Lemma User

You are operating inside an existing pod — use its resources (tables, files, functions, agents, workflows, connectors) to get work done for the user. You are not redesigning the pod; that's the lemma-builder skill.

This is the operator companion to lemma-builder: the runtime view of the same model. For the model itself, read lemma-builder/references/pod-model.md — this doc grounds in it and assumes it.

The model, from the operator's seat

(Grounds in pod-model.md.) A pod is one team's workspace under one permission boundary. What that means when you run commands:

• You act as a specific user. Whether a human at a terminal or an agent on someone's behalf, every call carries your identity. A workload (function or agent) runs under delegated identity — it acts as the user who invoked it, never as a service account. So /me and row visibility always resolve to that user.
• RLS scopes what you see. On an RLS table (the per-user default) you see and edit only your own rows — another member's row is invisible (a fetch returns 404, a list omits it). On a shared table (enable_rls: false) everyone sees the same rows. This holds for everyone, admins included; reading across all users' rows needs an explicit mode=ADMIN opt-in (admin-gated, not the default flow). The read-only query API enforces RLS the same way.
• /me is your private tree. /me/... resolves to your own file subtree (owner-only). Every other path is pod-shared — top-level folders like /knowledge, /contracts. There is no /pod prefix: a path is shared unless it's under /me. Folder grants cascade to everything beneath them.
• Missing access has two shapes. A human without the pod role gets a permission error; a workload missing a grant gets MISSING_WORKLOAD_RESOURCE_GRANT (naming the resource a builder must grant).

Put user-facing deliverables in /me (or the appropriate shared folder) — never leave the only copy in a local temp path.

Orient first

lemma pods list            # marks the currently active pod
lemma pods describe        # inventory: tables, agents, functions, workflows, schedules + a file tree
                           # (apps are NOT in it — use `lemma apps list`)
                           # tree shows 2 folder levels; --depth N / --full for more

Workspace sessions inject LEMMA_TOKEN, LEMMA_BASE_URL, LEMMA_ORG_ID, LEMMA_POD_ID (and LEMMA_WORKSPACE_URL) — use them; never invent bootstrap config. (Outside an injected workspace — e.g. running the CLI on a laptop — project-root .lemma.<server>.env files supply the same LEMMA_* values per server for that folder; injected/real env always takes precedence.) Default output is a compact, complete table/detail view (schemas included) — prefer it; it costs far fewer tokens than JSON. Use --output json only to pipe/save, and --full to expand folded fields. Pass payloads with --data '<json>' (-d) or --file path.json (-f); target another pod with --pod <id-or-slug>; add --yes for destructive commands in automation. CLI groups are plural (lemma files, lemma tables, lemma records, …), and most have a singular alias (lemma file, lemma table). Not all: tools, query, datastore, runtime, servers, auth, and config exist only as written.

For multi-step scripting prefer the Python SDK over chained CLI calls:

from lemma_sdk import Pod
pod = Pod.from_env()       # auth + pod from the environment

Files — the pod is a searchable knowledge base

This is the area you'll lean on most. Uploaded documents are auto-indexed — the pod is the RAG system. PDF/DOC/DOCX/ODT/RTF/Markdown/text/HTML/EPUB are extracted, chunked, embedded, and converted to page-marked markdown on upload. Data/binary (CSV, JSON, XLSX, images, email) are stored but never indexed — they won't appear in search. So: search to find, cat to read, child + view-image to see.

Because the pod auto-produces a document's markdown, page images, and figures, read those first (the commands below) — never re-parse a pod file. Reach for the liteparse-documents skill (lit) only for a document from outside the pod (e.g. a PDF an agent fetched from the web) or as a fallback when a pod file's derived artifact is missing or insufficient (scanned/OCR, bounding boxes).

Search — find the relevant passages

lemma files search "refund policy" --scope /knowledge                 # HYBRID, folder + all subfolders
lemma files search "termination clause" --scope /contracts --method VECTOR   # semantic only
lemma files search "invoice 4471" --scope /inbox --method TEXT --direct      # keyword, immediate children only

Results are ranked passages with page numbers, so you can jump straight to cat … --pages N. --scope + the default SUBTREE (folder and everything beneath) is your retrieval lever — scope a search to one knowledge folder to keep it tight. --method is HYBRID (default), VECTOR (semantic), or TEXT (keyword); --direct limits to a folder's immediate children — it only takes effect alongside --scope, and is ignored without one. Reach for search before reading whole files or guessing.

