You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The follow-ups from reviewing v0.7.0's warm-up thread, plus the schema contract
the MCP tools never declared. Minor rather than patch because harken::mcp::ServerInfo gained a public field, which breaks literal
construction — the rule for picking that digit is now written down in CLAUDE.md instead of re-derived per release.
transcribe_status no longer misreports a warm-up in flight. It derived
everything from a pure stat, so during the startup download it said the model
"is NOT cached: the first transcription call will download it" — both halves
wrong in the one window an agent is most likely to ask. The download is
already running, and a call does not start a second one, it waits on it. The
tool stays stat-only (that is what lets it answer "will a call stall?" without
ever risking a download); it now reads the warm-up state alongside the stat
and reports idle, downloading, ready or failed.
A failed warm-up is finally visible to a client. It used to exist only as
a stderr line, so an agent asking for status right after got a plain "not
cached" with no hint the download had already failed once. transcribe_status
now surfaces the recorded error.
All three MCP tools declare an outputSchema describing the structuredContent they already returned, plus readOnlyHint: true and openWorldHint: false. A client now has a machine-readable contract for the
structured payload and can tell these tools mutate nothing without reading
prose. A key-set drift guard ships with them — a declared schema with nothing
checking it is worse than no schema.
Fixed: a panic in the warm-up thread wedged every tool call for the life
of the process. The warm state was settled only on the happy path, so an
unwind left it pending forever and the gate blocked on it — the client saw initialize and tools/list answer normally, then a tools/call that never
returned and never errored. A drop guard now settles the state on unwind too,
so a panic behaves like a failed download: the call is let through and the
engine retries inline.
Fixed: the server outlived EOF until its warm-up download finished. A
client that spawns and kills MCP servers (probing a config, restarting after
an edit) left one live process per attempt, each streaming up to 466 MB with
nothing to stop it. The thread is now detached at EOF, bounding shutdown. An
abandoned download can never corrupt the cache — only a Content-Length- and
SHA-256-verified file is renamed into place — but it does leave its .partial
file behind, which is inert.
158 → 166 tests.
Breaking
harken::mcp::ServerInfo has a new public field, warm: Option<Warmth>.
Construct it with ..ServerInfo::default(), or pass Some(warmth) to have transcribe_status report the warm-up. CLI and MCP surfaces are unaffected —
no flag, tool or argument changed shape.
Install harken 0.8.0
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/montezuma-p/harken/releases/download/v0.8.0/harken-installer.sh | sh
The first-run and cold-start paths, which had the least test coverage and the
worst failure modes: every bug below ended in a wrong answer delivered
confidently — an empty transcript, a silent CPU run, a poisoned cache — rather
than in an error.
New subcommand: harken warm [--model X] — downloads the model into the
cache and exits. Retires the "pre-warm with any CLI run" workaround.
MCP mode no longer stalls on the first call. The server downloads an
uncached model on a warm-up thread while it already answers initialize/server/discover/tools/list; a tool call arriving mid-download
waits on that download instead of starting a second one. A failed warm-up is
logged and the call is let through to retry inline, so a transient network
failure at startup never wedges or kills the server. Measured on a fresh
cache: initialize at 0 ms and tools/list at 17 ms while 77 MB streamed.
New MCP tool: transcribe_status — reports model, cache state (path and
size), language, device and whether the whisper context is loaded, without
transcribing. It reads the cache with a pure stat, so it can never itself
trigger a download.
Fixed: a file whose every packet failed to decode (valid headers, garbage
payload) produced an empty transcript with exit 0 — over MCP, isError: false
with empty text, so an agent reasoned over a transcript that did not exist.
Undecodable packets are still skipped, but now counted: all of them failing is
an error, a partial skip warns about gaps, and ResetRequired truncation
warns instead of passing silently. "Silence" and "nothing decoded" are no
longer the same outcome.
Fixed: a malformed glob (harken '[bad') reported 0 done and exit 0,
indistinguishable from a valid pattern matching nothing. A pattern that cannot
compile is now an input error naming the pattern, with exit 2. A valid
glob matching zero files stays exit 0 — that distinction is pinned by a test.
