Asset C library CPU Notes
icx-v1.1.0-x86_64_v3-unknown-linux-gnu.tar.gz glibc 2.34+ (Ubuntu 22.04, Debian 12, RHEL 9 or newer) x86-64-v3 (AVX2: Intel Haswell / AMD Zen or newer) Recommended on mainstream distributions
icx-v1.1.0-x86_64-unknown-linux-gnu.tar.gz glibc 2.34+ any x86-64 Portable CPU baseline
icx-v1.1.0-x86_64_v3-unknown-linux-musl.tar.gz none (fully static) x86-64-v3 Any distribution, including Alpine and older glibc; passes all 9 icx benchmark gates
icx-v1.1.0-x86_64-unknown-linux-musl.tar.gz none (fully static) any x86-64 Most portable; HDC bind measures ~170 ns/op, above the 100 ns benchmark gate, with no visible effect on hook latency

The glibc archives were added after release by the release workflow (CI run 44843132910) and carry GitHub build-provenance attestations (gh attestation verify <archive>.tar.gz --repo doublegate/Infinity-Context); the musl archives were built locally at release time and are unchanged. SHA256SUMS covers all four archives and THIRD-PARTY-NOTICES.txt, which holds the license texts of the statically linked Rust crates. The glibc archives include that file; for the musl archives, which predate it, it is the separate download.

sha256sum -c SHA256SUMS --ignore-missing
tar xzf icx-v1.1.0-x86_64_v3-unknown-linux-gnu.tar.gz
install -m 0755 icx-v1.1.0-x86_64_v3-unknown-linux-gnu/icx ~/.local/bin/icx
icx init && icx doctor

The v3 builds exit with "illegal instruction" on CPUs without AVX2; use a baseline build there.

Performance of these binaries. The figures elsewhere in these notes and in docs/numbers.md were measured on glibc builds tuned to the measuring machine's CPU. On a 2,000-artifact workspace with a heavier hook workload than the published one (so compare rows, not absolute values), median PostToolUse latency was:

Build In-process (ICX_NO_DAEMON=1) Via icx daemon
glibc, CPU-native (reference) 112.7 ms 10.0 ms
glibc baseline CPU (local build, same configuration as the -gnu baseline archive) 140.4 ms 28.7 ms
musl x86-64-v3 (this release) 142.0 ms 10.7 ms
musl baseline (this release) 169.8 ms 29.5 ms

musl's allocator adds about 20-25% to an in-process hook, and the baseline CPU target roughly triples daemon latency; the glibc v3 archive was not separately measured. Through the daemon the musl v3 binary is within about 7% of the native reference. Run the daemon (ICX_DAEMON_AUTOSTART=1) with these binaries. For the best in-process latency, build from source on the target machine. Corresponding source for every binary is this tag (the source archives below).

From source (any platform with Rust 1.100.0):

git clone --branch v1.1.0 https://github.com/doublegate/Infinity-Context.git
cd Infinity-Context
cargo install --path crates/icx-cli   # Rust 1.100.0 (pinned in rust-toolchain.toml)
icx init && icx doctor

Verification

On the release commit: fmt and clippy (-D warnings) clean, 152 tests passing, line coverage 95.11% (gate 95%), docs build with -D warnings, all 9 benchmark SLO gates pass, cargo-deny, REUSE 3.3 and gitleaks clean. Redactor mutation score 74.7% (68.1% at v1.0.0).

Remediation release for the adversarial audit of v1.0.0 in docs/claude/.
Finding IDs (ICX-AUD-NNN) refer to that audit; per-finding status is in
docs/claude/12-v1.1.0-remediation-status.md.

Migration

  • Workspaces now live under $XDG_STATE_HOME/icx/workspaces/<project-hash>/ in a single
    icx.sqlite3. A repository-local .icx/ is ignored with a warning; adopt one you own with
    icx trust --dir <path>. v1.0.x JSON stores are migrated on first load without deleting them.
  • icx proxy requires the install token: use the base URL it prints
    (http://127.0.0.1:49380/t/<token>) or send x-icx-token. Non-loopback binds need
    --allow-remote.
  • icx deref frames output as untrusted data by default; use --raw for exact bytes.

