# thesa
`thesa` is a Rust CLI to archive GitHub and Bitbucket repositories, ML models, datasets, websites, free 3D assets, and reusable audio.
Repository refresh helpers live in `repoforge`, model target parsing lives in
`modelforge`, dataset scraping is powered by `baseforge`, website archiving is
powered by `siteforge`, license-aware 3D bundles use `assetforge`, and reusable
music and sound effects use `soundforge`.
Made by Trevor Knott for Knott Dynamics. The terminal experience is built with
Scrin and Aisling for the TUI, dashboards, loaders, and visual effects.
`thesa` is the operator-facing binary for the Knott Dynamics archive stack. It
keeps reusable parsing and repository refresh logic in focused crates, then adds
CLI/TUI orchestration, archive manifests, integrity sidecars, and terminal
workflows.
## What It Archives
- GitHub repositories by owner/organization and Bitbucket Cloud repositories by workspace.
- Existing Git repository archives via safe refresh/pull workflows.
- Hugging Face, Ollama, and CivitAI model targets.
- Hugging Face Datasets and Zenodo dataset records through BaseForge.
- Websites and documentation through Siteforge.
- Poly Haven, ambientCG, Sketchfab CC0, and Google Scanned Objects 3D assets through AssetForge.
- Wikimedia Commons music and game-oriented audio through SoundForge.
## What It Produces
- Local archive directories under `./archives` by default.
- `.thesa/manifest.json` and `.thesa/checksums.blake3` integrity sidecars.
- Scrin/Aisling TUI dashboards for interactive archive flows.
- Siteforge AI-readable website bundles with `AGENTS.md`, indexes, compact page artifacts, and raw crawl evidence.
- Dependency-complete, checksum-verified 3D bundles with AssetForge license manifests.
- SHA-1-verified original audio with SoundForge attribution and license manifests.
## Install
`cargo install --path .`
From crates.io after publishing:
`cargo install thesa`
On Windows, install with the Windows Rust toolchain from Git Bash, PowerShell, or
Command Prompt:
```bash
rustup default stable-x86_64-pc-windows-msvc
cargo install thesa --locked
```
Git Bash can run the installed Windows `thesa.exe`, but a binary installed inside
WSL is a Linux binary and will not run from Windows Git Bash. If Git Bash reports
a missing DLL, check the DLL name: `VCRUNTIME*.dll` means the Microsoft Visual C++
Redistributable or Build Tools runtime is missing; `libgcc_s_seh-1.dll` or
`libwinpthread-1.dll` usually means an MSYS2/MinGW Rust toolchain is being used,
so switch to the rustup MSVC toolchain above or ensure MSYS2's `mingw64/bin` is on
`PATH`. `thesa` uses Rustls for its direct HTTPS client and Siteforge uses bundled
SQLite, so default installs do not require external OpenSSL or SQLite DLLs.
Then run:
`thesa`
Or force the interactive TUI explicitly:
`thesa --tui`
Short form:
`thesa -T`
## Usage
`thesa [TARGET] [options]`
New command form:
- `thesa archive site <URL>`
- `thesa archive url <URL>`
- `thesa archive dataset <TARGET>`
- `thesa archive asset <TARGET>`
- `thesa archive audio <TARGET>` (`sound`, `sounds`, and `soundforge` aliases)
- `thesa archive refresh [OWNER]`
- `thesa archive list`
- `thesa archive inspect <ARCHIVE_ID>`
- `thesa archive search <QUERY>`
- `thesa archive export <ARCHIVE_ID> --format <FORMAT>`
- `thesa archive pack <ARCHIVE_ID>`
- `thesa archive verify <ARCHIVE_ID>`
- `thesa archive resume <ARCHIVE_ID>`
- `thesa archive diagnostics [ARCHIVE_ID]`
- `thesa verify <ARCHIVE_DIR>`
Run with no arguments for interactive mode.
