@lotics/cli 0.181.1 → 0.183.0

package/docs/cli_reference.md

@@ -41,10 +41,10 @@ Per-command syntax, flags, contracts, and gotchas for the public `lotics` CLI. S
 | `lotics app deploy [--prune] -m <message>` | `-m` is OPTIONAL — omitted, the deploy derives the version message from what it actually pushed; pass `-m` when you have a reason worth recording. npm run build; tar source + dist; POST /v1/apps/{id}/versions multipart. **One command ships everything**: before the bundle moves, a deploy pushes every binding the project has ahead of the app — an edited workflow body or declaration, edited agent prose, a changed query — through `set_app_query`, then `set_app_workflow`, then `set_app_agent`, and fails the release if any push is refused. That order is required: an agent declares the query and workflow aliases it may call, so pushing it before its own new query is refused. A workflow's `description` is part of that push and is compared against the recorded baseline, not the live app — it lives on the workflow ROW, which `getApp` does not carry. It never AUTHORS a binding itself — those verbs stay the single writers — and each push carries the fingerprint the project last saw live (`lotics.synced`), so a stale checkout is refused rather than overwriting another author's edit. `package.json` means the same thing for both artifacts: editing `lotics.agents.<alias>.inputs`/`outputs` is pushed exactly like the workflow equivalent (only those two fields — `set_app_agent` merges, so everything the manifest does not model is left untouched). It also regenerates `.lotics/app_fields.ts` before building, since the build INLINES it and a stale copy would ship ids that no longer name what the source thinks they do. `lotics app check` reports the same set without pushing; neither has a `--strict`. What the version RECORDS as the aliases it calls — the set `remove_app_workflow` / `remove_app_query` / `remove_app_agent` consult to refuse unbinding one the served version still reaches — is read by the SERVER out of the source archive this deploy uploads, not reported by the deploy. That matters because the deploy is also what unbinds: a client supplying the evidence used to refuse its own removal cannot be checked by it. After a successful deploy it warns about any alias the source CALLS that is NOT bound, and **names the inverse** — but as a TRANSITION, not a state: bindings this bundle *stopped* calling, compared against what the previous deploy's bundle called (`package.json#lotics.bundle_calls`, which a deploy records). It does NOT remove them: **`--prune` does, and only when passed.** The distinction is what makes the report readable. A static scan sees the bundle's call sites and an agent's `query_aliases`/`workflow_aliases`; it cannot see `lotics run run_app_query`/`run_app_workflow`, whose whole contract is that the alias is bound server-side, or chat's call under `app:use` — and the capability catalog publishes EVERY declared alias to both. So an alias the bundle never called is the normal shape of an agent-facing binding, not a dead one, and reporting it fired on every app built to be driven by an agent while pointing at the flag that deletes it. An alias that WAS called and is not any more is different: that is a call site the author removed, which is compile-, check- and deploy-clean while the binding keeps serving. **No baseline ⇒ nothing reported** — a manifest from before this field, or one a `pull` rebuilt from the server row, costs one quiet deploy and then self-heals, because silence is the only honest answer with nothing to compare. The baseline is STICKY: it advances only once nothing is outstanding, so the `--prune` this warning names still finds the transition on a later run instead of reporting ✓ over a binding that still serves. Pruning runs AFTER the version is live, because the removal tools refuse an alias the SERVED version still declares — so doing it first is refused by the guard that makes it safe. `--prune` is skipped ENTIRELY (with a warning, never a failure) when the source computes an alias at run time, since the scan cannot tell which binding that reaches and pruning "the rest" would be guessing with a deletion. **A removal DELETES the local declaration too** — `package.json#lotics.<kind>.<alias>` and its `synced` baseline — because leaving it would undo the prune: the manifest is what the next plain deploy pushes FROM, so the binding came straight back. That makes the act destructive rather than merely reversible, so what it deleted is written to `.lotics/pruned/<kind>/<alias>.json` and the ✓ names that file plus the `set` verb that re-binds it, on the same line. (These trees are never committed, so git is not the fallback; `lotics app pull --from-version <apv_…>` is the only other route back.) After a successful prune the generated companions are regenerated from the narrowed manifest — the `.d.ts` set and, when a query was pruned, `.lotics/app_fields.ts`, whose table set is derived from the surviving query ASTs. A table named ONLY by the pruned query leaves `F`/`OPT`, which is reported: if your source still addresses it, add the table id to `package.json#lotics.codegen.tables`. A local write that fails at any of this reports what could not be written and which aliases were already unbound server-side; it never fails the release, which is already live. A binding that will not unbind is reported and does NOT fail the release: the version is live and correct — and the server refuses to unbind a WORKFLOW this workspace has actually run (a recorded execution means a caller the source cannot name), which surfaces here as `✗ could not unbind …` with the date it last ran. After a successful deploy it also REFRESHES the `.lotics/workflows/<alias>.globals.d.ts` of any alias whose `// lotics:declaration` stamp says this deploy moved its declaration (only those — refreshing every bound alias would cost one round trip each on every deploy to fix something only ever wrong right after a manifest edit), from the manifest declaration, re-wrapping the SAME on-disk body (never re-fetching it, so local edits survive). A deploy is the moment the manifest becomes real, so it is also the moment the local types stop matching it — and the author's next act is usually `workflow set`, whose body would otherwise be typechecked against the declaration as it stood before this deploy. Non-fatal: the release already shipped, and stale types never fail it. |