Security

  • Repository-shipped .icx/ workspaces are no longer loaded into agent context; workspaces are
    owner-checked, 0700/0600, sealed with a per-install key and MAC-verified (ICX-AUD-010, -013).
  • The proxy authenticates clients and binds loopback only by default. The curl header files that
    carry upstream API keys are written to the private runtime directory ($XDG_RUNTIME_DIR/icx)
    and stale ones are swept on start (ICX-AUD-012, -020).
  • icx structural markers anywhere in untrusted input, including the latest message, are escaped
    unless they are byte-identical to a block icx emitted; deref output carries a
    trust=untrusted_data provenance header (ICX-AUD-011).
  • Bounded hook stdin, MCP lines (4 MiB, oversized lines drained without buffering) and fold
    depth (64) (ICX-AUD-014).
  • Only structural tool-call/-result blocks for the icx pager are exempt from marker
    neutralization; tool output that merely contains "name": "icx" is not.
  • Reopening a store resets widened icx.sqlite3, -wal and -shm modes to 0600.
  • install.key is checked (regular file, owner, mode) before it is read, so a FIFO is refused
    instead of blocking; only a missing key is created.
  • icx encode-ahtl refuses an untrusted workspace instead of printing a handle it never stored.
  • icx trust adopts a store's content but not its emitted-block digests, so a planted digest
    cannot exempt forged markers from escaping. MCP grep results carry the same
    trust=untrusted_data provenance header as deref and slice.
  • icx mcp-serve and icx proxy re-verify the workspace per request, so a later icx trust
    takes effect and a workspace whose ownership, mode or seal changed is no longer used.
  • Proxy curl header files go to a directory verified to be ours and private
    ($XDG_RUNTIME_DIR/icx, else <tmp>/icx-<uid>) and stale ones are swept every minute.
  • Redactor: URL userinfo passwords for every scheme, SendGrid, Slack app/webhook, GitLab,
    DigitalOcean, Shopify and Twilio formats, more key names, multi-line values, single-line PEM,
    and bounded unterminated key blocks (ICX-AUD-015, -016). A key with no inline value redacts
    only a following scalar line: nested mappings, comments and prose ending in key: are kept.
    Token-only URL userinfo (https://<token>@host, 16+ mixed or 32+ token characters) is
    redacted. A quoted continuation value
    is treated as a nested key only if its closing quote precedes the separator. Private-key banners are matched
    within one -----BEGIN ...----- banner, and text around a key block is kept.
  • agy transcriptPath opens non-blocking without following symlinks; the bus file refuses
    symlinks, foreign owners and permissive modes (ICX-AUD-018, -019).
  • Hooks never echo input fields back as decisions (Claude Code, Codex and agy PreToolUse; agy
    PostInvocation) and never force an agent to continue (agy Stop) (ICX-AUD-007, -009).

