feat: run a saved model as its real node when previewing (#2028)

[ralphstodomingo]

Closes #1875. Fixes #2028. Replaces #1894, which added a separate command; this takes the approach @mdesmet proposed in review instead.

Problem

Execute SQL sends the editor contents as dbt show --inline, which compiles them as an anonymous node (inline_<hash>). Any Jinja that reads the current node's identity — model.name, this, model.config, or a macro derived from them — resolves against a placeholder rather than the model in front of you, so the model fails to compile.

Same root cause, two reported symptoms: #2028 (a model.name guard macro rejects the model) and #1875 (--inline leaves selected_resources empty, so packages like upstream-prod misbehave).

Approach

The earlier attempt added an opt-in Execute dbt Model command plus a hint pointing at it. That works, but the user still has to fail once and then press something else. Review preferred routing the existing path, which is what this does — no new command, no new keybinding, nothing to discover. For the reported bug, Cmd+Enter now simply works.

Execute SQL decides in this order:

1. Something is selected → run the selection inline. Unchanged.
2. Untitled buffer → run inline. Unchanged.
3. Saved, no unsaved edits → run it as the real node (dbt show --select). No prompt: the buffer is byte-identical to disk, so the SQL is the same either way and the only difference is that identity now resolves.
4. Unsaved edits → run inline, exactly as today. The edits are nearly always what you want previewed, so there is no interruption. Only if that run fails for lack of node identity is the alternative offered — that failure is the one reliable signal that running the saved node would have behaved differently.

Case 3 is the only silent behaviour change, and it is safe precisely because the content is identical.

Integration capability

Whether an integration can supply the real node as context comes from canInjectNodeContextInSqlSnippetExecution() on the integration, not from testing dbt.dbtIntegration. Which modes qualify belongs to the library and will change as integrations gain the capability, so the extension does not hardcode mode strings.

The Python bridge reports false today — it compiles in-process under a temporary id and has no API for supplying the node. Those users always run inline and never see the offer. AI-8139 tracks fixing that at the source, after which it stops being an exception with no change here.

Saving

We never save silently. Case 4's offer is a prompt; the document is written only after the user picks Save and run as model, and a failed save is reported rather than swallowed. Every other path leaves the file untouched.

Requires

AltimateAI/altimate-dbt-integration#148, which supplies the three things this routing needs: the useNodeContext option on executeSQL, the canInjectNodeContextInSqlSnippetExecution() capability check, and the compile-error surfacing that produces the marker case 4 keys on. This PR is red until that merges and a release cuts, since it compiles only against those APIs. Verified locally against a build of that branch: typecheck clean, 642 tests passing.

Tests

6 new unit tests, plus updates to two container assertions for the widened signature.

• canRunAsModel: allowed for a known model on a supporting integration; refused when the integration cannot supply node context; refused when the file is not a manifest node
• the failure handler: offers the saved-node run when the failure was identity-related; stays silent for an ordinary SQL error; passes successful results through untouched

Verification

The behaviour was verified end-to-end in code-server against a project whose macro asserts on model.name, on the corecommand path — before, an inline preview of that model died with a bare TypeError; after, it returns node_name = stg_identity_probe with identity resolved. A scripted fixture lives in test-fixtures/inline-preview-loses-node-identity/.

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: cd87055c-be51-4a93-9b4c-5f8a72b79ac8

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[ralphstodomingo]

Docker E2E — routing verified in the running extension

Built from this branch with AltimateAI/altimate-dbt-integration#148 linked, deployed to codercom/code-server, driven against test-fixtures/inline-preview-loses-node-identity/ — a dbt project whose check_name macro asserts on model.name. Integration corecommand.

Every case below is evidenced by dbt's own invocation log, not inferred from the UI, so each shows the exact command the extension produced.

Case 3 — saved, no unsaved edits → runs as the real node

This is the reported bug (#2028), and it is gone. Cmd+Enter returns rows where master dies with TypeError: Cannot read properties of undefined (reading 'data'):

dbt show --log-level debug --select stg_identity_probe --limit 500 --output json --log-format json

node_name = stg_identity_probe — identity resolved, macro passes. No error, no hint, no second keystroke.

Case 1 — selection → still inline, unchanged

With lines 2-4 selected (excluding the macro call):

dbt show --log-level debug --inline select
  '{{ model.name }}' as node_name,
  '{{ this.identifier }}' as this_ident X
 --limit 500 ...

model.name resolves to inline_query — proof this took the inline path, not --select. Only the selected lines were sent. The stray X is an artifact of driving the editor headlessly and is incidental.

That run also had unsaved edits present, so it doubles as proof that rule 1 takes precedence over rule 4: a selection runs inline and raises no offer.

Case 4 — unsaved edits → inline, then offer

Runs inline as intended, and fails on identity:

dbt show --log-level debug --inline {{ check_name() }}\nselect ...

The offer then appears:

'stg_identity_probe' did not compile because previewing unsaved SQL runs it as an anonymous query, so model.name, this and model.config do not resolve. Saving lets it run as the real model. — [Save and run as model]

The button works. Clicking it, with the file dirty:

	before	after
disk	as wthis_identq	as ewthis_identq
dbt invocation	show --inline …	show --select stg_identity_probe …
tab state	dirty	clean

It saves the document and re-runs as the real node — only after the user chooses it. Nothing is ever saved silently.

Case 5 — integration without node-context support

The true path is proven in the product: --select routing only happens when supportsExecuteModel() returns true. The false path (Python bridge) is covered by unit tests asserting canRunAsModel returns false; it needs core mode and a restart to drive in the UI, which was not done.

Two findings worth a reviewer's attention

1. A real bug, found and fixed here. The identity marker only attached when dbt reported the compile error on stdout. dbt-core reports it on stderr, where processJSONErrors produced a plain Error — so on that path a consumer could not distinguish an identity failure from any other SQL error, which is the marker's entire purpose. Both streams are marked now. Because the stream handling lives in runShowCommand / runFusionShowCommand, which executeModel also uses, the fix threads an explicit isInlinePreview flag rather than inferring it — a --select run cannot fail for lack of node identity and must never carry the marker. Six tests pin this: both integrations × both streams, plus negatives proving executeModel failures stay unmarked.

2. A timing caveat worth being deliberate about. Immediately after a window reload or container restart, previews route inline until the manifest is parsed into the extension's memory — canRunAsModel fails safe when it cannot confirm the node is real. That is the correct conservative behaviour, but it does mean previews in the first seconds of a session can route differently from later ones.

Suite

310 library tests, 642 extension tests, typecheck clean, 0 lint errors.

Note on merge order

This PR is red until AltimateAI/altimate-dbt-integration#148 merges and a release cuts — it compiles only against executeModel, supportsExecuteModel and the error marker. Verified locally against a build of that branch.

---

Note

Case 4's quoted prompt above is superseded. Review on AltimateAI/altimate-dbt-integration#148 found that the marker this hint keys off reports that the failed run was an inline preview, not why it failed — so the old wording asserted node identity as the cause even when a preview died on an unreachable warehouse. The prompt now reads:

Previewing 'X' failed. Unsaved SQL is previewed as an anonymous query, so model.name, this and model.config do not resolve — saving lets it run as the real model.

Routing is unchanged, so cases 1-3 and the case-4 trigger below still hold as evidenced. The re-run confirming the new wording renders is pending and this note will be replaced with it.