 | `lotics app versions [app_id]` | `GET /v1/apps/{id}/versions` — print deploy history newest-first (version number, timestamp, deployer name, build status, the `-m` message; `*` marks the currently-served version). app_id from the local manifest, or pass one to inspect any app without pulling it. Admin-only server-side (mirrors deploy + source download). Answers "what shipped, when, by whom" — e.g. whether a fix was live at an incident's time. Title → stderr, table → stdout (pipeable). |
 | `lotics app codegen [path]` | Regenerate `.lotics/*` from the manifest + workspace schema **without a deploy**. The three `.d.ts` companions (`app_{workflows,queries,agents}.d.ts`) are always rewritten (synchronous, no network). When credentials resolve, also rewrites the **runtime** `.lotics/app_fields.ts` — a real `.ts` exporting `F` (table→field→`"fld_…"`) + `OPT` (table→select-field→option→`"opt_…"`) keyed by display-name aliases, for the tables the app's queries reference (+ optional `package.json#lotics.codegen.tables` allowlist). **There is one form, and that is what makes a starter's source portable**: the keys are slugified DISPLAY NAMES and a starter carries its labels verbatim, so running codegen in a copy's own workspace emits the same keys pointing at that workspace's ids — no binding fetched at load, no prebuilt bundle to keep in step. Also refreshes each bound workflow's `.lotics/workflows/<alias>.globals.d.ts` + re-wraps its EXISTING `src/workflows/<alias>.ts` body in the current envelope (strips + re-wraps; never re-fetches the body, so local edits survive). **`.lotics/` is reconciled to the manifest, not merely added to** — a `<alias>.globals.d.ts` whose alias the manifest no longer declares is DELETED. Only that exact filename shape is removed; anything else in the directory is left alone. The reconcile runs before the credential branch, so it happens offline too. The authored counterpart is never deleted — a `src/workflows/<alias>.ts` the manifest does not declare is NAMED instead (`check` and `set` both take their alias set from the manifest, so editing an undeclared body is a silent no-op). A getApp / binding / schema / dts-fetch failure is non-fatal (warns, keeps the last-generated files). **Re-silvers `package.json#lotics.agents`** from the live app row whenever its `inputs`/`outputs` disagree, then rewrites the agent `.d.ts` from the refreshed block: that block is a mirror AND the offline seed for `useAgentRun` typings, so a stale copy types the app against an agent that does not exist. The write is surgical and order-preserving, so it changes only the fields that actually differ. A hand edit to that block is therefore reverted — it never changed the agent anyway; to change one, `set_app_agent`. |
-| `lotics setup <starter_id> [path] [--email <addr>] [--json]` | **The whole first run, in one command.** Creates an account when this machine has no credential (the same call `auth signup` makes — `--name` and `--timezone` apply), then does exactly what `starter init` does, then prints the one-time sign-in link. It exists because the two-command form has a seam where the FIRST command exists only to produce a credential for the second, and a caller pasting a prompt has to get both right. **`--email` is only for creating an account**: with a credential already resolvable it is REFUSED rather than obeyed, because the two can name different organizations and preferring either one silently copies a starter into an org the caller did not name — the message says how to do each thing on purpose. Without it, `setup` copies into the account you already have and is a pure alias for `starter init`. **`--json` prints one object on stdout and nothing else** — `organization_id`, `workspace_id`, `app_id`, `project_dir`, `signin_url`, and `created` — which NAMES what landed (`tables`, `templates` and `knowledge_docs` are alias arrays; `sample_records` is a row count, since rows are not named things). Aliases rather than counts because the next question is about a particular artifact: a copied template carries the publisher's wording and a copied knowledge doc describes how they work, so "which of these should be mine?" is the conversation a copy starts, and a count cannot begin it. Plus a `warnings` array carrying everything the prose form would have said out of band — an unbindable knowledge doc, a sign-in link that could not be minted, and the notice naming whose build is about to run on this machine. A warning is never merely silenced: when the command fails with an error before it can emit, the ones it had collected go to stderr alongside it. npm's and vite's own output is captured under `--json` and surfaced only if the build FAILS, where it rides out in the error. Reachable with no install: `npx -y @lotics/cli start …`. |
