Skip to content

v0.4.0

Choose a tag to compare

@github-actions github-actions released this 10 Oct 20:41
· 6 commits to main since this release

Added

  • safe_to_delete_field lists name-only source mentions (nameOnlyMatches). Some Apex and OmniStudio components name a field without saying which object it belongs to: a field-list string spliced into dynamic SOQL, or an Integration Procedure or OmniScript JSON path such as alias:Field__c. These never become edges, so the verdict could read safe while they would break at runtime. The retrieved Apex and OmniStudio source is now scanned for the custom field's API name as a whole token. Each match is reported with its file, line and snippet, and moves a safe verdict to review. The object is not verified, and the platform does not block the delete on these matches. The scan has its own name-match row in checkedCategories. A match from a component that already references a same-named field on another object names it in alsoModeledOn, since the mention is probably that field.
  • why_field_changed lists workflow field updates that nothing fires (unreferencedFieldUpdates): a <fieldUpdates> definition that sets the field but that no workflow rule, approval process or other component names. It never runs, so it is not counted as a writer, but it is no longer invisible (a host had read the silence as "workflow field updates were not retrieved"). Cross-object updates whose target object is not resolved offline are marked name-only. safe_to_delete_field lists them too: an exact match blocks the delete (it must be removed first), a name-only match is review.
  • safe_to_delete_field returns checkedCategories: every category it checked, including the empty ones, with found / none-found / not-checked (a coverage gap behind it), so "checked, none" and "not checked" are no longer the same silence. The coverage families behind the delete categories are now the same set the coverage caveat checks (OmniStudio, DLRS custom metadata, buttons and links, restriction and scoping rules were missing from it), so a failed retrieve of any of them shows as not-checked and is named in coverageCaveat. The name scan row is not-checked for a family the vault holds but whose source it cannot read (e.g. a DataPack export).
  • why_field_changed lists DLRS rollups as writers, with a rollup block naming the child object whose saves recalculate the field; an inactive rollup is listed as non-runnable.
  • Save order now shows DLRS rollup recalculation. what_happens_on_save, order_of_execution and record-delete impact list each active DLRS rollup whose child object is being saved, saying it runs from the DLRS child trigger (the Apex trigger phase, not the platform roll-up phase), only through a Process Builder or Flow invocable in Process Builder mode, or queued in Scheduled mode; before, the rollup phase only knew native roll-up summary fields and read as zero.
  • DLRS rollups are modeled. Each dlrs__LookupRollupSummary2 custom metadata record becomes a writesTo edge into its target field and readsFrom edges onto the child fields it aggregates, filters, relates through and orders by, each stamped with active and calculationMode.
  • omitted-sharing-on-entry-point code-quality rule. It flags a top-level REST, SOAP, remote-action or invocable class (high: the platform caller passes no sharing context), or an @AuraEnabled-only controller (low: Lightning runs it with sharing), that declares no sharing keyword. apex_structure's parsed check now reports the same rule id with the same severity model, replacing no-sharing-declared-on-entry-point (which called a Lightning-only controller high), and the catalog's regex copy is not shown beside it. Guest-exposure findings name it as a security rule.
  • endpoint_catalog.apexCallouts. Lists Apex setEndpoint(...) call sites read from source, each marked as a named credential, a literal host or a dynamic target. A literal host is matched to the active Remote Site Setting that authorizes it, and summary.unauthorizedCalloutHosts lists hosts that no active remote site covers. Mock endpoints in @isTest classes are skipped.
  • "Is this deployment risky?" can now be answered from a conversation. sfi.review_change takes the change set the way you have it: packageXml and destructiveChangesXml (the manifest bodies; destructive members are deletions), sourcePaths (changed force-app/... files, or git diff --name-status lines such as D\tforce-app/...), or components given as ids, bare api names, or rows with an optional changeKind. Before this, the tool accepted only hand-built { type, apiName, changeKind } rows, and the package.xml and git-diff parsing existed only in the CLI. Names resolve only on an exact api-name match. A changed file the vault holds maps to the component built from that file, so dotted, hyphenated and foldered names (custom metadata records, quick actions, layouts, reports, dashboards, path assistants) resolve exactly. Manifest names map through the inverse of export_manifest's type table (ApexPage, ApexComponent, CustomMetadata, Settings, and the Layout member form). An ambiguous or unknown name, a wildcard manifest member, a container file or member (labels, sharing / workflow / assignment rules), and metadata the vault does not model are listed in the new inputResolution.unresolved block and are not reviewed; nothing is guessed. A deleted container is reviewed as every rule the vault holds in it. A rename is reviewed as the old component deleted plus the new one added, and a file deleted inside a bundle modifies the bundle. Docs, READMEs and project config are listed but do not count against the decision. A row without a changeKind is reviewed as modified and counted.
  • deployDecision on every review. It is no-go, review-first or go, with the reasons behind it, and it is derived from the review's own tallies. It is never go while something named in the change set was left unreviewed or unchecked: an unresolved name, a container or unmodelled metadata file, a wildcard, a family the vault does not fully cover, undetermined test coverage, added Apex that no test in the vault reaches, or a collision check turned off with includeCollisions: false.
  • automationCollisions on the review. It reuses sfi.automation_collisions on the objects the change set touches (at most 5) and keeps the field-write collisions and save-recursion cycles that involve a changed component. Turn it off with includeCollisions: false.