Version 4.5.1 hardens interactive and archive boundaries: EOF at a prompt is an
error rather than a blank retry, generic Thesa manifests reject symlink roots and
entries, and failed SoundForge IDs never receive incomplete Thesa sidecars.
Sketchfab requires a nonblank token before discovery, including dry runs; blank
submission or cancellation stops that archive route without printing the token.
Use `--tui` or `-T` to force interactive mode. If `TARGET` is also passed, the
TUI flag takes precedence and opens the TUI prompt instead.
The TUI landing screen offers six archive routes: GitHub/Bitbucket repositories,
ML models, BaseForge datasets, Siteforge website archives, and AssetForge free
3D assets, and SoundForge reusable audio. Selecting BaseForge
prompts for a dataset target and archives into `./archives/datasets` by default.
Selecting Siteforge prompts for a URL and then shows a preflight options screen
for full-site vs single-page archives, max depth, max pages, archive output
profile, and JavaScript rendering mode. TUI Siteforge runs archive into
`./archives/sites` by default.
Selecting AssetForge offers Poly Haven, ambientCG, Sketchfab CC0, or Google
Scanned Objects, then prompts
for an id, search, category, top, or latest target. Results default to the top 20,
are ranked before selection, and archive into `./archives/assets`.
Selecting SoundForge first offers Music, Game Audio, or Any Audio and then prompts
for a Commons file, search, category, top, or latest target. Audio archives default
to `./archives/audio`; Commons access is anonymous and never opens a token screen.
Repository `TARGET` can be an owner/workspace, `owner/repo`, GitHub URL, or
Bitbucket URL. Bare targets use `--repo-provider github|bitbucket`; provider URLs
and SSH remotes are detected automatically.
Examples:
- `thesa torvalds/linux --repo-provider github --dry-run`
- `thesa workspace/repo --repo-provider bitbucket --dry-run`
- `thesa https://bitbucket.org/workspace/repo --dry-run`
For model mode, `TARGET` can also be presets:
- `top` (most downloaded)
- `latest` (most recently updated)
For model mode, set `--mode models` and choose one provider:
- `--provider hf` (Hugging Face, default)
- `--provider ollama` (local Ollama instance)
- `--provider civitai` (CivitAI)
In interactive/TUI model mode, provider selection appears before the model target
prompt.
For dataset mode, `TARGET` can be:
- `owner` for a Hugging Face dataset namespace.
- `owner/dataset` or `dataset:<id>` for one Hugging Face dataset.
- `search:<query>` for a Hugging Face dataset search.
- a Hugging Face dataset URL such as `https://huggingface.co/datasets/squad`.
- a Zenodo record id or `https://zenodo.org/records/<id>` when using `--dataset-provider zenodo`.
- `top` or `latest` presets.
For dataset mode, set `--mode datasets` and choose one provider:
- `--dataset-provider hf` (Hugging Face Datasets, default)
- `--dataset-provider zenodo` (Zenodo records)
The command form is also available:
- `thesa archive dataset dataset:squad --provider hf --dry-run`
- `thesa archive dataset "climate data" --provider zenodo --dry-run`
For free 3D asset mode, provider targets include:
- a provider id or canonical provider URL
- `search:<query>` or `category:<name>`
- `top` or `latest`
Bundle controls:
- `--asset-format auto|gltf|fbx|blend|usd|archive` and `--asset-resolution auto|original|1k|2k|4k|8k` in top-level mode
- `--format` and `--resolution` under `thesa archive asset`
- `--asset-limit <1-100>` or `--limit <1-100>` controls ranked discovery (default `20`)
- `--asset-max-download-bytes` or `--max-download-bytes` to bound each selected bundle
- `--skip-existing` verifies every local size and BLAKE3 digest, plus provider MD5 when supplied
Poly Haven assets are CC0, but access to `api.polyhaven.com` is governed by
separate terms and is free only for reasonable non-commercial use unless Poly
Haven grants a commercial API license. Thesa displays required Poly Haven
attribution and the API terms URL next to API-derived content.
ambientCG assets are CC0 and use its official API. Sketchfab searches are limited
to downloadable CC0 models; downloads require `--asset-token` or
`SKETCHFAB_API_TOKEN` and preserve creator plus Sketchfab attribution.