+| `lotics setup <starter_id> [--email <addr>] [--json]` | **The whole first run, in one command.** Creates an account when this machine has no credential (the same call `auth signup` makes — `--name` and `--timezone` apply), then does exactly what `starter init` does, then prints the one-time sign-in link. It exists because the two-command form has a seam where the FIRST command exists only to produce a credential for the second, and a caller pasting a prompt has to get both right. **`--email` is only for creating an account**: with a credential already resolvable it is REFUSED rather than obeyed, because the two can name different organizations and preferring either one silently copies a starter into an org the caller did not name — the message says how to do each thing on purpose. Without it, `setup` copies into the account you already have and is a pure alias for `starter init`. A path positional is accepted and IGNORED with a warning — nothing is written to disk any more — so a prompt written for an older CLI still runs. **`--json` prints one object on stdout and nothing else** — `organization_id`, `workspace_id`, `app_ids` (alias → id), `apps` (each app's `version_number`, or its `error`), `signin_url`, and `created` — which NAMES what landed (`tables`, `templates` and `knowledge_docs` are alias arrays; `sample_records` is a row count, since rows are not named things). Aliases rather than counts because the next question is about a particular artifact: a copied template carries the publisher's wording and a copied knowledge doc describes how they work, so "which of these should be mine?" is the conversation a copy starts, and a count cannot begin it. Plus a `warnings` array carrying everything the prose form would have said out of band — an unbindable knowledge doc, a sign-in link that could not be minted, the publisher's-code disclosure, an app that landed without a version. A warning is never merely silenced: when the command fails with an error before it can emit, the ones it had collected go to stderr alongside it. Reachable with no install: `npx -y @lotics/cli setup …`. |
 | `lotics starter list` | **Works with no account**, and that is the point: whether to copy a starter or build from scratch is decided before one exists, so requiring a key would mean signing up to learn the answer was no. Unauthenticated it lists what Lotics publishes (`GET /v1/starters/official`, public). Authenticated it is the org shelf — The starters this organization can copy — Lotics-reviewed ones plus its own, each with at least one released version. Deliberately NOT a catalogue of everything published: the server returns exactly what a copy would be allowed to take, so the list can never offer something that then refuses. Admin-only. |
-| `lotics starter show <starter_id>` | One starter's name, description, current version and trust standing (`official` — reviewed by Lotics; `your organization's own`). Read it before copying a starter you did not publish. |
-| `lotics starter init <starter_id> [path]` | **Copy a starter into this workspace.** Server-side it scaffolds the tables and fields, creates the document templates and knowledge docs, inserts the sample records, creates a BESPOKE app and materializes its queries, workflows and agents onto it. Then it stops: a starter ships **source only**, with no prebuilt bundle, so the app source is downloaded here, hydrated against the live app (the same steps `app pull` runs), built, and deployed — which is why this needs node and a few minutes, and why the app is not servable until the deploy lands. **What you get is yours outright**: an ordinary app plus ordinary tables, with no link back to the starter, nothing pinned, and nothing to upgrade. Edit any of it. **It builds the publisher's code on your machine** — the copy has to build where your workspace's field ids are, so `npm run build` runs their build script and vite loads their `vite.config.ts` in Node, as you. Dependencies install with `npm ci --ignore-scripts` — the lockfile's exact tree (so a starter published months ago resolves the same packages today), and no lifecycle scripts (the path that fires before you run anything). Neither changes the build itself. That is why provenance is the gate: **copyable only if the starter is Lotics-reviewed or your own organization published it** — copying runs the author's code (its bundle, its workflows, its agents) under YOUR authority, so provenance is the gate, enforced server-side. **Refuses a workspace that already has tables** unless `--adopt`: scaffold matches an entity by DISPLAY NAME, so a starter declaring `Contacts` would bind to yours. `--json` prints one object on stdout instead of progress (the shape is under `lotics setup`) and captures npm/vite output unless the build fails. `--no-sample-data` skips the sample records; with them, the created record ids are reported — they are ordinary records, delete them whenever. Resolves and ANNOUNCES its workspace first (`lotics → <org> / <workspace>` on stderr). Admin-only. Authoring the registry (`starter publish/release/unpublish`) stays operator-only. |
+| `lotics starter show <starter_id>` | One starter's name, description, current version, shelf tile and trust standing (`official` — reviewed by Lotics; `your organization's own`; otherwise `not copyable from this organization`), plus the date it was unpublished once it has been. Read it before copying a starter you did not publish. Admin-only, and readable by id from any org — but an unpublished starter 404s for every org except the one that published it. |