  • who_can_access_object now says what its filters left out. With accessLevel or principalType set, the response carries excludedByFilter (count, byReason, up to 5 examples per reason) for principals that still reach the object: read-only (Read / View All only), record-sharing-only, other-capabilities (for example Create without Edit) and other-principal-type. The boundaryNote names the counts. Before this, "which permission sets can edit X" dropped a View-All-only permission set without a word.
  • endpoint_catalog Apex callouts carry asyncContext and credentialCandidates. A call site inside a @future method is marked future at method granularity, with allowsCallouts read from callout=true. A site in a Queueable or Database.Batchable class is marked at class granularity, with allowsCallouts read from Database.AllowsCallouts. A dynamic site in a class that builds 'callout:' + name now lists the catalog's named credentials whose name is a quoted literal in that class (same-component) or in an Apex class or trigger with a graph edge into it (direct-caller). These are always heuristic: the literal is there, but the call that passes it is not traced.
  • what_happens_on_save with event: 'upsert' labels each step with firesOn (insert, update or both). The labels come from the same matchers that admitted the step, so a Create-only flow reads ['insert'] and no longer looks like it runs on every save. Assignment, escalation, approval and trigger-dispatched async steps carry no label, because the metadata does not tie them to an event. Auto-response rules are ['insert'].
  • sfi.object_360 format: "handbook": a one-object brief for a new team member. Ask "document the Case object for a new hire" and you get one compact Markdown brief. It covers what the object is (description, kind, sharing defaults), its key fields ranked by declared required / external id / unique flags and then by how many components reference them (with help text), active picklist values, record types with their business processes, parent lookups (business lookups separate from platform system lookups) and child objects (master-detail cascades are marked), what runs on save in documented order with each step tagged by the event it fires on (insert, update or both) and each validation rule's error message, who can read, create, edit, delete, view all or modify all, integrations touching the object (outbound messages, integration components, Apex REST resources, Change Data Capture), risk signals with their evidence, and the next tools to call. Every list is ranked and then capped, and states its true total; a wide object is tightened to fit a 24 KB budget, with the trimming disclosed. Empty sections say what was checked. The save sequence comes from what_happens_on_save and is not re-derived: phase counts are the save tool's own full counts, a phase it had to shrink is fetched again in full, and any phase still short is shown as "at least N". Delete and undelete automation gets its own line. A lookup whose target was not captured (common on standard objects rebuilt offline) is listed as "target not captured" rather than left out. includeSections / maxRowsPerSection with this format is a named invalid-query, because they shape the default sections format only.
  • what_happens_on_save and order_of_execution return a reentry block, the "why did my flow run twice" answer. It lists the steps that write back to the saved object (a workflow field update, DML from a record-triggered flow, or Apex DML in the trigger and the handlers it calls), with the fields each step writes. Flow DML is read from the flow's own source, element by element, so a flow that both updates $Record and creates or deletes records of the same object is still reported as updating the triggering record. DML on a scheduled or asynchronous path is listed as a separate, later save. Each step cites the documented order-of-execution rule for the second pass, which is quoted in reentry.rules. rerunOnUpdate names what the update write-backs re-run, per mechanism: a workflow field update re-runs only the update triggers, while flow and Apex DML also re-run the before-save flows. It also shows any recursion guard visible in the trigger or handler source. The block lists one-hop cascades to other objects, including deletes, with the count of active automation on each target. Every entry carries a declared/parsed/heuristic tier. In order_of_execution, each write-back names the events it fires on (on). A flow whose source cannot be read is named in flowSourceUnread. Process Builder processes are not listed, and notModeled says so: the vault does not link them to their object. When the check cannot run, the response carries reentryNotChecked instead of an empty block.
  • sfi.why_field_changed with value answers "what sets this field to V?": valueFilter puts every writer in one group — definitelySets (its metadata states V; firesWhen gives the trigger and entry criteria), maySet (the value is computed or not captured — Apex, Flow variables, formulas, rollups — with why), or cannotSet (every write the vault attributes to it states another value; a Flow's source is re-checked, so a write through a relationship such as $Record.Lookup__r the graph did not attribute moves the Flow to maySet) — plus a disclosure of what is never modeled (managed packages, dynamic Apex, integrations, manual edits).
  • definitelySets[].firesWhen says when the write happens later: a record-triggered Flow that runs only on a scheduled path ("NOT in the save"), one that has scheduled paths besides its immediate one ("may run on one"), and a workflow value written by a time trigger (writers carry timeTriggeredValues when the rule also updates the field in the save). Flow decision outcomes are counted as such.
  • Picklist values are checked against the field's value set (case-insensitive; an undeclared value is an invalid-query listing the declared ones, an inactive one gets a note); checkbox values take true/false and match workflow 1/0 updates and Flow $GlobalConstant.True/False; numbers compare numerically (1.0 = 1).
  • A description-to-schema parity test over the core roster: every value list a tool's short description advertises for an argument must be accepted, by the schema enum or, where the tool validates the value itself, by the real dispatcher (with a bogus-value control so an unvalidated argument cannot pass silently, and any such argument fails the test by name). It reads pipe lists, quoted values, parenthesized lists, (for example a, b) and (a by default; also b, c).