Google Scanned Objects uses the official Gazebo Fuel API without a token. The
collection is CC BY 4.0 and requires retained Google Research attribution, which
AssetForge records with the license and source metadata. Thesa prints the exact
attribution text `Google Research via Gazebo Fuel` for both archive runs and dry
runs.
For audio mode, targets include bare searches, `search:<query>`,
`category:<Commons category>`, `file:<Commons file title>`, Commons file URLs,
`top`, and `latest`. Use `--audio-kind music|game-audio|any`, `--audio-limit
<1-50>`, and `--audio-max-download-bytes` in top-level mode. The command form uses
`--kind`, `--limit`, `--max-download-bytes`, `--archive-root`, `--skip-existing`,
`--filter`, and `--dry-run`.
SoundForge uses Wikimedia Commons' official Action API and original
`upload.wikimedia.org` files; it does not scrape Commons pages. Its
commercial-capable allowlist is CC0, verified public domain, CC BY, and CC BY-SA.
Thesa retains creator attribution, the source page, normalized license, and
license URL even where attribution is not legally required. Copyright or
public-domain status does not remove trademark, privacy, publicity, personality,
cultural, or other rights, so review each source page for the intended use.
Asset and audio dry runs list selected archive candidates. With
`--skip-existing`, provider manifest and checksum verification happens during the
actual archive operation, so a candidate shown by a dry run may then be safely
skipped. AssetForge and SoundForge collection discovery and ranked result
selection remain in the fullscreen TUI. Filters are applied before the picker;
exact and single-result targets proceed directly. Command and non-TUI runs retain
the numbered selector.
Common options:
- `--depth <DEPTH>`: clone with commit depth
- `-c, --concurrency <N>`: number of repos cloned in parallel (default `4`)
- `--filter <PATTERN>`: filter repository names by substring
- `--skip-existing`: skip repos that are already present
- `--token <TOKEN>`: GitHub API token (or use `GITHUB_TOKEN` env)
- `--bitbucket-token <TOKEN>`: Bitbucket access token (or use `BITBUCKET_TOKEN` env)
- `--repo-provider <PROVIDER>`: provider for bare repository targets (`github`, `bitbucket`)
- `--hf-token <TOKEN>`: Hugging Face API token (or use `HF_TOKEN` env)
- `--civitai-token <TOKEN>`: CivitAI API token (or use `CIVITAI_TOKEN` env)
- `--hf-mirror`: use `hf-mirror.com` for Hugging Face API calls
- `--provider <PROVIDER>`: model provider for model mode (`hf`, `ollama`, `civitai`)
- `--dataset-provider <PROVIDER>`: dataset provider for dataset mode (`hf`, `zenodo`)
- `--asset-provider polyhaven|ambientcg|sketchfab|google-scanned-objects`: free 3D asset provider
- `--asset-token <TOKEN>`: user Sketchfab API token (or `SKETCHFAB_API_TOKEN`)
- `--asset-format <FORMAT>` / `--asset-resolution <RESOLUTION>`: AssetForge bundle selection
- `-T, --tui`: force the interactive Scrin interface
Git archive refresh options:
- `thesa archive refresh`: scan `./archives/**` for GitHub/Bitbucket worktrees and run `git pull --ff-only`
- `thesa archive refresh <NAMESPACE>`: refresh only archives whose slug owner/workspace matches `<NAMESPACE>`
- `thesa archive update` and `thesa archive pull`: aliases for `refresh`
- `--archive-root <DIR>`: root to scan recursively (default `./archives`)
- `--owner <NAMESPACE>` / `--workspace <NAMESPACE>`: provider namespace filter
- `-c, --concurrency <N>`: number of repositories refreshed at once (default `4`)
- `--filter <TEXT>`: refresh only matching owner/repo slugs or paths
- `--dry-run`: print discovered archives without pulling
Refresh only touches actual Git worktrees. It derives `owner/repo` from recognized
GitHub or Bitbucket `origin` URLs and falls back to the archive path layout,
such as `archives/<owner>/<repo>`. Dirty worktrees are skipped with an error so
local edits are not overwritten silently. The discovery, slugging, clean-check,
and `git pull --ff-only` execution are provided by the `repoforge` crate; thesa
adds archive manifests and terminal reporting around that core.