+| `lotics starter init <starter_id>` | **Copy a starter into this workspace.** Server-side it scaffolds the tables and fields, creates the document templates and knowledge docs, inserts the sample records, creates every app the starter carries and materializes each one's queries, workflows and agents onto it — then deploys each app from the dist the starter was published with, rewriting the publisher's sentinel field keys to this workspace's. No build runs anywhere, nothing is written to this machine, and nothing here needs node: the apps are live when the command returns. **What you get is yours outright**: ordinary apps plus ordinary tables, with no link back to the starter, nothing pinned, and nothing to upgrade. Edit any of it — `lotics app pull <app_id>` is how an app's code is edited afterwards. **The publisher's code runs in your workspace as you** — its apps, workflows and agents — which is why provenance is the gate: **copyable only if the starter is Lotics-reviewed or your own organization published it**, enforced server-side; the disclosure is printed (and carried in `--json`'s `warnings`) whenever the starter is not your own. **Refuses a workspace that already has tables** unless `--adopt`: scaffold matches an entity by DISPLAY NAME, so a starter declaring `Contacts` would bind to yours. An app whose deploy failed is reported by name with its reason and the exit is non-zero, but the copy is complete around it — the tables, the records and the app row exist — so it must not be run again; the publisher fixes the starter and it is copied into a fresh workspace. The sign-in link lands on the app when there is one, else on the workspace's app list. `--json` prints one object on stdout instead of progress (the shape is under `lotics setup`). `--no-sample-data` skips the sample records, and a copy that ADOPTS an existing table writes none either — that table already holds real rows, and the fixture set links to itself, so it is all-or-nothing; with them, how many landed is reported. They are ordinary records, delete them whenever. Resolves and ANNOUNCES its workspace first (`lotics → <org> / <workspace>` on stderr). Admin-only. Authoring the registry (`opctl starter publish/unpublish`) stays operator-only. |
 | `lotics starter fixtures capture [--entity <alias> ...] [--limit <n>]` | **(authoring)** Write this app's live records into the project as `fixtures/<entity-alias>.json` — the sample data a starter carries, so a copy lands with something in it. Run from an app project; the app id comes from its manifest. The alias-keyed shape is produced server-side, because the aliases are minted when the starter is extracted and exist nowhere a project can read them. `--entity` is repeatable and comma-separated; omitted, every table the app declares is captured. **Capture a linked set in ONE call** — a link between two rows only resolves within a single capture, so taking companies and contacts separately drops the edge between them (it says so when it happens). `--limit` bounds rows per table (default 10, max 200). **READ WHAT IT WROTE before committing**: these rows are created verbatim in every workspace that copies the starter, so a real customer name, price or address captured here is published. Files, formulas, rollups, lookups and autonumbers are never captured — the platform writes those. Admin-only; writes nothing to the workspace. |
 | `lotics upgrade` | Update this CLI in place. Runs the same installer a person would, chosen by how THIS copy arrived: an npm install upgrades through npm, a script install re-runs the script — the runtime knows which (the executable is compiled, the npm bin runs under node), so nobody has to. It downloads nothing itself; resolving a version, verifying the checksum and replacing a running executable already exist in the installers, and a second copy of that inside the binary would be a second thing to get right. Replacing the binary while it runs is safe — a rename leaves the running image mapped on unix, and on Windows the installer moves the old aside precisely because the file is in use. Already current is a no-op that says so. Needs no auth. |
 | `lotics docs` \| `lotics docs <area>` | The index of the reference docs, **resolved out of the packages installed beside this project** — never carried by this CLI. **Both levels are discovered by looking**: every `@lotics/*` package carrying an `AGENTS.md` or a `docs/` in any `node_modules/@lotics` from the current directory UPWARD (nearest wins, so a hoisted root copy never shadows the one a project's own imports resolve to), and within each, every area it actually ships. Titles come from each file's own `# heading` and the version from the installed `package.json`, so a doc OR a whole package added upstream appears with no change to this CLI, and a skewed install is visible rather than reassuring. A package's index is named after the package (`lotics docs ui`), never `index`. `@lotics/app-sdk`, `@lotics/ui` and `@lotics/cli` sort first as a reading ORDER, not a filter. Both the index and `<area>` print to **stdout** — the index is the payload of a bare `lotics docs`, so `lotics docs | grep -i excel` works — with only the provenance line on stderr, so `lotics docs ai > ai.md` is the doc alone; a name two packages share is refused with both qualified forms (`lotics docs ui/templates`) rather than resolved silently. Needs no auth. Outside a project only `@lotics/cli`'s own resolve, and it says so. |
@@ -57,7 +57,7 @@ Per-command syntax, flags, contracts, and gotchas for the public `lotics` CLI. S
