docs(ai): document the chassis runtime mode

[keydunov]

Check List

• Tests have been run in packages where changes have been made if available
• Linter has been run for changed code
• Tests for the changes have been added if not covered yet
• Docs have been added / updated if required

Description of Changes Made

The agent runtime config accepts a third value, chassis, alongside plain
and reasoning — it's already in the agents.yml JSON schema (surfaced as
autocomplete in the Semantic Model IDE) and shown as a selectable value in the
Agent Settings UI, but the Runtime docs only covered plain and reasoning.

This adds chassis to:

• the runtime property description in the Properties table
• the Runtime mode table, with a short description of its behavior (a lean,
  semantic-model-focused toolset that places query results inline in the
  response instead of as separate tool calls)
• the experimental-status warning already covering reasoning, since
  chassis looks similarly early-stage from the surrounding code (heavy,
  active iteration in the days before this PR)

No new page — this follows the existing table-row pattern used for reasoning.

---

🤖 Generated with Claude Code

https://claude.ai/code/session_013ZRDPkg9or671PqULs9kPh

---

Generated by Claude Code

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated
cubed3	🟢 Ready	View Preview	Sep 21, 2026, 12:59 AM

💡 Tip: Enable Automations to automatically generate PRs for you.

[claude]

Claude finished @keydunov's task in 1m 28s —— View job

---

Review complete — no blockers; 2 low findings (docs clarity, preview gating)

Full review

• Read the diff
• Verify chassis against the source
• Check docs page consistency
• Handle prior review threads
• Post findings

Findings

#	Location	Severity	Concern
1	docs-mintlify/admin/ai/index.mdx:121	low	The chassis row describes implementation, not a decision criterion — unlike the plain/reasoning rows beside it. "Places query results inline… instead of as separate tool calls" is internal mechanics; the user-visible effect on chat output is the part a reader needs.
2	docs-mintlify/admin/ai/index.mdx:124	low	The experimental warning doesn't say whether chassis is gated. docs-mintlify/CLAUDE.md requires preview features to direct the reader to Cube support for activation — if chassis is flag-gated, someone who sets it from this table and silently gets plain has no way to diagnose that.

Both posted inline.

Verification limits

chassis appears nowhere in this repo outside the file this PR edits (rg -i chassis → 1 file). The agents.yml JSON schema, the Agent Settings UI, and the runtime implementation the PR description cites all live outside the OSS monorepo, so the described behavior — the lean toolset, the inline result placement, and the experimental status — could not be checked against source here. It rests entirely on the author's reading of a private codebase; a maintainer with access should confirm the wording before merge. The PR description itself hedges ("chassis looks similarly early-stage from the surrounding code"), which is exactly the claim that needs a second pair of eyes.

What's correct

• Both touch points updated consistently — the Properties-table pointer at :54 and the mode table at :121; no other docs page enumerates runtime values (multi-agent.mdx:72 refers to the shared property list by link, so it needs no edit).
• Follows the docs-mintlify convention of extending an existing table over creating a page.
• Scoped tightly: 3 additions, 2 deletions, no padding.

Review threads

Listing returned 0 of my own threads (single page, hasNextPage: false) — no prior threads. 0 resolved, 0 re-affirmed.

· branch [`claude/gallant-ramanujan-3qukdv`](https://github.com/cube-js/cube/tree/claude/gallant-ramanujan-3qukdv)

[claude]

Claude finished @keydunov's task in 1m 36s —— View job

---

Review complete — 1 medium, 2 low; the medium blocks on a factual confirmation, not on the writing.

Full review

• Read the diff and the full changed file
• Verify chassis against the repo
• Check prior review threads
• Post inline comments

Findings

#	Location	Severity	Concern
1	docs-mintlify/admin/ai/index.mdx:121–125	medium	Both the behavior claim and the experimental label are inferred, and neither is verifiable from this repo
2	docs-mintlify/admin/ai/index.mdx:121	low	"result cards" is terminology used nowhere else in the docs
3	docs-mintlify/admin/ai/index.mdx:124	low	The warning's "chassis toolset" no longer matches the reworded table row

On finding 1 — rg -i chassis over cube-js/cube hits only the line this PR adds. The agent runtime and agents-schema.json live in Cube Cloud, so nothing here confirms either that chassis renders results inline or that it's experimental. The PR body is candid that the experimental label came from the mode "look[ing] similarly early-stage from the surrounding code" — recent commit churn isn't a stability signal, and if chassis is actually flagged preview, docs-mintlify/CLAUDE.md wants the preview callout rather than an inline <Warning>. One confirmation from the chassis owner settles both.