Namespace-scoped refresh is the fastest way to update one provider archive:
- `thesa archive refresh tknott95 --dry-run`
- `thesa archive refresh tknott95 --concurrency 8`
- `thesa archive pull --owner tknott95 --filter agent`
Archive output directories are ignored by git and excluded from Cargo packages.
Website archive options:
- `-o, --output <DIR>`: output root for website archives (default `./archives/sites`)
- `--input <FILE>`: seed URLs from a file, one per line (`#` comments allowed); may be used with or without a positional URL
- `--full-site`: follow in-scope site links instead of archiving one page
- `--allow-cross-domain`: permit cross-domain links
- `--max-depth <N>`: limit crawl depth
- `--max-pages <N>`: limit page count
- `-c, -j, --concurrency <N>`: number of concurrent fetch workers
- `--delay-ms <N>`: politeness delay between requests per worker
- `--archive-id <ID>`: stable archive directory id
- `--include <GLOB>`: include URL glob pattern, repeatable
- `--exclude <GLOB>`: exclude URL glob pattern, repeatable
- `--archive-profile <full|agent|lean>`: choose Siteforge artifact volume
- `--max-page-size-bytes <N>`: maximum bytes fetched per HTML page
- `--max-asset-size-bytes <N>`: maximum bytes fetched per asset
- `--max-total-archive-size-bytes <N>`: maximum bytes written by the archive
- `--render-js`: use Siteforge browser rendering for JavaScript pages
- `--render-js-auto`: render only thin JavaScript app shells
- `--browser-command <COMMAND>`: browser command template for JavaScript rendering
- `--browser-profile <DIR>`: browser profile directory for JavaScript rendering
- `--render-wait-ms <N>`: milliseconds to wait before extracting rendered DOM
- `--header 'Name: Value'`: extra request header, repeatable
- `--cookie <COOKIE_HEADER>`: cookie header value for authenticated/public crawls
- `--user-agent <VALUE>`: Siteforge fetch/render User-Agent
- `--timeout-secs <N>`: HTTP timeout per request
- `--retry-count <N>`: retry attempts for failed frontier items
- `--ocr`: enable Siteforge OCR for code-like downloaded assets
- `--dry-run`: print the Siteforge crawl plan without fetching pages
Avoid storing secrets in docs or shell history when using `--header` or
`--cookie`; prefer a short-lived shell variable or a local wrapper script outside
the repository.
Siteforge archives write an AI-readable bundle under the archive id directory.
The default `agent` profile avoids duplicate YAML Markdown and plain-text copies:
- `AGENTS.md`: first file to open for agent-readable instructions/context
- `agent-index.json`: machine-readable index for agents/tools
- `readable/basic_markdown`, `readable/json`: default extracted page content
- `readable/markdown`, `readable/text`: additional representations in explicit `full` mode
- `chunks/chunks.jsonl`: chunk stream for downstream indexing
- `manifest.json`, `checksums.json`, `crawl.db`: Siteforge crawl metadata
- `.thesa/manifest.json`, `.thesa/checksums.blake3`: thesa integrity sidecars
Siteforge management commands default to `./archives/sites`. Use
`--archive-root <DIR>` when archives live elsewhere.
Siteforge export formats include `markdown`, `basic-markdown`, `jsonl`, `json`,
`html`, `warc`, `zip`, `tar`, `agent-tar`, and `all`.
