ZeroWidth

Stops a run that is still PENDING / RUNNING / SCORING — the brake on a run that's spending more than expected or was started by mistake. The queue stops at once; an item already handed to the flow finishes on its own timeout. The run settles as FAILED with a 'cancelled' reason and keeps the items it completed. A run that already finished returns run_not_live. May return `needs_confirmation`.

One run: status, overall score, and by default the ten lowest-scoring items with their per-criterion scores and the judge's reasoning (`items: all` for every item, `none` for totals only). THE ANSWER to 'why did the score drop' — read this, then name the criterion that slipped and quote the reasoning on the lowest items. Works for runs Caliper ran and for runs submitted from CI (triggeredBy 'ci', usually labeled with a pull request number). Still no polling loops; read it when the user asks.

What the flow actually DID on one run item: every step in order (what it said, which tools it called with what arguments, what came back) and the final answer, plus status, duration, and cost. Read this when a low score needs explaining beyond the judge's reasoning — a wrong tool call or an empty tool result is usually the cause, and the fix is different from a prompt fix. Pass the run item's `id` (NOT `itemId`) from caliper_evals_runs_get. Items whose outputs were supplied from outside (external evals) have no trace and return not_found. Steps are capped for transport; the run page in Caliper…

Recent runs for one eval, newest first: status (PENDING/RUNNING/SCORING/DONE/FAILED), overall score once DONE, label, and timestamps. THE CHECK-BACK for caliper_evals_run: when the user asks how the run went, read this — cite the run id and score, and compare against the PREVIOUS run's score for the delta. Still don't poll in a loop; check when the user asks or when reporting.

Changes an eval's title, description, visibility, schedule (interval + time anchor), runOnPublish (queue a run whenever the flow publishes), regressionThreshold (score drop vs the previous run that triggers an alert; null = off), or scenarioId (the scenario runs use). Pass only what changes. The dataset, rubric snapshot, and target are fixed at creation — create a new eval to change those. Scheduled and on-publish runs spend credit on their own, so state that plainly when turning them on. May return `needs_confirmation`.

How a Workbench flow is ACTUALLY doing, with receipts: every Caliper eval targeting the flow, recent runs with scores, the latest run decomposed into per-criterion averages, the score delta vs the previous run, and the worst-scoring items WITH the judge's reasoning. Use this BEFORE claiming a flow works or proposing changes — and cite the runId + scores when you do. The worst items are diagnostic: failures clustered around missing company facts suggest a knowledge gap (consider proposing a Compass interview with the workflow owner) rather than a prompt problem.

Creates the scoring rubric an eval's LLM judge uses — 1-10 criteria, each scored on a numeric scale (default 1-5), judged pass/fail (kind pass_fail), or checked in code with no judge (kind check: contains, not_contains, matches a regex, valid_json, max_chars, equals_expected). Write criteria about the FLOW'S JOB (accuracy to source material, tone, refusal behavior), not generic 'quality'.

Deletes a rubric. Existing evals keep their snapshot of it and keep running; nothing new can bind to it. May return `needs_confirmation`.

One rubric with its criteria — each criterion's name, what the judge looks for, and the score scale. Read this to explain a score (which criterion slipped and what it asks for) or before caliper_rubrics_update. Get the rubricId from caliper_rubrics_list or an eval's rubricId.

Lists the rubrics in the active workspace (id, name, description, criterion count). Check here BEFORE caliper_rubrics_create — reuse an existing rubric's id in caliper_evals_create when one already scores the same job. Read a rubric's criteria with caliper_rubrics_get.

Changes a rubric's title, description, visibility, or replaces its criteria wholesale (pass the FULL list — criteria get fresh ids). Evals snapshot the rubric when they're created, so an existing eval keeps scoring with the criteria it started with; say so, and offer to create a new eval when the criteria change materially. May return `needs_confirmation`.

Datasets this source keeps adding to as conversations arrive: which dataset, for what (review, eval, spec), the pick it matches, sampling, how many have landed, and whether it's still running or why it stopped.

Stops a source from adding new conversations to a dataset. What it already added stays. Get the feedId from caliper_source_feeds_list. May return `needs_confirmation`.

Apps sending their agent's traces to Caliper — from their own code, OpenTelemetry, or a published Workbench flow — busiest first: name, how it sends, traces in the last 14 days (per day), and the tools its agent calls most with the share of traces using each. Start here when the user asks how their agent behaves on real traffic; then caliper_sources_over_time for one source.

One source day by day: traces, failures, tool calls, lookups that found nothing, tokens, cost, and speed (meanMs; p50/p95 as 'answered within' bucket edges); each tool with its calls and the share of each day's traces that used it; the documents lookups landed on most; models used. Use it to answer 'how often is it calling web search and is that changing' — compare the first and last weeks and name the day it moved. Follow with caliper_traces_list on that day or tool to show the conversations behind it.

Saves the conversations behind a pick (a day, a tool, a document, failures — the same narrowing as caliper_traces_list) into a new dataset (newDatasetName) or an existing one (datasetId), newest first up to `limit`. `purpose` decides what each item keeps: review = the agent's reply, tool calls included, for people to rate; eval = the reply becomes the expected answer; spec = only the questions, for people to answer. Set keepAdding to make it live: new matching conversations keep arriving (every one, or 1 in 10 / 1 in 100), up to 1,000. Then offer the next step — caliper_evals_create, or a revi…

Copies one starter into the workspace as a new dataset and a new rubric the user owns and can edit. Returns both ids; the next step is caliper_evals_create binding them to the flow (or target 'external'). Needs caliper:datasets:write and caliper:evals:write.

The starter sets Caliper ships: prompt injection, system-prompt leaking, over-refusal, personal-data handling, bias under ambiguity. Each is a small original dataset (single messages AND multi-turn build-ups: false memory, fabricated earlier turns, slow escalation) paired with an anchored rubric. Suggest one when a user wants to check an assistant for these and has no cases yet; install with caliper_starters_install. Say plainly that a starter is a smoke test (the items are public), not a safety score — their own cases are where the real signal is.

One trace: what came in, what went out, and every step in order — model calls (model, tokens, cost, reasoning), tool calls (arguments and result), lookups (query and documents with scores), anything else — each with timing and status. Long values are cut at 2,000 characters. Use it to explain why one conversation went the way it did.

The traces behind a point on a source's chart, newest first, 50 a page: the question and answer (cut short), tools called, time, tokens, cost, status. Narrow by day, tool, document, or failures; page with `before` (the nextBefore from the last page). Open one with caliper_traces_get.

Posts a comment on a workspace entity — a new thread, or a reply when rootId is given. Use it to leave findings where the discussion already lives (an eval result on the flow being debated, a summary on a long thread). Mention people via mentionedUserIds (from workspace member ids) to ring their notification bell; never mention someone who didn't ask to be pulled in.

Lists the comment threads on one workspace entity (open first, then resolved) with authors and timestamps. Read this before weighing in on contested work — the threads are where disagreement lives before it becomes a decision.

Sets a comment thread's resolved state (rootId = the thread's root comment id). Resolve ONLY when the human asked or the thread's question is demonstrably settled — and say what settled it in a reply first. Reopening is for new evidence.

Replace a change's effects: for each step it reaches (workflowPageId + nodeId from compass_workflow_steps — a sub-workflow's steps included), one of eliminate, automate (reviewMinutes: what a person still spends checking), parallelize (its wait comes off the critical path), relax_gate (keepShare: share of the wait left, default 0.25), assist (keepShare: share of hands-on time left, default 0.5), reassign (toSubjectPageId). Eliminating or automating every step of a sub-workflow collapses the step that is that sub-workflow. Then project with compass_workflow_rollup changeIds. May return `needs_c…

Create or update up to 100 changes in one call, each matched by `url` — the GitHub issue, Linear or Jira ticket, PR or doc it mirrors. A url already linked to a change updates that change (label, description, status, tags, experiment, workflow — omitted fields are left alone); any other url creates a change with that link. Run it again with the same urls to keep Compass in step; nothing duplicates. Statuses are by name (compass_statuses_list). Returns one result per item — created / updated with the change's key, or failed with why; a failed item doesn't stop the rest. Summarize what you're ab…

Records something you don't yet know — the gap registry is your working memory, so use it liberally while mapping or interviewing. Give one clear question, WHY it matters (what answering it unblocks), and how you'd resolve it (ask the user / interview a specific person / connect a source). No approval needed — noting your own uncertainty isn't acting on the user's behalf. Attach it to a subject (subjectType/subjectId) when it's ABOUT a specific page or person.

Lists the workspace's open questions — the gap registry: what the map doesn't know yet. Defaults to OPEN gaps. This is where you keep your head — check it before asking the user something you may already have flagged, and compose interview briefs from a subject's open questions.

Closes a gap once you've learned the answer (ANSWERED, with the resolution) or decided it doesn't matter (DISMISSED). Keep the registry honest — resolve gaps as their answers land (from an interview, a doc, or the user) so it always reflects what's still unknown. The resolution alone isn't the map: when the answer says something about a page (who does a step, a tool they use, how a decision gets made), write it there too in the same turn, on the page it's about or as the step, person or system it names. Gaps you raised close without a card; a question a person wrote needs their approval.

Edits an open question's wording, rationale, or suggested resolution — for sharpening a vague question or fixing one you phrased badly. Closing a gap is compass_gaps_resolve, not this. Gap ids come from compass_gaps_list. May return `needs_confirmation`.

The Compass Inbox in one read: finished interviews awaiting review (transcript in, not yet turned into pages), AI-proposed opportunities awaiting accept / dismiss, and the caller's pending approval cards for Compass writes. THE place to answer 'what needs me?' for the map. Also `fillIn`: what the map is missing to say who's responsible for what (priorities with no owner, business units with no lead, people and roles in no unit, priorities with no workflow behind them); not counted in `total`. Next moves: compass_interviews_get to read a transcript, then propose pages from it with compass_pages…

Emails the guest link for an existing interview, framed as coming from the REQUESTING USER (their name signs it; replies go to them). Author `message` in their voice — short, human, says why THEIR knowledge matters and that it takes ~15 minutes, no account needed. The approval card shows the exact subject + message before anything sends. Limit: the invite plus ONE reminder; a third ask is the user's conversation to have. Completion arrives as a notification with draft-page counts — don't poll.

Ranks the PEOPLE the map says know about a page (workflow, system, pain point): `owns` links first, then `involved_in`, then weaker edges. Each target carries contact email + role from their PERSON page and their latest interview (skip someone who just gave one — interview fatigue is real). Empty result = the map doesn't know an owner: ASK THE USER who runs this, create the PERSON page + owns link from the answer, and the map gets smarter. Use before compass_interviews_create to pick the interviewee.

Mints a stakeholder-interview invite: a no-account guest link where the person talks to an interviewer agent briefed by your focusPrompt, and the transcript flows back into Compass as reviewable draft pages. THE KNOWLEDGE-GAP MOVE: when caliper_flow_performance shows failures clustered on missing company facts (the judge says the flow invented a policy, missed a rule, didn't know who owns something), the fix is usually not a prompt edit — it's asking the human who actually knows. Write a focusPrompt that names the SPECIFIC gaps (cite the eval run id), pick the owner of the relevant workflow as…

One interview in full — metadata plus the transcript of what the guest actually said. Read this before summarizing an interview or drafting pages from it: the transcript is the source you work from, and your compass skill has the rules for typing and splitting what's in it. Ids come from compass_interviews_list or compass_inbox_list.