 | `lotics app workflow check [alias]` | Check the editable workflow bodies locally, no auth / no network, in the **server's own order** — parse, then type-check. **Parse** runs `parseWorkflowJs` from `@lotics/shared` (the SAME module `verifyWorkflow` calls, never a second implementation) over the stripped body `set` would upload, with `toolNames: undefined` (the CLI ships no tool registry, so tool-name resolution stays a server check while every shape/scope rule runs here). A body the subset rejects reports **that error alone** and skips the compiler — it never reaches the server's compiler either, so tsc's opinion of it is noise. **Type-check** then builds an **isolated** `ts.Program` per alias from exactly that alias's `{body, globals}` pair — mirroring the server, which verifies one body at a time — so the per-alias ambient `trigger` never collides and `trigger.app_workflow.inputs` is checked against the right alias. All aliases run in ONE node process (N programs, not N `tsc` spawns), with the SAME compile options the server uses at set-time verify (lib `es2022` with no DOM, target ES2022, strict, NodeNext, `types:[]`, skipLibCheck) and the app's OWN `typescript` (resolved from its `node_modules`, never bundled into the CLI). What the compiler sees is the **checked source**, not the file: `rewriteAccumulatorAppends` from `@lotics/shared` — the SAME transform the server applies before its set-time compile — is applied in memory, so a pulled body's canonical `out = concat(out, [item])` accumulator checks green here exactly as it saves there, and the body on disk is never rewritten. Reports `<file>:<line>:<col> - <TS####\|subset>` at the **physical** line in `src/workflows/<alias>.ts`, so an editor jump lands on the offending code (these are deliberately NOT `set`'s body-relative numbers — `set` prints no file path, so there is no format to agree with); exits non-zero if any alias fails. Green is honest but not total: `set` additionally resolves names, lints and structurally validates against the live workspace — passes that need its tables and tool schemas, so they cannot run offline, and the success line says so. A bound alias with no body file yet warns + skips; a body with no globals errors (naming `lotics app codegen`, which refreshes types WITHOUT touching the body — a pull would overwrite it). **It also keeps the types honest.** Each alias's `.lotics/workflows/<alias>.globals.d.ts` carries a `// lotics:declaration <hash>` stamp of the manifest declaration it was rendered from; `check` compares it to `package.json#lotics.workflows.<alias>` and, when they differ, re-renders that alias's dts from the LOCAL declaration before compiling. Without it the verdict was confidently wrong in the exact case an author needs it — declare an input, run `check`, and get `TS2339: Property 'x' does not exist` pointing at your body for a schema the types have never been told about. The server renders a dts from a SUPPLIED declaration, so this works before the manifest has ever been deployed, which is when it matters (the order is edit → check → set). This is the ONE thing `check` uses the API for: it is skipped entirely when the stamps match (the common case, so `check` stays instant and offline), and with no credentials or a failed fetch it WARNS and checks against the older types rather than blocking. A file written before the stamp existed reads as unknown, never as matching, so a pre-existing checkout heals on its first run. |
 | `lotics app subdomain <new-subdomain>` | Rename the app's public `<slug>.lotics.app` address via `PUT /v1/apps/{id}/subdomain`. app_id comes from the local `package.json` manifest; the chosen slug must be a valid DNS label and free; the old address stops resolving. |
 | `lotics app rename "<new name>"` | Change the app's display name (launcher/title) via the `update_app` tool. app_id comes from the local `package.json` manifest; the public address (`subdomain`) and code (`deploy`) are unchanged. |
-| `lotics app dev [path] [--port=N] [--vite-port=N] [--view-as=<member_id>]` | Spawn Vite dev server + an RPC-forwarding HTTP server. The wrapper page embeds the iframe with `sandbox="allow-scripts allow-same-origin"` matching production; postMessage ops (query / workflow / members / context / upload / openExternal / urlState / agentRun) are forwarded to api.lotics.ai using the CLI's API key — file bytes move in **both** directions through the dev server's own relays, never browser↔storage: dev runs against the PROD bucket, whose CORS admits `https://*.lotics.app` and not `http://localhost:<port>`, so a direct browser transfer is blocked — no upload could complete and no preview engine (PDF/Word/Excel all FETCH the bytes) could read a file. `upload` mints a presigned URL and PUTs it **to `PUT /_upload/<file_id>`** (`dev/upload_relay.ts`) from the wrapper page — same-origin, so no preflight and no CORS — and Node forwards it on; every presigned `url`/`thumbnail_url`/`preview_url` on a **file object** in an RPC result is rewritten to **`GET /_file/<token>`** (`dev/file_relay.ts`, absolute — the iframe would resolve a relative path against Vite), which streams the bytes back with `Range` passthrough (206s intact, so PDF seeking works) and an `Access-Control-Allow-Origin` for the Vite origin (the one cross-origin hop left is OUR response to allow). Neither relay ever takes a destination from the client — it gets a `file_id`/token and transfers only to/from a URL it minted or observed itself, so there is no client-controlled target and no SSRF surface. A URL in a record's own text cell is NOT rewritten. Production is unchanged (direct-to-storage, no bytes through the API server); `openExternal` and `urlState.get/set` are handled locally (the latter read/write the wrapper page's own address bar — `set` writes in place via `replaceState` and browser back/forward broadcast a `url-state` message back, so `useUrlState` survives refresh and is shareable in the dev loop; in-app *routing* is the app's own (the iframe owns its url via `@lotics/app-sdk/router`), and the wrapper bakes the saved screen (`_loc`) into the iframe src on load so a refresh restores it, mirroring production); `agentRun` (streaming) is proxied through `POST /_agent_run`, which opens the run's SSE with the CLI key and pipes chunks back to the iframe (`stream-chunk`* → `stream-end`), so `useAgentRun` works in the dev loop just like production; `context` resolves