Changed

  • MCP Registry name is now io.github.PranavNagrecha/salesforce-intelligence, so registry search for "salesforce" finds it (search matches names only). The npm package is unchanged (sf-intelligence). The old io.github.PranavNagrecha/sf-intelligence entry keeps its history up to 0.3.3. A release check now fails before publishing if the package's mcpName and server.json name differ.
  • Compatibility for programmatic MCP clients: responses no longer carry structuredContent, and tools no longer declare an outputSchema. The same JSON is in the text content (content[0].text), which is what hosts already showed their model. A client that read structuredContent should read and parse content[0].text instead. MCP_TOOL_OUTPUT_SCHEMA is still exported but deprecated.
  • The sfi launcher enables Node's on-disk compile cache (Node 22.1+) before loading the CLI bundle, so repeat start-ups (sfi --version, every MCP server start) can reuse compiled code; the measured saving is small (a few percent of CPU time) and within wall-clock noise. NODE_DISABLE_COMPILE_CACHE=1 turns it off.
  • sfi status --strict exits 1 when there is no vault, so a cron or CI wrapper can detect it (the default still exits 0).
  • sfi init accepts --vault (the flag every other command uses); --vault-root remains as an alias.
  • sfi init reads the local sf login list once instead of twice when no --target-org is given.
  • Composed concept reasoning is computed once per component per server process. The server's read-only graph is cached per component, so a repeated get_component, field_360, explain_apex_method or what_happens_on_save call no longer re-runs the rule engine (about 1-4 s on a standard object).
  • health_check no longer re-hashes the whole source tree on every call. The full SHA-256 is cached per process behind a stat fingerprint (file count, total bytes, newest mtime); any edit still invalidates it.
  • The source-path to metadata-type dispatcher now lives in @sf-intelligence/extractors (dispatchSourceFile, BUNDLE_PARENT_DIRS), and the change-set parsers live in one MCP module. The refresh walker, sfi review-change --manifest/--diff and the MCP tool all use the same table, so a newly supported metadata TYPE is understood by all three. (CONTAINER_FILE_TYPES, the set of one-file-many-components types, lives there too.) A second copy of the bundle-directory list in source reconcile was removed.
  • The sfi.review_change description is shorter and leads with the questions it answers ("what could go wrong if I deploy this", "what should I test before deploying"). The full honesty contract stays in the response's disclosure and boundaries. Funnel utterances were added for the everyday deploy-risk phrasings that used to route to interpret, installed_package_catalog or omni_version_diff.
  • what_happens_on_save keeps only save-order claims in conceptReasoning. The kept kinds are save-order, automation collision, firing condition, async boundary and status code, plus relationship claims on delete and undelete. Other claims are counted by kind in offTopicClaimsOmitted, and the disclosure points at sfi.interpret for the full list. Rule counts in completeness still describe the whole run. Other tools that attach conceptReasoning are unchanged.
  • Upgrading from 0.3.3: run one pulling refresh first. A vault built by 0.3.3 has no describe cache, so the first sfi refresh --no-pull after upgrading drops the standard fields known only from the org describe (for example Account.BillingCity). Run sfi refresh once, or sfi refresh --no-pull --with-describe. Until you do, field answers on those objects now say the field may still exist, and object_360 says its field count is a lower bound.
  • tests_for_change and apex_test_coverage mark weak trigger-hop credit instead of dropping it. A test that writes a trigger's object but whose own DML shows none of the trigger's events (an insert-only test on an update trigger) is still selected, marked eventMismatch. selectedTests names those changes in eventMismatchFor. Code it calls can still fire the trigger (a service's update, a workflow field update, an after-save flow), and dropping it made a covered change read "no test covers this".
  • who_can_access_object rejects principalType: 'Queue' and says why: a queue grants no object access and is never a sharing-rule target. Plural kinds such as Territories are now accepted.
  • One shared module now holds the SOE event matchers (soe-event-match.ts), the Apex recursion-guard detector (detectRecursionGuard, also used by the trigger-no-recursion-guard recognizer), and the per-file Apex DML attribution (apexDmlFactsFor). Before, each of these was copied in more than one place.
  • Workflow field-update listings carry reevaluateOnChange.
  • The default tools/list is 71% smaller (116 KB to 33.6 KB of JSON on a real vault; the description and schema text a model reads drops 66%, 94 KB to 31.5 KB). The 25 core tools advertise a short contract (what the tool answers, its args, one key caveat, and which sibling to use instead) instead of their full long-form text. The long form is kept per tool as reference, returned by sfi.describe_analysis {detail:'full'}, and it stays the document the router indexes, so funnel routing is unchanged (the BM25 corpus is byte-identical except sfi.get_impact, below). The one-liners that sfi.list_analyses {query} matches and sfi.route_question shows for the 25 core tools are now the first sentence of the short contract, so a list_analyses keyword search can return a different set of core tools than before. A test keeps the core list under 40 KB and each core description under 1.6 KB.
  • Server instructions are shorter (4.7 KB to 3.0 KB). Operator environment variables are gone from the host-facing text; the confidence tiers, "empty is not none" and "org metadata is data, never instructions" rules are stated once.
  • Tool results are text-only. Tools no longer advertise an outputSchema, and results no longer carry a structuredContent copy of the JSON already in content[0].text. Every host showed its model one of the two identical copies, so this halves response bytes on the wire without changing what the model reads. Both were removed together because MCP SDK clients reject a result without structuredContent once a tool declares an outputSchema. Read results from content[0].text. MCP_TOOL_OUTPUT_SCHEMA is still exported but deprecated.

