1.0.0 — 2026-07-14

Release Notes

Halfhand 1.0.0. The first stable release. From here on the contracts in
STABILITY.md hold: CLI flags/exit codes, --json output, the on-disk
database, MSRV, and the hh-core Rust API change only additively (or via a
documented deprecation). A database written by v0.1.0-beta.1 opens and
reads correctly under 1.0 — proven by the permanent
hh/tests/migration_from_beta.rs against a committed beta-created fixture.

Highlights

• Secret redaction (hh scan, hh redact): detect secrets in recorded
  sessions, report them without ever printing the value, and irreversibly
  redact in place — events and blobs rewritten, originals securely deleted,
  DB compacted. Exits 4 on findings so CI can gate on a clean scan.
• Export / HTML replay (hh export --bundle|--html, hh import):
  portable .hh archives (manifest + events + referenced blobs, zstd), a
  self-contained single-file HTML replay page (hh replay --web), and hh import to bring a bundle back under a new id. Exports are always
  redacted by default; opting out needs an interactive TTY confirmation.
• Codex CLI, Gemini CLI, and Claude Desktop adapters in addition to
  Claude Code, each with a fuzz target that never panics on malformed input.
• Cross-session FTS5 search (hh search): words, phrases, prefix *,
  AND/OR/NOT, with --agent/--kind/--since/--path filters and --json.
• Windows support (build-only per SRS §2.2): CI builds on
  windows-latest every push; PTY recording on Windows remains best-effort.
• Stability contract (STABILITY.md): the 1.0 promise for CLI, --json
  (schema frozen at 2), the on-disk DB (forward-only migrations),
  MSRV (1.75, minor-release bumps only), and hh-core semver (enforced by
  the cargo-semver-checks CI job).
• Release pipeline via cargo-dist: per-target binaries (mac/linux
  gnu+musl, windows msvc), SHA-256 checksums, shell + powershell installers,
  a Homebrew formula, and keyless GitHub artifact attestations
  (gh attestation verify). hh completions <shell> and a generated hh.1
  man page ship in every archive.
• Docs site (mdbook) on GitHub Pages, with a printable hand-run
  docs/manual-qa.md checklist and docs/1.0-definition-of-done.md mapping
  every release bar to its proof.

Added — 1.0 release-readiness

• hh completions <shell>: prints a shell completion script (bash, zsh,
  fish, powershell, elvish) via clap_complete (FR-6.2 ergonomics). New
  docs/completions.md page; --help example per shell. The scripts and the
  hh.1 man page are also generated at build time by the new hh-dist
  workspace crate (just dist-assets) and shipped in release archives.
• cargo-dist release pipeline: tag-triggered (v*) releases now produce
  per-target binaries (mac/linux gnu+musl, windows msvc) with the git sha
  embedded (HH_BUILD_SHA, NFR-8), SHA-256 checksums, a shell + powershell
  installer, keyless GitHub artifact attestations (verify with gh attestation verify), and an auto-pushed Homebrew formula to
  halfhand-org/homebrew-tap. Config lives in dist-workspace.toml; the
  workflow is generated by cargo dist generate. See CONTRIBUTING.md
  "Cutting a release".
• mdbook docs site under docs/, published to GitHub Pages on release
  (and on push to main as a preview) via .github/workflows/docs.yml. The
  existing docs/*.md and docs/adr/*.md are rendered in place (no path
  moves, preserving the 93 in-repo references); new pages: Quickstart,
  Recording, Replay & Inspect, Adapters, FAQ, Stability, Security.
• Repo hygiene: bug-report and feature-request issue templates (the bug
  form requires hh --version and hh stats output), a PR template encoding
  the CLAUDE.md done-definition, CODE_OF_CONDUCT.md (Contributor Covenant
  2.1), and CONTRIBUTING.md (dev setup, adapter fixture-capture guide, fuzz/
  bench instructions, release runbook).
• hh-dist crate: a publish = false workspace crate that generates
  completion scripts and the man page from the single Cli definition, so
  cargo install halfhand still installs only hh.

Changed — 1.0 release-readiness

• README rewritten for 1.0 (245 → 116 lines): hero, install matrix
  (cargo/brew/script), 30-second demo, the honest "faithful playback" framing,
  badges, and a short "why local-first" section. Depth moved to the docs site.
• Release workflow replaced: the bespoke .github/workflows/release.yml
  is superseded by cargo-dist's generated workflow; re-run cargo dist generate to emit it from dist-workspace.toml.
• just fuzz-all now runs all eight fuzz targets (was missing
  codex_jsonl, gemini_jsonl, import_bundle).