the viewer (`member_id` from `cli/whoami` + `comments_enabled` from the local manifest) and fetches the installation's stored `config` live from the app row, so `useConfig()` renders the same values as production. `--view-as` (global flag; also `LOTICS_VIEW_AS`) threads `x-view-as-member-id` so `is_current_member` + `context` resolve to that member — **admin key only** (the server 403s a non-admin), writes stay attributed to the key owner. Hot reload via Vite; full DevTools / Playwright access via plain localhost. **Every forwarded op logs one line naming its ALIAS** — `[rpc] query applicants 231ms` — and `query applicants (count)` for a count request, which is a SECOND full execution of the same query rather than a cheap lookup. When requests overlap the line carries `· N in flight`. That number is the one to watch: the server bounds how many app queries run at once, so requests past the bound wait and the wait lands inside each request's own duration — a burst reads as "every query got slower", which looks like a slow database and is not one. A screen firing its list plus three facet counts on one keystroke shows up here as eight lines over one or two aliases; see `@lotics/app-sdk` `docs/data_fetching.md` (`useCount`, and handing `usePaginatedQuery` a `total`) and `docs/queries.md` §10 for collapsing them. **Holds no realtime connection** — push belongs to the product frontend, so an app previewed here never updates on an external write (a CLI run, another tab, an agent): reload to see it. Deliberate rather than missing, since the alternative is a second implementation of the channel in the wrapper page, and a blanket poll here would hide an app whose queries do not declare their tables — the one mistake the real host punishes. The startup banner says `realtime: off` so this is visible without reading this table. The dev-optimizer pre-bundle list (`optimizeDeps.include`, load-bearing for dev) is imported from `@lotics/ui/vite` (`loticsOptimizeDeps`) rather than hardcoded in the scaffold, so it tracks the installed `@lotics/ui` and can never go stale. Binds **loopback only** (`127.0.0.1`) — `/_rpc` dispatches with the developer's API key, so a socket on every interface would hand anyone on the network full read/write on the workspace. |
+| `lotics app dev [path] [--port=N] [--vite-port=N] [--view-as=<member_id>]` | Spawn Vite dev server + an RPC-forwarding HTTP server. The wrapper page embeds the iframe with `sandbox="allow-scripts allow-same-origin"` matching production; postMessage ops (query / workflow / members / context / upload / openExternal / urlState / agentRun) are forwarded to api.lotics.ai using the CLI's API key — file bytes move in **both** directions through the dev server's own relays, never browser↔storage: dev runs against the PROD bucket, whose CORS admits `https://*.lotics.app` and not `http://localhost:<port>`, so a direct browser transfer is blocked — no upload could complete and no preview engine (PDF/Word/Excel all FETCH the bytes) could read a file. `upload` mints a presigned URL and PUTs it **to `PUT /_upload/<file_id>`** (`dev/upload_relay.ts`) from the wrapper page — same-origin, so no preflight and no CORS — and Node forwards it on; every presigned `url`/`thumbnail_url`/`preview_url` on a **file object** in an RPC result is rewritten to **`GET /_file/<token>`** (`dev/file_relay.ts`, absolute — the iframe would resolve a relative path against Vite), which streams the bytes back with `Range` passthrough (206s intact, so PDF seeking works) and an `Access-Control-Allow-Origin` for the Vite origin (the one cross-origin hop left is OUR response to allow). Neither relay ever takes a destination from the client — it gets a `file_id`/token and transfers only to/from a URL it minted or observed itself, so there is no client-controlled target and no SSRF surface. A URL in a record's own text cell is NOT rewritten. Production is unchanged (direct-to-storage, no bytes through the API server); `openExternal` and `urlState.get/set` are handled locally (the latter read/write the wrapper page's own address bar — `set` writes in place via `replaceState` and browser back/forward broadcast a `url-state` message back, so `useUrlState` survives refresh and is shareable in the dev loop; in-app *routing* is the app's own (the iframe owns its url via `@lotics/app-sdk/router`), and the wrapper bakes the saved screen (`_loc`) into the iframe src on load so a refresh restores it, mirroring production); `agentRun` (streaming) is proxied through `POST /_agent_run`, which opens the run's SSE with the CLI key and pipes chunks back to the iframe (`stream-chunk`* → `stream-end`), so `useAgentRun` works in the dev loop just like production; `context` resolves the viewer (`member_id` from `cli/whoami` + `comments_enabled` from the local manifest) and fetches the app's stored `config` live from the app row, so `useConfig()` renders the same values as production. `--view-as` (global flag; also `LOTICS_VIEW_AS`) threads `x-view-as-member-id` so `is_current_member` + `context` resolve to that member — **admin key only** (the server 403s a non-admin), writes stay attributed to the key owner. Hot reload via Vite; full DevTools / Playwright access via plain localhost. **Every forwarded op logs one line naming its ALIAS** — `[rpc] query applicants 231ms` — and `query applicants (count)` for a count request, which is a SECOND full execution of the same query rather than a cheap lookup. When requests overlap the line carries `· N in flight`. That number is the one to watch: the server bounds how many app queries run at once, so requests past the bound wait and the wait lands inside each request's own duration — a burst reads as "every query got slower", which looks like a slow database and is