← All transcripts

GitHub Trending Today #45: claudish-to-english, openanalytics, deepseek-harness, human-review, ha.mr Transcript, AI Summary & Key Points

Github Awesome · 13 hours ago · Science & Technology · 14:55 · EN-US

🧠 AI Summary

Thirty-five trending open-source GitHub projects are presented, with a strong focus on AI agents, local-first workflows, developer tooling, privacy, and safety. The projects include tools for simplifying Claude Code output, improving agent skills, reviewing agent work, managing memory and token costs, enforcing safer execution, researching with citations, and coordinating multiple agents. The roundup also covers analytics, 3D reconstruction, terminal interfaces, media generation, infrastructure, and WebGL.

🔑 Key Points

  • Claudish to English rewrites Claude Code responses into plainer language using a display hook and a local model by default.
  • OpenAnalytics is a self-hostable analytics stack that measures traffic and revenue without cookies, fingerprinting, or cross-site profiles.
  • DeepSeek Harness treats every part of an agent harness as a plug-in and is built on the Cordis composability framework.
  • Human Review lets people edit HTML or Markdown, comment on text and images, delete sections, and send all feedback to an agent as one batch.
  • Legal Skills provides three skills for US utility patent work, including a 316-item pre-filing audit and simulated USPTO examination.
  • Agent Safe Pipeline uses immutable hashed intents, an independent decision policy layer, approvals tied to exact intents, and single-use grants for downstream credentials.
  • Mole verifies extracted quotes against original sources, checks contradictions, returns citations, and reserves every model call against a chosen budget.
  • Several projects keep data local, including TokenTab, Lore, MCP Memory, Pi Transcribe, and Mole, while adding local indexes, models, or searchable records.

✅ Actionable items

  • Use Claudish to English to simplify Claude Code responses without changing the agent or saved transcript.
  • Use Human Review to edit documents, comment on phrases or images, delete sections, and send the complete review as one batch.
  • Convert technical books or documentation folders into agent skills with Book to Skill; use OCR for scanned PDFs.
  • Use ASD-STE100 Skill to detect ambiguity, passive voice, complex tenses, long noun clusters, hedging, and sentences containing multiple actions.
  • Use TokenTab to inspect local Claude Code session logs and calculate costs by model, project, day, and activity.
  • Use Legal Skills' Python-based checks for dates, claim counts, and fees rather than relying on model counting.
  • Use OSS PR Reviewer to review pull requests from metadata and patches without cloning or executing the code.
  • Use MCP Memory to store agent memories as readable Markdown files with provenance, tags, verification status, and staleness metadata.
  • Use Pi Transcribe for local voice dictation in the Pi coding agent; audio remains on the machine after the speech model is downloaded.
  • Use PGBot to create local database baselines and compare changes such as slower queries or newly unused indexes.
  • Use Mole to set a model-call budget, verify quotations against source pages, check contradictions, and keep private datasets local.
  • Use Termdom to build terminal interfaces with HTML, CSS, DOM APIs, flexbox, forms, scrolling, and input events.

