Skip to content

v0.1.2

Choose a tag to compare

View all tags
@github-actions github-actions released this 27 Sep 22:14
· 5 commits to main since this release
v0.1.2
c3af3e6

SparseRead v0.1.2 — Codex and Pi integrations

Codex and Pi now join the existing Claude Code, OpenCode, OpenClaw and NanoBot
adapters. Framework adapters reuse SparseRead core: reader algorithms and gate
thresholds are unchanged; package version metadata is synchronized to 0.1.2.

Download and install

Download sparseread-source-installer-v0.1.2.zip (Windows or any OS) or
sparseread-source-installer-v0.1.2.tar.gz (macOS/Linux), verify it against
the attached SHA256SUMS, and extract it. Open a terminal in sparseread-v0.1.2.
The source bundle is not an offline binary: dependency installation needs network
access. Prerequisites: uv, Python 3.11+ (commands select 3.12), Node 22+, and the
target agent CLI. Tested host versions: Codex CLI 0.158.0-alpha.2 and Pi 0.87.1.

uv run --python 3.12 python scripts/install_sparseread.py \
  --platform codex --workspace /absolute/path/to/project --doctor

uv run --python 3.12 python scripts/install_sparseread.py \
  --platform pi --workspace /absolute/path/to/project --doctor

Restart/reload in the target project. Review project extensions and Codex hooks
before trusting them. If Pi reports an untrusted project, review its files and
explicitly rerun with --pi-approve: approval applies only to that install
command. SparseRead does not add it automatically. Use --sparseread-mode advisory
for non-blocking guidance; --reader-extras none omits PDF/XLSX dependencies.

The individual Codex/Pi archives are developer/plugin resources, not standalone
Python runtime installers. The source-installer bundle is the recommended end-user
download. This release attaches npm tarballs and Python distributions to GitHub;
it does not claim that these versions were published to npm or PyPI.

Fixes and verification

  • Workspace-local wheel runtime, six tools, conservative once-only native-read
    redirection, native fallbacks, and project configuration preservation.
  • Codex cache-safe reinstalls; Pi persistent JSONL lifecycle, UTF-8, cancellation,
    timeout and restart handling; installation symlink overwrite protection.
  • Stale raw references and core protocol errors are failed tool calls, not
    successful evidence. Raw ranges use Unicode character offsets, not bytes/lines.
  • Actual installed Codex MCP and Pi SDK extension replay five local record
    scenarios. Broad long-log focus required a fixed follow-up to retrieve all
    expected details; JSONL remains native rather than an unsupported sparse reader.
  • Clean source-bundle installation succeeds in a Chinese/space-bearing project
    path. Full regression, Node 22/24 tests and independent Luna/max audit are
    documented in the linked reports.

No private record bodies or full local traces are distributed. Deterministic
tool replay is not an autonomous model quality/token benchmark; real Windows
agent hosts and older host versions remain unverified.
Evidence-only reduction is not total response savings: in these samples, short
logs/summaries returned more JSON characters than the original file. Prefer
native/bounded reads for short or full-fidelity tasks; see the overhead table.

Installation and recovery ·
Real-record validation ·
Independent review

Assets 21
Loading