Fixed: a truncated download was renamed into the cache and read as a cache
hit forever, failing context init on every later run with no hint. The written
byte count is now checked against Content-Length, the .partial file
carries a PID+counter nonce so concurrent cold starts cannot interleave into
one file, and a guard removes it on any error path. The context-init error
also names the model path and says to delete it if corrupt.
Downloads are verified against HuggingFace's SHA-256 (X-Linked-Etag,
read off the redirect response — the CDN's plain ETag is a different hash).
A mismatch refuses to cache. This is a corruption and mirror guard, not
provenance: the server states its own file's hash.
Fixed:--device cuda ran at CPU speed reporting nothing. No shipped
binary compiles a GPU backend, so the flag found none to bind and ggml quietly
proceeded on CPU. It now warns on stderr and runs on CPU — a warning rather
than an error, so scripts passing --device optimistically keep working.
README and docs/ARCHITECTURE.md no longer imply GPU offload is available.
Dependencies: ureq 2 → 3 (the migration Dependabot could not do), which
drops the url/idna/icu_* chain — Cargo.lock went 172 → 150 crates
before sha2 was added for the hash check. That chain was half the reason for
the 1.88 MSRV; the floor is now let-chains alone, so rust-version is
unchanged.
src/model.rs split into cached_path (pure stat), ensure_downloaded
(filesystem + ureq, hence Send) and resolve_model (the CLI wrapper), with
progress behind a new ProgressSink trait — BarSink keeps indicatif for the
CLI, StderrSink prints one plain line per 10% in MCP mode, where a drawn bar
in a piped stderr is escape-code noise.
Transcriber gained is_loaded(), defaulted to false so existing
implementors need no change. mcp::serve() keeps its signature; main() uses
the new serve_with_info().
Tests: 137 → 158, including the first coverage of src/audio.rs and src/model.rs. All still offline.
Install harken 0.7.0
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/montezuma-p/harken/releases/download/v0.7.0/harken-installer.sh | sh
Commits now follow Conventional Commits (<type>(<scope>): <subject>,
small and atomic), enforced by a new commits CI job. The convention is
documented in .github/CONTRIBUTING.md and CLAUDE.md.
The MCP server now speaks revision 2026-07-28 alongside the initialize
handshake of 2025-06-18 and older. The current revision is stateless: it
dropped the handshake, so each request carries its protocol version and client
capabilities in _meta. Both eras are served from one dispatcher,
discriminated per request with no state kept. New: server/discover (which is
also the stdio backward-compatibility probe), -32022 for an unsupported
version, resultType and _meta.serverInfo on modern results, and ttlMs/cacheScope on tools/list. Replies to legacy clients are
byte-identical to before, which a test enforces.
Fixed: a single non-UTF-8 byte on stdin ended the whole MCP session. It
now gets -32700 like any other unparseable frame and the server stays up.
Fixed: a tools/call with no name reported Unknown tool: instead of
naming the actual fault.
src/mcp.rs became the src/mcp/ module directory (jsonrpc, era, tools, whatsapp_tool). No behavior change from the split itself.
Dependencies: zip 2 → 8 and indicatif 0.17 → 0.18. The indicatif bump
drops number_prefix, clearing RUSTSEC-2025-0119 — cargo audit now reports
one allowed advisory instead of two. ureq stays on 2 pending the Response::header / into_reader migration (#31).
Breaking (internal API):harken::mcp::handle_line is no longer public.
It was an implementation detail of serve(), never used outside the crate,
and its signature has to change for the protocol work. serve() is
unaffected and remains the entry point.
Install harken 0.6.0
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/montezuma-p/harken/releases/download/v0.6.0/harken-installer.sh | sh
The README and crate metadata now lead with the MCP server: harken's primary
surface is "give your agent ears", with the batch CLI and the Claude Code
skill as the other two faces of the same binary.
Relicensed from MIT to MIT OR Apache-2.0 — the Rust ecosystem's dual
license, adding Apache-2.0's explicit patent grant. LICENSE became LICENSE-MIT, with LICENSE-APACHE alongside it.