The rest of the change is right: adding a table row instead of a page matches the reasoning precedent and docs-mintlify/CLAUDE.md's "edit existing pages" rule, and line 54's Properties-table update keeps the two lists in sync. No other page lists runtime values, so nothing else needs updating.

Review threads

The listing returned 0 of my own threads — none to resolve, none re-affirmed.

· [branch `claude/gallant-ramanujan-3qukdv`](https://github.com/cube-js/cube/tree/claude/gallant-ramanujan-3qukdv)

[claude]

Claude finished @keydunov's task in 1m 34s —— View job

---

Review: docs-only, no blocking issues — 0 high, 0 medium, 1 low (already tracked on an open thread)

Full review

• Read PR diff
• Read docs-mintlify/admin/ai/index.mdx in full
• Check docs conventions (docs-mintlify/CLAUDE.md)
• Check for other pages needing the same update
• Handle prior review threads
• Post review

Findings

Low — stale "toolset" wording in the warning (docs-mintlify/admin/ai/index.mdx:124)

The <Warning> closes with "...or the chassis toolset", but 111351a rewrote the table row to drop any mention of a toolset — line 121 now reads "Optimized for semantic-model exploration. Returns query results inline...". The warning points at a concept the page no longer defines. "the chassis runtime" matches the surrounding vocabulary.

Not posted inline: already carried by an open thread on the same line with the same fix direction.

Not posted — unverifiable behavior claim (docs-mintlify/admin/ai/index.mdx:121)

chassis appears nowhere in cube-js/cube — the runtime lives in cubejs-enterprise, so the inline-results claim can't be checked from this checkout. The author cited resolve-runtime.ts and resolve-runtime.spec.ts in the review thread; taking that at face value.

Checks that came back clean

• Scope is complete. runtime is enumerated in exactly one place (index.mdx:54 and the mode table). multi-agent.mdx:72 links to /admin/ai#properties rather than re-listing values, and no other page enumerates runtime modes — nothing else needs the third value.
• Preview-feature rule (docs-mintlify/CLAUDE.md): correctly not applied. The author verified chassis isn't activation-gated, so the required "reach out to support to activate" callout doesn't apply; the experimental <Warning> is the right treatment.
• Terminology: "tool-call results" matches reference/embed-apis/chat-api.mdx, which uses "tool calls" throughout. The earlier "result cards" appeared nowhere else in the docs — b315a16 was the right call.
• Conventions: table-row extension over a new page, no H1, no padding — all consistent with the concision rule.

Review threads

Listing returned 3 of my own threads. Resolved 1 (PRRT_kwDOCOH19M6kNMdX — the "result cards" wording, applied verbatim in b315a16). Re-affirmed 2 as still live: PRRT_kwDOCOH19M6kNMoa (the chassis toolset wording at line 124, unchanged in the tree) and PRRT_kwDOCOH19M6kNMhb (the experimental/unverifiable claims, outdated but the <Warning> it questions is still in the file).

• [`claude/gallant-ramanujan-3qukdv`](https://github.com/cube-js/cube/tree/claude/gallant-ramanujan-3qukdv)

[claude]

Claude finished @keydunov's task in 1m 40s —— View job

---

Docs change is accurate and well-scoped — 1 low finding (wording), posted inline

Review details

• Read diff and changed file in context
• Verify chassis claims against source
• Check terminology consistency with neighboring docs
• Handle prior review threads
• Post findings

Finding

docs-mintlify/admin/ai/index.mdx:124 — the warning's decision criterion doesn't discriminate. "Stick with plain unless you specifically need extended reasoning or semantic-model exploration" tells a reader to always pick chassis: every agent use is semantic-model exploration, plain included. The table row one line up carries the actual differentiator (query results inline in the answer); the rewording in 31e63f5 dropped it from the warning. Posted inline with a suggestion.