🤖 AI in practice

Used for

Rewrite Claude Code responses into plainer language while preserving the agent and saved transcript. 00:11
Give coding agents concrete interface-motion and visual-design guidance. 00:35
Review and edit an HTML or Markdown document before sending the complete set of human feedback to an agent. 01:54
Convert a technical book or documentation folder into a compact agent skill that loads only the material relevant to a question. 02:22
Rewrite aerospace instructions so one AI agent can pass them to another without clarification. 02:46
Track Claude Code token usage and turn recorded counts into cost reports by model, project, day, and activity. 04:00
Operate a personal bot that uses a browser, memory, scheduled routines, and a Linux desktop to carry out tasks. 04:25
Prepare and adversarially test US utility patent filings with agent skills. 04:50
Run an autonomous software-factory workflow from a GitHub issue to a draft pull request. 06:06
Generate full-length songs from lyrics and a music description. 08:15
Control whether an AI agent's proposed actions are allowed, escalated, or blocked before execution. 08:39
Generate a structured Markdown code review for a GitHub pull request without executing or cloning its code. 09:03
Give coding agents persistent, searchable project memory across sessions. 09:28
Dictate code or instructions directly into the Pi coding agent. 09:51
Constrain a DeepSeek harness's initial tool access, then expand it after the agent establishes a durable trajectory. 10:17
Diagnose PostgreSQL health and provide database evidence to scripts or AI agents. 10:40
Run multiple terminal coding agents side by side while recording their messages, decisions, approvals, and terminal actions. 11:07
Research a question with verified evidence, contradiction checks, citations, and controlled spending. 13:12
Give coding agents a persistent repository map for targeted search and dependency tracing. 13:40
Reduce the context cost of large MCP tool catalogs while preserving on-demand access for an agent. 14:30

Agents

  • Rakazo — Operate a persistent personal bot with a browser, memory, scheduled routines, and a Linux desktop. 2 held 04:25
  • Eve software factory — Turn a labeled GitHub issue into a reviewed draft pull request. 2 held 06:06
  • Agent Safe Pipeline — Execute AI-agent actions only after independent policy checks and, for risky requests, approval. 2 held 08:39
  • Mole — Perform evidence-based deep research within a chosen spending budget. 2 held 13:12

Advice

  • Use explicit, simplified technical-English instructions when one AI agent must pass work to another, and preserve all conditions and scope during rewriting. for Developers building multi-agent workflows
    Ambiguous words, passive voice, complex tenses, long noun clusters, hedging, and multiple actions in one sentence can require clarification.
  • Keep dates, claim counts, and fees in deterministic code rather than relying on a model to count them. for Developers building legal or document-review agents
    The transcript states that models cannot count reliably.
  • Use an independent reviewer from a different model vendor and expose it only to the pushed branch when reviewing agent-generated code. for Teams deploying autonomous software-development workflows
    This prevents the reviewer from relying on the implementer's reasoning and makes the review independent.
  • Reserve the cost of every research-model call against a budget before allowing it to run. for Developers building research agents
    This makes research spending a hard boundary rather than a suggestion.

What it could not do

  • Models cannot reliably count dates, claim counts, and fees. — The legal-skills workflow uses Python for these values and reports unchecked items as not reached rather than treating them as passed.
  • Scanned PDFs require OCR before book-to-skill can process their contents. — The book-to-skill workflow supports PDFs, EPUBs, office documents, and plain text, but OCR is required for scanned PDFs.
  • TokenTab's cost accuracy depends on a manually maintained price table, and its activity labels use deterministic heuristics. — The tool derives costs and activity categories from local Claude Code logs.
  • Local research data in Mole stays on the machine, with only aggregates allowed out. — This restricts what can be sent outside the local environment.

🧰 Tools & AI usage

AI is used for

  • Simplifying generated text — Claudish to English sends Claude Code responses to a local model and rewrites the displayed output into plainer language.00:11
  • Generating agent skills — Book to Skill converts technical books and documentation into compact skills, chapter notes, glossaries, patterns, and reference rules.02:22
  • Agent execution — DeepSeek Harness provides a plug-in-based agent harness with a local web UI.01:28
  • Human-agent collaboration — Human Review packages edits and comments into a batch for an agent to process.01:54
  • Agent safety — Agent Safe Pipeline separates agent intent submission from policy decisions and credential use.08:39
  • Code review — OSS PR Reviewer asks OpenAI for findings from pull request metadata and patches, then validates and filters the results.09:03
  • Persistent coding context — MCP Memory stores structured project memories in readable files and provides local search and checkpoint information.09:28
  • Research — Mole decomposes questions, searches sources, verifies quotations, checks contradictions, and returns citations within a chosen model-call budget.13:12

🔗 Links mentioned

📄 Transcript

Searchable transcript of GitHub Trending Today #45: claudish-to-english, openanalytics, deepseek-harness, human-review, ha.mr — Github Awesome (14:55). Search for a phrase, then click its timestamp to jump straight to that moment in the video.