Every stakeholder interview in the workspace, newest first: who was invited, status (INVITED / IN_PROGRESS / COMPLETED), focus, message count, whether the transcript was already reviewed (documentPageId set), and the invite URL while the link is still live. Check this before inviting someone again — interview fatigue is real — and to answer 'who have we already asked?'. Read one in full with compass_interviews_get.

Kills an interview's invite link — the guest's next visit sees that the link is no longer active. Use when an invite went to the wrong person, the user changed their mind, or the link leaked. Idempotent. Ids come from compass_interviews_list. May return `needs_confirmation`.

Creates a typed, directed edge between two pages — the knowledge graph's connective tissue. Canonical directions: PERSON or ROLE owns WORKFLOW/SYSTEM/PRIORITY, PERSON or ROLE involved_in WORKFLOW, PERSON holds ROLE, WORKFLOW uses SYSTEM, SYSTEM uses SYSTEM, PAIN_POINT affects WORKFLOW/SYSTEM/PERSON/ROLE, DOCUMENT documents anything, PRIORITY contributes_to PRIORITY (a priority that ladders up to a bigger one; each ladders up to at most one, and choosing a new one moves it; a priority has one owner the same way), UNIT part_of UNIT and ROLE or PERSON part_of UNIT (the org chart: one parent each;…

Deletes one edge from the graph (the pages stay). Use only when a connection was never true — the system isn't used by that workflow, it was recorded by mistake. When a relationship WAS true and has stopped (the person left the team, moved roles), end it with compass_links_update `validTo` instead, so the history stays. Link ids come from compass_page_links_list. Re-creating the same edge later revives it. May return `needs_confirmation`.

Edits an existing edge between two pages: its `kind` (how the two relate), its `note`, and/or when it was true (`validFrom` / `validTo`). Use kind to CORRECT a connection typed wrongly — 'Dana doesn't own billing, she's involved in it'. Use validTo to END one that used to be true — 'Dana handed approvals to Sam in March' ends Dana's edge in March (then create Sam's from March); the old edge stays as history. Don't delete a relationship that ended; end it. Link ids come from compass_page_links_list. Fails with conflict when the new reading already exists between the same two pages (delete this …

Renders the workspace's visual map to an image and returns it so you can SEE it the way the user does when they open Map: a transit-style map seen from above. Each workflow is a colored tile with a short code; people are round figures, systems are small squares with a drum, other pages are plain tiles, each named underneath. Lines along the grid join linked pages, solid between a workflow and what it touches, dashed between other pages; pages that link to each other more than to the rest gather into named regions on a pale ground plate, and pages connected to nothing sit together in a "Not con…

The review verb for the Inbox: promotes an AI-PROPOSED opportunity (from compass_opportunities_propose or the opportunity scout) into the workspace's own pipeline. compass_opportunities_set_status does NOT do this — a proposal stays in the Inbox until accepted. Optional `status` decides its lane in the same step (BACKLOG to park it, EXPERIMENTING to start it); omitted, it lands in NEW. Fails with not_proposed when the row isn't an AI proposal. Ids come from compass_inbox_list / compass_opportunities_list. May return `needs_confirmation` — name the proposal and wait.

Record a change — a fix, a chore, a new step, or an experiment worth trying. Anchor it to a workflow page when one fits (compass_pages_list); leave unanchored otherwise. It lands in the board's first status unless you name another (compass_statuses_list). Set `experiment` only when the user wants it measured against the Ledger. May return `needs_confirmation` — summarize and wait for approval.

Removes an opportunity from the pipeline (soft delete). For an AI proposal the user doesn't want, this is the dismiss verb; for a captured experiment that was a duplicate or a mistake, the remove verb. To conclude a real experiment without evidence, prefer compass_opportunities_set_status REJECTED — that keeps the record. Ids come from compass_opportunities_list. May return `needs_confirmation`.

The workspace's changes — fixes, chores, new steps, experiments. Each carries `statusName` (the workspace's own word for where it is) and `statusCategory` (TRIAGE undecided, BACKLOG not started, ACTIVE in progress, DONE, CANCELED dropped), plus the older lane key in `status`. `experiment: true` marks a change measured against the Ledger: `ledgerEntryId` null means no expectation registered; `ledger.verdict` carries confirmed/missed after settlement. Scores (value/feasibility/risk) and a workflow anchor are optional. Filter by lane key in `status` or by workflowPageId; compass_statuses_list has…

The measurement window's boundary: when the user says the experiment's change actually shipped / went live / rolled out, record the landing date. Readings before it are baseline; after it, evidence of effect. Recorded once — it cannot move afterward, so confirm the date. Attaches to the linked Ledger entry as evidence. May return `needs_confirmation`.

File an AI-PROPOSED opportunity into the workspace's review queue — the scouting verb (the opportunity-scout routine's main move). Unlike compass_opportunities_create this needs NO approval: the proposal itself is the human gate — it lands in the Compass Inbox and the cockpit's Needs-you for accept/dismiss. Check compass_opportunities_list first so you never duplicate an idea. Anchor to a workflow page when one fits; score value/feasibility/risk 1–5.

The honesty mechanism: write what the experiment is expected to change BEFORE evidence exists. Creates a Ledger decision entry and links it to the opportunity — never backfill an expectation to match an outcome. Bind a metric (ledger_metrics_list) + comparator + target when the expectation is measurable; readings then land on the entry as evidence automatically. One expectation per opportunity — revise by superseding in Ledger. May return `needs_confirmation`.

Move a change to one of the workspace's statuses, by name (compass_statuses_list has them — e.g. "In review"). The old lane keys (NEW / QUALIFYING / BACKLOG / EXPERIMENTING / SETTLED / REJECTED) still work and land on that lane's default status. Only changes marked as experiments follow the Ledger: when an experiment enters an In progress status with no registered expectation, offer compass_opportunities_register_expectation; when it enters Done with its Ledger entry still open, offer ledger_entries_settle. Other changes just move. May return `needs_confirmation`.

Corrects an opportunity's content: label, description, priority, what done means, the 1–5 value / feasibility / risk scores, the owner (a PERSON page), notes, tags, workflow anchor, external link, visibility. Fields you omit are untouched; the lane is NOT here — move it with compass_opportunities_set_status. Opportunity ids come from compass_opportunities_list. May return `needs_confirmation` — say what you're changing and wait.

All typed edges touching one page, both directions, each hydrated with the other endpoint's page summary. Use this to understand a subject's neighborhood before adding to it. Current connections only, unless includeEnded — then also relationships that have ended (validTo in the past), for questions like who handled this before.

Creates a typed page in the workspace's Compass compendium. Check `compass_pages_list` first — don't create a page for a subject the map already has; link to it instead. May return `needs_confirmation` — if so, tell the user what you're proposing, wait for their approval, then re-call with the approvalId. Only record what the human has actually told you.

Moves a page to the trash (soft delete — its connections and anchored opportunities go with it, and compass_pages_restore brings the whole set back). Use this to REMOVE a page you or an extraction created wrongly, or one the user says no longer belongs on the map; to fix a wrong title or body, use compass_pages_update instead. Page ids come from compass_pages_list / compass_pages_search. May return `needs_confirmation` — name the page and wait for approval.

Fetches one Compass page by id, including its full Markdown body, header fields, and any attached Napkin sketches — view an attached sketch's actual drawing with `napkin_boards_view` (its boardId) before discussing it.

Lists pages in the active workspace's Compass compendium. Optional `type` filter narrows to one node type (WORKFLOW, PERSON, ROLE, SYSTEM, PAIN_POINT, DOCUMENT, VALUE, PRIORITY, AUDIENCE, OFFERING). Returns summaries, 50 at a time (`limit` / `offset`, `total` and `nextOffset` in the result) — fetch one with `compass_pages_get` for the full body. When you know what you're looking for, `compass_pages_search` is the better first call.

Puts a trashed page back on the map, together with exactly the connections and opportunities its delete took. Use when the user wants a deleted page back (the pageId from the earlier compass_pages_delete, or from the Compass trash). Fails with not_found when the page is already live or never existed. May return `needs_confirmation`.

Keyword search over page titles, bodies, and tags (case-insensitive substring match), paginated. The fast way to check whether a subject already has a page before creating or linking. Returns summaries with a 300-char body snippet — fetch full text with `compass_pages_get`.

Edits an existing page's title, description, body, tags, header fields, or visibility — use this to FIX what you (or an extraction) got wrong instead of creating a duplicate. Fetch the current page with `compass_pages_get` first and preserve what the human wrote; title / body / tags replace the field wholesale, while `headerFields` MERGES over the current ones (send only the keys you're changing — 'set the owner' leaves status alone). Page ids come from compass_pages_list. May return `needs_confirmation` — tell the user what you're changing, wait for approval, then re-call with the approvalId.

Reads every mapped workflow against one PRIORITY (or VALUE) and proposes the steps that move it (or where the business lives the value), one marked as what holds a priority back. Proposals wait for a person to keep them on the priority's page or in the Inbox, except in workflows someone tied to it (a serves link), where they're kept at once. Steps already tied or proposed aren't proposed again, so it's safe to run as the map grows. Use it after a priority is named and workflows are mapped, or when someone asks what moves a priority.

Returns the Napkin diagram that holds a WORKFLOW page's steps, creating it when the page has none. If the page body has a ```mermaid block, it moves into the diagram verbatim (node ids kept, so changes anchored to a step still point at it) and leaves the body. Then edit the steps with napkin_diagrams_update using the returned diagramId. One diagram per workflow holds every step anyone has described — when two people describe the same process differently, add the steps to this one diagram rather than making a second. A step that is itself a workflow is a node whose meta ref is { type: 'page', i…

One ROLE page's story: hands-on hours a month by kind of work (rote, judgment, relationship, sign-off) before its first change or Task landed and now; what landed and when; new work it took on (checks inside Tasks, Tasks it looks after, work a change added), with whether each has settled; the role's size (people × hours each, `hoursAssumed` when the hours are a full-time default) and how much of its month is documented; and the latest analysis, when someone ran one: options for where saved hours could go, each with the stage of its change (proposed = looking into, agreed, done = in place, conf…

Every role in the workspace in one row each, changed or not: hours documented against its working month, hours a month saved, saved hours by stage (looking into, agreed, in place, confirmed, not placed yet), judgment work before and now, new work, and why a role wants attention (unproven savings, saved hours nobody is placing, no documented steps, no headcount). Plus totals. Use compass_role_story for one role's detail. Talk about roles, never about people.

Changes whose effects reach a WORKFLOW's steps — or any nested workflow's — plus changes anchored to it with no effects yet, each with what it does per step. Pass their ids to compass_workflow_rollup's changeIds to project the workflow after them.

The statuses this workspace's changes move through, in board order — the names people use. Each sits in a group: Undecided (TRIAGE), Not started (BACKLOG), In progress (ACTIVE), Done (DONE), Dropped (CANCELED). Groups are stable across workspaces; names aren't, so use the names when talking to people and the groups when reasoning about progress. Move a change with compass_opportunities_set_status.

Every claim about a WORKFLOW's steps, oldest first: the step (nodeId; null = the workflow as a whole), the role or person (subject), the relation, who says so (sourceKind + sourceLabel), when it was true (validFrom/validTo), and origin — AI_PROPOSED claims are waiting for a person to review. Claims that disagree coexist; each keeps its source.