New subcommand: harken mcp — an MCP (Model Context Protocol) server
over stdio, exposing transcribe_file and transcribe_whatsapp_export to
any MCP client (Claude Code, Claude Desktop, Cursor, Windsurf). Protocol
version 2025-06-18, hand-rolled JSON-RPC 2.0 on serde_json — zero new
dependencies, no async runtime. Model/language/device are per-server flags
with the CLI's defaults; one whisper context serves the whole session. The
WhatsApp tool returns transcripts directly as content and writes no files.
This is the one mode where stdout is written to (protocol frames only; logs
stay on stderr).
The build is now reproducible across machines: rust-toolchain.toml pins the
compiler to 1.97.1 (excluded from the published crate, so cargo install is
unaffected), and cargo-audit/cargo-machete are version-pinned. clippy -D warnings on a floating stable meant a new Rust release could redden the tree
with no change of ours.
CI gained two jobs: msrv, which checks the declared 1.88 floor still builds,
and deps, which runs cargo audit and cargo machete — both were
local-only gates before.
The release path is now rehearsable: a pull request builds all five target
platforms without publishing (pr-run-mode = "upload"). Until now the only
way to exercise the cross-platform build was to push a v* tag, which is
exactly how v0.4.0 died on the Windows build.
Install harken 0.5.0
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/montezuma-p/harken/releases/download/v0.5.0/harken-installer.sh | sh
The crate is now edition 2024 and declares an explicit MSRV of 1.88
(rust-version). No behavior change: the code already used the 2024 form of unsafe extern and has no unsafe fn bodies, so the only practical effect is
that let-chains became available — three nested ifs in src/whatsapp.rs
collapsed into them. 1.88 is the floor twice over: let-chains are not available
before it, and the dependency graph (icu_*, via url) already required it.
Older toolchains now get a clean MSRV message instead of a failure deep in a
transitive dependency.
Fixed: a WhatsApp export with a bad date in a message header crashed
instead of exiting cleanly. The header patterns validate shape, not the
calendar, so 31/02/2026, 99/99/2026 or a non-ASCII digit panicked while
parsing untrusted zip content. Such a line is now treated as what it is — not
a header — and folded into the previous message as a continuation. It is also
excluded from the day-first/month-first inference, where a single bad line
used to be able to flip the date order of every valid message in the chat.
Fixed: an unreadable chat entry inside an otherwise valid zip panicked
instead of exiting 2 like every other input error. An attachment that
cannot be read is now a warning and the rest of the batch continues.
New --format md: a transcript meant to be read, with the source name as
the title and one [hh:mm:ss] text line per segment.
Install harken 0.4.2
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/montezuma-p/harken/releases/download/v0.4.2/harken-installer.sh | sh
Transcription now goes through FFI bindings kept in this repo (src/ffi.rs)
against a vendored whisper.cpp pinned to v1.7.6 (vendor/whisper.cpp
submodule). whisper-rs is gone. Output is byte-identical to v0.3.1 on the
same audio.
Source builds no longer need libclang: bindgen left the build graph, and
whisper.cpp no longer goes through CMake. The CMAKE_POLICY_VERSION_MINIMUM
note now applies only to opus, on hosts without a system libopus.
The build targets an explicit AVX2/FMA/F16C floor (Haswell 2013 and newer)
instead of inheriting the build machine's instruction set, so released
binaries stop depending on whatever CPU the CI runner had. HARKEN_NATIVE=1
opts a source build into the host's full ISA. Measured within 3% of the v0.3.1
engine on the same machine (five interleaved A/B pairs).
New opt-in smoke test that loads a real whisper context through the FFI: cargo test --test ffi_smoke_test -- --ignored. The 78-test suite runs
against a fake engine and cannot catch an FFI mistake.
Building from a git checkout now needs --recurse-submodules (or git submodule update --init --recursive).
The published package ships only the vendored sources the build compiles —
upstream's examples, bindings and tests are excluded (1.3 MiB crate).
(v0.4.0 was tagged but never released: MSVC spells the C++ standard option /std:c++17 and silently ignored the unix spelling, so the Windows build
compiled as C++14 and failed on std::filesystem.)
Install harken 0.4.1
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/montezuma-p/harken/releases/download/v0.4.1/harken-installer.sh | sh