Captions sourced from the original video on YouTube, published by Github Awesome. The video, its captions and all related intellectual property remain the property of their respective owners; AINotes claims no ownership. Provided for research, accessibility and search — see the Transcript Notice and Copyright Policy.

00:00 Welcome back to GitHub awesome. This is GitHub trending today number 45. 35 trending open- source projects on GitHub right now. Let's get into it. Clottish to English rewrites Claude code responses into planer language without changing the agents or saved transcript. A display hook buffers the stream, sends the response to a local model by default, then appends or replaces what appears on screen.

00:29 An optional hook can create simplified markdown copies or overwrite selected files. Emil Kowalsski's skills gives coding agents concrete design judgment for interface motion instead of vague instructions like make it feel better. The collection can build animations with deliberate easing and duration, audit existing motion, find places where animation helps, and warn against unnecessary movement.

00:52 It also covers UI library selection, rapid prototypes, and Apple inspired web design guidance. Open analytics is a self-hostable web analytics stack that measures traffic and revenue without cookies, fingerprinting, or cross-sight profiles. Visitor identity uses a salted hash that rotates nightly. Raw IP addresses aren't stored, and privacy signals are honored before events are written.

01:17 Beyond page views, it handles funnels, web vitals, Stripe revenue attribution, exports, live presence, and an MCP server. Deepseek harness is Deepseek's own agent harness. And its organizing idea is that everything is a plug-in, including the parts most harnesses hardcode. It's built on Cordis, an existing composability framework with a paper behind it rather than a bespoke architecture.

01:42 Node and npx get you a local web UI on port 380. The readme is refreshingly bare. No feature list, no benchmark table, no claims. Human review opens an HTML or markdown file in your browser as an editable document. Fix a sentence yourself. Select a phrase and comment on it. Click an image and comment on that. Delete a section outright. Then hit send and your agent receives every edit and comment as one batch rather than a paragraph of instructions it has to parse.

02:15 Commandclick links to review several pages before sending. Runs locally. No account, no API key. Booktoskill converts a technical book or documentation folder into an agent skill that loads only the chapter needed for a question. A local extractor cleans the source. Then the generator builds a compact skill file, chapter notes, glossery, patterns, and quick reference rules.

02:37 It supports PDFs, EPUBs, office documents, and plain text with OCR required for scanned PDFs. ASD ST100 skill adapts aerospace simplified technical English for instructions that one AI agent must pass to another without clarification. It flags ambiguous words, passive voice, complex tenses, long noun clusters, hedging, and sentences that hide multiple actions, then rewrites them without dropping conditions or scope.

03:07 Strict mode fits tools and errors, while a looser mode handles technical pros. Anti-slop is an opinionated oxalant plugin that rejects TypeScript and JavaScript patterns which discard type evidence. It catches chained assertions, widened values later cast back, broad unknown contracts, unsafe dictionary types, reflective access, and undocumented assertions.

03:28 It also bans module mocking in favor of explicit dependency seams. PDFC CN gives React developers a Shad CN style component registry for generating PDFs without starting every document from raw layout primitives. You copy components into your codebase, customize themes and props, then compose tables, forms, charts, invoices, and reports with either tkumi or form.

03:52 The documentation renders live previews so changes are visible before export. Token tab reads local session logs from claude codeex then turns their recorded token counts into costs by model project day and activity. You can inspect tables, emit JSON or open a local hostonly dashboard that rereads logs on every request. Nothing needs an account or leaves your machine.

04:18 Prices come from a manually maintained table and activity labels use deterministic heruristics. Rockazo is an open- source take on a grost style personal bot. Each bot gets one thread, its own Linux computer with a browser, memory, and scheduled routines reachable from web, desktop or mobile. The pitch is that you bring the model through open router and choose where the computer runs.