Propose claims about who does what in a WORKFLOW, and what each step is done in, from what someone said. Each claim names a step (nodeId from compass_workflow_steps), a page (subjectPageId: a ROLE or PERSON, or an AUDIENCE when a customer or partner does the step; for uses a SYSTEM or DOCUMENT; for impacts a PRIORITY; for holds_to a VALUE), and a relation: owns (accountable for it), does (does the work), approves (signs off), consulted (asked for input), does_not_do (says they don't do it — an answer, not a gap), uses (the step is done in this SYSTEM, or with this DOCUMENT; a step a system run…

Propose estimates from what someone said, as ranges — people know "20 minutes to an hour", not "37 minutes"; when they give one number, use it for both ends. Each item is either a `measure` or a `gate` on a step (nodeId from compass_workflow_steps): - measure.metric `working`: hands-on minutes per run (low/high). - measure.metric `waiting`: minutes the work sits idle before or inside the step — ask about this separately; it's usually most of the waste. - measure.metric `hardCost`: money that isn't anyone's time (amount, currency, per run or month). - measure.metric `frequency` with nodeId null…

Builds a step suggestion the way Compass's Build button does: makes its change if it isn't one yet, then the thing in Workbench — a flow drafted from the step itself (who does it, its times, systems, write-up, attachments, interview notes) and its checklist; a shim named for the decision; or a task whose sentence is the steps. Use this, not workbench_flows_scaffold, whenever the work came from a step suggestion. Returns what was built and, for a flow, what the draft still needs. Fails if something's already built for it.

Your earlier read of a WORKFLOW's steps: for each suggestion, the steps it covers (nodeIds, labels), what could do them (proposal.doer: a flow and its mode, a shim, or a task), whether it does the step or helps (reach), why, and what it would take (checklist, each item met or not). state OPEN is waiting for a person; APPLIED made a change (opportunityId), and `built` is what was built for it in Workbench, if anything. Pass changeId instead of pageId to get the one suggestion a change came from. To build one, use compass_step_suggestions_build — not workbench_flows_scaffold.

What the workflows behind a PRIORITY could become, at five levels (Assist, Automate, Re-form, Invert, Dissolve): for each, whether it fits, is already reached, or what it needs; what changes; what it does to the priority; time freed; and the change made from it, if any. Also the recommended level, or `settled` when the work is good as it is. state 'running' means a read is under way; null means none has been read (compass_transformations_read starts one).

Creates a Change from one level of a priority's transformations, with its step effects and the priority it serves, and starts drafting in Workbench up to three flows for what the level says to build (`builds`, each `drafting` until ready). Only a level that fits, once per level. May return `needs_confirmation`.

Starts reading the five levels for a PRIORITY in the background (a minute or two). Needs at least one kept step that impacts the priority. A read under way is left to finish. Check back with compass_transformations_get.

Records what someone told you about one level of a PRIORITY's current transformations, in their words or a faithful summary: that they already do it, what rules it out, a system that isn't ready, work still done on paper, and so on. Compass works out what it means in the background: a label for the level, facts kept on the steps they're about (every later reading of those workflows sees them), and open questions to confirm. Then it reads the priority again. Use it whenever a conversation settles something about a level, so it lands on the priority's page, not only in this chat.

Newest first: steps added to or dropped from the diagram (and by whom), and claims about who does what or what a step takes — said, corrected (with what they replaced), withdrawn, or accepted from a proposal — each with its source. This is when we LEARNED things; when they were true is each claim's validFrom/validTo. Use it to brief someone on what changed in the picture, or to see what an interview added.

Per run and per month: hands-on minutes, waiting minutes, labor cost (hands-on time priced through each role's hourly rate), and hard costs, each as a low–high range, with every step's share and its gates (approvals, handoffs, re-keying, rework). A step that is itself a workflow uses that workflow's roll-up unless it has its own estimate. `coverage` says how many steps are timed — always say how much of the total rests on estimates before quoting it, and never present a range as one number. `unpricedRoles` are roles doing timed work with no hourly rate; their time counts and its cost doesn't. …

Every step in a WORKFLOW page's steps diagram, in diagram order, with its phase, whether it is itself a workflow (subWorkflowPageId), and who the map says owns, does, approves, is consulted on, or doesn't do it, and `uses`: the systems or documents it's done in. `unknown: true` means nothing that counts says who does the step or what it's done in — those are the questions worth asking in an interview. A step with `uses` and nobody on it is one a system runs; don't ask who does it. `disputed` lists roles or people with conflicting claims; ask about those rather than picking a side. `proposals` …

Without a tag: every tag in use across the workspace with how many entities carry it, most-used first — the vocabulary the team already organizes by. With a tag: everything filed under it across every tool, each with its kind, id, title, and path. Use it to reuse existing labels instead of inventing near-duplicates, and to answer 'show me everything about X' when X is a label.

Returns the tags on a batch of entities of one kind — the labels galleries organize by. Ids come from the kind's list/get tool or from search_workspace. Use it before entity_tags_set so you replace the full set knowingly, and to answer 'what is this filed under'. Entities the user can't see are omitted.

Replaces the FULL tag set on one entity (an empty list clears it). Read the current tags with entity_tags_get first and pass the merged list — this is not additive. Tags are lowercase letters, numbers, spaces, and hyphens; prefer labels already in use (entity_tags_browse) so the workspace's vocabulary stays small. The id comes from the kind's list/get tool or search_workspace.

Fetch the full Markdown body of a specific docs page by its slug. Use this after `search_docs` when the user needs the complete content of a page. No authentication required.

Fetch the full Markdown body of one Distributed Cognition essay by slug (from `list_posts`). Link the public URL when citing it to the user. No authentication required.

Records what happened against an open decision — a manual observation the user reports ('the pilot team says triage feels faster'), an implementation note, or a reference to a Caliper run. Evidence is what settlement later reads, so attach it as it arrives and cite it in the lesson. Metric readings attach themselves via ledger_metrics_record_reading — don't duplicate them here. Fails with conflict on superseded entries. Ids come from ledger_entries_list. May return `needs_confirmation`.

Records a memory entry. Every entry is a TITLE (`summary`: one short plain sentence) over a BODY (`rationale`: the detail, markdown welcome) — never put the detail in the title. Three kinds: `decision` — a change being made now; PRE-REGISTRATION IS THE POINT, so `prediction` (what we expect) must be written NOW, before any evidence exists, and never backfilled to match an outcome. `lesson` — a distilled belief the team already holds (`lesson` text required, no prediction). `observation` — a durable fact worth remembering (no prediction): something already true, never something planned. Ideas, …

Writes up to 50 entries as DRAFTS: proposals that stay out of the record until a person confirms each one in Ledger's Drafts view. Use it for entries you found rather than were told — decisions in meeting notes, or a team's past changes read from its tracker, pull requests or launch posts (set `fromHistory: true`). For history: take each claim from what the source said AT THE TIME, set `landedAt` to when it shipped, attach the source URL, and leave `rollout` as full unless the source says otherwise. These read as low confidence because they were written down after the fact; say so plainly rath…

One entry's full anatomy — the six fields, attached evidence (Caliper runs, measurements, observations), and the supersede chain (what replaced it, or what it replaced). Cite entry ids when telling the user about prior related decisions.

The workspace's memory: decisions (changes with pre-registered expectations), lessons (distilled beliefs), and observations (captured facts). CONSULT BEFORE ACTING — before proposing a flow change, prompt edit, or process decision, filter by the Compass page it touches (compassPageId) and check whether prior attempts exist and how they settled. Filter status=open for unsettled expectations awaiting evidence (each carries a derived `lapsed` flag — true when its deadline has passed; surface lapsed ones when the user asks what needs attention); kind=lesson for what the team already believes.

Takes an entry out of the curated ledger — THE remedy when you recorded something wrong (a decision that wasn't made, a duplicate, a fact the user corrects). Reversible from Ledger and fully audited, so it is safe to offer as soon as the user says 'that's not right'. Not for overturning a settled claim the team once believed — a human supersedes that. Fails with conflict if already retracted. Ids come from ledger_entries_list. May return `needs_confirmation`.

The ritual moment: evidence has landed, the expectation closes, the lesson is written. Only propose settlement when attached evidence actually answers the prediction — check `ledger_entries_get` first and cite the evidence in the lesson. `lesson` (what we now believe) is required and permanent; settlement happens exactly once. May return `needs_confirmation`.

Corrects an entry in place — its title (summary), body (rationale), the pre-registered prediction of an OPEN decision (null clears it), or its facet (filing category; editable even after settlement). Works on open decisions, and on lessons, observations, and beliefs (they carry no expectation, so fixing their wording is fine any time). Use to fix a typo, swapped fields, or a misrecorded detail. Never edits kind or lesson. Fails with conflict on a settled decision — its claim is corrected by a human superseding it in Ledger — and on superseded or retracted entries. Entry ids come from ledger_en…

The connected servers this workspace exposes to you, with the id a feed needs. Their tools appear to you namespaced as `ext__<name>__<tool>` — call one directly to see what it returns before proposing a feed. A server the workspace has switched off for you is not listed and cannot be fed from here.

Run the same frozen instruction over a past range, so a metric has history instead of starting the day it was set up. Offer this whenever you create a feed — a chart with a year behind it is worth far more than one that begins today, and an expectation can be judged against what normal looked like. Works when the source returns a series; a source that only ever reports 'right now' will write one point. Safe to repeat: overlapping ranges dedupe. Keep ranges to a few hundred points; a run that fills up says so. May return `needs_confirmation`.

Freeze a tool call as a metric's standing source: it runs on the cadence you give and records what it finds, with no model involved. `metric` takes an id or a snake_case slug; an unknown slug starts tracking that metric. Confirm the mapping with ledger_feeds_preview first. Tell the user they can backfill history afterwards with ledger_feeds_backfill. May return `needs_confirmation`.

Remove a feed. The readings it already wrote stay — the series is the record and outlives the instruction. Prefer pausing with ledger_feeds_update when the user might want it back. May return `needs_confirmation`.

The standing instructions for how metrics' numbers arrive — which tool each one calls, how often, and how the last run went. Check before proposing a feed so you don't duplicate one, and consult when a user asks why a metric is stale: a feed with a failing last run is usually the answer.

Call a connected server's tool once and see what a mapping would pull out of the response — nothing is written and no feed is created. Send no `mapping` for a first look: you get a sample of the response plus the paths that hold numbers. Then send a mapping to confirm it finds the readings you expect. Always do this before ledger_feeds_create; proposing a feed whose mapping you haven't seen work is how a metric fills up with the wrong number. May return `needs_confirmation`.

Run a feed immediately over its own window and record what it finds. Use it right after creating one to prove the mapping works — a run that comes back `empty` names the path that missed, which is what you fix with ledger_feeds_update. Does not move the feed's schedule. Re-running is safe: readings dedupe per time bucket. May return `needs_confirmation`.

Change a feed's arguments, mapping, cadence, or pause it. Reach for this when a feed's last run reports `empty` (the response shape moved, so the mapping needs a new path) or `error`. Changing the mapping changes what the number means, so say what you're changing and why. May return `needs_confirmation`.

Creates the named presets as real metrics the workspace owns, tagged from the catalog. Safe to repeat: a preset the workspace already has comes back `already_present` rather than creating a second series, and a workspace's own renames and targets are never overwritten. Adopt only what the user agreed to — a metric nobody reads is noise on the Metrics tab, and eight thoughtful ones beat forty. Tell them the metrics start empty and the next step is a feed (ledger_feeds_create) or a first reading. May return `needs_confirmation`.