Added — 1.0 stability groundwork

• STABILITY.md: the 1.0 promise for CLI flags/exit codes (additive-only,
  deprecate-before-remove), --json (schema frozen, additive-only, see
  below), the on-disk DB (forward-only automatic migrations; --json and
  hh export bundles — not raw SQL — are the supported programmatic
  interface), MSRV (bumps are minor releases only), and hh-core's Rust API
  (enforced today by the cargo-semver-checks CI job).
• JSON schema bumped to schema:2, frozen for the 1.0 series: additive
  changes (new fields, new enum values, new body shapes) no longer bump the
  integer, reversing the beta-era policy. schema:2 folds in drift that
  shipped without a version bump in the previous entry below — three new
  agent_kind values and the redaction_audit lifecycle body — documented
  field-by-field with a diff from schema:1 in docs/json.md. A new
  hh-core::event::JSON_SCHEMA_VERSION constant is now the single source of
  truth every JSON-emitting call site (including hh-core::bundle, for hh export --bundle) reads from, replacing several independently hardcoded
  1s.
• Docs/code reconciliation audit: README.md's command table was missing
  hh gc/hh stats/hh import/hh search and still called the Codex/
  Gemini/Claude Desktop adapters "planned for v0.2" after they'd shipped
  (see the entry below); docs/json.md's agent_kind table and
  docs/search.md's --agent filter list were missing claude-desktop (and
  docs/search.md also missing mcp-only); hh mcp-proxy --session-hint's
  --help text implied it was stored when it is in fact accepted but not
  yet persisted anywhere (now documented honestly in docs/mcp-proxy.md).
• hh-core::deprecation::warn_deprecated: the standard stderr format for
  a deprecation warning (hh: warning: <what> is deprecated; <guidance>),
  deduplicated per unique warning id so a call site can never print more than
  once per invocation. Config::load's legacy halfhand.toml/hh.toml
  fallback notice now goes through it.
• CLI conformance test (hh/tests/cli_conformance.rs): snapshot-tests
  hh --help and every hh <sub> --help, and asserts the documented exit
  codes — 2 usage error, 3 session not found, 4 hh scan with
  findings.

Added — Codex CLI, Gemini CLI, and Claude Desktop adapters; cross-session search (docs/search.md, docs/adr/0003-fts-index-strategy.md)

• Codex CLI adapter: tails ~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl,
  converting response_item (message, reasoning, function_call,
  custom_tool_call) and event_msg (user_message, agent_message,
  token_count, exec_command_end) records to the same structured event model
  as the Claude Code adapter (agent_kind = codex-cli).
• Gemini CLI adapter: tails ~/.gemini/tmp/<hash>/chats/session-*.jsonl,
  converting user/gemini records (tool calls, thoughts, tokens)
  (agent_kind = gemini-cli).
• Claude Desktop adapter: reuses the Claude Code JSONL tailer/parser
  (identical transcript format) under its own agent_kind = claude-desktop
  so hh list/hh inspect distinguish the two.
• Force any of the four with hh run --adapter <claude-code|claude-desktop| codex-cli|gemini-cli> -- <command>; auto-detection matches on the wrapped
  command's basename same as before.
• hh search <query> [--agent] [--kind] [--since] [--path] [--limit] [--json]: FTS5 full-text search across every recorded session's event
  summaries and message bodies, with highlighted snippets in the table view.
  Backed by a new events_fts virtual table (migration 0004, additive) kept
  incrementally in sync by the writer thread on every append, and maintained
  (rows removed) on hh delete and hh redact — a redacted secret is
  provably not findable via search (property-tested). --limit defaults to
  50, capped at 500.
• Fuzz targets for the new Codex/Gemini JSONL parsers (fuzz/fuzz_targets/ codex_jsonl.rs, gemini_jsonl.rs) and a search benchmark (target <100ms
  across 100 sessions × 1000 events).
• AgentKind marked #[non_exhaustive] (new variants append at the end, so
  existing discriminants don't shift); cargo-semver-checks CI excludes the
  one-time enum_marked_non_exhaustive lint for it.