Checked, no finding

• Placement and pattern. Follows the existing reasoning precedent — one table row plus the runtime property description in the Properties table, no new page. Matches docs-mintlify/CLAUDE.md's "default to editing existing pages" rule.
• Preview-feature callout. docs-mintlify/CLAUDE.md requires an activation <Warning> for preview features. Per the author's own verification against resolve-runtime.ts in cubejs-enterprise, runtime: chassis is a per-agent opt-in that needs no support activation, so the preview boilerplate would be wrong here. The experimental warning is the right call.
• Behavioral claim. chassis appears nowhere in this repo (grep -rni chassis hits only the edited file) — the implementation is in cubejs-enterprise, so the inline-results claim is not verifiable from here. Taking the author's source citation.
• tool-call results wording (index.mdx:121). Commit b315a16's rationale says this matches analytics-chat.mdx; that file contains no occurrence of "tool call". The term lives only in docs-mintlify/reference/embed-apis/chat-api.mdx, a stream-protocol reference. Since the Runtime table is itself developer-facing config docs, the term is defensible — noted, not posted inline.

Review threads

The listing returned 0 unresolved threads on the PR (none of them mine), so nothing to resolve and nothing to re-affirm. The three prior keydunov review comments are already-resolved or non-thread entries.

· branch [`claude/gallant-ramanujan-3qukdv`](https://github.com/cube-js/cube/tree/claude/gallant-ramanujan-3qukdv)

[claude]

Claude finished @keydunov's task in 2m 9s —— View job

---

Review complete — 1 medium, 1 low. Docs-only; chassis row omits the reduced-toolset trade-off.

Full review

• Read the PR diff
• Read the changed file in full context
• Verify claims against the repo (schema, terminology, related pages)
• Check prior review threads
• Post findings

Findings

Medium — docs-mintlify/admin/ai/index.mdx:121 (posted inline)

The row documents only the presentation difference (inline results) and no longer says chassis runs a reduced toolset. The original wording in 5c0c117 ("a lean, semantic-model-focused toolset") carried that; 111351a replaced it with "Optimized for semantic-model exploration", which reads as tuning, not lost capability.

Per the round-5 review comment's own source reading, chassis's tool list has no MCP tool support, no scheduled tasks, no dbt sync, no sheet tools, no workbook creation. MCP Connectors is a documented Premium/Enterprise feature (docs-mintlify/admin/ai/mcp-connectors.mdx:8) whose page promises its tools "become available to the agent in Analytics Chat" with no runtime caveat. Setting runtime: chassis silently drops them; "not yet feature-complete" in the warning is too vague to be read as "your connectors stop working". Suggested replacement text is in the inline comment, plus a matching caveat on mcp-connectors.mdx.

Low — docs-mintlify/admin/ai/index.mdx:124 (not posted inline)

The warning's escape hatch ("unless you specifically need … query results returned inline in the answer") and the row's lead ("Optimized for semantic-model exploration") give two different reasons to pick chassis, three lines apart. Both are defensible individually; the churn across 111351a → 31e63f5 → 7be489a left them out of sync. Resolving the medium finding above likely settles this too, so it isn't worth its own thread.

Note on b315a16's rationale — the commit message says "The docs already call this 'tool calls' … (analytics-chat.mdx)". analytics-chat.mdx contains neither "tool call" nor "result card". The term does exist in the docs, but only in docs-mintlify/reference/embed-apis/chat-api.mdx (e.g. :580, :582) and api-reference/chat.yaml:149. The wording landed on is fine for this developer-facing page — the justification for it was not.

Verification limits

The chassis implementation (resolve-runtime.ts, chassis/Graph.ts, the agents-schema.json enum, the runtimeChassisDescription UI string) lives in cubejs-enterprise, not this repo — grep -ri chassis across cube-js/cube matches only the file under review. Every behavioral claim about chassis in this PR rests on the author's quotes from that repo; I could not independently confirm them here. What I did verify in-repo: the docs terminology audit above, and the MCP Connectors page's unqualified promise.

Other focus areas (security, performance, testing, comments) don't apply — the diff is 3 added / 2 removed lines of MDX prose, no code, no schema, no navigation change needed (admin/ai/index is already registered).

Review threads

The listing returned 0 of my own threads (in fact zero unresolved threads of any author). Nothing to resolve, nothing re-affirmed.