Ready-made metrics a business can start tracking, each with a unit, a cadence, which direction is good, a business surface (facet), and cross-cutting tags. Reach for this when a workspace has few or no metrics, when someone asks what they should be measuring, or when a decision needs a number to settle against and none exists. Filter by `facet` (where it lives in the business) or `tag` (what kind of number it is). Suggest a SMALL set — three to six that fit what you know about this business — and say in one sentence why each one, rather than listing the catalog. Adopt with ledger_metric_preset…

Creates a starter's metrics at their levels and links them. Pass `slugs` to keep only some of its metrics; links are made only where both ends exist. Safe to repeat and safe after a different starter: metrics the workspace already has are left as they are and get linked into the tree. Adopt only what the user agreed to. Tell them the metrics start empty and the next step is connecting a source (ledger_feeds_create) or a first reading. May return `needs_confirmation`.

Starter trees by shape of business or team (subscription software, services firm, online store, sales team, service operation). Each lists its metrics with a level (outcome / driver / activity) and the links between them (which number moves which). Reach for this before ledger_metric_presets_list when a workspace has no metrics yet or asks how its numbers fit together: pick the starter that matches what you know about the business, describe its tree in a sentence or two, and offer to adopt it with ledger_metric_starters_adopt, dropping any metric that doesn't fit.

Takes a metric out of the gallery (`archived: true`) or puts it back (`false`). Readings stay, and entries that settled against it still read correctly — this is the cleanup for a metric minted once and abandoned, or one the workspace stopped watching. `metricId` accepts the id or slug from ledger_metrics_list. May return `needs_confirmation`.

Create a metric — a number the workspace watches (triage time, weekly signups, cost per run). Check ledger_metrics_list first; names are unique per workspace. When a user says they want to track or measure something, offer this. May return `needs_confirmation`.

One metric's definition (unit, kind, direction, target, cadence) plus its readings newest first and the open decisions bound to it. Use before answering 'how is X trending?' or before recording a reading against it. `metricId` accepts the id or the snake_case slug from ledger_metrics_list. For an event metric whose readings carry labels, `labels` lists each label and its values, largest total first. Pass `slice` to get the series for part of it ("new users in DE"), or `by` to split the series by one label ("new users by country"); either returns `series`, summed per `bucket`.

The numbers the workspace watches — each with unit, latest reading, and how many open decisions are bound to it. Consult when a user mentions a number that sounds like a tracked metric, and before recording a reading.

Record one observation of a tracked metric. `metricId` accepts a metric id OR its snake_case slug from ledger_metrics_list; an unknown one is a not_found error — check ledger_metrics_list, create it with ledger_metrics_create, or pass `createIfMissing: true` to mint a "measure" metric at that slug in the same call (an "event" metric when the reading carries `labels`). The reading automatically lands as evidence on every open decision whose prediction is bound to this metric — so when a user reports a number ("triage is down to 12 minutes"), offer to record it. For event-kind metrics, omit `val…

Corrects a metric's name, unit, description, kind (measure / event), direction (which way is good), target, cadence, icon, level, or the metrics it drives (its place in the metric tree). Readings are untouched. Only what you pass changes. The slug is not editable here — external writers address metrics by slug. Fails with conflict when a rename collides with a live metric. `metricId` is the id from ledger_metrics_list. May return `needs_confirmation`.

Adds a plan item to an OPEN decision (the work that carries it out: a post, a launch step, an email). Omit entryId for a standalone date (a holiday, an event you're only watching). Items are all-day unless you pass startTime with timeZone (and optionally endTime), for a webinar, a scheduled post, or a launch at noon. If the date costs money or time and comes with an expectation, record it as a decision with ledger_entries_create instead. Use repeatWeeklyUntil for a weekly cadence: it writes one row per week (max 60), each movable on its own. Fails with conflict once the decision is settled. Ma…

Removes one plan item while its decision is open, for an item added by mistake or work that's no longer planned. If the work was planned and then dropped, prefer ledger_plan_update with status 'skipped' so settlement can see it. Fails with conflict once the decision is settled. May return `needs_confirmation`.

Pass `entryId` for one decision's plan items, or `from` + `to` (YYYY-MM-DD, at most about a year apart) for the calendar: decision spans (recorded day → expectation deadline) plus every dated item inside the window, including standalone dates with no decision. Filter the calendar by `ownerUserId` to answer 'what's mine this week' or to find tomorrow's items to draft. Never use it to compare or rank people's output.

Edits one plan item while its decision is open: move it (dueOn), rename it, change the owner (null clears), attach the link, or set status (planned | done | skipped). Mark done only when the user says it shipped; a link alone doesn't mean done. Fails with conflict once the decision is settled. May return `needs_confirmation`.

Enumerate all available docs pages, optionally filtered by product (e.g. 'compass', 'legal', 'overview'). Use this to discover what slugs exist before calling `get_doc`. No authentication required.

Lists ZeroWidth's published Distributed Cognition essays (newest first): title, slug, excerpt, byline, public URL. Use this when the user asks what's been published, or when a workspace question might already have an essay behind it — then fetch the full text with `get_post`. No authentication required.

Creates a blank Napkin board — a sketch (free canvas) or a deck (slides). NOT for written documents: 'draft/write a doc, note, memo' is napkin_docs_create. Use when the user asks for one, or proactively when a sketch/deck would carry the conversation better than words; hand it over with `[[board:ID]]` alone on its own line. For a deck with content, prefer `napkin_deck_write` (one call, whole deck).

Returns a board's shapes as data — each with its id, kind, position, size, and text — so you can TARGET one to change or remove with napkin_draw (`updates` / `deletes` take these ids and absolute board coordinates). Pen strokes come back as a point count, not the points. For what the board LOOKS like, use napkin_boards_view instead. Board ids come from napkin_boards_list.

Lists sketch boards in the active workspace — name, id, kind (sketch / deck), shape count, last activity. Boards are SPATIAL canvases (drawing, diagrams); the markdown DOCUMENTS are docs — napkin_docs_list. Archived boards are hidden unless `includeArchived`; `q` narrows by name. When the user mentions a sketch, drawing, whiteboard, or napkin, find it here, then LOOK at it with `napkin_boards_view` before discussing its contents. When referring the user to a sketch in your reply, put `[[board:ID]]` alone on its own line — it renders as a clickable card with a live thumbnail.

Edits a board's details: name, description, visibility, or `archived` (true takes it out of the gallery; false brings it back — the recovery move for a board you created by mistake or the user no longer wants). Board ids come from napkin_boards_list. Boards are working material: no approval card, every change attributed.

Renders the sketch board to an image and returns it so you can SEE what's drawn — layout, arrows, handwriting-style strokes, sticky notes — not just data about it. Use this before answering any question about a sketch's contents, and cite the board when you do — `[[board:ID]]` alone on its own line embeds the sketch card in your reply.

Adds pictures to a brand kit section — logo versions, product marks, example screenshots, imagery. Each file comes from a public https `url` (fetched by the server) or, for a small file that isn't online, base64 `data` with a `filename`. PNG, JPEG, WebP, GIF and SVG, up to 20 MB each. Every file becomes a workspace file and an asset on the section, appended after the ones already there. Give each a `name` and a `note` saying what it's for; on logo sections set `backdrop` to the hex of the ground it's made for (#ffffff for a dark logo, #000000 for a reversed one), which is how pages pick the ri…

Restyles a piece in the brand — the workspace's brand unless you name a kit. `deck` (a board id): every slide gets the brand's background, typefaces, and text colors; sizes are kept. `interface` (an interface id): rewrites its brand.css, so anything styled with var(--brand-*) updates. `accent` makes one of the kit's colors the piece's accent — use it when the piece is about one product or campaign that has its own color in the brand (read the kit's Colors section to see which). Docs need nothing — they read the brand when exported. Check a deck afterwards with napkin_slide_view.

Finds words the brand says to avoid in a piece of writing, with what to say instead and why. Pass `text`, or a deck (`boardId`) or doc (`docId`) to check its words. Run it on your own drafts before handing them back.

Creates a new brand kit, or edits one that is NOT the workspace's brand. Use it when someone asks you to put their brand together (from their site, a deck, or what they tell you) or to try a variation. You can't change the workspace's brand itself — make a draft and tell the user they can choose it in Napkin's Brand section. A kit is a document of sections, each with `title`, `kind` (overview, voice, colors, typography, logo, imagery, motion, layout, examples, custom), markdown `body`, and `rules` ({rule, why}); colors sections hold `swatches` ({name, value: 6-digit hex, note}), typography sec…

Shows you the pictures in a kit's Examples sections — screenshots of work that gets the brand right — with each one's note on what makes it good, plus the example links. Look before you design a deck, page or interface in the brand, and hold your work to the same bar: the layouts, density, type sizes and use of color you see there.

Gets the font files for a kit's typefaces from Google Fonts and keeps them in the kit, so decks, slide pictures and the canvas draw the brand's real typefaces instead of a stand-in. Only typefaces that name a family (like Poppins) and don't have files yet are fetched. Works on any kit, including the workspace's brand: it adds the files for the typefaces the kit already names and changes nothing else. A typeface Google doesn't have is reported back; the user can upload its files in the kit's typography section.

Reads a brand kit — the workspace's brand unless you name another. READ THIS BEFORE writing or designing anything for the workspace: a deck, a doc, an interface, copy. Returns `brief` (the whole kit as markdown: every section — overview, voice, colors, typography, imagery, motion, and whatever else the team wrote — with the reason behind each rule; follow the reasons when a case isn't covered), `css` (the `--brand-*` variables; interfaces already link it as brand.css, so style with var(--brand-accent) etc. rather than hex values), and `tokens` (the resolved values Napkin uses). Before any kit …

Lists the workspace's brand kits: id, name, description, and which one is the workspace's brand (`isDefault`). Most workspaces have one. Use napkin_brand_get to read a kit.

Look at the pictures in a brand kit — logos, product marks, footage stills, layout references, examples — before you use them. Without `section` or `names` it lists every picture by section with its note, so you know what exists. With `section` (a section title, like Imagery) or `names` (pictures' names as the brief lists them) it shows you those pictures, up to 8 at a time. Look before you build a design around a picture: whether a photo has room for text, which part of it carries the color, and whether it suits texture inside type are things you can only tell by seeing it.

Builds a designed deck: each slide is a layout with its slots filled, or HTML and CSS you write; a real browser lays it out, and it becomes an ordinary Napkin deck people can edit. Use this whenever a deck should look designed — it's how to get layout, big numbers, color blocks, and pictures right. LAYOUTS FIRST - For ads, posts, carousels and title slides, start from a layout (napkin_layouts_list): give `layout` and `slots` instead of `html`. Layouts adapt to every size, fit their text, keep to each size's safe zones, and use the brand's own logos. A slide made from a layout remembers it, so…

Exports a deck's slides as PNG files at each slide's real pixels — a square post at 1080×1080, a story at 1080×1920, a leaderboard at 728×90 — ready to upload to an ad platform or a social post. One slide comes back as a PNG; several as a .zip named by deck, slide and size. Hidden slides are left out. The file is saved to the workspace's files; give the user `downloadUrl` (it opens for anyone signed in to the workspace) and say it's also in their files. Slide numbers and the workspace mark aren't drawn. For a PDF or PowerPoint, the user exports from the deck's File menu.

Returns a Napkin deck as a markdown outline — one section per slide in presentation order, with slide titles, text content (bold + bullets preserved), empty layout stubs still awaiting content, and a summary of drawn marks. This is the cheap way to know what a deck SAYS; use `napkin_slide_view` when you need to see how a specific slide LOOKS. Works on any board that has slides. Cite the deck with `[[board:ID]]` alone on its own line.