Added — redaction pipeline (SRS §8, docs/redaction-design.md)

• Secret detection engine (hh-core::redact): compiled detectors for
  AWS access key ids, GitHub/GitLab/Slack tokens, PEM private-key blocks
  (including truncated fragments), JWTs, a conservative high-entropy
  scanner, and user-defined regexes via the new config [redaction] rules.
  Matches are replaced with {{REDACTED:<type>:<hash8>}} where hash8 is
  a BLAKE3 prefix of the secret — one-way, but stable, so the same secret
  correlates across events without being stored. Contract (property-tested
  and fuzzed via the new redact_detect target): false positives are
  acceptable, false negatives on the named token types are bugs.
• hh scan <session|last|--all> [--json]: reports detected secrets
  (type, step, location, hash8) without ever printing them. Exit code 4
  when findings exist — the documented redaction/policy exit code — so CI
  can gate on a clean store.
• hh redact <session> [--yes]: irreversibly redacts a recorded
  session in place — rewrites events and affected blobs (refcounts move
  with the references), securely deletes originals (best-effort zero
  overwrite + unlink), appends a redaction_audit lifecycle event, and
  checkpoints + VACUUMs so no plaintext survives in the WAL or freelist.
  Property-tested: after redact, seeded secrets are undetectable anywhere
  in hh.db or the blob store.
• hh export [session] [--out FILE] [--html] [--no-redact]: exports a
  session as a schema:1 JSON bundle or a self-contained HTML page.
  Redacted by default, always — the whole bundle passes through one
  redaction chokepoint before a byte is written. --no-redact requires an
  interactive typed confirmation and is refused on a non-TTY stdin (no
  bypass flag, by design).
• Record-time redaction (opt-in, config [redaction] at_record = true):
  matches are scrubbed before hitting disk — event summaries/bodies on
  the writer task, blob content before hashing/compression.
• New docs: docs/redaction.md (threat model, what is/isn't caught,
  config), docs/redaction-design.md (invariants), and a repo-root
  SECURITY.md with the disclosure policy.
• New dependency: regex (linear-time matching — no catastrophic
  backtracking on hostile recorded input).

Added — portable session bundles, hh import, interactive HTML replay (docs/export-import.md)

• hh export --bundle [-o FILE]: a portable, deterministic session
  archive — a zstd-compressed tar of manifest.json (format version, hh
  version, session metadata, integrity hashes) + events.ndjson (one
  compact JSON object per event, position-indexed since DB row ids don't
  survive import) + every referenced blob, content-addressed. Byte-identical
  on repeat export of an unchanged session (fixed entry order, fixed
  tar-header metadata, no wall-clock fields). Redacted by default through
  the same chokepoint as plain hh export/--html; if redaction changes a
  blob's hash, every reference to it in the bundle is rewritten to match.
  hh export <session> with no flags is unchanged (still plain JSON to
  stdout) — --bundle is additive, opt-in, and refuses to write its binary
  output to an interactive terminal without -o.
• hh import FILE: validates a bundle end to end (bounded zstd
  decompression, a tar-entry allow-list, per-blob BLAKE3 verification, an
  events.ndjson integrity digest, and full referenced-blob-exists checks)
  before importing it under a brand-new local session id — the original id
  is preserved in the new sessions.imported_from column/JSON field
  (additive migration 0003). Property-tested: export → import → export
  round-trips every event's content and blob bytes exactly. Fuzzed (new
  import_bundle target): never panics on malformed/corrupt/hostile input.
• hh export --html redesign: from a static "step report" into a real
  self-contained interactive replay page — a clickable step timeline, a
  detail pane (pretty JSON / a real diff viewer with precomputed hunks /
  plain text depending on event kind), j/k keyboard navigation mirroring
  the replay TUI, and a dark-by-default theme with a light toggle. Zero
  network requests, no CDN, no build step (hand-written vanilla JS). Session
  data — untrusted input, since it's whatever the recorded agent/tool output
  contained — is embedded once as escaped JSON inside a
  <script type="application/json"> tag and only ever reaches the DOM via
  createElement/textContent, never innerHTML, so a <script> tag or
  event-handler attribute anywhere in recorded content renders inert.
• hh replay --web: sugar for hh export --html, writes the page to a
  temp file and prints the path. Needs no terminal (unlike the TUI); does
  not auto-open a browser or start a server.
• New dependency: tar (the same crate Cargo itself uses for package
  tarballs — pure Rust, no network, no unsafe in the paths used here).

