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.
• A workload's authority is the intersection, not the union: its own grants and what the invoking person could do themselves. A grant is a ceiling on the workload, never a promotion for the person driving it — a POD_VIEWER who invokes an agent granted datastore.record.write still cannot write through it, and someone removed from the pod can do nothing through any workload. So "the agent is granted this" is only half an answer; check the person too.
• 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 three shapes, and the refusal code says which. A human without the pod role gets INSUFFICIENT_PERMISSION. A workload missing a grant gets MISSING_WORKLOAD_RESOURCE_GRANT, naming the resource a builder must grant. A workload that holds the grant but is acting for someone who lacks the permission gets DELEGATION_EXCEEDS_INVOKER — granting the workload more is the one fix that cannot work; give the person the permission, or run it as someone who already has it.

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 (or --json) only to pipe/save, and --full to expand folded fields. Errors, warnings and progress lines go to stderr, so stdout carries the result and nothing else: lemma --json … | jq stays parseable even when the command fails, and you never need to redirect to keep the JSON clean. 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: query, datastore, runtime, servers, telemetry, auth, and config exist only as written. There is no lemma tools group — first-party tools are the in-process agent tools described below, not a CLI surface.

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. That allow-list is the whole of it: spreadsheets and tabular data (CSV, TSV, JSON, YAML, XLSX, ODS), presentations (PPTX, ODP), images (no OCR), email (.eml/.msg), audio, video and archives 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. So: "what does page 3 look like?" → files child …/pages/page_0003.jpg → view-image; "what does it say?" → files cat … --pages 3.

As an in-process tool, view_image takes exactly one of pod_file_path (the datastore, no download needed) or workspace_file_path (the sandbox) — it never infers the store from the path, and both or neither is an error. It handles images only; hand it a PDF and it points you at pod_view_document_pages, which renders that document's pages directly and skips the files child round trip. It rides on any vision-capable model regardless of configured toolsets, so you probably have it.

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 = still working, FAILED = will be retried, FAILED_PERMANENT = out of retries or unprocessable, and never re-driven; only a fresh upload of the content re-opens it).

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, and a value over 256KB or a record over 1MB is refused (put a document in a pod file and keep its path in a FILE_PATH column). 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.

query run returns {items, total, truncated}, and total is not a count of matches — it is how many rows came back, equal to the match count only when nothing was cut. Results stop at the deployment's row cap (1,000 by default); past it truncated is true and items is a prefix. Read truncated before you quote a number: a capped result looks exactly like a complete one, and reporting total as a total is how you tell somebody they have 1,000 orders when they have forty thousand. There is no cursor — aggregate in SQL (count(*)) or narrow the query. (The in-process pod_query tool is the same shape under other names: rows/row_count/truncated, plus a note when it was cut.)

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>

WAITING on a run means a person, and only a person. A run parked on an agent conversation, an async function, or a timer stays RUNNING — so RUNNING is not proof that anything is executing, and WAITING is not the general "blocked" state. runs get settles it via active_wait: wait_type is HUMAN, AGENT, FUNCTION, or TIME, alongside 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 form with no assignee is submittable by any member with execute access but belongs to nobody's queue, so an empty runs waiting does not mean no run is parked.

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 (the ceiling; requests above it are clamped). A snoozed conversation is healthy and needs nothing from you. The CLI does not distinguish the two — conversations get reports status and last_run_status but no wait reason — so tell them apart from the transcript: conversations approvals <id> lists an outstanding ask_user/approval, and an empty list on a WAITING conversation means it is asleep.

A notification is the third thing, and unlike the other two it comes looking for you: an agent or a workflow has asked you for something. They arrive wherever you already talk to the pod — Slack, Telegram, WhatsApp, email — and always leave a copy in your Lemma inbox, so nothing is only on a channel that can fail.

Each carries two independent states, and reading them as one is the usual mistake. status is about you: OPEN (still owed), RESPONDED, ACKNOWLEDGED (seen and dismissed, or sent as a pure FYI), EXPIRED (nobody answered in 72h), CANCELLED. delivery_status is about the channel: PENDING, DELIVERED, UNDELIVERABLE when no chat app or mailbox could carry it — usually because you have never messaged the pod's bot — or FAILED when a channel was chosen and the send raised. Only FAILED is a real delivery fault. UNDELIVERABLE is not: the notification exists and the inbox has it.