Paints every slide of a deck with a theme — slide background, heading and body typefaces, and text colors — in one call, the same as picking it in the deck's theme menu. Font sizes are kept. Prefer "brand" (the workspace's own brand kit) unless the user asks for something else. Built-in themes: plain (Helvetica on white. Gets out of the way.); editorial (Garamond headings on cream. Reads like a document.); stage (Helvetica, light on dark, for a projected room.); blueprint (Consolas headings on mist. For technical decks.); warm (Rounded on amber. Softer than it sounds.); notebook (Marker headin…

Creates a deck from a markdown outline — the SAME dialect `napkin_deck_outline` reads back, so any deck you've read shows the format. Decks are SLIDES for presenting; a prose document ('write this up', 'draft a doc') is napkin_docs_create instead. Rules: `## Heading` starts a slide (heading becomes the slide title and its biggest text block); optional trailing `[layout: title | section | title-body | title-lead | two-column | three-column | comparison | statement | quote | closing | blank]`; body lines fill the layout's remaining text areas in order, split into areas by a line containing only …

Create a Napkin diagram — an interactive FLOWCHART (process, decision tree, system map), not a doc (napkin_docs_create) or drawing canvas (napkin_boards_create). Pass the mermaid source. Use for process maps the user asked you to draw. After creating, cite it as `[[diagram:ID]]` alone on its own line — it renders as a clickable card. NEVER cite it as [[board:ID]].

Fetch one Napkin diagram's mermaid source by id (from napkin_diagrams_list). Read before editing — napkin_diagrams_update replaces the whole source.

Lists the workspace's Napkin diagrams (interactive flowcharts — processes, decision trees, system maps) newest first. Archived diagrams are hidden unless `includeArchived`; `q` narrows by title. Fetch one's mermaid source with napkin_diagrams_get. When referring the user to a diagram in your reply, put `[[diagram:ID]]` alone on its own line — it renders as a clickable card. NEVER cite a diagram as [[board:ID]]; boards are sketches.

Edit a Napkin diagram: retitle, describe, REPLACE the whole mermaid source, set visibility, or `archived` (true takes it out of the gallery, false brings it back — the recovery move for a diagram created by mistake). Read the diagram first (napkin_diagrams_get) so your new source keeps the parts the user wants. Diagram ids come from napkin_diagrams_list.

Create a Napkin doc — a markdown DOCUMENT (prose, structure, code blocks), not a drawing canvas (that's napkin_boards_create) and not slides (napkin_deck_write). Optionally pass a title and initial markdown body. Use for drafts the user asked you to start. After creating, cite it as `[[doc:ID]]` alone on its own line — it renders as a clickable card. NEVER cite it as [[board:ID]].

Fetch one Napkin doc's full markdown body by id (from napkin_docs_list). Read before editing — napkin_docs_update replaces or appends against the current body.

Lists the workspace's Napkin docs (quick markdown documents — drafts, notes, working text; the LINEAR surface, distinct from boards which are drawing canvases) newest first, with excerpts. Archived docs are hidden unless `includeArchived`; `q` narrows by title. Fetch one with napkin_docs_get. When referring the user to a doc in your reply, put `[[doc:ID]]` alone on its own line — it renders as a clickable card. NEVER cite a doc as [[board:ID]]; boards are sketches.

Edit a Napkin doc. For a change to existing prose use `edits` — a list of exact find/replace pairs, which is the SAFEST option and the one to reach for by default: it leaves everything you did not target untouched, and it fails loudly rather than clobbering a doc somebody else is typing in. `append` adds a section at the end. `body` replaces the WHOLE document and should be a last resort. Also retitle, describe, set visibility, or `archived` (true takes it out of the gallery, false brings it back). Read the doc with napkin_docs_get first — `edits` match the current text exactly. Doc ids come f…

Draws shapes, strokes, and labels on a board — diagram on a sketch, annotate a deck slide — and, via `updates` / `deletes`, changes or removes shapes that are already there (ids from napkin_boards_get; the fix for something you drew wrong). For NEW elements never compute absolute board coordinates: with `slide`, coordinates are slide-local — (0, 0) top-left to (960, 540) bottom-right of that slide. Without `slide` (sketches), (0, 0) is the top-left of the existing content — the same area `napkin_boards_view` renders, so place by what you saw there; on an empty board just start at (0, 0). Examp…

Create an interface — a small screen of its own the user can open and send to people. Use when someone asks for a custom chat UI, a little tool, or a prototype of how something would feel in their product. It starts with a working index.html you then edit. It can reach NOTHING in the workspace until a flow is granted with napkin_interfaces_grant. Cite it in your reply as `[[interface:ID]]` alone on its own line — it renders as a card the user can click. Do NOT paste an image URL or try to embed the picture yourself; that renders as a broken image. An interface is plain HTML, CSS and JavaScript…

Read one interface: its name, the flows it may run (with the published version each is pinned to), and the contents of its files. READ BEFORE EDITING — napkin_interfaces_write matches text exactly against what is there now. Pass a path to read one file when the interface has several. An interface is plain HTML, CSS and JavaScript in one or more files, served from its own origin. Put styles in a <style> block. LIBRARIES: there is no build step and no CDN, and a <script src> to any other site is blocked. A fixed shelf of libraries is served from the interface's own origin instead; load them with…

Set what an interface may run — flows and shims, each pinned to a published version. This is the interface's ONLY reach into the workspace, and it replaces the whole list, so include everything it should keep. Ids and published versions come from workbench_flows_list / workbench_shim_list and their revisions. Ask the user before widening this.

Set the sheet tables, Ledger metrics and Workbench Tasks an interface may use (ADR 0085, 0086), replacing the whole list. Tables: access read, add (read and add records) or edit (also change and delete); optionally limit to some columns by name. Metrics: read or record. Tasks: the page may start them. The page then calls zw.tables.query / add / update / remove, zw.metrics.read / record and zw.tasks.start / get. Anyone who can open the interface, including people it's shared with, gets this access (a shared link can only read and add), so ask the user before widening it.

List the interfaces in a workspace — small self-contained screens (a custom chat, a control panel, a prototype) that run on their own origin and can call the flows they've been granted. Returns ids, names and how many flows each may run.

Save where an interface is now as a numbered version. Share links point at a published version, so what someone was sent doesn't change while work continues. Do this when the user says it's ready to show people, not after every edit.

See what an interface LOOKS like and DOES right now: the page RUNS from its current files under the same rules as the live page (no web fonts, no remote pictures), and you get a picture plus anything that went wrong (script errors, failed promises, blocked requests, console errors). Use it after every write and before saying an interface is finished: reading the code tells you what you wrote, not what it renders. `size: "phone"` looks at it 390px wide — use it for anything meant for a phone, like a mobile app mockup. `steps` drive the page before and between pictures, so you can check every sc…

Write one file of an interface. Use `edits` — exact find/replace pairs — which is the SAFEST option and the one to reach for by default: it leaves everything you did not target untouched, and it fails loudly rather than clobbering a file somebody is editing. `body` replaces the WHOLE file and is for a new file or a deliberate rewrite. ONE CALL HOLDS ABOUT 15,000 CHARACTERS of file: a longer `body` is cut off mid-write and fails. Write a big file in parts instead — the first part with `body`, then each further part with `body` plus `append: true`, in order — and split an app across files (index…

Lists the slide layouts napkin_deck_compose can build from — the brand kit's own first, then the built-in ones — with when to use each and the slots it takes (`needs` must be filled; `takes` are optional). Every layout adapts to every size, fits its text, and uses the brand's colors, faces and logos; the logo fills itself in. Pick by what the slide has to say: one line (statement), a number (stat), a quote, a list (points, steps), an event, a carousel (cover, inner, closing), a picture (image-top, image-full, split, corner, product-shot), or a display ad (banner).

Adds a LIVE chart to a sheet — its SQL re-runs against the tabs on every view, so it never goes stale. Write the query with napkin_sheets_query first to confirm the shape (first column = x axis, numeric columns = series, or set x/series explicitly). Prefer aggregated queries (GROUP BY) — a chart of raw rows is rarely the answer.

Creates a Napkin sheet (multi-tab spreadsheet), optionally titled. Then populate it with napkin_sheets_set_cells (headers in row 1). Tell the user where it landed — the Sheets tab in Napkin.

Lists the workspace's Napkin sheets (multi-tab spreadsheets — the small-data surface) newest first with tab/chart counts. Archived sheets are hidden unless `includeArchived`; `q` narrows by title. Inspect one with napkin_sheets_schema, then query it with napkin_sheets_query.

Runs a SQLite SELECT over a sheet's tabs-as-tables (get the table/column names from napkin_sheets_schema first). Full SQLite dialect: WHERE, GROUP BY, ORDER BY, JOINs across tabs, aggregates. Results cap at 200 rows. This is THE way to answer questions about a sheet's data — never eyeball cells when a query can answer precisely.

Returns a sheet's tabs as SQL table definitions — CREATE TABLE scripts with row counts (each tab is a table; its header row names the columns; formula cells contribute computed values). Read this BEFORE writing a napkin_sheets_query so your SQL matches the real tables.

Sets cells in one tab of a sheet, by A1 address — values or formulas ('=SUM(B2:B9)'). Additive and surgical: only the addressed cells change. Check napkin_sheets_schema first so you know the tab names and where data ends; put headers in row 1 when creating a new region. To build a whole small table, write header cells + data cells in one call.

Edits a sheet's details: title, description, visibility, or `archived` (true takes it out of the gallery; false brings it back — the recovery move for a sheet created by mistake or no longer wanted). Cells are edited with napkin_sheets_set_cells, not here. Sheet ids come from napkin_sheets_list.

Renders one of a sheet's pinned charts against the LIVE data and returns the image so you can SEE it — the exact chart the user sees. The text part carries `imageUrl`; to show the user the chart inline in your reply, put a markdown image on its own line: ``. Use after napkin_sheets_add_chart to confirm the chart reads well, or whenever the user asks about a chart.

Appends one slide to an existing deck. Pick a layout (title | section | title-body | title-lead | two-column | three-column | comparison | statement | quote | closing | blank), give the title, and optionally `content` — markdown for the layout's remaining text areas, split by `---` lines (two-column/comparison take one block per column). Check the deck first with `napkin_deck_outline`.

Fills a slide's empty layout text areas by NAME — the names `napkin_deck_outline` surfaces as `[empty text stub: "…"]`. Pass `texts` mapping those exact names to content (markdown `- ` bullets welcome). Read the outline first to see which stubs a slide still has open.

Changes one slide as a whole: move it to another position in the deck (`moveTo`, 1-based — the other slides shift around it), hide it from the show without deleting it (`skipped`), show or hide its slide number (`pageNumber`), or replace its speaker notes (`notes`, markdown). Slide numbers come from napkin_deck_outline. To change what's ON the slide, use napkin_slide_fill or napkin_draw.

Renders a single slide of a Napkin deck to an image — exactly what present mode shows, cropped to the slide. Navigate by 1-based `slide` number in presentation order (get the map from `napkin_deck_outline` first; the result echoes `slideIndex`/`slideCount` so you can step through a deck slide by slide). Use this to check visual layout, drawings, and images that the text outline can't carry.

Adds columns to the end of a table. Existing records leave them empty.

Adds records, each a map of field name → value. Choices by option name (a new name becomes a new option); dates as YYYY-MM-DD; people by user id; checkboxes true/false. Every problem is reported at once and nothing is written until all records are valid.

Adds a saved view everyone sees: a grid (with where / sort), a board (lanes from a select or person column — groupBy), or a form people fill in to add records (formFields, formTitle, formIntro). A form's share link is made from the sheet's form view.