Read — cat is page- and mode-aware

lemma files cat /knowledge/handbook.pdf                 # auto: raw text for .md/.txt, converted markdown for PDF/DOCX/…
lemma files cat /knowledge/handbook.pdf --pages 3-7     # 1-based page slice over the converted markdown (great for long books)
lemma files cat /me/notes/log.md --lines 10-50          # 1-based line slice over raw text
lemma files cat /knowledge/handbook.pdf --mode markdown # force converted markdown (errors if not a document)
lemma files cat /scratch/data.csv --mode text           # raw bytes (binary → flagged, not dumped)

--mode is auto (default) / text / markdown. Output is capped at ~50,000 chars by default (matching the in-process agent tool); widen with --max-chars 0 (unlimited), --max-lines N, --max-tokens N, or --full, or narrow with --pages / --lines. The payload reports page_count, the returned range, and a truncated flag so you know when to page — page-range slicing is how you read a long document without blowing the budget.

lemma files download /knowledge/handbook.pdf ./handbook.md --markdown   # save converted markdown
lemma files download /knowledge/handbook.pdf ./handbook.pdf             # exact original bytes

See — child page images + view-image

A processed document exposes hidden child artifacts at <file-path>/<artifact>:

lemma files children /knowledge/handbook.pdf                          # list them
lemma files child /knowledge/handbook.pdf/document.md --pages 3-7     # page-marked markdown range
lemma files child /knowledge/handbook.pdf/pages/page_0003.jpg ./p3.jpg  # fetch a rendered page image

• …/document.md — page-marked converted markdown (<!-- PAGE n -->)
• …/pages/page_0001.jpg … — rendered page images (1-based)
• …/images/image_0.png … — extracted figures

Use view-image to actually see a pod file. Those rendered page JPEGs (and any uploaded image) are exactly what the view-image capability reads — fetch one with files child (or a URL with files url) and view it to see a chart, a signature, a scanned form, a layout. This also works on workspace files directly. So: "what does page 3 look like?" → files child …/pages/page_0003.jpg → view-image; "what does it say?" → files cat … --pages 3.

Write & transfer

lemma files mkdir /knowledge
lemma files upload ./report.md /me/reports/report.md          # documents auto-index
lemma files upload ./data.csv /scratch/data.csv --no-search   # skip indexing
lemma files write /me/notes/draft.md "first line"             # create/overwrite (or pipe via stdin)
lemma files append /me/notes/draft.md "next line"             # append (read-modify-write, last writer wins)
lemma files ls /knowledge ; lemma files tree /
lemma files stat /knowledge/handbook.pdf                      # metadata incl. indexing status
lemma files mv /me/notes/draft.md /me/notes/final.md
lemma files rm /scratch/data.csv

Indexing lags briefly after upload — stat shows status (COMPLETED = searchable, NOT_REQUIRED = stored but not an indexed document, PENDING/PROCESSING/FAILED).

Link to a file — pick by who opens it

lemma files url /reports/summary.pdf                           # app_url (in-app, signed-in member) + short-lived download url
lemma files share /reports/summary.pdf --ttl 3h --max-hits 50  # public, no-login, expiring + hit-capped

url returns an app_url deep-link for pod members (must be logged in) plus a short-lived raw download url. share mints a public link anyone can open without logging in — it expires (--ttl = 30m/3h/24h; default 3h, max 24h) and stops serving after --max-hits downloads (default 50, max 100), bounding egress if it leaks. Emailing/messaging someone outside the pod → share; pointing a member at a file in the app → url. (In a function or agent, the same via the SDK: pod.files.get_url(path) / pod.files.create_signed_url(path, …).)

Tables, records, query

lemma tables list
lemma tables get tickets                              # schema: columns, types, enums
lemma records list tickets --limit 20
lemma records get tickets <record-id>
lemma records create tickets --data '{"title":"New item","status":"new"}'
lemma records update tickets <record-id> --data '{"status":"done"}'
lemma query run "select status, count(*) as total from tickets group by status"

Read the table schema before writing — ENUM columns reject values outside options. Prefer query run (a read-only SELECT subset — one SELECT, no writes) for aggregates and joins instead of paging records; it reads across any tables, including RLS tables, where it returns only your own rows (RLS scopes every caller the same way). To read across all users' rows on an RLS table you'd pass mode=ADMIN — admin-gated, not the default, and agents never use it.

