NJ-Kit 2.4.4: nj-listen

New skill: nj-listen

Gives Claude ears. A meeting recording (audio or video, multitrack, minutes to hours) becomes a timed, speaker-labelled transcript, locally, free and open source, with no cloud call.

• Built for Arabic dialects (Jordanian, Levantine) with MSA and English words mid-sentence.
• Pipeline: Silero VAD → Whisper large-v3-turbo int8 on the Intel iGPU (OpenVINO GenAI) ∥ sherpa-onnx speaker diarization → names from each attendee's self-introduction (ERes2Net voiceprints).
• Measured on a laptop with no graphics card (i5-1334U + Iris Xe): 49.8 minutes of audio in 18.8 minutes.
• Checkpoints per stage, pause/stop/resume, multitrack (one track per speaker), --selftest with no models.

Install or update: /plugin update nj-kit.

NJ-Kit 2.4.3

• Dead-code sweep. Knip 5 (JS/TS) and vulture (Python, confidence ≥ 90) over the whole repo; every finding checked by hand before a change.
• pdf-typography: removed scripts/--help and scripts/--help.pages.json, a PDF and its page map written by running the renderer with --help as the output name; they had shipped since 2.0.0.
• nj-motion-deck template: Root.tsx takes the frame rate from theme.ts's FPS instead of a second literal 30; Slide.tsx's Props type is module-private (only that file uses it).
• nj-evidence/scripts/kb.py: two … if False else None statements that never ran removed from the self-test.
• nj-writer/reference/arabic-tashkeel/evidence/corpus.py: a re.sub whose replacement always returned the match (a no-op) removed.
• tests/test_v3.py: an unused import re as _re removed.
• Left as they are: arabic-english-proofreading/scripts/pdf-structure.py render_page(..., has_table) keeps its unused parameter (callers pass it; signature kept); the Knip notes from 2.4.2 still hold.

NJ-Kit 2.4.2

• nj-method and nj-kit: the delegation sections now state the plain working mode (2026-10-30) that replaced NJ Governance v3/v4: the session does the work itself on the normal model; subagents only for real parallel work or a crew a skill defines; nj-delegate, nj-agent-ops, nj-supervise and nj-director stay in the kit and run only when the owner names them.
• Knip 5 scan of every JS/TS file (temporary config, scripts treated as entry points, vendored code ignored): no unused files, no unused dependencies. Findings left as they are, with the reason: unlisted playwright / react / remotion / @receptron/laya imports are by design (each skill's scripts run inside the user's own project, which supplies them); unused entry exports are public functions the skill docs call or template examples; general-video/scripts/frame-packets.mjs imports ../../hyperframes/scripts/lib/frame-packets-core.mjs, which the pinned upstream hyperframes no longer ships — third-party code, reported upstream-side, not patched here.

NJ-Kit 2.4.1

• nj-method: the routing table sends commercials, promos, launch and brand films to nj-multimedia; other video still starts at hyperframes (or nj-style-film for a narrated style film). The checkpoint list names the crew folders a commercial adds before the script.

NJ-Kit 2.4.0

• New kit skill nj-multimedia (kit now 18 bundled skills; 45 with the installer): an agency-crew method for commercials, launch and brand films. Eight stages, each leaving files a gate checks: intake with a deliverables matrix (aspects × languages decided up front), creative and marketing crews in parallel (measured reference breakdown, style DNA, drawn metaphors, sound plan; positioning, three scripts, a claims.md with a source per claim), a director merge plus a devil pass, one owner stop on the proposal and keyframe boards, sound and motion crews in parallel (one take per voice line, fitted deterministically), review, render, release kit. Ready-to-fill crew briefs, a craft reference (stage, shot, motion grammar per object type, pacing profile, voice and captions, Arabic and RTL, 9:16, end card), review checklists and lessons from the reference runs.
• scripts/mm.py (stdlib + ffmpeg): init, gate (refuses a stage whose inputs are missing, a brief with empty required slots, a claim without a source, an unapproved storyboard, an unpassed review or render), fit (each line's spoken length against its window before any voice call), probe (duration ±1 frame, size, fps, audio, loudness, language-pair parity), sheet, selftest.
• scripts/frame_probe.mjs (Playwright): seeks a HyperFrames composition and fails on an element more than 1 px off its declared axis, a rendered face outside the house font (Chrome's platform fonts, not the CSS), Arabic on-screen text with tashkeel, tatweel, Persian letters, Latin punctuation or Arabic-Indic digits, or a page error. Fixture and self-test included.