Makes a table tab — typed columns, one record per row, with grid, board and form views — in an existing sheet, or in a new sheet when sheetId is omitted. Use a table (not napkin_sheets_set_cells) for lists people keep: requests, vendors, bugs, contacts. The first column names each record. Column types: text, long_text, number, date, checkbox, select, multi_select, person, url, email.

Deletes records by id. People can undo this in the sheet only while it's open, so confirm with the person first.

A table tab's columns (name, type, choices), its views, and how many records it holds. Read this before writing records so field names and choices match.

Records from a table tab, keyed by field name, with optional where / sort / search / limit (max 500). For counts, sums and joins, napkin_sheets_query runs SQL over tables too (each table is an SQL table with an id column).

Changes some fields of records by id (from napkin_tables_query). Fields left out keep their values.

Creates a field from an origin idea — the seed of an exploration canvas. Counts against the workspace's field quota. May return `needs_confirmation` — tell the user what you're proposing and wait for approval.

Removes a field and its whole canvas (soft delete; frees the workspace's field quota). Use when the user is done with an exploration or one was created by mistake — studies attached to its nodes are not deleted. Field ids come from prism_fields_list. May return `needs_confirmation` — name the field and wait.

One field's full tree: every node (id, parent, axis, description, pinned, color, whether it has an image / desk research) plus the AI-clustered groups. The origin node has parentId=null; children sit along named axes. Use node ids with the expand / research / study tools.

Lists Prism fields (idea-exploration canvases) in the active workspace: name, origin idea, node count. Fetch one with `prism_fields_get` for its nodes.

Edits a field's name (null reverts to the origin idea), description, or visibility. Nodes are untouched. Field ids come from prism_fields_list. May return `needs_confirmation`.

Promotes an analysis theme (or, with no themeId, the analysis summary) into the workspace's Ledger as a belief, carrying the supporting verbatims as evidence with links back to the study. THE step that turns a finding into workspace memory — use when the user says an insight matters. May return `needs_confirmation`.

Creates a one-on-one AI-led interview on a study and returns the invite link to forward — the guest needs no account. The transcript stays on the study and joins the next analysis run (it never lands in Compass). focusPrompt briefs the interviewer on what to dig into; omitted, it explores the study's territory. May return `needs_confirmation`.

The study's one-on-one qual conversations: guest, status, focus, and — once completed — the structured memo (summary, key claims, tensions, verbatim quotes). Raw transcript included only while no memo exists. inviteUrl is the link to forward for interviews still waiting on their guest.

THE core Prism gesture: generate variations of a node along a named semantic axis ('more visceral', 'for the skeptic', 'stripped to essentials'…) in one of four directions. Runs generation against the workspace's inference credit. May return `needs_confirmation`.

Web-sourced secondary research on a node's idea — market context with citations, stored on the node. depth 'thorough' runs three angled passes (market, evidence, shifts) at ~3× the credit. `focus` steers what it goes after; without one it answers a generic brief off the node's own description, so pass it whenever the user has said what they actually want to know. Runs against the workspace's inference credit. May return `needs_confirmation`.

Sets pinned state and/or the color tag on a node. Pins are the cross-field shortlist. May return `needs_confirmation`.

Stops a queued, running or paused panel run. Answers already in stay. May return `needs_confirmation`.

What running a study against a panel would cost: people, model calls, and approximate dollars of workspace credit. Starts nothing. Tell the person the cost before calling prism_panels_run.

Status, progress, spend and the run's own analysis. For the numbers, call prism_studies_results with panelRunId. Synthetic answers are never people's: always say so when you report them.

Saves a panel from a spec (usually from prism_panels_draft) and draws its people from population records in the background, which takes a few seconds. Visible to the workspace unless visibility is PRIVATE. May return `needs_confirmation`.

Turns a plain description of an audience ("100 Midwest senior citizens") into a panel spec: filters on Census and General Social Survey fields, a size, and stated shares for traits with no data behind them (assumed). Saves nothing. Show the person the spec and assumedNotes, then pass the spec to prism_panels_create. poolSize is how many people in the population sample match.

One panel with its current version and the share of people by every field (assumed fields flagged). Use it to confirm a draw looks like the audience the person meant.

Every synthetic panel the person can see: name, current version (people, status), runs so far, and fit (how far it sat from people the last time a study had both). Reuse an existing panel before drafting a new one.

Starts a background run: every person on the panel answers the study (or, for kind interview, sits with the study's interviewer). Spends workspace credit — give the person the estimate from prism_panel_runs_estimate first. Works on draft studies. Answers land as their own result set, labeled synthetic, never counted as responses. Follow with prism_panel_runs_get. May return `needs_confirmation`.

A cross-study digest: every study with responses — title, origin (standalone / canvas node / tracker wave), response count, latest analysis summary, top themes, and which insights were kept in Ledger. The starting point for 'what have we learned about X?' — follow up with prism_studies_get (insights) or prism_studies_results on the studies that matter.

One desk-research report in full: the brief (markdown) and the sources it cites. Its claims are what public sources say, not what anyone at the company told you; when you put them on the Compass map, cite sourceKind 'prism_report' with this report's id.

Every desk-research report on a study, newest first: id, title, the focus it was asked, depth, status (running / ready / failed) and an excerpt. Read one in full with prism_research_reports_get.

Web-sourced desk research kept as a report on a study, with citations. Steer it with `focus`: research accumulates, a broad sweep first, then narrower runs as the questions sharpen. depth 'thorough' runs three angled passes at ~3× the credit. To research a company, make a standalone study about it (prism_studies_draft_standalone) and run one report per area of the business, e.g. focus 'how finance closes the month and approves spend at <company>, from job posts, filings and press'. It plans questions, searches, reads the pages and keeps only findings it can quote from them, so it takes a few m…

Creates a research program that fields the SAME instrument on a schedule — brand tracking, a weekly pulse — each wave an ordinary study. The questions freeze once the first wave fields (that's the trendline), so get them right with the user first. Optionally bind a Ledger metric + score rule so every wave posts a reading (a `scorer` needs a `metricId`). Waves launch on schedule, or on demand with prism_series_launch_wave. May return `needs_confirmation`.

CLOSES the current wave (no more responses; it scores and posts to Ledger if bound) and fields the next one from the series' frozen instrument, readying its participant link. Fielding spends credit for follow-ups, closing chats, and analysis, and closing a live wave can't be undone — say both plainly before proposing. Manual launches don't move the schedule clock. May return `needs_confirmation`.

Recurring research programs (ADR 0031): name, cadence, enabled state, next wave time, wave count, and the metric + score rule when the series posts readings. Waves themselves are ordinary studies.

Edits a series: `enabled: false` pauses the schedule (true resumes), interval / scheduleTime / scheduleTimezone / dayOfWeek move the cadence, name and follow-up / auto-analyze settings change any time. The instrument (questions, scorer) is editable only while no wave has fielded — after that the server refuses with `invalid`, and the answer is a new series. Series ids come from prism_series_list. May return `needs_confirmation`.

Kicks the theme analysis (incremental — reads only responses and interviews since the last run; the previous themes carry forward). Returns immediately with the run's phase; the run continues server-side, so wait a moment and re-read prism_studies_get — findings are fresh once its insights stop reporting an active run. upToDate: true means there was nothing new (a healthy no-op, not an error). Runs against inference credit. May return `needs_confirmation`.

Server-computed banner cut with significance: a target question (scale/number → group means + NPS on 0–10; choice/multi → per-option shares) split by acquisition source or by any single-choice question, each group tested against its complement at 95% (Welch t / two-proportion z). vsRest says higher/lower/not_significant — or not_tested when either side is under the 30-response floor; never present not_tested as 'no difference'. Use this for every 'does X differ by Y?' question instead of eyeballing raw rows.

Permanently deletes a study with its responses, interviews, analysis, and share links. Right for a draft that won't be fielded or a duplicate; for a live study the user just wants to stop, prefer prism_studies_update closed: true — that keeps the data. Study ids come from prism_studies_list_workspace. May return `needs_confirmation` — name the study and its response count, then wait.

Generates a short study (5-8 questions probing the node's problem space — respondents never see the idea itself) and creates it as a DRAFT. The user reviews, previews, and opens it for responses from the study page (or via prism_studies_update publish). Runs generation against inference credit. May return `needs_confirmation`.

Qual-first design: drafts a survey GROUNDED in the study's completed interviews — the recurring claims become measurable questions, the tensions become the choices, in the interviewees' own words. Replaces the study's current questions with the draft (nothing fields until publish). Needs at least one completed interview (see prism_interviews_list). An optional steer biases the instrument ('focus on pricing'). Runs generation against inference credit. May return `needs_confirmation`.

Generates and creates a DRAFT study that isn't tied to any canvas node — brand tracking, workspace-level research, anything you can describe in a sentence. It appears on Prism's Studies page; the user previews and opens it for responses there (or via prism_studies_update publish). Runs generation against inference credit. May return `needs_confirmation`.

One study's questions, response count, open/closed state, and the STORED analysis (summary + themes with strength; insights.activeRun set = an analysis is running right now — wait for it before citing). participantUrl is the live share link to forward when the study is open and one has been minted. Cite theme strengths when reporting results.

Feedback studies attached to one Prism node: title, question count, response count, open/closed.

All studies regardless of origin — standalone, canvas-node, and series waves — newest first: title, question count, response count, draft / live / closed state, and the field / node / series it hangs off. The starting point for 'which studies do we have?'; prism_studies_list is the per-node view, prism_research_digest the findings view. Read one with prism_studies_get.

An AI research analyst writes the report's narrative layer — executive summary, key findings with their numbers, recommendations — from the study's computed record (funnel, per-question aggregates, themes, quotes). Stored on the study; the printable report at the study page renders it above the charts. Runs against inference credit. May return `needs_confirmation`.

Everything the Results tab shows, as numbers you can trust without counting raw rows: fielding funnel (opens → starts → completes, median completion time, sources, per-question drop-off), per-question aggregates (scale distributions + means + NPS on 0–10, choice/multi counts, rank first-place + mean ranks, MaxDiff set scores, word frequencies, text samples), and quality flags (speeders, duplicate participant ids). ALWAYS use this for quantitative questions about a study — never tally answers yourself.

The stored k-means segmentation over the study's numeric answers: named segments with size, share, and the distinguishing features (segment mean vs overall). Null when none computed yet — prism_studies_segments_compute discovers them (needs 30+ responses and 2+ numeric questions).

Runs k-means over the study's numeric answers (silhouette-picked k) and names the discovered segments from their computed profiles. Needs 30+ responses and 2+ numeric questions; recompute overwrites. Runs naming against inference credit. May return `needs_confirmation`.

publish: true takes a draft live (one-way — the participant link starts working). closed toggles whether a live study accepts responses. Also edits the study: title, the `questions` instrument (whole replacement — read prism_studies_get first; LOCKED once responses exist, the server refuses with `invalid` and the answer is a new study), `adaptiveFollowUps` (AI probes on curious answers), the closing chat (`endChatEnabled` + `endChatFocus`), `completionRedirectUrl` for a BYO panel (null clears), and `autoAnalyzeTarget` (null = manual only). Study ids come from prism_studies_list_workspace. May …

Cross-instrument findings for one subject (today: a Prism node) — attached feedback studies with response counts + theme counts, and whether desk research exists. (Interviews are study-level now — read them per study with prism_interviews_list.) THE PLACE TO LOOK before proposing new research: cite what already exists.