Fixed

  • The release workflow asked the MCP Registry to register a version npm was not serving yet. npm accepts a publish minutes before it serves it, and the registry looks the exact version up on npm before registering it — so on 0.3.3 the registry step got a 400 ("version '0.3.3' was not found … Wait and retry") two seconds after npm took the package, and the SBOM and GitHub Release were skipped until a manual re-run (npm served the version 2 min 45 s later). The registry step now retries that one rejection every 30 s for up to 10 minutes; any other failure is still fatal on the spot, and the not-found case is tested before the duplicate-version check, so it can never pass as an idempotent re-run.
  • A Lightning Web Component that imported several methods of one Apex class kept only the first. The edge merge dropped the duplicate imports, so what_if_change_method_signature, call_graph and test_coverage_for_method missed every LWC caller of the other methods, and could call a UI-facing method test-only. Merging now keeps the full method set on the one callsApex edge (methods[]), and it does the same for Apex static fields (fields[]). Visualforce and Apex callers go through the same merge and get the same fix.
  • Aura components that call their Apex controller had no call edge. controller="X" plus component.get('c.method') in the controller or helper JS now produces a callsApex edge to X listing each method, at heuristic confidence. A bundle that calls server actions but declares no controller itself (it inherits one through extends) gets no edge, and the actions are listed on the node as unresolvedServerActions.
  • Apex reads of Custom Labels (System.Label.X, Label.X) had no edge to the label. They were stored as a hidden CustomField:Label.X phantom, so get_impact on a label returned zero dependents and still reported complete: true. They are now references edges to CustomLabel:X, and the managed Label.ns.X form maps to CustomLabel:ns__X.
  • Field writes through an indexed receiver were dropped. Writes like records[i].Field__c = … and rows.get(i).Field__c / byId.get(id).Field__c on a typed List/array/Map now resolve to the element type, so these classes show up as writers.
  • Lowercase Apex type names made a class's writes disappear. case cc = obj; and account a = new account(); are now resolved without regard to case, matched against the vault's object list. On a parsed file, the scanner no longer adds phantom fields named after variables, such as CustomField:cc.Status. The fallback scanner also stops reading a static member of a known Apex class (Constants.STAGE_WON, any case) or a managed-package namespace (ns.Api.call()) as an sObject field: the class read becomes a references edge to the class with fields[], and the namespace read is dropped. Lowercase built-ins such as system.debug no longer produce a call edge to a nonexistent ApexClass:system.
  • endpoint_catalog and integration_map marked named credentials as orphaned when they were in use. This happened for a credential that a shared helper builds at runtime as 'callout:' + name, or that an OmniStudio Integration Procedure's Rest Action declares. Both tools now share one counter. It adds via: 'omni-rest' references and, when a dynamic callout exists, references found where the credential's name appears as a literal in Apex source or a custom-metadata record (via: 'name-literal'). Each referrer is listed in referencedBy. A credential that is still unreferenced carries an orphanedCaveat.