Functions, workflows, schedules

lemma functions list
lemma functions run score_ticket --data '{"ticket_id":"..."}'   # check output_data / status / logs
lemma functions runs list score_ticket                          # past runs (debug)
lemma functions runs get score_ticket <run-id>                   # NB: function AND run id

lemma workflows list
lemma workflows run intake --data '{"title":"..."}'             # WAITS for the run by default (--no-wait to fire);
                                                                # --data is submitted to the entry form
lemma workflows runs list intake
lemma workflows runs get <run-id>                               # status, current node, active_wait, step_history, errors
lemma workflows runs waiting                                    # form waits assigned to you (your approval queue)
lemma workflows runs submit-form <run-id> --data '{"approved": true}'  # complete the form the run is waiting on
lemma workflows runs cancel <run-id>                            # cancel a running/waiting run

lemma schedules list
lemma schedules pause <id> ; lemma schedules resume <id>

A run in WAITING is paused on a human form, an agent conversation, an async function, or a timer — runs get shows which via active_wait (wait_type, node_id, assignee, external reference, and the form schema for human waits). If a form wait is assigned to you (runs waiting lists them), runs submit-form --data with the form's fields completes it and advances the run. This is how you participate in human-agent workflows.

A conversation in WAITING is a different thing, and the difference matters before you go chasing it: it is either blocked on you (an ask_user question or an approval card — answer it and the agent continues) or snoozed, meaning the agent suspended itself and wakes on its own within 24 hours. A snoozed conversation is healthy and needs nothing from you. Conversation status carries wait_reason (HUMAN or SNOOZE) and, when snoozed, wakes_at.

Agents and chat

lemma agents list
lemma agents chat triage-agent "Summarize today's urgent tickets"   # interactive or one-shot
lemma agents run triage-agent "Classify this: ..."                  # waits + streams the result (--no-wait to detach)
lemma conversations list --agent triage-agent                       # an agent's runs (each run is a conversation)
lemma conversations list --parent-id <conversation-id>              # what a sub-agent was asked, and answered
lemma conversations messages <conversation-id>
lemma conversations send <conversation-id> "Continue with the next batch"

An agent acts under your delegated identity — it sees exactly what you'd see (your RLS rows, your /me, your connected accounts), plus only the resource grants its builder gave it.

Connectors

Two ways in, and which one you have depends on how the agent was granted:

• Direct tools (the CONNECTORS toolset) — in-process, no sandbox. These are deferred: they are not in your prompt prefix, so reach them with search_tools first, then search_connector_operations and run_connector_operation. Prefer this when you have it — no shell involved.
• The CLI (lemma connectors …, needs the workspace toolset) — same operations through a sandbox round trip. Use it when you are driving a shell anyway, or when you need the discovery views below.

Either way the authorization is identical: a connector:<name>:use grant per app, executed through the invoking user's connected account. Having the toolset is not having access to any particular app.

As direct tools

Once search_tools has surfaced them: leave auth_config unset and search by what you want to do — the search spans every installed connector and each hit names the auth_config to run it against, so you never have to guess which install does email:

search_connector_operations {"query": "send an email"}
  -> [{auth_config: "workspace-gmail", operation: "gmail_send_email", relevance_score: …}, …]

run_connector_operation {"auth_config": "workspace-gmail",
                         "operation": "gmail_send_email",
                         "arguments": {"recipient_email": "a@b.com", "subject": "Hi", "body": "…"}}

Wrong arguments come back as invalid_arguments with the operation's input_schema attached — correct and retry, don't go fetch the schema separately. Pass auth_config on search only to narrow to one install; describe_connector_operation only when you want the full schema up front; output_path on run to land a file result in the pod.