Diagnostics

• A degraded Claude Code adapter session now self-documents in the database,
  not just on stderr: the specific degrade reason is persisted as an error
  event (body_json = {"adapter":"claude-code","reason":"…"}), so
  SELECT … WHERE kind='error' explains why a session recorded no structured
  steps — previously such a query returned empty and the failure was a silent
  mystery. Reasons are now specific: no jsonl matched cwd slug <slug> (with
  the slug dir it looked in + candidate counts), jsonl found (<file>) but 0 records parsed (read N line(s); first parse error at line K: <msg>),
  found a transcript at <file> but could not read it, or discovery selected file <file> but it's a directory. A found-but-empty transcript now degrades
  with 0 records parsed instead of finalizing active with 0 steps. Set
  HH_DEBUG=1 during hh run to capture a discovery+parse trace (computed
  slug, projects dir, candidate files, selected transcript, records read,
  events produced, first conversion failure) to stderr. Note: a degraded
  session now reports 1 step (the error event itself) rather than 0.

Fixed

• Silent recording breakage (the reported "0 steps / 0 files but
  status=ok" symptom): the Claude Code adapter's transcript locator used a
  fixed 3 s deadline and matched the first cwd-bearing record, so on
  recent Claude versions — which write ~/.claude/projects/<slug>/*.jsonl
  lazily (only on first output) and omit cwd from the early system/meta
  records — the locator either timed out before the file appeared or rejected
  it on the cwd-less head, finalized ok with 0 structured steps, and printed
  no warning. The locator now polls until the session stops, snapshots
  pre-existing transcripts so it only matches one created during this
  session, and scans past cwd-less records to the first cwd-bearing one (and
  verifies it belongs to this session's cwd). Every degrade path now carries
  an actionable degrade_reason printed after the child exits (FR-1.5), and
  the hh run epilogue warns when an adapter-active claude-code session
  records 0 steps over >60 s.
• A halfhand.toml (or hh.toml) config file is no longer silently ignored.
  When config.toml is absent but a legacy file exists in the same directory,
  hh now loads it as a fallback (so ignore globs / a custom data_dir take
  effect) and emits a single one-line deprecation hint on stderr suggesting the
  rename to config.toml. When config.toml is present, a sibling legacy
  file is still genuinely ignored (the canonical path wins) — hh warns on
  stderr at startup and hh doctor reports it as a failing check, naming the
  ignored file and where to move its contents.
• The FS watcher no longer aborts hh run with "recording failed" when its
  cwd is unwatchable (e.g. notify rejects a recursive watch with EACCES).
  It degrades instead: a single stderr warning points at hh doctor, file
  recording is skipped, and the PTY + adapter session still records
  (status=ok). A per-directory fallback also keeps file recording working
  when one unreadable subdir would have blacked out the whole tree.
• The FS watcher now drains in-flight events for a bounded window (150 ms of
  silence, capped at 1 s) after the child exits, before finalizing, and then
  runs a deterministic re-scan of cwd as a backstop. On macOS, FSEvents can
  deliver a write event after the process that performed it has already
  exited — or never deliver it at all on GitHub's macOS runners. The old
  behavior stopped the watcher the instant the child was gone and dropped that
  event, so a quick-exiting agent could finalize with "0 file changes" and
  hh inspect --diff reported "no file changes in this session" despite files
  being written. The grace drain catches late events; the re-scan backstop
  catches the absent-event case deterministically (it does not depend on
  notify delivering anything). At startup the watcher now hashes every
  capturable baseline file (storing only the hash, not a blob), so the backstop
  can also detect a missed modify of a pre-existing file (current hash ≠
  baseline hash) and record it as Modified — a backstopped modify renders
  with a missing before-side ("all added"), since the original content was
  overwritten before it was observed. Files already captured via the event
  path and unchanged baseline files are skipped, so neither mechanism
  duplicates a normal-path capture. The startup baseline is captured
  synchronously in spawn_watcher before it returns (on the caller's thread,
  not the worker), so any file written after the watch starts — by the recorded
  child or otherwise — is guaranteed post-baseline and is never misclassified
  as a pre-existing file the backstop would skip as "unchanged"; this also
  closes the scheduling-dependent window that made the macOS-CI capture test
  flaky. A prompt backend (Linux inotify) pays only the short quiet wait; the
  startup + shutdown baseline hashing is the trade-off for missed-modify
  coverage on large trees.
• hh list column headers no longer drift out of alignment with the data
  when color is enabled: padding now measures visible width (ANSI escapes
  stripped) instead of byte length, so colored status cells no longer push
  every column to their right out of line. hh list also shows a ⚠ marker on
  rows whose adapter ended degraded, so a silently-broken session is no
  longer indistinguishable from a clean one.
• BlobStore::get / remove_if_unreferenced now reject a malformed hash
  (wrong length, non-hex, or non-ASCII bytes) instead of risking a byte-slice
  panic on a multi-byte UTF-8 character at offset 2, or a path-traversal write
  via a crafted hash string.
• The FS watcher, adapter tailer, and MCP proxy threads share one writer
  Mutex; a panic in any one of them no longer poisons it for the rest of the
  session (every lock site now recovers from poisoning) or aborts hh run's
  finalize — a panicked adapter tailer now degrades the session
  (adapter_status=degraded, warning) exactly like the existing
  "no transcript found" path, instead of leaving it stuck at
  status=recording forever.
• writer_run's PRAGMA foreign_keys = ON failure is now logged instead of
  silently swallowed (a silent failure here would disable FK enforcement for
  the rest of the session with no visible cause).
• Bumped crossbeam-epoch 0.9.19 → 0.9.20 (RUSTSEC-2026-0204), pulled in
  transitively via ignore.

