The defect was reproduced through the installed speak-clipboard → speak-text
chain, substituting clipboard input and capturing spd-say arguments without
producing audio. A private 23,083-character Markdown fixture reached Dispatcher
with URLs, heading/table markers, HTML entities, image wrappers, citation labels
and backticks still present. The old cutoff also missed the truncation warning
on this fixture.
An independent six-package comparison selected Mistune 3.4.0 with its table-enabled AST and a speech renderer. Its experimental adapter passed fourteen targeted checks; production validation is separate. The investigation also exposed loss of structural newlines, literal code entity/URL content, and punctuation following prose URLs in provisional renderers. These findings were handed to a separate fixer and independently checked at both the normalizer and actual Dispatcher-command boundary.
The independent final suite passed 126 tests with no skips on Python 3.14.9; the parent repeated the full suite on Python 3.15.2, also 126 passed. The complete private fixture normalizes to 17,965 characters with 249 newlines; actual Dispatcher submission is its exact 10,000-character whole-word prefix, with 144 newlines and one truncation notification. All 41 inline-code spans and six code blocks retain their content under the speech whitespace policy.
The private article and captured speech remain outside the public repository. Public regressions use small generic fixtures. The normalizer CLI's retained raw-input limit is 1,048,576 Unicode code points; the spoken limit remains 10,000 characters after cleanup, ending at a whole-word boundary and reporting truncation.
Final native packages passed actual pacman, apt/dpkg and dnf/rpm installation,
runtime checks and removal in isolated Arch, Debian 13 and Fedora 45 roots. Each
installed speak-text was exercised with the code/entity/URL/punctuation and
newline regressions, including long hidden destinations and cancellation before
stdin EOF. All three package formats contain identical regular-file payloads,
including both private dependencies, licenses and the nested research report.
An independent rebuild of the frozen validation source was byte-identical in
all three formats. Independent read-only package-database/filesystem checks
confirmed removal and preservation of the HOME fixture.
After transactional user-local installation, the original installed clipboard capture passed with no checked unwanted markup, all checked useful terms present, 144 structural newlines, the exact independently reviewed speech prefix, and one notification. All ten installed commands and four modules matched source bytes; the installer manifest was consistent. Provider/voice settings, credentials, desktop shortcuts, Dispatcher configuration and saved speed were hash-checked for preservation. The selected provider remained Cartesia. This fix was verified by capturing submitted text; live synthesis and subjective pronunciation were not retested. The automated release additionally runs fresh Docker native-package validation before publishing.
The implementation went through separate verifier → fixer → verifier rounds. The final complete offline suite passed 94 tests with no skips on both Python 3.11.18 and 3.15.2. This includes 59 backend/installation tests, 23 release tests, 3 dependency consistency tests, and 9 native packaging tests:
.venv/bin/python -m unittest discover -s tests -vCoverage includes real local WebSocket servers, actual mpv, an isolated Speech Dispatcher instance, and fresh/existing user-local installations. CI runs the same suite on Python 3.11 and 3.15; see the workflow results for the current commit's status.
- Arbitrary Unicode input chunks could split words. Both backends now assemble complete words and preserve the full accepted text exactly once.
- Early provider completion could silently discard unread text. Unexpected terminal frames are failures, including terminal audio frames delayed by a slow player.
- mpv's cache-pause behavior could stall resumed audio despite
--cache=no. Real-player tests assert playback resumes while stdin remains open. - An absolute final deadline could cut off healthy long generation. The deadline now measures provider inactivity separately from playback backpressure.
- Local receive timing could falsely imply that a Cartesia context had expired while the player was draining audio. Independent regressions cover continuing input and EOF during a slow drain.
- Double-quoted Speech Dispatcher placeholders allowed shell expansion of text. Both templates now preserve literal quotes, dollars, backticks and command substitutions. An execution sentinel remains absent in the actual sd_generic regression; an 8,815-character Unicode fixture is passed intact.
- Cancellation, queued-stop races, missing audio, safe error messages, character limits, installation rollback and modified-file preservation are covered by the offline suite.
The installed backend files were hash-compared with the reviewed source. Tests against Cartesia Sonic 3.6 / Sergei then verified:
- Real PCM playback through an isolated PipeWire/PulseAudio sink before stdin EOF, with input deliberately split inside a UTF-8 codepoint and a word.
- Continued playback after a five-second input gap using the same mpv process.
- Registered Plasma speed and stop actions, cancellation exit status and IPC cleanup; the saved playback speed was restored.
- A running OpenCode server using the separately reviewed
opencode-talkPR: one backend process and an exact match between submitted speech text and the completed response, with no final replay. In that sample, stdin reached the command 48 ms after the first delta and 248 ms before the final text event.
First audible playback in the paused-input Cartesia sample was about 1.4 s from launch. This is one observation, not a latency guarantee or benchmark. The separate plugin suite passed 290 tests across 23 files, with typecheck and smoke checks; its upstream PR remains a separate contribution.
Capture-only native checks, repeated after loading the hardened plugin, verified
that existing and newly created child sessions produce no narration, while the
root response is delivered once. An
API-created reviewer without parentID reproduced unwanted background speech:
it is an independent root session. Reviewer creation was corrected to preserve
ancestry. Separate regression tests harden the plugin's metadata lookup so
unknown, malformed, or mismatched session information cannot claim speech.
Independent review and local fixtures covered main-push authorization, deterministic versions, conflicting tags, interrupted draft uploads, published asset verification, and multi-commit pushes. A failing package validation for any commit prevents publication of the entire push. Actionlint passed.
Two builds from the same exported fixture commit produced byte-identical source, Debian, RPM, and Arch archives, plus checksums and release notes. This local reproducibility check used build-only execution; the production release command additionally requires native installation/removal validation before publication.
The native packages passed actual pacman, apt/dpkg, and dnf/rpm installation, unprivileged runtime checks, and removal in official Arch, Debian 13, and Fedora 45 root filesystems. Checks covered the private dependency, TLS trust, both local WebSocket protocols, real mpv, shipped Speech Dispatcher templates, mixed-install stopping, unrelated-process survival, and preservation of user settings. An independent reviewer checked package-database and filesystem cleanup. These local runs used user namespaces and bwrap; the GitHub release gate runs fresh Docker containers and must independently succeed before publishing.
Live checks cover the tested Linux desktop and account. They do not establish universal pronunciation quality, all audio devices, or provider availability. ElevenLabs protocol/error/cancellation behavior was tested offline in this release cycle; live synthesis was not repeated because account quota was limited.