## Forge Crates
`thesa` keeps reusable archive logic in small crates so it can stay focused on
CLI/TUI orchestration:
- `repoforge`: discovers Git worktrees under archive roots, derives GitHub
`owner/repo` slugs, rejects dirty worktrees, and refreshes clean clones with
`git pull --ff-only`.
- `modelforge`: normalizes Hugging Face, Ollama, and CivitAI target strings into
provider-neutral target enums before thesa calls provider APIs.
- `baseforge`: discovers Hugging Face/Zenodo datasets, inspects downloadable
files, and archives them with bounded concurrent downloads.
- `siteforge`: captures websites into AI-readable local archives and provides
website archive management commands.
- `assetforge`: discovers ranked license-aware models through official Poly Haven,
ambientCG, and Sketchfab APIs plus CC BY 4.0 Google Scanned Objects through
Gazebo Fuel, and writes bounded, BLAKE3-verified archives.
- `soundforge`: discovers allowlisted reusable audio through the official
Wikimedia Commons Action API and writes bounded SHA-1/BLAKE3-verified archives.
The thesa binary combines those crates with Scrin and Aisling to provide the
interactive command deck, repo/model/dataset pickers, clone dashboards,
Siteforge option screens, AssetForge discovery, and archive progress displays.
## Package Safety
The Cargo package is explicitly whitelisted. Releases include only the manifest,
lockfile, README, license, changelog, source, and helper scripts. Archive output,
sessions, notes, logs, build directories, and local scratch files are not shipped.
Included package files:
- `Cargo.toml`
- `Cargo.lock`
- `README.md`
- `LICENSE`
- `CHANGELOG.md`
- `src/**`
- `git.sh`
- `scripts/**`
Repo picker controls (TTY only):
- `Space`: check or uncheck a repo
- Mouse: hover a visible row to focus it, left-click to check or uncheck it
- `a`: check all visible repos
- `c`: clear filter and checks
- `b`: open batch-size overlay (`1`, `2`, `3`, `6`, `8`)
- `PageUp`/`PageDown`: move cursor by one viewport page
- `Home`/`End`: jump to first/last visible result
- `Enter`: clone checked repos only, or visible repos when nothing is checked
Asset/audio ranked picker controls (fullscreen TUI):
- `Up`/`Down`: move through ranked results
- `Space`: toggle the active result
- `a`: select all ranked results
- `n`: clear all results
- `Enter`: confirm the current selection; an empty selection is a no-op
- `Esc`: cancel the route
Clone dashboard controls (TTY only):
- `t`: toggle repository detail lines in the live batch display
- `q`: request stop after current batch completes
- `m`: motion effect
- `b`: Burn effect
- `w`: Wipe effect
TTY clone runs use a Scrin dashboard. Aisling renders the top Matrix deck,
effect pane, and TQDM-style loader rows for concurrent repository pulls. The TUI
is built against newer Scrin interaction support and remains keyboard-first for
SSH reliability.
Mode, provider, repository, and model picker screens also accept mouse hover and
left-click actions when the terminal supports mouse reporting. Dataset selection
uses a compact numbered prompt after BaseForge discovery.