04:42 Take control of the desktop to log into a site and the bot keeps the session after you let go. Legal skills is three agent skills for US utility patent work. a 316 item pre-filing audit, a simulated USPTO examination that runs until two independent examiners clear the same draft, and an adversarial design around test. The discipline is what stands out.

05:06 Models can't count, so dates, claim counts, and fees run in Python, and an item nobody checked reports as not reached rather than a quiet pass. HAMR shrinks links without storing them in a backend database. Instead, it compresses the URL itself, stripping common patterns, matching domains against Huffmancoded dictionaries, then packing the remainder into a compact character set.

05:30 For QR codes, it switches to QR's alpha numeric mode to reduce encoding overhead. Everything runs in the browser, so there are no accounts or link records. Lore turns the session files, codecs, and clawed code already leave on your machine into a searchable desktop archive. Readonly adapters connect each conversation to its repository and git history, then store the index in local SQLite without modifying agent files.

05:56 It even distinguishes commits recorded during a session from repository state observed later. Foreman is a template for an autonomous software factory. Label a GitHub issue and it runs the work through four agent stations. classify, plan, implement, review, and hands you a draft pull request to merge. The stations are isolated, and the point worth studying is the reviewer, which runs on a different model vendor and sees only the pushed branch, not the implementers's reasoning.

06:29 So, the review is actually independent. Cinematic zoom flies a camera from orbit to a chosen landmark, then exports the move as an MP4. Five shot presets animate distance, bearing, pitch, lens, and roll independently, while the app solves sunlight, and exposure for the location. Recording is frame locked, so stream geometry can settle before each frame.

06:51 You'll need a cesium or Google tile key, WebGL2, and webc hardware encoding. TTFX brings terminal text effects to Rust as one dependencyfree static binary. So terminal animations can live in shell pipelines without starting a Python interpreter. It preserves the original effects, options, cursor behavior, and even implementation quirks with CI comparing generated frame streams against a pinned upstream release.

07:19 Existing commands mostly need only the binary name changed. Bluey puts your iPhone's messages on a Linux desktop without a Mac sitting in a closet acting as a relay. That's the part that kills every other option. No Apple login, no cloud service, no subscription, just a paired phone over Bluetooth. It rides MAP, PB, and Aensis, the same profiles your car stereo uses.

07:45 So there's no hidden Apple protocol underneath. Apple Liquid Glass WebGL recreates translucent interface surfaces as a framework free WebGL renderer instead of a stack of CSS filters. Signed, distance shapes drive refraction, variable blur, chromatic dispersion, Fresnell highlights, shadows, and accurate hit testing with separate frosted and clear optical models.

08:07 It handles live canvas or video backdrops and respects reduced transparency preferences. Minia Music 3 generates 5inut songs from lyrics and a music description, not short clips you stitch together. An 8B global model tracks structure across the whole song, while a 0.6B local model fills in frame level acoustics. Synthesis fuses their continuous hidden states rather than decoding from discrete tokens, which is where vocal detail survives.

08:39 Agent safe pipeline keeps an AI agent from approving its own actions. The agent submits an immutable hashed intent. An independent desion policy layer allows, escalates, or blocks it. And risky requests can require approval tied to that exact intent. A sealed executive validates parameters and consumes a single-use grant before using downstream credentials.

09:03 OSSPR reviewer turns a GitHub pull request into a structured markdown review without cloning or executing its code. It fetches metadata and patches, batches large diffs deterministically, asks OpenAI for findings, then validates, dduplicates, and filters them by severity. Repository rules come from the trusted base branch, so a pull request can't rewrite its own review policy.

09:28 MCP memory gives coding agents persistent project memory stored as readable files rather than an opaque service. Each record becomes structured markdown with provenence, tags, verification status, and staleness metadata, while a local SQLite index handles search and key lookup. Namespaces, separate contexts, and checkpoint tools surface where the last session stopped.

