# Rollup example commands Runnable examples for inventory, digest generation, summary routing, the local web UI, and local tooling. **The rollup** is the weekly Markdown + HTML digest Rollup writes to `--output-dir` (default `~/Documents/rollup-outputs`). Each run also copies `rollup_logo.png` and `favicon.ico` beside the HTML file, and by default runs every output writer (xteink / txt / json / epub). See [OUTPUT_WRITERS.md](OUTPUT_WRITERS.md). See [README.md](../README.md) for setup, safety guarantees, and configuration defaults. Optional sticky config: [CONFIG.md](CONFIG.md). Product shape: [CONTRACT.md](CONTRACT.md). Position: [COMPARISON.md](COMPARISON.md). Roadmap: [ROADMAP.md](ROADMAP.md). **Default deploy:** [DOCKER.md](DOCKER.md). Most recipes below use the host CLI (`python -m rollup` / `rollup`). With Docker, run the same commands via `docker compose exec rollup …` (paths inside the container are under `/data/*` — see [DOCKER.md](DOCKER.md)). Run from the project root with the virtualenv active (host Python path only): ```bash cd /path/to/rollup source .venv/bin/activate ``` **Default digest mode** needs no Ollama server and makes no network calls unless you pass `--linkedin` (or enable `[linkedin]` in TOML), `--reddit` (or enable `[reddit]` in TOML), webpage queue fetches (on by default; `--no-webpage` to skip), `--scholar-mode detailed` (or `[scholar].mode = "detailed"`), or `--ollama`. Pass `--ollama` only when you want LLM summaries from a local Ollama instance. If you pass summary flags (for example `--summary-profile`) without `--ollama`, Rollup ignores them and prints a warning. Mode shorthand: `--dry-run` parses only and writes nothing; preview-summary mode is the normal no-Ollama digest; `--ollama` enables entry LLM summaries; `--final-review` runs whole-digest QA and can use Ollama even without `--ollama`; `--final-review-mode report` writes advisory QA only, while `apply` applies validated summary-only fixes. ## Install ### Default: Docker ```bash mkdir -p ~/.config/rollup ~/Documents/rollup-outputs ./state ./logs touch ~/.config/rollup/config.toml ~/.config/rollup/env chmod 600 ~/.config/rollup/env cp docker-compose.override.yml.example docker-compose.override.yml docker compose up -d --build open http://localhost:8765 ``` Full guide: [DOCKER.md](DOCKER.md). ### Advanced: host Python / venv | Method | Command | |--------|---------| | PyPI (when published) | `pip install rollup` | | Git checkout | `pip install .` | | Editable + tests | `pip install -e ".[dev,web]"` | | Web UI only | `pip install 'rollup[web]'` | | LiteLLM providers | `pip install 'rollup[llm]'` | | EPUB writer | `pip install 'rollup[epub]'` | | uv | `uv sync --extra dev --extra web` | ## Inventory Discover mbox folders and message counts (read-only; no body parsing): ```bash python -m rollup inventory python -m rollup inventory --root tests/fixtures/Newsletters.sbd python -m rollup inventory --json-out ./output/inventory.json # After setting root in ~/.config/rollup/config.toml, bare inventory uses that path: python -m rollup inventory ``` ## Digest without Ollama (default) Preview and generate digests with no Ollama server and no network calls (unless LinkedIn, Reddit, or webpage fetch is enabled). `--no-ollama` is optional — it is the default when neither `--ollama` nor `--no-ollama` is passed. ```bash python -m rollup digest python -m rollup digest --root tests/fixtures/Newsletters.sbd python -m rollup digest --lookback-days 7 python -m rollup digest --dry-run --root tests/fixtures/Newsletters.sbd python -m rollup digest --folder tech --exclude-folder hoops python -m rollup digest --include-seen-undated python -m rollup digest --cron --root tests/fixtures/Newsletters.sbd python -m rollup digest --no-grouping --root