· branch [`claude/gallant-ramanujan-3qukdv`](https://github.com/cube-js/cube/tree/claude/gallant-ramanujan-3qukdv)

[claude]

Claude finished @keydunov's task in 2m 10s —— View job

---

Review complete — docs-only change, no blockers. 0 high, 2 medium, 1 low.

Full review

• Read changed files and diff
• Verify claims, links, and docs-mintlify/CLAUDE.md conventions
• Check prior review threads
• Post findings

Findings

Sev	Location	Issue
Medium	docs-mintlify/admin/ai/index.mdx:121	The em-dash parenthetical reads as the complete gap list, but by the PR's own source review chassis also drops scheduled tasks, dbt sync, sheet tools, eval tooling, and VCS/web tools. Row is also ~3× the neighbouring rows' length. Posted inline with a suggestion.
Medium	docs-mintlify/admin/ai/mcp-connectors.mdx:19	New gray <Note> sits back-to-back with an existing one; with the plan-availability note at line 6 the page opens with three. docs-mintlify/CLAUDE.md reserves gray <Note> for plan gating precisely so it stays scannable. Posted inline.
Low	docs-mintlify/admin/ai/index.mdx:124	"Stick with plain unless you specifically need extended thinking or query results returned inline in the answer" offers a presentation detail as the reason to accept an experimental mode. The earlier round's "semantic-model exploration" was the stronger criterion. Not posted inline — this wording has already been through two reviewer-directed passes (b315a16, eb55d0f), so it's your call, not a defect.

Verified

• /admin/ai#runtime resolves: ## Runtime at docs-mintlify/admin/ai/index.mdx:113, and admin/ai/index.mdx maps to /admin/ai.
• No other docs page enumerates runtime values, so admin/ai/index.mdx:54 and :121 are the only places needing the third value. docs-mintlify/admin/ai/multi-agent.mdx:72 links to #properties rather than repeating the list — nothing to update there.
• No docs.json change needed (no new page), and the table/frontmatter conventions hold.
• Not verifiable from this checkout: the behavioural claims about chassis (reduced toolset, inline results, no MCP connector tools, no workbook creation) live in cubejs-enterprise, which isn't part of this repo. Taken as accurate on the strength of the UI-copy and resolve-runtime.ts citations in your review replies.

Review threads

Listing returned 0 of my own threads — no prior threads, so nothing resolved and nothing re-affirmed.

· branch `claude/gallant-ramanujan-3qukdv`

[claude]

Claude Code is working…

I'll analyze this and get back to you.

View job run

[claude]

Claude finished @keydunov's task in 2m 21s —— View job

---

Review complete — docs-only change, accurate; 1 low finding

Full review

• Read PR diff and both changed files in full
• Verified every new internal link resolves to a real page
• Checked for other docs that list runtime values or make unconditional agent-capability claims
• Resolved stale review threads / deduped findings
• Posted review

Verification

All four new links resolve to real pages, with titles matching the product terms used:

Link	Target
/docs/explore-analyze/analytics-chat	analytics-chat.mdx
/docs/explore-analyze/workbooks/workbook-agent	workbook-agent.mdx (title: Workbook Agent)
/docs/explore-analyze/dashboards/dashboard-agent	dashboard-agent.mdx (title: Dashboard Agent)
/docs/explore-analyze/dashboards/widgets/ai-summary	ai-summary.mdx

/admin/ai#runtime matches the ## Runtime heading at docs-mintlify/admin/ai/index.mdx:113.

docs-mintlify/admin/ai/index.mdx:54 and the table row at :121 agree on the value list, and multi-agent.mdx:72 defers to #properties rather than repeating the values, so nothing else went stale. A repo-wide sweep for every agent / all agents / any agent found mcp-connectors.mdx:36 as the only unconditional capability claim — so the asymmetry of carving out chassis there and nowhere else is justified, not an oversight.

Finding (low, non-blocking)

docs-mintlify/admin/ai/index.mdx:127 — the new paragraph tells readers runtime applies to the Workbook Agent, but docs/explore-analyze/workbooks/workbook-agent.mdx:21-32 still lists "Create and save reports", "Build and edit dashboards" and "Publish dashboards" unconditionally. Per this PR's own source check, workbook creation is exactly what chassis's reduced toolset drops — so an admin who sets runtime: chassis loses authoring with no signal on the page that documents it. This is the same contradiction mcp-connectors.mdx:36 was fixed for in 2a78f90, newly reachable because this PR is what points readers from the runtime section at the Workbook Agent. Posted inline as discussion_r4058631361. Non-blocking: the runtime section itself is correct, and the fix lands in a file outside this PR's scope.