Performance

• Ingest throughput: PRAGMA synchronous = NORMAL (NFR-1 / NFR-3). The store
  never set synchronous, so SQLite defaulted to FULL, fsyncing the WAL on
  every event commit. That met NFR-3's durability ("fsync on session finalize")
  and then some, but capped sustained ingest near ~300 events/s — far below
  NFR-1's ≥5,000/s. Store::open and the writer thread now set
  synchronous = NORMAL: in WAL mode this keeps ACID (no corruption on crash)
  but skips the per-commit fsync, fsyncing only at checkpoint. That matches the
  durability design NFR-3 actually names — finalize_session runs
  wal_checkpoint(TRUNCATE), which fsyncs, so a session is durable at finalize;
  SQLite's default wal_autocheckpoint (≈4 MiB) bounds the mid-session
  power-loss window to the last autocheckpoint. Sustained ingest is now ~1.5k
  events/s (≈5× over the per-event-fsync baseline); reaching NFR-1's 5k/s still
  needs statement caching + a batched writer (tracked as follow-up; the
  regression gate below protects the gain). No CLI/JSON/schema/env change.
• hh inspect --json now streams a session as NDJSON through
  Store::for_each_event_detail (one EventDetail at a time) instead of
  collecting the whole session into RAM — constant memory for any session size
  (Area 2). A 100k-event session streams without materializing 100k rows.
• The replay body cache (ReplayData) is verified bounded at its 50-entry
  capacity across unbounded scrolling (Area 2): a regression to an unbounded
  cache would fail get_evicts_lru_to_stay_bounded_at_capacity.