not one. A screen firing its list plus three facet counts on one keystroke shows up here as eight lines over one or two aliases; see `@lotics/app-sdk` `docs/data_fetching.md` (`useCount`, and handing `usePaginatedQuery` a `total`) and `docs/queries.md` §10 for collapsing them. **Holds no realtime connection** — push belongs to the product frontend, so an app previewed here never updates on an external write (a CLI run, another tab, an agent): reload to see it. Deliberate rather than missing, since the alternative is a second implementation of the channel in the wrapper page, and a blanket poll here would hide an app whose queries do not declare their tables — the one mistake the real host punishes. The startup banner says `realtime: off` so this is visible without reading this table. The dev-optimizer pre-bundle list (`optimizeDeps.include`, load-bearing for dev) is imported from `@lotics/ui/vite` (`loticsOptimizeDeps`) rather than hardcoded in the scaffold, so it tracks the installed `@lotics/ui` and can never go stale. Binds **loopback only** (`127.0.0.1`) — `/_rpc` dispatches with the developer's API key, so a socket on every interface would hand anyone on the network full read/write on the workspace. |
 | `LOTICS_UI_SRC=<abs path to packages/ui/src>` (env, not a command) | Dev-link `@lotics/ui` to a monorepo checkout for the length of ONE command, **for every tool at once**. The app's `vite.config.ts` gets its whole `resolve` block from the kit (`resolve: loticsResolve()` — `@lotics/ui/vite`), which reads the variable at call time and adds the `@lotics/ui/*` → working-copy alias, so kit edits go live under `lotics app dev` (HMR) and bundle under `lotics app deploy`. In the same breath, every command that regenerates types (`create`/`pull`/`dev`/`deploy`/`codegen`, all via `writeAppDts`) writes **`.lotics/tsconfig.link.json`** — the matching `paths`, which the app's `tsconfig.json` `extends` — so `tsc`, vitest, eslint and your EDITOR resolve the same copy Vite does. Unset ⇒ every one of them goes back to `node_modules`, and the generated file is rewritten inert. **Why `paths` and not `npm link`:** the kit ships un-built `.tsx`, so a kit file outside `node_modules` resolves its OWN `react`/`react-native` from the monorepo — two copies in one program and every shared type stops matching ("Two different types with this name exist, but they are unrelated"). The generated file therefore also pins every peer @lotics/ui declares to the APP's copy, types-package first (`react` → `@types/react`; pinning the runtime package instead strands tsc on a `.js` with no declarations). The pin set is derived from the installed kit's `peerDependencies`, so it tracks the kit rather than rotting. **Nothing hand-written is touched** — the generated file lives in `.lotics/` (the CLI's own dir) and no config is edited by regex. Identical for a monorepo app and an EXTERNAL one (e.g. `~/lotics_apps`). `app deploy` still warns whenever the variable is set — that the bundle carries kit code from your working copy, or that the app's config predates `loticsResolve()` and never reads it, so the PUBLISHED kit is going out. An app whose `tsconfig.json` already `extends` something else is told rather than rewritten: add `./.lotics/tsconfig.link.json` to the array yourself. |
 | `lotics xlsx <subcmd>` | Local .xlsx read/write/edit using the bundled `@lotics/xlsx` engine (no auth, no network). `write` takes its JSON inline, as `@file`, or piped on stdin, ingested exactly as `run` ingests tool args. 14 named subcommands (read, write, set-cell, clear-range, merge, unmerge, add-sheet, delete-sheet, rename-sheet, insert-rows, delete-rows, insert-cols, delete-cols, set-style) + `batch` for applying multiple of the same 14 ops in a single parse/export cycle. `read` also takes `--sheet <name>` (limit output to one sheet — unknown name fails with the available list) and `--range <sheet>!<A1:G60>` (limit to a cell window; the `<sheet>!` prefix is optional when `--sheet` supplies the sheet, a single cell like `S1!B2` is a 1×1 window) to trim a large workbook's JSON — the output shape is unchanged, only the `sheets` array and each sheet's `cells` map are filtered. **`read` reports formatting back, so a generated file is verifiable through this path** rather than by unzipping OOXML: each cell carries `numFmt` when the file gave it one, and `--with-format` adds the resolved `style`. The asymmetry is deliberate — a parsed cell's style is *never* absent (every cell resolves to at least a font — size, name, colour), so emitting it by default would put three noise keys on every plain cell and make “is this styled?” unanswerable by presence; `numFmt` is genuinely absent on an unformatted cell, so it needs no flag. **`write` takes sheet-level `colWidths` (`{"A":34}`) and `rowHeights` (`{"1":44}`)** — without them every column is the default width and a human-facing workbook is unreadable no matter what the cells say. Both are written *pinned* (`customWidth`/`customHeight`), so Excel does not auto-fit them away, and both apply to a row/column that holds no cells (a spacer row's height survives). Keys are a bare column letter and a bare row number, bounded by Excel's grid (`A`…`XFD`, `1`…`1048576`): a key outside it, or a cell ref like `A1` where a column letter belongs, is **rejected** rather than resolved to something adjacent — past the grid the reference is written into the file verbatim, addressing a cell that cannot exist. Unknown **sheet** properties are rejected on the same terms as unknown cell properties — a silently-ignored `columnWidths` typo is a file that looks written and is not. **A `--flag` a subcommand does not know is refused by name** (`--with-formats` would otherwise read as proof the file carries no