Review threads

The unresolved-thread listing returned 0 threads (mine or anyone's) — nothing to resolve, nothing re-affirmed as a duplicate.

Notes

No lint or test run: the change is .mdx content only, with no build step touched. yarn dev in docs-mintlify was not run — Mintlify's dev server needs network access to render, and the change introduces no new components or frontmatter.

· branch [`claude/gallant-ramanujan-3qukdv`](https://github.com/cube-js/cube/tree/claude/gallant-ramanujan-3qukdv)

[claude]

Claude finished @keydunov's task in 2m 19s —— View job

---

Docs-only change, reads accurate against the PR's own cited sources — 1 medium, 1 low.

Review details

• Read the PR diff
• Read the changed files in full context
• Verified internal links resolve and are registered in docs.json
• Checked claims for internal consistency
• Handled prior review threads
• Posted findings

Findings

Medium — docs-mintlify/docs/explore-analyze/workbooks/workbook-agent.mdx:34 (posted inline, r4058639750)

"These authoring capabilities are not available when the agent runs the chassis runtime" is categorical over all four bullets above it — create/save reports, search reports, build and edit dashboards, publish dashboards. The tool list cited in this PR's own threads (chassis Graph.ts) registers dashboard-app tools and openInCubeTool with the dashboard kind implemented, so dashboard building/publishing plausibly still works; only report/workbook creation is evidenced as missing. Either narrow it to the capabilities actually absent, or hedge it the way admin/ai/index.mdx:121 already does ("some agent features are unavailable").

Low — docs-mintlify/admin/ai/mcp-connectors.mdx:15 (not posted inline; outside the diff hunk, so not anchorable)

The page intro still says connector tools "become available to the agent in [Analytics Chat]" unconditionally, ~20 lines above the chassis exception the PR adds at line 36. Not wrong enough to block — the Concepts bullet is the normative statement — but worth a word of hedging next time that paragraph is touched.

Verified

• All four new links resolve to existing pages and are registered in docs.json: docs/explore-analyze/analytics-chat, docs/explore-analyze/workbooks/workbook-agent (docs.json:57), docs/explore-analyze/dashboards/dashboard-agent (docs.json:125), docs/explore-analyze/dashboards/widgets/ai-summary (docs.json:119).
• Properties table (index.mdx:54) and Runtime table (index.mdx:121) now agree on the three accepted values.
• No preview callout required per docs-mintlify/CLAUDE.md — the earlier thread established runtime: chassis isn't support-activation-gated, and the existing experimental <Warning> is the right treatment.
• "such as" framing on the chat-surface list (index.mdx:127) correctly avoids reading as exhaustive.
• The chassis implementation lives in cubejs-enterprise, not in this repo, so behavior claims were checked for internal consistency against the source excerpts quoted in this PR's threads rather than re-read from source.

Review threads

Listing returned 0 of my own threads (single page, hasNextPage: false) — no prior threads. 0 resolved, 0 re-affirmed.

· branch [`claude/gallant-ramanujan-3qukdv`](https://github.com/cube-js/cube/tree/claude/gallant-ramanujan-3qukdv)

[claude]

Claude finished @keydunov's task in 2m 18s —— View job

---

1 medium, 0 high — docs-only change, links and anchors all resolve

Review details

• Read PR diff and commits
• Verify linked pages exist (dashboard-agent, ai-summary, workbook-agent, analytics-chat)
• Check claims against source / prior review threads
• Resolve stale threads, post findings

Review threads

The listing returned 0 of my own threads (paged to the end — page 1 had 0 nodes and hasNextPage: false, and the follow-up cursor call also came back empty). Nothing to resolve, nothing to re-affirm. The 17 prior review comments on this PR are all keydunov-authored threads, which aren't mine to close.

Findings

Medium — docs-mintlify/admin/ai/index.mdx:121: the chassis row names no gaps and links nowhere.
Posted inline. "some agent features are unavailable" leaves a reader configuring runtime with no way to learn what they'd lose. Both known gaps are documented (mcp-connectors.mdx:36, workbook-agent.mdx:34) but both link back to /admin/ai#runtime — so this row is the inbound landing point for the question it declines to answer.