tests/fixtures/Newsletters.sbd python -m rollup digest --grouping-report --root tests/fixtures/Newsletters.sbd ``` After a digest, browse the indexed archive (requires `pip install 'rollup[web]'`): ```bash python -m rollup web --open ``` ## LinkedIn content searches (opt-in network) Author-list (`fromMember`) content-search URLs become digest sections. Default layout is **`feed`** (one `linkedin:feed` section). Set `layout = "per_search"` to keep each named search as `linkedin:` (needed for `--folder linkedin:general` below). Fetch uses LinkedIn Voyager with **your** browser session. Link posts also fetch the linked article body by default (external blogs, Pulse). Reference: [CONFIG.md](CONFIG.md#linkedin-content-searches-optional). ### 1. Save a fromMember search URL 1. Log in at [linkedin.com](https://www.linkedin.com) in a desktop browser. 2. Open **Search**, switch the results type to **Content** (not People or Jobs). 3. Filter by the people you want (`fromMember`). The address bar must contain `/search/results/content/` and `fromMember=` with `ACo…` ids. 4. Copy that URL. Save it on the web **LinkedIn** page (`/linkedin`) with a display name, or in TOML. Store URLs only — never cookies. ```toml # ~/.config/rollup/config.toml (or ./rollup.toml) [linkedin] enabled = true article_fetch = true # default; set false to skip linked-article HTTP layout = "per_search" # named section per search; omit for default feed (`linkedin:feed`) [linkedin.searches.general] url = "https://www.linkedin.com/search/results/content/?origin=FACETED_SEARCH&datePosted=%5B%22past-week%22%5D&fromMember=%5B%22ACoAAAMN5aEBk7L5BGyjHbFsDr40zYqwuSB7tlw%22%2C%22ACoAAA5GcN4BlMrjuK1OVX4Q63rShHLMZuQ1Qyg%22%5D" display_name = "General" enabled = true ``` Keyword-only search, company pages, follows, and mentions are not supported yet. ### 2. Copy session cookies (li_at and JSESSIONID) Voyager needs two cookies from the **same** logged-in browser pane. They are session secrets: put them in the process environment only. Never TOML, Settings, git, or screenshots you share. | Cookie | Environment variable | Role | |--------|----------------------|------| | `li_at` | `ROLLUP_LINKEDIN_LI_AT` | Session | | `JSESSIONID` | `ROLLUP_LINKEDIN_JSESSIONID` | Voyager CSRF (`ajax:…`) | **Chrome / Edge / Brave** 1. Stay logged in on `https://www.linkedin.com`. 2. Open DevTools (`Cmd+Option+I` on macOS, `F12` elsewhere). 3. **Application** (Chrome/Edge) or **Application / Storage** → **Cookies** → `https://www.linkedin.com`. 4. Find `li_at`. Copy the **Value** column (a long `AQE…` string). 5. Find `JSESSIONID`. Copy the value. It usually looks like `ajax:123…` or `"ajax:123…"`. Surrounding quotes are optional; Rollup accepts either. **Firefox** 1. DevTools (`Cmd+Option+I`) → **Storage** → **Cookies** → `https://www.linkedin.com`. 2. Copy `li_at` and `JSESSIONID` values as above. **Safari** 1. Enable **Develop** menu (Settings → Advanced → Show features for web developers). 2. **Develop → Show Web Inspector → Storage → Cookies** for `linkedin.com`. 3. Copy `li_at` and `JSESSIONID`. They rotate together. After a 401, refresh **both** from the same pane. Treat a new `li_at` with a stale `JSESSIONID` as expired. ### 3. Export and run In the same shell you use for `rollup` (do not commit these lines): ```bash export ROLLUP_LINKEDIN_LI_AT='AQE…' # li_at value export ROLLUP_LINKEDIN_JSESSIONID='ajax:…' # JSESSIONID value ``` Dry-run first (no LinkedIn HTTP; warns if cookies are missing): ```bash python -m rollup digest --linkedin --dry-run ``` Then fetch. `--linkedin` is required unless `[linkedin].enabled = true` in TOML: ```bash python -m rollup digest --linkedin --folder linkedin:general python -m rollup digest --linkedin --lookback-days 7 python -m rollup digest --linkedin --linkedin-refresh # bypass listing cache python -m rollup digest --no-linkedin # mail only, even if TOML enables LinkedIn python -m rollup digest --linkedin --no-linkedin-article-fetch # posts only, no article HTTP ``` A successful fetch logs `Fetching LinkedIn fromMember feed (N authors) via Voyager`. With `layout = "per_search"` (as in the TOML above), posts land in `linkedin:` (here `linkedin:general`) as **standalone** entries. Default `feed` layout uses `linkedin:feed` instead. Link posts use the article title as the subject when Voyager exposes one, and append the fetched article body after the commentary teaser. Mute a noisy author: ```bash python -m rollup sources disable li:member:ACoAAA5GcN4BlMrjuK1OVX4Q63rShHLMZuQ1Qyg ``` For launchd/cron, `~/.config/rollup/env` is loaded at CLI startup. You can also put the same two variables in the job environment (plist `EnvironmentVariables`). See [CRON.md](CRON.md#environment-variables-network-sources). If fetch fails, see [TROUBLESHOOTING.md](TROUBLESHOOTING.md#linkedin-fetch-failed-401--429--checkpoint). ## Reddit subreddits (opt-in network) Public listings via a transport ladder (optional OAuth JSON → public JSON → www RSS → old.reddit RSS). No Reddit account is required. Sections land in `reddit:feed` or `reddit:<sub>` depending on `layout`. Listings reuse a SQLite cache within `fetch_ttl_hours` (default 24). Reference: [CONFIG.md](CONFIG.md#reddit-subreddits-optional). ### TOML ```toml [reddit] enabled = true layout = "feed" # feed | per_source sort = "hot" [reddit.subs.python] enabled = true [reddit.subs.machinelearning] enabled = true mode = "posts" limit = 5 ``` Or use **Configuration Centre → Reddit** (`/reddit`) to add subs, toggle inclusion, and set per-sub mode/sort/cap (saved in TOML). The Reddit page and Run Studio show the estimated fetch wait (70s between subs; ~1 request/minute). ### CLI Dry-run first (no Reddit HTTP): ```bash python -m rollup digest --reddit --dry-run ``` Then fetch. `--reddit` is required unless `[reddit].enabled = true` in TOML: ```bash python -m rollup digest --reddit --lookback-days 7 python -m rollup digest --reddit --folder reddit:feed python -m rollup digest --reddit --reddit-refresh # bypass listing cache python -m rollup digest --no-reddit # mail only, even if TOML enables Reddit ``` Mute a noisy subreddit: ```bash rollup sources disable reddit:sub:python ``` Optional script-app OAuth (never TOML) goes in `~/.config/rollup/env` — see [CONFIG.md](CONFIG.md#optional-oauth-environment-only). Failures: [TROUBLESHOOTING.md](TROUBLESHOOTING.md#reddit-fetch-failed-429--missing-sub). ## Google Scholar alerts Scholar emails already in Thunderbird. Default mode keeps them as item-list mail (no extra network). Detailed mode fetches each paper and summarises it. Reference: [CONFIG.md](CONFIG.md#google-scholar-alerts-optional-detailed-mode). ```toml [scholar] mode = "detailed" max_papers_per_email = 8 max_fetches_per_run = 40 ``` ```bash python -m rollup digest --scholar-mode default python -m rollup digest --ollama --scholar-mode detailed python -m rollup digest --scholar-mode detailed --dry-run # parse papers, no HTTP ``` Or set mode and caps in **Settings → Google Scholar**. Registry source detail shows a banner when the source looks like Scholar. ## Webpage articles (opt-in network) HTTPS article URLs live in SQLite (`webpage_queue`), not TOML. Add URLs from the web **Articles** page (`/articles`), the Firefox **Add to Rollup** add-on, or enqueue programmatically. The next digest fetches pending rows once into folder `webpage:queue` and caches the body. Later digests reuse that cache and include an article when it was **saved within the lookback window** (same rule as email dates). Failed fetches stay **failed** for retry. Pass `--no-webpage` to skip. Reference: [CONFIG.md](CONFIG.md#webpage-articles-optional). ### Web UI ```bash python -m rollup web --open # Articles → add HTTPS URL → Run Studio → run digest ``` ### Firefox extension Load [`extension/firefox`](../extension/firefox/README.md) as a temporary add-on (`about:debugging#/runtime/this-firefox`). Copy the capture token from **Articles**, paste it into add-on options, then use the toolbar or context menu **Add to Rollup**. The add-on POSTs to `/articles/capture`; `rollup web` must be running. Capture is ingest-only (the next digest fetches the URL). ### CLI Digest with fixture mail plus any pending queue URLs (network fetch for pending rows): ```bash python -m rollup digest \ --root tests/fixtures/Newsletters.sbd \ --mail-root tests/fixtures \ --lookback-days 3650 \ --no-ollama python -m rollup digest --no-webpage # skip webpage queue even when pending rows exist python -m rollup digest --folder webpage:queue --no-ollama # queue only ``` Mute a noisy site in the source registry: ```bash python -m rollup sources disable web:host:example.com ``` ## Doctor ```bash python -m rollup doctor --root tests/fixtures/Newsletters.sbd python -m rollup doctor --json --root tests/fixtures/Newsletters.sbd python -m rollup doctor --full --root tests/fixtures/Newsletters.sbd ``` Inspect merged TOML + profile: ```bash python -m rollup config print python -m rollup --config ./rollup.toml config print --profile daily ``` ## Cron helpers (launchd preferred on macOS) ```bash python -m rollup cron print-launchd --python "$(which python)" --workdir . python -m rollup cron print-crontab --python "$(which python)" --workdir . python -m rollup cron status ``` See [docs/CRON.md](CRON.md) for weekly non-AI digest scheduling. ## Digest with Ollama (recommended full run) Requires a running local Ollama server (`--ollama` enables network calls to loopback by default). **Recommended full run** (all folders, 7-day lookback, per-type model routing): ```bash python -m rollup digest --ollama --summary-routing-report ``` `--ollama` alone enables type routing by default. Use `--summary-routing-report` to print which profiles and models were used. ### LiteLLM / external providers Requires `pip install 'rollup[llm]'` and provider credentials in the environment (for example `OPENAI_API_KEY`). `--ollama` still enables LLM calls; choose the transport with `--llm-provider`: ```bash python -m rollup digest --ollama \ --llm-provider litellm \ --llm-model openai/gpt-4o \ --summary-routing-report ``` Optional OpenAI-compatible / proxy base (CLI only, never sticky): ```bash python -m rollup digest --ollama \ --llm-provider litellm \ --llm-model openai/gpt-4o \ --llm-api-base http://127.0.0.1:4000 ``` Do not use LiteLLM model strings that route native Ollama (`ollama/…`); use `--llm-provider ollama` instead. Cron/launchd snippets may include `--llm-provider` / `--llm-model`; secrets must be present in the scheduler environment. ### Inspect profiles and routes ```bash python -m rollup digest --list-summary-profiles python -m rollup digest --list-newsletter-types ``` ### Routing modes Single profile for the whole digest: ```bash python -m rollup digest --ollama --summary-profile standard python -m rollup digest --ollama --summary-profile deep ``` Explicit per-type routing (same as `--ollama` default): ```bash python -m rollup digest --ollama --summary-type-routing --summary-routing-report ``` Disable per-type routing and use the `standard` profile for every message: ```bash python -m rollup digest --ollama --no-summary-type-routing --summary-routing-report ``` Compare multiple profiles side by side (writes one rollup set per profile): ```bash python -m rollup digest --ollama --summary-variants rough,standard,deep --summary-routing-report ``` Variant mode writes files such as: - `output/2026-07-02-newsletter-digest.rough.md` - `output/2026-07-02-newsletter-digest.deep.html` - `output/rollup_logo.png` - `output/favicon.ico` ### Smoke tests and cache control ```bash python -m rollup digest --ollama --folder tech --lookback-days 7 --summary-routing-report python -m rollup digest --ollama --rebuild-summaries --folder tech --lookback-days 7 python -m rollup digest --ollama --rebuild-summaries --summary-routing-report ``` Re-run without `--rebuild-summaries` to confirm cache hits in the stats block. ### Effort presets (machine power) Swap the whole summary ladder and companion defaults in one flag: ```bash python -m rollup digest --list-efforts python -m rollup digest --ollama --effort light --lookback-days 7 python -m rollup digest --ollama --effort high --lookback-days 7 --summary-routing-report python -m rollup doctor --ollama --effort high --network ``` `balanced` matches today's built-in defaults. Override models inside an effort with `[efforts.high]` (and friends) in TOML or Settings. Explicit `--ollama-model` / `--final-review-model` / `--max-chars-for-llm` still win over the preset. Do not combine `--effort` with `--summary-profile-set`. Use one model for every profile on a single run (not sticky): ```bash python -m rollup digest --ollama --effort balanced --single-model qwen2.5:7b --lookback-days 7 python -m rollup doctor --ollama --single-model qwen2.5:7b --network ``` ### XTEINK e-ink output Write a second, device-optimized digest (short lines, no URLs) beside the normal files via the **xteink** output writer: ```bash python -m rollup digest --xteink --lookback-days 7 python -m rollup digest --output xteink --lookback-days 7 python -m rollup digest --xteink --ollama --effort high --lookback-days 7 ``` Outputs look like `…-newsletter-digest.xteink.md`. Skipped under `--dry-run`. See [XTEINK_USAGE.md](XTEINK_USAGE.md). ### JSON, TXT, and EPUB outputs By default these run automatically with every digest. To write only a subset (or Markdown/HTML alone): ```bash python -m rollup digest --output json --output txt --lookback-days 7 python -m rollup digest --output none --lookback-days 7 # EPUB needs: pip install 'rollup[epub]' python -m rollup digest --output epub --lookback-days 7 ``` Same run stem as the core digest with `.json` / `.txt` / `.epub` extensions. See [OUTPUT_WRITERS.md](OUTPUT_WRITERS.md). ### Custom profile sets Export built-in profiles, edit model names to match your local Ollama library, then run: ```bash python -m rollup digest --export-summary-profile-set ./output/summary_profiles.json python -m rollup digest --ollama --summary-profile-set ./output/summary_profiles.json --summary-routing-report ``` Each profile supports Ollama generation fields: | Field | Default | Notes | |-------|---------|-------| | `num_predict` | `2048` | Max generated tokens (`options.num_predict` in the Ollama request) | | `think` | `false` | Top-level Ollama `think` flag: `false`/`true` for Qwen3; GPT-OSS needs `"low"`/`"medium"`/`"high"` (booleans are ignored) | | `options` | `{}` | Additional Ollama model options (temperature is set separately via `temperature`) | Example fragment after export — adjust model names and generation settings: ```json { "schema_version": 1, "default_profile": "standard", "profiles": { "max": { "provider": "ollama", "model": "qwen3.6:27b", "prompt_style": "deep", "temperature": 0.2, "num_ctx": 65536, "timeout_seconds": 600, "num_predict": 2048, "think": false, "enabled": true, "description": "Long essays", "options": {} } }, "type_routes": { "essay": "max" } } ``` Do **not** put `think` or `num_predict` inside `options` — Ollama ignores `think` there on `/api/generate`, which causes empty summaries on thinking models. For