Adds a Workbench flow to a room as a member, so people can @mention it and it can work on things with you there. The flow must be shared with the workspace and published. Use it after building a specialist flow for a room's work, or when someone asks for a flow to join. May return needs_confirmation; say which flow and why, and wait for approval.

Creates a room: a shared space where people work together with you, and the person asking is always in it. Add people by name or email (members of this workspace); `private` keeps it to them, otherwise everyone in the workspace can open it. `opening` is your first message in it: why the room exists and the concrete next step, written for everyone in it. When the room continues this conversation, put what they need from it in `opening`; nothing else carries over. Check rooms_list first so you don't make one that already exists. May return needs_confirmation; say in a sentence what you're making…

The workspace's rooms: where groups of people work together with the assistant and flows. Returns each room's id, name, people and flows, linked folders, open sidebar count, and when it last moved. Use it to find the room a user means before rooms_read or rooms_post.

Posts a message into a room's stream as the assistant, where everyone in the room will see it. Use it when a user asks you to share something with a room, or when a routine's instruction says to deliver its result to a room other than the one the routine already posts in. Write for the room, not for the person who asked: say what it is and why it's here in the first line. To get someone's attention, write @ and their name as rooms_list shows it; they're told even if they only hear mentions. May return needs_confirmation; summarize the post and wait for approval.

What a room has been saying: its catch-up summary (and whether messages have landed since), its sidebars with their outcomes, and the most recent messages as plain lines with who said them. Use it when a user asks what happened in a room, what a group decided, or to catch them up. For older talk, pass `before` (a message id: the page ahead of it) or `from` (a time: the page starting then); those return just the messages, with hasEarlier and hasLater. rooms_search finds where something was said. Quote people by name; never present a summary as a decision unless a sidebar settled it.

Searches the messages in the rooms the person can see, or in one room, newest first. Each hit is a line with who said it, when, the room, and the message id. Use it when someone asks what a group said about something, when it was decided, or who brought it up, and the answer may be older than what rooms_read shows. To read around a hit, call rooms_read with that room and `before` set to a later message id, or `from` set to a time shortly before the hit.

Starts the assistant and flows in a room working on something together in a sidebar: they take turns without a person between each, and the result comes back to the room's stream. The flows must already be in the room (rooms_add_flow). Use it when someone asks for agents to work something out, or to hand a specialist flow a job. Anyone in the room can read along or stop it. May return needs_confirmation; say what they'll work on and who, and wait for approval.

Create a routine: something you keep doing for the user on a schedule, one turn per run ("every morning, read these forums, add new questions to the Questions table, and post the count in #support"). Each run has full workspace context and your machine where the workspace has machines. Use a routine when nobody but you acts on each run. Use a Task instead (workbench_tasks_create) when a person acts on each item before anything else happens, or the work is a process mapped in Compass. Name in writes the tools each run writes with on its own; other writes wait for approval on every run. A site r…

List the assistant's routines — scheduled natural-language instructions run by the assistant. Returns each routine's id, name, instruction, cadence, delivery, enabled state, and last-run outcome. Check this before creating a routine so you don't duplicate one that already exists.

Sets the full list of sites a routine's machine may reach on each run (example.com, or *.example.com for its subdomains; up to 20), for a routine that has to use a site like a person would because read_webpage can't get it. It replaces the list, so read routines_list first (allowedSites) and pass every site to keep. Always waits for approval, and only a workspace admin's approval applies it; for anyone else it says who can. The next run reaches the new list.

Update a routine's instruction, cadence, delivery, the writes each run makes on its own, or its enabled switch (pause/resume). Only pass the fields being changed; `writes` replaces the whole list. Cadence changes recompute the next fire time. When people now act on each item, or it has become a process mapped in Compass, build a Task instead (workbench_tasks_create).

Search ZeroWidth product documentation. Returns matching pages with title, slug, public URL, and a query-relevant snippet. Use this when the user asks about a ZeroWidth product (Compass, Workbench, Caliper, Prism, Ledger, Napkin, the assistant), an API behavior, or a policy. No authentication required — the docs corpus is public.

Finds anything in the workspace in one call — uploaded files (searching the text inside them), folders, Workbench flows, shims, knowledge bases and tasks, Compass pages and changes, Caliper datasets, evals, rubrics, reviews, specs and sources, Ledger entries, Napkin sketches, decks, docs, sheets, diagrams and interfaces. The same search the person uses in Home. Use it FIRST whenever the user mentions something without saying where it lives ('the SOW', 'the dispatch interviews', 'that invoice'). The query is words in any order: each word must appear in the thing's name, its tags, the names of t…

The flow's stack traces: recent executions with per-node timelines — what each node received, produced, how long it took, and the exact error when one failed. USE THIS when a run misbehaves instead of guessing: read the failing node's inputs/error, then propose a fix (workbench_flows_edit_text) grounded in what actually happened. Pass executionId to inspect one run, or just flowId for the most recent runs. Node inputs/outputs are truncated for transport — the full record is in the flow's dev drawer.

READ THIS FIRST before writing or editing any raw flow orchestration JSON. Covers the document shape, the settings-vs-input-ports rule (temperature, max_tokens, tools, response_format are PORTS, not settings — constants reach ports via value nodes), exact link format, plugin links, and three complete worked examples.

Attaches a knowledge base to an AGENT flow's draft, the same way picking it in the agent editor does: a docs knowledge base becomes a search tool the model calls when a question needs it; a tabular one also puts its tables in the prompt and adds a query tool. The flow needs a system prompt (scaffold it with one). Attaching one that's already there changes nothing. Do it BEFORE workbench_flows_publish: a published version is what interfaces and schedules run, so attach, try it with workbench_flows_run, then publish — and if something already pins an older version (an interface grant), re-grant …

Deletes a flow. Its schedules stop, guest share links die, and evals targeting it can no longer run (their binding shows flowOk: false) — check caliper_flow_performance for evals and workbench_flows_schedules_list before proposing, and prefer workbench_flows_update with archived: true when the user just wants it out of the way. May return `needs_confirmation`.

Applies ONE precise text replacement to a flow's DRAFT orchestration — Edit-tool semantics: `find` must be the EXACT current text, copied character-for-character from workbench_flows_get, and must occur exactly once anywhere in the flow (system prompts, node settings, metadata, a node's type). `find` is text inside ONE value, not JSON: to change a model, find the model slug alone (`anthropic-claude-haiku-4-5`) and replace it with another `llm` slug from workbench_node_catalog_get. It can't add or remove nodes or links. Zero or multiple matches return an error instead of guessing. This is the i…

Copies a PUBLIC flow from another workspace (a template) into this one as a new draft the user owns. Identify it by its flowUuid (the id on Workbench's public template pages and in `workbench_flows_get` output); optionally pin which published version to copy. Flows already in this workspace can't be forked — open them instead. Counts toward the plan's flow cap. May return `needs_confirmation`.

One flow. Default `view: summary` lists its nodes and links so you can pick one; `view: node` with a nodeId returns that node's full settings and prompt — read THAT before workbench_flows_edit_text, and copy `find` text from it character-for-character, never from memory. `view: full` returns the whole orchestration body.

Lists every flow in the active workspace the caller can see. Returns summaries (id, name, visibility, updatedAt) — fetch one with `workbench_flows_get` for the full orchestration body.

Snapshots the flow's current DRAFT as a named published version — the version schedules (workbench_flows_schedule), guest share links (workbench_flows_share_create), public-API runs, and 'published'-stage evals execute. The draft keeps evolving after this; runs on the published side don't change until the next publish. Publishing does NOT run the flow, but any eval set to runOnPublish starts a run (that spends credit — mention it when one exists). Refused when the draft hasn't been run since it last changed: try it first with workbench_flows_run on a few real questions, including one it should…

The flow's revision history, newest first: published versions carry a `version` label and note; entries with version null are draft autosaves. Use it to answer 'is this published?' (any entry with a version), to find a revision id for caliper_evals_create's flowRevisionId, or to see when the draft last changed.

Execute a flow and return its outputs synchronously. Runs the draft by default; pass source:"published" (optionally a version) to run the live published version. Spends workspace LLM budget, bounded by the token's cost cap. `input` is the flow's input envelope, e.g. {"kind":"chat","messages":[…]} or {"kind":"form","values":{…}}.

Creates a runnable first-draft flow from a spec you author: pick the simplest pattern that fits (classifier for read-and-bucket, structurer for transform/extract/draft-for-review, agent for genuinely conversational), write a production-quality system prompt grounded in what the user told you, and mark anything stubbed with [STUB: ...] markers plus stubNotes. The draft opens in Workbench's simple editor at /w/<workspace>/flows/<id> — give the user that path. May return `needs_confirmation`; show the user what you're proposing and wait for their approval, then re-call with the approvalId. When t…

Saves a named scenario: values for the flow's variables, e.g. one demo customer. It shows in the test pane's Scenarios menu, and Caliper evals can run as it. Use the person's details from what the user shared (a CRM screenshot, a profile) and say which ones you made up. May return `needs_confirmation`.

Deletes one scenario. A design session running as it goes back to no scenario. May return `needs_confirmation`.

A flow's variables (each {{token}} in an agent's prompt, with its type) and every saved scenario with its values. Read this before creating or updating a scenario, so the values use the flow's real variable names.

Renames a scenario or changes its values. `values` MERGES into what's saved: pass only the variables that change, and null to clear one. Get the scenarioId from workbench_flows_scenarios_list. May return `needs_confirmation`.

Set a specific built flow to run automatically on a cadence — the deterministic counterpart to a routine. Each run executes the flow's latest PUBLISHED revision with the fixed `input` envelope and routes the output per `delivery`. USE THIS when the user wants a flow they've built to run on a schedule ("run my digest flow every morning"). Do NOT create a routine for this — a routine runs a free-form instruction, not a built flow. The schedule runs the PUBLISHED flow, so the flow must be published (check workbench_flows_revisions_list; publish with workbench_flows_publish if it isn't). `input` i…

Removes a schedule for good. Prefer workbench_flows_schedules_update with enabled: false when the user might want it back. May return `needs_confirmation`.

Every schedule on one flow: cadence, whether it's paused (`enabled: false`), the next fire time, delivery, and the last run's status. Read this before workbench_flows_schedule (don't create a duplicate) and to get the scheduleId for workbench_flows_schedules_update / workbench_flows_schedules_delete / workbench_flows_schedules_run_now.

Runs one iteration of a schedule immediately through the real scheduled pipeline (same input, same delivery — the email or channel it normally posts to) without moving its cadence. Works on paused schedules. Use it to test a schedule the user just set up. Spends workspace credit and delivers for real, so it sits behind the approval gate — may return `needs_confirmation`. The run is fire-and-forget: check workbench_flows_schedules_list for lastStatus, or workbench_executions_get for the trace.

Changes one schedule: `enabled: false` pauses it (configuration kept), `enabled: true` resumes and recomputes the next fire time from now; `interval` / `scheduleTime` / `scheduleTimezone` / `scheduleDayOfWeek` change the cadence; `label`, `delivery`, and `input` can change too. Pass only what changes. Get the scheduleId from workbench_flows_schedules_list. May return `needs_confirmation`.

Creates a no-account guest link where anyone can chat with the flow's latest PUBLISHED revision at workbench's /s/<token> page — the fastest way to put a working flow in a stakeholder's hands ('here, try it'). The flow must have a published version (workbench_flows_revisions_list shows it; publish with workbench_flows_publish if not). Conversations are capped per guest; the link can expire, and you can list links with workbench_flows_shares_list and kill one with workbench_flows_shares_revoke. May return `needs_confirmation` — say who the link is for and wait.