This isn't a re-opening of the earlier duplication finding: 62f204e stripped the examples because 83d5c44 had made the row the single source by deleting the per-page notes — but 2a78f90 and e3936f5 restored those notes. With them back, naming the examples here duplicates nothing, and as it stands nothing enumerates the gaps at all. The suggested wording keeps the list explicitly non-exhaustive.

Verified, no issue

• All four new link targets exist as files: docs/explore-analyze/dashboards/dashboard-agent.mdx, docs/explore-analyze/dashboards/widgets/ai-summary.mdx, docs/explore-analyze/workbooks/workbook-agent.mdx, docs/explore-analyze/analytics-chat.mdx. The /admin/ai#runtime anchor matches the ## Runtime heading.
• No new pages, so no docs.json nav update is required.
• Product terms match docs-mintlify/CLAUDE.md — "Workbook Agent", "Analytics Chat", "Dashboard Agent", "AI summary".
• The MCP caveat is attached to the right bullet: mcp-connectors.mdx:35 is exactly the "every agent can use the tools you enable" claim that chassis contradicts.
• No preview callout needed — prior rounds established from resolve-runtime.spec.ts that runtime: chassis is not support-activation-gated, so docs-mintlify/CLAUDE.md's preview-warning rule doesn't apply.

Not raised

The Warning at index.mdx:124 frames the reduced toolset as something a reader would "need" while the row frames it as a limitation. That exact wording was chosen at the reviewer's request in c9f6478 after two rounds of churn (da7fb58, 01813e3); it's a matter of taste, not a defect, and re-litigating it would just restart the cycle.

Not run

The behavioural claims (chassis's tool list, the AI-widget runtime exclusion) live in cubejs-enterprise, which isn't checked out here — chassis appears nowhere in this repo outside these three docs files. I took the source citations in the prior review threads at face value rather than re-verifying them. No build or lint was run; these are content-only .mdx edits with no nav or frontmatter changes.

· [`claude/gallant-ramanujan-3qukdv`](https://github.com/cube-js/cube/tree/claude/gallant-ramanujan-3qukdv)

[claude]

Claude finished @keydunov's task in 2m 19s —— View job

---

Docs change is accurate; 2 low findings (1 posted inline, 1 on an untouched file)

Review details

• Read PR diff and the three changed files in full
• Verified every new link target resolves (/docs/explore-analyze/analytics-chat, /docs/explore-analyze/workbooks/workbook-agent, /docs/explore-analyze/dashboards/dashboard-agent, /docs/explore-analyze/dashboards/widgets/ai-summary, /admin/ai/mcp-connectors, anchor #runtime ← ## Runtime) — all exist
• Checked other docs pages for claims chassis contradicts
• Prior review threads
• Posted findings

Findings

1. docs-mintlify/admin/ai/index.mdx:121 — table cell carries prose that belongs under the table (low, posted inline)

~350 characters, three clauses, two inline links, next to one-line plain/reasoning rows. Suggested split: short row + the gap list as a sentence after the <Warning>, alongside the surface-scope paragraph already at line 127. Keeps both constraints earlier rounds landed on ("reduced toolset" so the row discriminates against plain; inline results).

2. docs-mintlify/docs/explore-analyze/scheduled-tasks.mdx:111-133 — chassis caveat missing where the same class of claim appears (low, not posted inline — GitHub only accepts inline comments on changed lines and this file is untouched)

The PR adds the "unavailable under chassis" caveat to mcp-connectors.mdx:35 and workbook-agent.mdx:34. scheduled-tasks.mdx:113-121 ("the agent in a scheduled run can … create and update reports, workbooks, and dashboards") and :126-133 ("create and manage Scheduled Tasks conversationally in Analytics Chat") make the same unconditional claim about tools your own source check (thread at 00:19:05 — chassis Graph.ts has no scheduled-task, workbook-creation, or dbt-sync tools) found chassis lacks. With two of three pages caveated, the silent one now reads as "supported under chassis."