GPT-OSS, set `"think": "low"` (not `false`) and keep enough `num_predict` headroom for reasoning + the visible summary. Inspect loaded profiles: ```bash python -m rollup digest --list-summary-profiles ``` Example model pulls for the built-in profile set: ```bash ollama pull llama3.2:3b ollama pull qwen2.5:7b ollama pull gpt-oss:20b ollama pull qwen3.6:27b ``` ## Live-run workflow Incremental checks before a full live digest: ```bash python -m rollup inventory --root tests/fixtures/Newsletters.sbd python -m rollup digest --root tests/fixtures/Newsletters.sbd # Prefer sticky paths in ~/.config/rollup/config.toml, then: python -m rollup inventory python -m rollup digest --folder tech python -m rollup digest --profile weekly python -m rollup digest python -m rollup digest --ollama --folder tech --summary-routing-report python -m rollup digest --ollama --summary-routing-report ``` ### Migrate from hardcoded paths ```toml # ~/.config/rollup/config.toml root = "~/email/gmail/Newsletters.sbd" mail_root = "~/email/gmail" [folders.tech] emoji = "💻" accent = "#4a7fd4" ``` ```bash rollup config print rollup doctor rollup digest ``` Explicit `--no-ollama` is equivalent to omitting both `--ollama` and `--no-ollama`. Optional gitignored local mail copy: ```bash cp -R "$HOME/email/gmail/Newsletters.sbd" ./fixtures/Newsletters.sbd python -m rollup digest --root ./fixtures/Newsletters.sbd ``` ## Ollama validation sequence ```bash python -m rollup digest --list-summary-profiles python -m rollup digest --list-newsletter-types python -m rollup digest --ollama --folder tech --lookback-days 7 --summary-routing-report python -m rollup digest --ollama --summary-profile standard --folder tech --lookback-days 7 python -m rollup digest --ollama --rebuild-summaries --folder tech --lookback-days 7 python -m rollup digest --ollama --summary-routing-report ``` Stop Ollama and re-run a smoke command to confirm preview fallback without crashing: ```bash python -m rollup digest --ollama --folder tech --lookback-days 7 ``` ## Final review (editorial QA) Run a whole-digest editorial QA pass after assembly. Report-only by default: writes a JSON sidecar and does **not** change digest content. Advisory only. ```bash python -m rollup digest --root ./fixtures/Newsletters.sbd --final-review python -m rollup digest --root ./fixtures/Newsletters.sbd --final-review --final-review-profile concise python -m rollup digest --root ./fixtures/Newsletters.sbd --final-review --final-review-report ./output/review.json python -m rollup digest --root ./fixtures/Newsletters.sbd --final-review --no-final-review-cache ``` Final review does not require `--ollama` on the digest summarisation path (it uses Ollama independently when enabled). Apply mode (`--final-review-mode apply`) applies validated summary-only patches linked to `safe_auto_fix: true` issues; missing fingerprint echoes skip the whole set. Under `--cron`, apply also requires `--final-review-allow-cron-apply` and conservative whole-set caps. When enabled, a short QA summary also appears in the digest’s collapsed “Digest generation details” section at the end. ```bash # Apply mode (interactive / manual) # mail_root must contain the newsletter root (omit --mail-root to infer the .sbd parent) rollup digest --root tests/fixtures/Newsletters.sbd --final-review \ --final-review-mode apply \ --output-dir /tmp/rollup-out --state-dir /tmp/rollup-state # Group-level LLM summaries (opt-in; requires Ollama) rollup digest --root tests/fixtures/Newsletters.sbd --ollama --group-summaries \ --output-dir /tmp/rollup-out --state-dir /tmp/rollup-state ``` ## Source registry Persistent per-newsletter controls (see [SOURCES.md](SOURCES.md)): ```bash # After a digest has observed senders rollup sources list --state-dir /tmp/rollup-state