Fixed

  • Hook latency no longer grows with workspace size: SQLite store with append-only rows, persisted
    HDC index snapshot, and a bit-sliced bundler. 338 artifacts: 940 ms to 14 ms per hook in
    process; about 3 ms median through the resident daemon (ICX-AUD-001, -047, -050).
  • Storage is lossless after redaction: CRLF, trailing newlines and literal delimiters survive;
    escaping applies only to model-facing views (ICX-AUD-021).
  • 64-bit handle collisions are refused instead of overwriting, in both the CAS and the DAG,
    including Hermes carried-AHTL lines and subagent summaries (ICX-AUD-022, -023).
  • The daemon flushes and releases the workspace lock at most 1 s after the first unsaved change
    even under continuous hook traffic. A failed flush keeps the unsaved state in memory, releases
    the lock and retries every second; after 5 consecutive failures it logs an error and drops the
    changes so the daemon can exit. Requests that never reach a handler (malformed, wrong
    protocol) do not trigger a save.
  • Concurrent first use of the install key, and concurrent prepare of one workspace, no longer
    collide on temp files or observe a directory before its seal is written.
  • A save refuses to seal a snapshot older than the store on disk (another writer committed
    since it was loaded). If the tables do not hold exactly the in-memory rows (rows removed or
    planted behind the process), they are rewritten from memory before sealing. A new
    workspace's unsaved state survives a first flush that failed after creating the store file. The daemon logs an error when another writer's newer
    store replaces its unsaved changes.
  • Emitted-block digests are stored in an append-only, seal-covered table instead of being
    rewritten inside the state blob on every save.
  • Epoch headers are hex everywhere and epoch ids are u32 (ICX-AUD-024); invalid slice_hint
    values are neither rendered nor parsed (ICX-AUD-025); clause chains are parsed iteratively and
    capped at 64 links (ICX-AUD-026).
  • Associative recall of a unique token in 1,000 log documents: top-1 5.1% to 99.0%, top-5 5.6%
    to 99.7% (ICX-AUD-049); one hit per artifact (ICX-AUD-042).
  • Claude Code: shape-preserving updatedToolOutput, plain-text PreCompact instructions,
    last_assistant_message on SubagentStop, and no model-facing output on PostCompact or
    SubagentStop (ICX-AUD-004). Codex: no duplicated additionalContext (ICX-AUD-005).
  • Anthropic proxy placement: client breakpoints are preserved, icx's breakpoint goes on a
    content block, at most 4 in total, and 1h TTLs are inherited (ICX-AUD-003).
  • The bus reports dropped publishes and reader loss instead of returning success
    (ICX-AUD-054).
  • FastCDC uses spread masks (mean chunk 13.2 KiB to 10.6 KiB at an 8 KiB target) (ICX-AUD-027).

Added

  • icx daemon (resident workspace over a private Unix socket), icx trust, ICX_NO_DAEMON,
    ICX_DAEMON_AUTOSTART, ICX_STATE_HOME and ICX_CONFIG_HOME.
  • MCP dual-era support: initialize negotiation for 2024-11-05 through 2025-11-25, plus
    server/discover and per-request versions for 2026-07-28 (ICX-AUD-008).
  • Optional wall-clock event times, enabled for CLI workspaces (ICX-AUD-048).
  • SeqlockTelemetryBus::dropped_publishes and poll_since_with_loss (ICX-AUD-054).
  • WireProxyTransformer::serve_http_listener_gated with a per-request WorkspaceGate.
  • proptest invariants for the codec, cargo-fuzz targets (fuzz/), and a cargo-deny gate
    (ICX-AUD-039, -040).

Changed

  • The benchmark's cache line is a byte-weighted prefix-reuse proxy over all turns, seal turns
    included (95.46% on the synthetic transcript), gated at 90%. Compression and CAS fixtures are
    labelled and the CAS fixture is no longer a single chunk (ICX-AUD-002, -027, -028).
  • CI actions are pinned to commit SHAs; the toolchain is pinned to 1.100.0 (ICX-AUD-040).
  • icx-proxy and icx-adapters declare GPL-3.0-or-later AND AGPL-3.0-or-later, matching
    REUSE; LICENSING.md's description of GPL section 13 is corrected (ICX-AUD-034).
  • Specifications, threat model, deployment guide, README and the AHTL PDF volume describe v1.1.0
    behaviour; canonical figures are in docs/numbers.md (ICX-AUD-017, -030,
    -031, -046, -051).

Removed

  • 10 arXiv PDFs whose license does not permit redistribution; INDEX.md links to arXiv
    (ICX-AUD-036).
  • Unused proxy dependencies icx-codec and serde (ICX-AUD-041); icx-cli depends on
    serde and rustix only on Unix.
  • Personal paths and an agent conversation id from tracked files (ICX-AUD-038).

Known limitations

  • Fractal suffix depth is still inferred from distance and can report false positives
    (ICX-AUD-052).
  • Codex does not yet apply updatedMCPToolOutput; under Codex icx indexes output but does not
    shrink it. The Codex rollout tailer does not read .jsonl.zst files and is not wired into the
    CLI (ICX-AUD-005).
  • No Hermes Python shim ships with icx (ICX-AUD-006).
  • The rest of the derived state (DAG, epochs, HDC event log, graph) is still rewritten in full
    on each save, so flush cost grows with history. The daemon flushes after any dispatched
    request, including hooks that change nothing.
  • Licensing and patent governance text (PATENTS.md, CLA.md, COMMERCIAL.md) is unchanged by owner
    decision (ICX-AUD-035).