styles). `xlsx` and `docx` own their whole tail: a global flag's NAME means nothing there, so `xlsx delete-rows f.xlsx S1 5 3 --force` is refused rather than run, and `xlsx set-cell f.xlsx S1!A1 -v` writes the value `-v`. Subcommands whose trailing arg is CONTENT (`xlsx set-cell`, `docx replace-text`, `docx append-paragraph`/`insert-paragraph`) are deliberately exempt: a value may legitimately begin with `-` or `--`, and there a typo is indistinguishable from data. `--help` is the one spelling still reserved everywhere. They are covered instead by arity — **every fixed-shape subcommand refuses an argument past the last one it reads**, whatever it looks like, because the likeliest source is a flag the caller believes exists and these commands write in place. Arity rather than a leading `--` is the discriminator, since a sheet name may legitimately begin with one. **A cell VALUE is read as the type the caller stated, on both JSON surfaces.** `write`'s `cells` and `batch`'s `set-cell` `value` take the same union — a bare `string | number | boolean | null`, or a `{value, formula, numFmt, style}` object — and honour it: a JSON string writes a text cell, digits and all, so `"0071000512345"` (MST, số tài khoản, số vận đơn) keeps its leading zeros and `"1234567890123456789"` keeps its last two digits, neither of which survives being re-read as characters. The one reading applied to a BARE string is a leading `=`, which is a formula — the only way to write one in the shorthand form; `{"value": "=SUM(A1)"}` is the stated literal, and how a cell that must hold the text `=x` is expressed, on either surface. `value` is a literal on the object form of BOTH surfaces — `formula` is the key that says otherwise — and a literal beginning with `=` is **written as asked and named in a stderr warning**, since it is the one literal indistinguishable from a mistake: it renders in a viewer exactly like the formula the caller probably meant, computes nothing, and is skipped by every SUM over the column. The object form also carries `numFmt` and `style` per cell in `batch`, the same as in `write`. The `set-cell` POSITIONAL is different because a shell argument carries no type: there the characters are read for what they denote (`TRUE` → boolean, digits → number), stopped by two things — the target cell's number format being Text (`@`), and a zero-padded digit string, which stays text whatever the target format says (a deliberate divergence from Excel: losing a leading zero is unrecoverable, while a text cell in a number column is visible). Every subcommand that can introduce a formula (`write`, `set-cell`, `batch`) **evaluates it and writes the cached value**, so a generated formula does not read back blank: Excel and Sheets recalculate on open, but parsers — including this CLI's `read` and the rest of the platform — take the cached `<v>`. A formula the engine cannot evaluate still gets written, with a stderr warning naming the cells, rather than silently leaving a hole where a number belongs. Atomic in-place write (temp file + rename). |
 | `lotics docx <subcmd>` | Local .docx read/write/edit using the bundled `@lotics/docx` engine (OOXML round-trip surface only — no ProseMirror baggage). `write` takes its JSON inline, as `@file`, or piped on stdin, ingested exactly as `run` ingests tool args. Subcommands: read, write, append-paragraph, insert-paragraph, delete-block, replace-text, batch. A legacy `.doc` (Word 97–2003 OLE2 binary) is detected in `loadFile` and routed through `@lotics/ooxml`'s `loadDocxFromBuffer` (which re-emits it as real OOXML) before reading — so `lotics docx read` works on a `.doc`, not just a `.docx`. Opaque blocks (tables, custom XML) preserved verbatim. Atomic in-place write. **`replace-text` matches across run boundaries** — Word splits a run at every formatting change, so a `{{marker}}` routinely lands split — and reads straight THROUGH marks that occupy no place in the sentence (`w:proofErr`, `w:footnoteReference`, endnote/comment refs + ranges, `w:bookmarkStart`/`End`, `w:lastRenderedPageBreak`). `w:proofErr` is the one that decides whether this works in practice — Word brackets every word its dictionary rejects, so on non-English text it lands between nearly every pair of runs. It still refuses to join across anything that occupies space in the text — `w:br`, `w:tab`, `w:sym`, a drawing, or any tag not on that allowlist — because the joined string does not represent the glyph and a match there would rewrite text the caller never saw. The SAME rule applies inside a table cell as outside it — both run one `replaceInParagraph` over paragraphs found at any depth, so a marker split by a line break is refused in both rather than rewritten in the cell and skipped in the body under a success message. Zero matches is always a hard error, never a silent no-op, and when the words ARE on the page the error names the block and the splitting mark (`The text IS present at block 1, split by w:br …`) rather than claiming the text is absent. |

package/docs/knowledge_docs.md

@@ -221,7 +221,7 @@ statement about discovery, so it cannot silently break an app that depends on a
 Reach for it instead of `rm` whenever the material still matters: last year's tariff schedule, a
 handbook a newer one replaced, the raw source a curated doc was written from.
 
-## Package-managed knowledge
+## Knowledge from a starter
 
 A knowledge doc can also arrive with a **starter** — a published snapshot of a workspace setup
 that carries a corpus of docs (and document templates) along with it. Copying a starter creates

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lotics/cli",
-  "version": "0.181.1",
+  "version": "0.183.0",
   "description": "Lotics SDK and CLI for AI agents",
   "type": "module",
   "bin": {