rollup sources show from:alerts@github.com --state-dir /tmp/rollup-state --json # Disable a noisy source rollup sources disable from:noisy@example.com --state-dir /tmp/rollup-state # Force a daily newsletter into weekly sender batches rollup sources set list:news.example.com --grouping sender_batch --priority 80 \ --state-dir /tmp/rollup-state # Always surface undated notifications from a source (still within lookback) rollup sources set from:alerts@github.com --always-surface --state-dir /tmp/rollup-state # Override a misclassified type rollup sources set from:editor@daily.example --type essay --state-dir /tmp/rollup-state # Backup / restore overrides rollup sources export --out /tmp/sources.json --state-dir /tmp/rollup-state rollup sources import --from /tmp/sources.json --state-dir /tmp/rollup-state ``` ## Local web UI **Preferred:** Docker Compose — see [Docker](#docker) below and [DOCKER.md](DOCKER.md). ### Advanced: host Python Browse indexed rollups, rate emails, and review newsletter quality. Binds to **loopback only** (`127.0.0.1` by default). Pass `--allow-non-loopback-bind` only for Docker port mapping. Requires the optional Flask extra. Bring-up (install → digest that indexes into state → start the UI): ```bash pip install 'rollup[web]' # or from a checkout: pip install -e '.[web]' python -m rollup digest --root tests/fixtures/Newsletters.sbd python -m rollup web --open ``` **Configuration Centre** (`/settings`) edits the real digest TOML (paths, profiles, writers, folder themes). **Run Studio** (`/run`) previews the effective run, dry-runs discovery, and starts a digest without composing CLI by hand. **Articles** / **LinkedIn** / **Reddit** manage those sources. See [WEB.md](WEB.md) and [CONFIG.md](CONFIG.md). Variants: ```bash python -m rollup web python -m rollup web --host 127.0.0.1 --port 8765 --open python -m rollup web --host 0.0.0.0 --port 8765 --allow-non-loopback-bind # Docker only python -m rollup web --config ~/.config/rollup/config.toml --open python -m rollup web reindex --state-dir ./state --output-dir ./output ``` See [WEB.md](WEB.md) for security model, Settings / Run Studio / Articles / LinkedIn / Reddit, quality score, and backup notes. Reader-body cache (same SQLite; Admin can prune/backfill too): ```bash python -m rollup bodies stats python -m rollup bodies check python -m rollup bodies backfill --dry-run ``` ## Docker Default deploy: web + digest in one container while sharing the same config, state, output, and mail paths as a host CLI. See [DOCKER.md](DOCKER.md). ```bash mkdir -p ~/.config/rollup ~/Documents/rollup-outputs ./state ./logs touch ~/.config/rollup/config.toml ~/.config/rollup/env chmod 600 ~/.config/rollup/env cp docker-compose.override.yml.example docker-compose.override.yml docker compose up -d --build open http://localhost:8765 ``` The image binds with `--allow-non-loopback-bind` so Flask can listen on `0.0.0.0` inside the container. Compose publishes `127.0.0.1:8765:8765`; access from the host uses `http://localhost:8765` (Host-header loopback checks unchanged). Override the port mapping only if you intentionally need LAN access. Do not run native `rollup web` and Docker web at the same time. Fixture smoke test without real mail: ```bash docker compose -f docker-compose.yml -f docker-compose.dev.yml up --build curl -fsS http://127.0.0.1:8765/rollups ``` ## Benchmark local models Compare local Ollama-compatible models on fixed prompts: ```bash python scripts/benchmark_ollama_models.py \ --models llama3.2:3b,qwen2.5:7b,gpt-oss:20b,qwen3.6:27b \ --runs 2 \ --num-ctx 16384 \ --out benchmarks/ollama_benchmark.json \ --markdown-out benchmarks/ollama_benchmark.md ``` ## Tests and fixtures ```bash python -m pytest tests/ -v python tests/generate_fixtures.py ```</sub>