sf-intelligence 0.8.0: the save-order answer is right on framework orgs, dead-code verdicts stop inventing confidence, and the regex router is gone

0.7.0 finished an external review. 0.8.0 started with a harder test: a plain Claude Code session reading the raw metadata of a real sandbox, against the same session with sf-intelligence attached, on eight questions an admin actually asks. Plain Claude answered all eight correctly. With the product, six. The two it lost were the product’s own fault. This release fixes those, then fixes the same class of mistake everywhere else we could find it. Nothing was added.

1. “What happens when I save this record?” is right on a framework org

Many orgs run their triggers through a framework that registers handler classes in a custom metadata type or a custom object. The vault can see the classes but not the registration rows, so the save-order answer silently left them out. Now every class the framework could run is listed as a conditional step, with the reason. Separately, a flow step used to show the flow’s first inner decision as if it were the condition for running at all. It now shows the flow’s real entry criteria, or says plainly that there are none and counts the decisions inside. On the test org, five of nine flow steps had shown a decision as their gate; now none do.

Re-run after the fix: eight of eight, both arms, and the product answered the save question in 56 seconds against 91 for grep.

2. Dead-code and deletion verdicts stop inventing confidence

Four kinds of use were invisible: a field read by a Lightning page’s visibility rule or placed on a record page, a field listed as a report-type or field-set column, a class loaded dynamically by name, and a trigger subscribed to Change Data Capture. Each is a real reference now. On the test org that removed 2,590 dangling report-type references and gave a field that drives two record pages’ visibility its two referrers back.

The verdict itself changed too. A custom field is never definitely_dead on heuristic evidence; it is likely_dead, and the answer names what was not checked (reports the vault did not retrieve, formula text, dynamic queries). Every usage row carries its referrer’s status, so an obsolete flow never reads as a second live writer. And the product no longer invents component ids: calls to Type, ApexPages or Site used to mint edges to classes the vault can never hold, 236 of them on the test org; they are gone.

3. Caveats only when the check could not run

A caveat is for what the vault truly cannot see. Several tools printed theirs unconditionally. The save-order tools now count Process Builder processes, approvals and time triggers before mentioning them. The permission-risk report reads the org’s security and session settings into its evidence instead of warning that they might matter. “Who can see these records?” shows the sharing rule’s actual filter; the placeholder (predicate not extracted) appeared 13 times in one answer on the test org and now appears zero times. Summary sentences such as “19 of 20 remote sites are active” are computed from the rows they summarise.

4. The assistant routes; the product advises

The regex router that tried to pick a tool from your wording is deleted: 7,854 lines, its offline mode and the test that kept every tool reachable through it. Your assistant already knows how to choose; what it needs is a good shortlist and honest signals. The semantic funnel gives both: it stops only on real ambiguity, labels weak evidence low, tolerates a one-letter typo, and finds the deletion and custom-metadata tools for the questions that ask for them. The analysis catalog matches any word of your question instead of all of them, which is why a nine-word question used to return nothing.

5. First-run fixes

A newcomer installed the package from a tarball and followed the README. What they hit is fixed: sfi init writes a relative vault root so a copied project refreshes, ignores the org-data files it should never have let into git, and its .gitignore is a deny-list; next actions name a command you can run (sfi refresh) rather than a slash command your editor may not have; the doctor speaks plainly; a dependency advisory on fresh installs is gone.

Upgrading

npm install -g [email protected]
sfi refresh

The refresh matters this time: the new references are read when the vault is rebuilt. If a script of yours read route_question’s route.intent, read toolCandidates instead. Everything else keeps its shape.