NJ-Kit 2.3.3

• nj-writer voice gate: a product name registered in voice-styles.json under latin_brands (an entry lists the forms of one name, e.g. QA Wrap and QA-Wrap) may stay Latin in tts_text of an Arabic narration line, so it is said exactly as in English. voice_style_gate.py (tts_fails(n, vis, tts, brands=())) exempts such a name from the spell-it-out rule (R-TASHKEEL-008); an unregistered Latin word still fails, and a missing Arabic word still fails. Four self-test asserts added.
• nj-writer/reference/arabic-tashkeel/kb/rules/tashkeel.md: R-TASHKEEL-008 names the exception.
• Known, not changed: voice_style_gate.py --selftest fails in this repo on the placeholder brand in the sample lines (the visible text says Acme, the tashkeel sample still has the original name); it passes on the author's machine. CI does not run it.

NJ-Kit 2.3.2

NJ-Kit 2.3.2

A small follow-up to 2.3.1. Still 27 plugin skills, 44 with the kit installer.

• nj-jira bug template: the evidence rules now state the Arabic file-name rule that 2.3.1 added to nj-qa and nj-jira/SKILL.md: <المشكلة>-<الخدمة>-<التاريخ>-<الوقت>, made by nj-qa/scripts/evidence_name.py. The template still told readers to name files by what they show.
• nj-qa self-check: test_nj_qa.py now covers the evidence-name rules: the name format, forbidden characters, the 150-character limit, -2 repeats, refusals that name the file, report-check and preview-check, and NJQA_LEGACY_EVIDENCE=1. It was missing from 2.3.1.
• Not changed: nj-qa/scripts/creds.py stays the single-account version. The newer multi-account version needs a schema file and default accounts from the author's review hub, which this repo does not ship.

Gates: scripts/publish-check.py (and --selftest), scripts/skill-table.py --check, site/_build/build.py --check, compileall and the kit's unit tests all pass.

NJ-Kit 2.3.1

NJ-Kit 2.3.1

Still 27 plugin skills (44 with the kit installer).

Evidence names are Arabic (nj-qa, nj-jira)

Every screenshot and video is named <المشكلة>-<الخدمة>-<التاريخ>-<الوقت>: a short specific Arabic phrase for the defect, the service's Arabic portal name, YYYY-MM-DD, then HH.MM in 24 hours (Windows has no :). At most 150 characters, -2 for a repeat in the same minute, no English except a ticket code such as PROJ-5494 after the issue.

• New nj-qa/scripts/evidence_name.py makes the name (CLI and import) and checks names (--check).
• nj_qa.py report-check and preview-check refuse any other picture or video name. A run from before the rule sets NJQA_LEGACY_EVIDENCE=1 and keeps its old names.
• nj-jira SKILL.md and references/bug-template.md carry the same rule for Bug evidence.
• Forward only: existing files are never renamed.

nj-qa

A blocked or cannot-test note can carry block_cause (data, system or tool). The review check refuses any other value, and refuses it on any other verdict.

agent-halo and nj-chart

Both arrived in 2.3.0. They are unchanged in 2.3.1 and now appear in the website's skill list, which was missing them:

• agent-halo: a coloured border and the agent's type at the top of every browser window an agent opens. Cosmetic; screenshots and page text never see it.
• nj-chart: charts, dashboards and KPI rows that read at a glance, in Arabic or English, light or dark, on Microsoft's Flint chart engine (MIT).

Website

The grouped skill list, its counts and llms.txt are now generated from the plugin's SKILL.md files (site/_build/build.py), so they cannot go stale. The build fails when a plugin skill has no group or no Arabic line, when the kit list differs from nj-kit/bundled, or when a page states a different skill count. CI runs build.py --check. A stale "23 installed by the plugin" now reads 27. The site uses system fonts only.

Release gate

python scripts/publish-check.py reports 0 hits: no credentials, no private project data, no machine paths, no font files.

NJ-Kit 2.3.0: nj-chart, nj-qa, agent-halo

What's new

• nj-chart: charts, dashboards and reports that read at a glance, in Arabic or English, light or dark. Built on Flint (Microsoft, MIT).
• nj-qa: an agent-crew method for testing a ticket's notes on a live system, with script-enforced guards (preflight gate, proof on every verdict, traffic-patch scan, budget stop).
• agent-halo: a coloured border and the agent's type on every browser window an agent opens.
• Updates to existing skills; the plugin now has 27 skills.
• scripts/publish-check.py: a release gate that fails on secrets, project data, personal paths and font files.
• The README skill table is generated from each SKILL.md; CI fails when it goes stale.

Install

See the README: add the marketplace, then install the nj-kit plugin in Claude Code.

Licences

MIT. Third-party credits in THIRD_PARTY.md (Flint by Microsoft, MIT).

NJ-Kit assets v1

Files that NJ-Kit's skills use, downloaded on demand by nj-kit/scripts/fetch_assets.py and checked by sha256:

• sources.db: the nj-researcher source database
• tashkeel-evidence-runs.zip: evidence runs for the Arabic tashkeel knowledge base

Fetch everything: python nj-kit/scripts/fetch_assets.py --all