Examples:
- Start TUI and pick target/repo interactively:
- `thesa`
- `thesa --tui`
- `thesa -T`
- Clone one repo:
- `thesa torvalds/linux --output ./repos`
- Clone all public repos for a user/org with shallow history:
- `thesa octocat --output ./repos --depth 1`
- Clone with explicit concurrency:
- `thesa octocat --concurrency 8 --output ./repos`
- Refresh existing GitHub archives:
- `thesa archive refresh --dry-run`
- `thesa archive refresh tknott95 --dry-run`
- `thesa archive refresh --archive-root ./archives --concurrency 8`
- `thesa archive pull --filter octocat`
- `thesa archive pull --owner tknott95 --concurrency 8`
- Archive a website:
- `thesa archive site https://example.com --full-site --max-pages 100`
- `thesa archive url https://example.com/docs --max-depth 2`
- `thesa archive site https://docs.example.com --archive-profile agent --render-js-auto`
- `thesa archive --dry-run url https://example.com/docs --full-site`
- `thesa archive url https://example.com/docs --dry-run --include '**/docs/**' --delay-ms 500`
- `thesa archive url https://example.com/docs --input seeds.txt --archive-id docs`
- `thesa archive url --input seeds.txt --full-site --archive-id docs`
- Manage Siteforge archives:
- `thesa archive list`
- `thesa archive inspect docs`
- `thesa archive search "install" --archive-id docs --limit 5`
- `thesa archive export docs --format agent-tar --output ./exports/docs-agent`
- `thesa archive pack docs --max-tokens 80000`
- `thesa archive verify docs`
- `thesa archive resume docs`
- `thesa archive diagnostics docs`
- Verify a thesa archive directory:
- `thesa verify ./archives/example`
- `thesa verify ./archives/example --json`
- Preview actions before cloning:
- `thesa org-name --dry-run`
- Models and providers:
- `thesa --mode models --provider hf meta-llama --dry-run`
- `thesa --mode models --provider hf meta-llama/Llama-3.1-8B-Instruct --dry-run`
- `thesa --mode models --provider ollama llama3 --dry-run`
- `thesa --mode models --provider hf top --dry-run`
- `thesa --mode models --provider hf latest --dry-run`
- `thesa --mode models --provider ollama top --dry-run`
- `thesa --mode models --provider ollama latest --dry-run`
- `thesa --mode models --provider civitai top --dry-run`
- `thesa --mode models --provider civitai latest --dry-run`
- `thesa --mode models --provider civitai 827184 --dry-run`
- Datasets and providers:
- `thesa --mode datasets --dataset-provider hf dataset:squad --dry-run`
- `thesa --mode datasets --dataset-provider hf search:"code data" --filter python --dry-run`
- `thesa --mode datasets --dataset-provider hf openai --concurrency 8 --skip-existing`
- `thesa archive dataset dataset:squad --provider hf --archive-root ./archives/datasets`
- `thesa archive dataset "climate data" --provider zenodo --dry-run`
- Free 3D assets:
- `thesa --mode assets --asset-provider polyhaven --asset-resolution 1k Shelf_01 --dry-run`
- `thesa --mode assets --asset-format gltf --asset-resolution 2k search:"wooden chair"`
- `thesa archive asset category:furniture --format fbx --resolution 1k --dry-run`
- `thesa archive asset Shelf_01 --format gltf --resolution 1k --skip-existing`
- `thesa archive asset search:chair --provider gso --dry-run`
- Reusable audio:
- `thesa --mode audio --audio-kind music top --dry-run`
- `thesa archive audio search:laser --kind game-audio --limit 10 --dry-run`
- `thesa archive sounds category:"Sound effects" --kind any --skip-existing`
## Scripts
You can compile and run with helper scripts:
- Run from repo root (interactive target prompt):
- `./run.sh`
- Pass a target directly:
- `./run.sh <target> [output-root] [concurrency] [extra thesa flags]`
- Compile only:
- `./scripts/compile.sh`
- Compile and run (default concurrency `4`):
- `./scripts/run.sh <target> [output-root] [concurrency] [extra thesa flags excluding --output]`
- Publish release to crates.io:
- `./scripts/publish.sh --allow-dirty`
- `./scripts/publish.sh --dry-run` (safe checks only)
`run.sh` creates an output folder from the target name (for example `owner` or,
for `owner/repo`, just `repo`) and runs clones inside that folder.
## Notes
- `--concurrency` sets the concurrent clone batch size.
- private repositories are intentionally skipped/blocked.
- Cargo packages exclude archive output, session files, logs, target directories,
and temporary test archives. Do not put secrets in project files.