Every guest link on one flow, newest first: label, whether it's still live, the shareUrl (live links only — dead ones have no URL), expiry, and how many guests and messages it has seen. Use it to answer 'who has access', to recover a link the user lost, and to get the shareId for workbench_flows_shares_revoke.

Kills one guest link immediately — anyone holding it gets a closed page from then on. Conversations already had stay in the flow's history. Get the shareId from workbench_flows_shares_list. May return `needs_confirmation`.

Changes a flow's name, description, visibility, tags, or archived state — the metadata around the flow, NOT its orchestration (prompts and nodes change through workbench_flows_edit_text). Pass only what changes. Archiving hides the flow from the default gallery but leaves it runnable, scheduled, and shared; pass archived: false to restore. May return `needs_confirmation`.

Adds documents to a knowledge base and starts ingestion (chunk + embed; graph KBs also extract entities; tabular KBs load CSV as tables). Text goes inline (Markdown, plain text, CSV — encoding utf8); binary files (PDF, .docx) go base64-encoded with their mimeType. Up to 50 sources per call, ~10 MB each. Ingestion runs in the BACKGROUND — the result carries a runId; check it with workbench_kb_ingestion_status when the user asks, don't poll. workbench_kb_search works once it finishes. Content from <attached-file> blocks is ideal source material. Check workbench_kb_documents_list first so you don…

Creates an empty knowledge base. Recipe 'docs' (default) for reference material an agent searches at runtime; 'tabular' for spreadsheet-style data queried with SQL; 'graph' for entity/relationship extraction. After creating, add content with workbench_kb_add_documents, then give an agent flow a search tool over it with workbench_flows_attach_knowledge_base (before publishing the flow). May return `needs_confirmation`.

Deletes a knowledge base and everything in it. Flows that search it will find nothing — check which agents use it (the user knows; the KB page in Workbench lists them) and say so before proposing. May return `needs_confirmation`.

Removes one ingested document and all its chunks; searches stop returning it at once. Get the documentId from workbench_kb_documents_list. To replace a document, remove it then workbench_kb_add_documents the new version. May return `needs_confirmation`.

Every document ingested into the KB's current content: id, name, type, size, chunk count. Empty for a KB with nothing ingested yet (or one whose ingestion is still running — check workbench_kb_ingestion_status). Use it before workbench_kb_add_documents to avoid re-adding a document, and to get the documentId for workbench_kb_document_delete.

One knowledge base: name, recipe (docs / tabular / graph), status, document and chunk counts, embedding model, and ingestion settings. Read it to confirm a KB has content before pointing a flow at it. Get the kbId from workbench_kb_list.

The state of one ingestion run started by workbench_kb_add_documents: PENDING / RUNNING / DONE / FAILED, chunks written, and the error when it failed. Read it when the user asks whether their documents are in yet, or before a search that needs them — don't poll in a loop; ingestion of a few documents takes under a minute. Pass the runId the add-documents call returned.

Lists every knowledge base in the active workspace the caller can see. Returns summaries (id, name, description, status, doc/chunk counts, embedding model). Use an id with `workbench_kb_search` to retrieve content.

Retrieves the most relevant chunks from one knowledge base — the core RAG primitive. `mode:"semantic"` (default) embeds the query and ranks by meaning (spends a small amount of workspace credit); `mode:"keyword"` is a free case-insensitive substring match. Each hit carries the chunk text, its document, and a score (cosine similarity 0-1 for semantic, match count for keyword). Get a kbId from `workbench_kb_list`.

The ground truth for what nodes exist and what their ports actually are — verify against this instead of recalling. Search by keyword/category for summaries; pass `slug` for one node's full detail (inputs, outputs, settings). Inputs are PORTS fed by links; settings live on the node — see `workbench_flow_authoring_guide`. The catalog is the WORKSPACE'S: model nodes the workspace's inference policy forbids come back with `allowed: false` — never author with those; pick an allowed model. Pass `flowId` to include the flow's pinned imports as nodes.

Adds one answer (a label and, optionally, when it applies). The shim needs three examples of it before it builds again with the new answer. May return `needs_confirmation`.

Adds example inputs to one answer. Write inputs a real person would actually type in the shim's setting — varied in length, register and specifics, each one clearly this answer and not another — never paraphrases of the answer's name. Read the shim first so new examples fit alongside the existing ones and land where recall is low. Exact duplicates are skipped. The shim rebuilds on its own afterwards; read it again for the new report. May return `needs_confirmation`.

Creates a shim from a name, the question it answers, and its answers. Add examples afterwards with workbench_shim_add_examples — three or more per answer and it builds on its own. May return `needs_confirmation`.

One shim in full: the question, where its inputs come from, every answer with its description and examples, what the shim still needs before it can build, and the newest build's report — held-out accuracy, recall per answer, which answers get mistaken for which, and the compiler's notes. Read it before writing examples: match the setting and the existing examples' register, and put new examples where recall is low or the compiler says an answer is thin.

Every shim the user can see: id, name, the question it answers, how many answers and examples it has, and the published version's held-out accuracy (null until something is published). A shim is a small decision model that runs inside an app with no model call. Start here to get a shimId.

Records THIS user's verdict on a run that is waiting on them — the labels come from the gate (workbench_tasks_run_view shows them via the plan; a wrong label errors listing the real options). Echo the verdict + note back to the user in one line BEFORE calling; that echo is the confirmation. Refuses when the gate is assigned to someone else.

Runs the compiler over a sentence and returns the PROPOSED plan (or repair-grade problems — fix by refining the sentence with the user, not by guessing). Nothing persists. Present the proposal in plain terms: which clauses run flows, which you will handle, who decides, what the compiler noted. `computer.sites` lists sites the plan names and whether the Task allows them; workbench_tasks_set_sites asks an admin to allow the rest. Then iterate or freeze. Compile may take ~15s.

Creates a DRAFT Task from a name + sentence. A Task is for work with people in the loop (someone approves, decides or acts on each item before anything else happens) or a specific process mapped in Compass; its runs are measured. When nobody but you acts on it and you'd just do the same thing each time on a schedule, make a routine instead (routines_create). Use this once the user's sentence feels settled after you've refined it together — then compile it. The sentence should be THEIR words: trigger first ('When a claim arrives…'), then steps, branches, and who decides what. Name rooms, sheets…

Deletes a task. Runs still in flight are stopped (a gate waiting on someone is closed), and its schedule trigger stops firing; past runs stay in history. Check workbench_tasks_get for live runs and say so before proposing. May return `needs_confirmation`.

Freezes a reviewed plan as the Task's next version and activates it. Pass the EXACT plan object a compile returned — never hand-edit it (change the sentence and recompile instead). Freeze ONLY after the user has seen the proposal and said go; echo what you're freezing in one line first. In-flight runs keep their pinned versions. After freezing, offer to start a first run.

One Task in full: its sentence, status, the latest saved plan (the clauses — which run flows, which the assistant handles, who decides at each gate) with the flow and people names the plan refers to, and its recent runs, plus the sites its machine may reach (allowedSites) and the sites its plan names or past runs were refused (computer.sites). Read this before editing a task's sentence (workbench_tasks_compile with taskId) and when the user asks what a task does. Get the taskId from workbench_tasks_list.

Lists the workspace's Tasks — sentence-orchestrated processes — with their sentence, status, and recent runs (id, status, who a waiting run is on). Use this to find a task the user names ('claims intake') and to see what's in flight; workbench_tasks_get reads one in full.

The run's current state rendered as its sentence: what ran (with timings), which path each gate took, what's waiting on whom and for how long, what's coming up. Use whenever the user asks where something stands, and after acting on a gate.

Sets the full list of sites a Task's machine may reach (example.com, or *.example.com for its subdomains; up to 20), for steps that have to use a site like a person would. It replaces the list, so read workbench_tasks_get first (allowedSites, and computer.sites for what the plan names) and pass every site to keep. Always waits for approval, and only a workspace admin's approval applies it; for anyone else it says who can. The latest saved version runs with the new list.

Kicks off a run of a Task and returns at once; the run carries on by itself and is live on the task's page. When the task expects an input document (its trigger has an input), pass the user's material as `input` — pasted text, extracted attachment text, or JSON. Tell the user it has started and what it does next. Check where it stands with workbench_tasks_run_view when they ask; don't poll it in a loop.

Changes a task's name — the gallery label only. The sentence and plan are versioned and change through workbench_tasks_compile + workbench_tasks_freeze, never here. May return `needs_confirmation`.

The workspace's inference-credit position: plan, this month's grant and how much of it is spent, and the purchased balance runs draw on past the grant. Check this before proposing spend-heavy work (eval runs, batch expansions) when credit has come up — and cite the numbers. Read-only; top-ups and plan changes happen in accounts, by a human.

Read an uploaded file (PDF, Word, PowerPoint, spreadsheet, text) as markdown. Find file ids with workspace_folders_browse (a folder's items of kind workspace_file) or search_workspace. For anything long, call with `outline: true` first: it returns the headings with their character offsets and the total length. Then read only the sections you need with `offset` and `maxChars`. Spreadsheets come back as tables. A file that can't be read as text says so; don't guess at its contents.

Save a picture the user attached in THIS conversation to the workspace's files, in the Home folder "From chat", and get back its file id. Use it when the user wants a picture they sent you used somewhere — an interface shows it as <img src="file:<fileId>"> (CSS: url(file:<fileId>)). Pick the picture by `name`: the name in its `[attached image: <name>]` marker. Omit `name` only when the latest message with pictures has exactly one. Saving the same picture again returns the file it was already saved as. Only pictures from this conversation can be saved; a picture on a website can't be fetched, s…

Sorts the files sitting directly in a folder into its subfolders, making a subfolder when two or more belong together; with `folderId: null`, files loose at the top go into the best folder anywhere. The same reading that files uploads on arrival: it runs in the background, and Home shows what moved, why, and an undo for each. Reads up to 60 files a call. Use it when someone asks to tidy or organize a folder; to put specific things somewhere, use workspace_folders_file. May return needs_confirmation.

Without a folder: every folder in the workspace as a tree, each with its id, full path, and how many things are filed directly in it. With a folder (by id or by name): what's inside it — subfolders, then the items themselves with their kind, id, title, tags, and path. Folders are the engagement axis: a client or project gathers work from every tool, so a folder is how you answer 'what do we have for Acme' when the answer spans flows, pages, and datasets. Reach for this when the user asks what's in a folder or wants the whole shape; to find one named thing, search_workspace (which also matches …

Makes a folder in Home, at the top or inside another folder. Folders gather a client's or a project's work from every tool. Check workspace_folders_browse first so you don't make one that exists. May return needs_confirmation.

Moves things into a folder, or out of every folder with `folderId: null`. Anything from any tool can go in: pass each one's `kind` and `id` as search_workspace or workspace_folders_browse gave them. A thing lives in one folder, so filing it somewhere moves it. Things the person can't see are skipped. May return needs_confirmation.

Renames a folder, or moves it (with everything in it) into another folder or to the top. May return needs_confirmation.

Creates folders for the work someone described, rooms linked to those folders (with your first message in each, the concrete next step), and starting pieces filed inside. Records what the workspace is for, which orders what home offers first. Use it once you know what they're working on, typically near the end of setting a workspace up. Say in one sentence what you're setting up and call it; the approval card lists the folders, rooms, and pieces, so don't list them in words too. Every workspace already has a General room, so don't recreate it. May return needs_confirmation; then wait for appro…