Answer it in the app, or just reply on the surface it arrived on — the agent handling that thread records your answer either way, and the asker sees the same result. A notification with responds_through_action is a workflow form: it is answered by runs submit-form, which validates against the node's schema, not by free text. It resolves exactly once: a second answer, from another device or after an expiry swept it, is refused rather than overwriting the first.

Agents and chat

lemma chat "What can you do in this pod?"                           # one-shot to the DEFAULT pod agent
lemma chat                                                          # interactive, default pod agent
lemma agents list
lemma agents chat triage-agent "Summarize today's urgent tickets"   # named agent; interactive without a message
lemma agents run triage-agent "Classify this: ..."                  # waits + streams the ANSWER (--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"
lemma conversations approvals <conversation-id>                     # gated tool calls / ask_user awaiting a decision
lemma conversations approve <approval-id> -c <conversation-id>      # --deny to reject; --session for the whole run

lemma chat takes an agent name, a message, or both, and tells them apart by whitespace: one quoted multi-word argument is a message to the default pod agent, one bare word is an agent name and opens an interactive session. Be explicit with --agent/--message when a name could be mistaken for prose.

conversations approve with no approval id acts on every pending approval in the conversation — it prints the list and asks you to confirm first, and --yes skips even that. Name the id whenever you mean one decision.

An agent acts under your delegated identity — it sees exactly what you'd see (your RLS rows, your /me, your connected accounts), and its grants are bounded by your own access, never added to it (see "Missing access has three shapes").

An agent can come and find a person

With the MESSAGING toolset an agent reaches a pod member who is not in the conversation — that is where the notifications above come from. It is not a paused wait: message_user returns immediately, the agent finishes its turn, and it is given a fresh turn in the same conversation once the last person answers, where check_messages reads the replies. Nothing sleeps and nothing polls.

• list_pod_members — resolve a name to an id or exact email (a name will not resolve on its own), and read reachable_on: the channels that can actually carry a message to that person right now.
• message_user — to, a Markdown message, an optional background_instruction (never shown to them; it tells the agent handling their reply what counts as an answer), an optional title, and an optional channel — email, slack, teams, telegram, whatsapp. Omit channel unless you have a reason: the default reaches them where they last spoke. A named channel is honored or refused, never quietly swapped, so check reachable_on first. Keep the returned notification_id.
• check_messages — status by notification id. RESPONDED is the only status that means somebody answered; call it when a turn opens saying replies landed, not in a loop.

To answer the person you are already talking to, just reply — ask_user is for a question that must block this turn, and neither is message_user.

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. (CONNECTORS is not alone behind search_tools: POD, SUBAGENTS, MESSAGING and SNOOZE are deferred the same way. Not seeing a tool in your prefix is not the same as not having it — go looking before concluding you cannot do something.)
• The CLI (lemma connectors …, needs the WORKSPACE_CLI 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. --metadata-only strips large HTML body fields from the response — reach for it when listing mail or documents, where the bodies are most of the payload and none of the answer. 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. Read the refusal code before you do anything. 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>. But DELEGATION_EXCEEDS_INVOKER means the opposite: the grant is there and the person is short. Adding grants will not move it — the person needs the permission, or the run needs a different invoker.
• Resource not found. Confirm the active pod (lemma pods list) and exact names (lemma pods describe; lemma apps list for apps).
• "No such command" / "has no option". You may be on an older CLI than the server. lemma doctor diagnoses version skew and duplicate installs; lemma update upgrades this CLI in place (--version to pin an exact one). Errors print on stderr, so a --json … | jq pipeline shows you an empty stdout rather than a parse failure — read the terminal, not just the pipe.
• 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 (spreadsheets, presentations, images and email are stored but never searchable); FAILED_PERMANENT means retrying will not help.
• A query's numbers look too round. Check truncated on the query run payload before believing total — 1,000 rows exactly is usually the row cap, not the answer.
• Workflow stuck. runs get <run-id> → active_wait shows what it's blocked on; step_history shows the failing node, its input, and error. Remember the run's own status is RUNNING for every wait except a human form, so RUNNING is not evidence of progress. 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 "…"

--category is one of cli, skill, platform, docs, other — which part of Lemma the report is about.

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