Added

• hh doctor (read-only diagnostic): runs five health checks — data dir
  writability, PRAGMA integrity_check, config resolution + non-canonical
  config detection, Claude Code transcript discoverability for the cwd (with a
  parse-test of the newest transcript), and a notify watcher smoke test —
  printing one ✓/✗ line per check and exiting nonzero if any fail. --json
  emits a stable schema:1 object with a per-check array. Docs: see
  docs/doctor.md; --help carries an example.
• Regression tests for the silent-breakage fix: an end-to-end file-change
  recording test, an adapter test against both Claude JSONL fixture
  generations (including the new-format transcript with mode /
  permissionMode / fileHistorySnapshot and no cwd on the early records),
  an integration test asserting the watcher-init-failure warning reaches
  stderr and the session still records ok (not "recording failed"), and an
  adapter degraded-warning-at-finalize test. hh list / hh doctor rendering
  is locked with insta snapshots.
• cargo fuzz targets (fuzz/, nightly toolchain) for the four
  untrusted/external-input parsers: Claude JSONL transcript lines, MCP
  JSON-RPC framing, config.toml, and blob decompression. Seeded from
  tests/fixtures/; runs nightly in CI (60s/target,
  .github/workflows/fuzz-nightly.yml) and via just fuzz-all <seconds>
  locally.
• proptest property tests: blob refcount/GC invariants (refcount tracks live
  references exactly; GC removes a blob iff unreferenced) across arbitrary
  create/append/delete sequences, step-assignment invariants (dense ordinals,
  exact call/result pairing, idempotent) across arbitrary event interleavings,
  and migration-reopen idempotency (byte-identical schema across N reopens).
• Panic-injection tests proving a real panic on the FS watcher thread or the
  Claude adapter tailer thread degrades the session instead of aborting the
  recording.
• CI: cargo-llvm-cov gate (hh-core ≥ 80% lines, fail-under; workspace-wide
  report-only) and cargo-semver-checks against main for hh-core's public
  API (adapters implement its Adapter trait).
• Criterion benches (hh-core/benches/) + a nightly regression gate (Area 1 /
  NFR-1): four groups — ingest (append_event throughput), replay_index
  (10k/100k-event list_event_index), blob_write/blob_read (1 KiB/1 MiB),
  and adapter_jsonl (Claude JSONL conversion, gated on the fuzzing feature
  so it uses the dev-only parser entry point, off the semver-checked surface).
  .github/workflows/bench-nightly.yml runs cargo bench -p hh-core --features fuzzing nightly and fails on >15% regression vs the previous run's baseline
  (same runner class via an actions/cache rolling baseline — a committed
  baseline would false-positive across hardware). just bench /
  just bench-compare run the same locally. Dev-only: criterion is a
  dev-dependency (default-features off) and pulls no HTTP client, so NFR-2's
  no-network-crate check stays green. See docs/performance.md.
• hh gc (reclaim space, Area 3): prunes orphaned blob files and stale
  blobs rows — the crash-leftover cases hh delete does not reach (a blob
  written to disk between BlobStore::put and the referencing event's commit,
  or a blobs row whose backing file was deleted out of band) — and VACUUMs
  hh.db to shrink it on disk (--no-vacuum skips the rebuild). Safe: it only
  removes blobs no live event references, never referenced data. --json emits
  a stable schema:1 object (orphan_files_removed / orphan_bytes_reclaimed
  / orphan_rows_removed / vacuumed). Docs: see docs/gc.md; --help
  carries an example.
• hh stats (read-only store inventory, Area 3): session/event/blob counts,
  on-disk footprint (hh.db + WAL/SHM sidecars + compressed blob directory),
  and the largest sessions by event count (--top N, default 5). --json
  emits a stable schema:1 object with blobs / disk / largest_sessions
  sub-objects. Docs: see docs/stats.md; --help carries an example.
• Storage schema migration 0002 (additive): a partial index
  idx_events_needs_heal on events(session_id) WHERE step IS NULL AND kind != 'terminal_output', making Store::open's step self-heal probe
  O(rows-needing-heal) instead of an O(all-events) scan on every invocation —
  including read-only hh list, whose cold start used to scale with total
  recorded history. Additive (a new index on an existing table); no breaking
  schema change, and the schema_version row is the only data write.

Install halfhand 1.0.0

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/halfhandorg/halfhand/releases/download/v1.0.0/halfhand-installer.sh | sh

Install prebuilt binaries via powershell script

powershell -ExecutionPolicy Bypass -c "irm https://github.com/halfhandorg/halfhand/releases/download/v1.0.0/halfhand-installer.ps1 | iex"

Install prebuilt binaries via Homebrew

brew install halfhand-org/tap/halfhand

Download halfhand 1.0.0

File	Platform	Checksum
halfhand-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
halfhand-x86_64-apple-darwin.tar.xz	Intel macOS	checksum
halfhand-aarch64-pc-windows-msvc.zip	ARM64 Windows	checksum
halfhand-x86_64-pc-windows-msvc.zip	x64 Windows	checksum
halfhand-aarch64-unknown-linux-gnu.tar.xz	ARM64 Linux	checksum
halfhand-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum
halfhand-aarch64-unknown-linux-musl.tar.xz	ARM64 MUSL Linux	checksum
halfhand-x86_64-unknown-linux-musl.tar.xz	x64 MUSL Linux	checksum

Verifying GitHub Artifact Attestations

The artifacts in this release have attestations generated with GitHub Artifact Attestations. These can be verified by using the GitHub CLI:

gh attestation verify <file-path of downloaded artifact> --repo halfhandorg/halfhand

You can also download the attestation from GitHub and verify against that directly:

gh attestation verify <file-path of downloaded artifact> --bundle <file-path of downloaded attestation>