Two consistent ways out: add the same one-liner there, or go back to the earlier round's decision that the runtime table is the single source and drop the per-page notes. Either is fine; the current half-and-half is the problem. Fix this →

Non-findings, checked

• mcp-connectors.mdx:15 states unconditionally that connector tools become available in Analytics Chat, 20 lines before the caveat at :35. Same page, caveat present — not worth another wording round.
• runtime is correctly listed as plain, reasoning, or chassis in the Properties table (index.mdx:54), consistent with the Runtime section.
• No new page, so no docs.json change needed — correct per docs-mintlify/CLAUDE.md's "prefer editing an existing page".
• chassis appears nowhere in this repo's source; the behavior and tool-gap claims rest on the cubejs-enterprise verification recorded in the PR threads, which I cannot re-check here.

Review threads

The unresolved listing came back empty — no prior threads of mine on this PR, so none resolved and none re-affirmed.

· branch `claude/gallant-ramanujan-3qukdv`

[claude]

Reviewing

• Gather PR context and diff
• Read changed files in full
• Verify claims against source
• Reconcile prior review threads
• Post findings

View job run · claude/gallant-ramanujan-3qukdv

[keydunov]

Re: the scheduled-tasks.mdx finding from the latest review pass (not postable inline — that file is untouched by this PR) — good catch that the per-page-caveat approach doesn't scale: the same "the agent can do X" pattern chassis contradicts also appears there, and potentially on other capability pages across the docs site that I haven't audited (evals, certified-queries, skills, etc.). Chasing each one down is unscoped for a PR documenting one config value.

Went with the reviewer's other offered option instead: reverted the two per-page notes (mcp-connectors.mdx, workbook-agent.mdx) added earlier in this PR, and kept admin/ai/index.mdx's Runtime section as the single source of truth for what chassis doesn't support, with links out to the two pages it affects most. Pushed in f9a0ebc.

---

Generated by Claude Code

[claude]

Claude finished @keydunov's task in 2m 20s —— View job

---

Review complete — 1 low finding (docs consistency), no blockers

Details

• Gather context (diff, changed file)
• Verify link targets resolve
• Verify factual claims against available source
• Reconcile prior review threads
• Post findings

Posted inline (1)

docs-mintlify/admin/ai/index.mdx:135 — the new sentence "AI summary widgets always use their own runtime, regardless of this setting" contradicts docs-mintlify/docs/explore-analyze/dashboards/widgets/ai-summary.mdx:35, which tells readers to override the agent per widget "when you need a particular agent's tooling, model, or guardrails." Since runtime is what the table says determines the toolset, per-widget agent choice is partly inert for tooling, and neither page acknowledges the other.

Checked and clean

• All five internal links resolve to existing pages: admin/ai/mcp-connectors.mdx, docs/explore-analyze/analytics-chat.mdx, docs/explore-analyze/workbooks/workbook-agent.mdx, docs/explore-analyze/dashboards/dashboard-agent.mdx, docs/explore-analyze/dashboards/widgets/ai-summary.mdx. Root-relative form matches the repo convention.
• No new page, so no docs.json change needed — correct per docs-mintlify/CLAUDE.md's "default to editing existing pages".
• Table row is well-formed; runtime enum at line 54 and the mode table at 117–121 now agree.
• No <Warning> preview callout needed: the author's cited reading of resolve-runtime.ts (per-agent runtime: chassis works without the tenant flag) means this isn't support-activation-gated, which is what that convention covers.

Not posted

• index.mdx:121 / :124 / :127 state the reduced-toolset point three times (table row, Warning, following paragraph), which brushes against docs-mintlify/CLAUDE.md's "Say it once." Withheld deliberately: this exact wording has ping-ponged across several earlier rounds, and another suggestion would restart that rather than settle it. Worth one final human pass if you want the callout trimmed.

Verification limits

The behavioral claims (chassis's tool list, resolve-runtime.ts, the useChassisRuntime flag) live in cubejs-enterprise, which is not part of this repo — I could not independently confirm them and am relying on the source excerpts quoted in the PR's own review threads.

Review threads

The unresolved-thread listing returned 0 of my own threads (no prior threads). Nothing to resolve, nothing re-affirmed; 1 new thread opened this round.

· branch [`claude/gallant-ramanujan-3qukdv`](https://github.com/cube-js/cube/tree/claude/gallant-ramanujan-3qukdv)