09:51 PI Transcribe adds local voice dictation directly to the PI coding agent. Press its terminal shortcut, speak, and it captures microphone audio, runs a downloaded speech model through transcribe.cpp, CPP then inserts the text at your cursor. The settings command handles models, language, microphone selection, shortcuts, and Chinese script variants. Audio stays on your machine after the model download.

10:17 Anchored standard gives DeepSeek harness a two-phase tool strategy instead of choosing between a constrained agent and a useful one. The first request gets the minimal prompt with only shell and file reading, anchoring the model's initial trajectory. After the first durable tool call or reply, the full standard catalog appears and survives resumes. Pgbot diagnoses Postgra SQL from its own statistics without installing an agent or requesting right access.

10:45 Its static binary produces deterministic health findings, stores local baselines, and highlights what changed between runs from slower queries to newly unused indexes. JSON and MCP modes feed the same evidence to scripts or AI agents, while optional explanations send scrubbed context to Gemini. Multi-agent workbench runs clawed code, codec CLI, and other terminal agents side by side in one browser room, each bound to a YAML roll card.

11:14 Every message, decision, approval, and terminal bite lands in an appendon event log, and the interface is derived from it rather than stored. So a crashed session rebuilds from the log. The orchestrator blocks execution until you resolve a decision. Zasfort open sources the infrastructure engine behind ZStack Zosphere platform, giving you a management plane for virtual machines rather than another desktop hypervisor.

11:41 It coordinates hosts, clusters, images, virtual networks, storage, snapshots, permissions, alarms, and migration through asynchronous services and a plug-in architecture. A workflow engine tracks multi-step operations and rolls them back when failures occur. Colite turns a folder of overlapping photos into a full 3D reconstruction pipeline in C++ 17.

12:04 It can produce coal map compatible sparse models, dense colored point clouds, posson or delane meshes and textured OBJ files with optional CUDA acceleration for the expensive matching and stereo stages. Stage manifests let interrupted jobs resume. Ember builds color pallets for terminals, charts, heat maps, and interfaces that still separate cleanly after an aggressive warm screen filter collapses much of the color gamut.

12:31 It ships profiles for moderate and deep red shift exported for popular terminals plus CSS, JSON, Python, and Mattplot Lib. The deeper profiles deliberately use fewer distinct categories instead of faking variety. Termdom lets web developers build terminal interfaces with the same HTML, CSS, and DOM APIs they already know. It renders real DOM nodes into character cells, supports the box model, flexbox, tables, forms, scrolling, and input events.

13:02 then diffs each frame into an updates because mutations trigger redraws. Vanilla JavaScript and front-end frameworks can drive the UI naturally. Mole is a local deep research agent that treats evidence, spending, and private data as hard boundaries rather than polite suggestions. It decomposes a question, searches sources, verifies each extracted quote against the original page, checks contradictions, and returns citations.

13:28 Every model call is reserved against your chosen budget before it runs. Local data sets stay on your machine with only aggregates allowed out. Graph gives coding agents a persistent map of your repository so each new task doesn't begin with the same GP and open file archaeology. It builds a deterministic treeitter graph of symbols, imports, and call relationships.

13:53 then exposes targeted search and dependency tracing through its CLI or MCP server. An optional model pass adds plain English subsystem summaries. Lumabri distributes mixture of experts inference across ordinary machines instead of requiring every participant to hold the entire model. A client keeps dense weights locally then fetches missing model blocks or sends routed expert activations to peers caching verified data locally.

14:20 Content hashes, signatures, and optional encrypted transport help reject tampered bytes and protect traffic. MCP tune keeps large MCP cataloges out of an agent's context by turning tool access into ordinary CLI calls. Server schemas stay in a shared local configuration. Tools launch only when invoked and results can return in compact tune or reduced schema formats instead of verbose JSON. It also checks outputs for prompt injection and credential leaks.