From the CLI

Third-party connector operations — run does the whole thing in one call: it resolves the connector, picks the operation, and executes. Still never guess a payload; let --dry-run hand you the schema.

lemma connectors run gmail "list recent emails" --dry-run    # resolves + prints the input schema
lemma connectors run gmail gmail_list_messages -d '{"max_results": 5}'
lemma connectors run gmail gmail_send_email \
  -d '{"recipient_email": "a@b.com", "subject": "Hi", "body": "..."}'

The first argument is the connector id you already know from the task (gmail, slack); it resolves to that connector's install. The second is an operation id, or plain English — the resolved id is printed so you can name it exactly next time. --dry-run, or simply omitting --data on an operation that needs input, prints the input schema instead of failing. --account accepts an account id or the connected email. An operation that CHANGES data and was inferred from text rather than named is refused without --yes — matching is lexical, so a read intent can land on a write.

When you want the wider picture rather than one call:

lemma connectors overview             # installed connectors: auth-config name, kind, connected accounts
lemma connectors status               # installed apps + your connected accounts
lemma connectors describe gmail       # per-connector usage guide, per kind
                                      # (kinds: package, composio, http, sql, mcp)
lemma connectors operations search "send email"                    # searches EVERY installed connector
lemma connectors operations search gmail "send email" --limit 5    # scoped; hits include their input schema

Workloads execute operations via the invoking user's connected account (delegated) — they never touch raw credentials. If no account is connected, create a connect request and hand the link to the user: lemma connectors connect-requests create gmail --auth-config-id <id>.

Workspace execution notes

• Long-running processes (dev servers, watchers, REPLs): keep one persistent interactive session and reuse it; one-shot shell commands for everything else.
• Local services are http://127.0.0.1:<port> inside the container. There is no user-constructible public preview URL for a workspace port — port access is a signed, expiring link the platform mints. To show someone a running app, deploy it (lemma apps deploy) rather than sharing a sandbox port.
• To keep web sources, use the browser skill's save-webpage <url> --formats markdown,pdf; upload durable artifacts to /me or a shared folder.
• Network errors (Could not resolve host, ENOTFOUND, TLS timeouts): check curl -sS "$LEMMA_BASE_URL" once, retry once, then report — don't loop.

Troubleshooting

• Row not visible / empty list / 404 on an RLS table. You only ever see your own rows — an absent row usually belongs to another member, not a missing record. Confirm with the owner or, if you have the admin role and the feature warrants it, the mode=ADMIN read path. Don't assume data loss.
• Permission denied / resource not visible. As a human you may lack the pod role — they ladder up: POD_VIEWER reads; POD_USER also writes records and runs agents/functions/workflows; POD_EDITOR also creates/updates tables and writes files; POD_ADMIN also deletes and manages members. As an agent, MISSING_WORKLOAD_RESOURCE_GRANT names a missing workload grant — a builder must add it (it never silently grants itself). lemma pods doctor lists every workload in the pod holding no grants at all, which is the usual cause; the fix is lemma agents permissions add <name> <resource>:<perms>.
• Resource not found. Confirm the active pod (lemma pods list) and exact names (lemma pods describe; lemma apps list for apps).
• ENUM rejected on a record write. Read lemma tables get <table> and use one of the listed options.
• Fresh upload not in search. Indexing lag — files stat for status, retry shortly. NOT_REQUIRED means it isn't an indexed document (CSV/JSON/XLSX/images/ email are stored but never searchable).
• Workflow stuck. runs get <run-id> → active_wait shows what it's blocked on; step_history shows the failing node, its input, and error. A human wait needs runs submit-form.

Report what got in your way

Hit a CLI, skill, or platform problem worth reporting — a confusing error, a flag that didn't do what it says, information you had to discover by trial and error, or something these skills got wrong? One command, and it is the only way any of that gets fixed:

lemma feedback --category cli --subject "…" \
  --issue-encountered "…" --expected-behavior "…" --actual-behavior "…"

See also

• The model → lemma-builder/references/pod-model.md
• Build/restructure a pod → the lemma-builder skill
• Inline live views over pod data → the lemma-widget skill
• Drive a browser → the browser skill; test a pod app systematically → lemma-app-qa
• Run a source-backed investigation → the lemma-research skill
• Perform quantitative analysis → the lemma-data-analysis skill
• Package established content into a durable file → the lemma-artifact-author skill
• Local parsing/OCR of ad-hoc files → the liteparse-documents skill