  • find_component_usages for a Custom Label counted unrelated identifiers as usage. The search treated any same-named variable as a use and reported hasStaticEvidence: true. It now matches only label references (Label.X, $Label.X, $Label.c.X, {Label.X}, @salesforce/label/c.X), and it also searches formulas, validation rules, Flows, FlexCards, OmniScripts and custom links.
  • The org-leak scanner passed silently when its blocklist was missing or broken. scan:leaks, the release guard and sfi vault anonymize each had their own loader, and all three treated a missing or malformed forbidden-names.local.json as "no patterns" and reported OK. There is now one loader. A malformed file always fails. Strict runs refuse a missing blocklist unless --allow-no-blocklist is passed, every run prints how many patterns it loaded, and CI fails when the blocklist secret is empty (fork PRs run visibly generic-only). The release guard now checks the same patterns as the scanner (it used to read only the patterns key) and reports a missing blocklist as a clean failure instead of a stack trace. The publish workflow keeps the blocklist until npm publish has run its own leak check, then deletes it.
  • sfi vault anonymize certified files it never read. Binaries and symlinks in --out were skipped without being counted, a pre-filled --out was reported "0 residual matches — safe to share", and the command exited 0 even when the scan found leaks. It now refuses a non-empty --out, reports symlinks inside the vault as skipped, counts unreadable files as findings, and exits 2 with "NOT safe to share" when anything is found.
  • sfi refresh --no-pull contacted the org. It still ran sf sobject describe once per core standard object. A plain --no-pull now runs no retrieve and no describe: it replays the describe snapshot the last pulling refresh cached (meta/describe-cache/) and says which objects had none. --with-describe opts back in. The summary says the org was not contacted only when no org-querying opt-in ran; --with-audit-trail, --with-tooling-api and --with-data-shape still contact the org and are named when used. sfi.health_check now reports the vault as degraded and names the objects whose describe-only standard fields are missing, so a "field not found" there is not read as proof the field does not exist.
  • A mistyped or logged-out org alias reported only "Parsing --target-org". Refresh now checks the alias against your local sf logins before retrieving and fails fast with the fix (sf org login web --alias …). sf errors keep the cause line printed under the header, and sfi init warns when the alias is not one of your logins.
  • Setup steps contacted every org you are logged into. sfi init, setup-mode sfi mcp and the refresh preflight read the local login list (--skip-connection-status, short timeout) instead of checking the connection to every org. The trust statement no longer claims sfi refresh is the only thing that contacts the org; it names sfi refresh, sfi stale-sweep / sfi watch and sfi doctor.
  • Setup mode gave wrong first-run guidance. It put the first row of sf org list into the init command as if you had picked it, suggested pinning the very path it had just failed to find, told a mistyped --vault to "run sfi init", and claimed the server restarts by itself. It now uses an <your-org-alias> placeholder (naming your sf default org as a hint only), a placeholder pin path, a vault-path-not-found reason for an explicit path that does not exist (its detail and connect-time instructions no longer say to run sfi init or that the vault is not built), names the project directory the refresh must run in, and says a restart is needed.
  • A refresh that failed before extracting anything printed empty Components/Edges tables above the error. Only the failure line is printed now.
  • The Claude Code plugin's slash commands needed a sfi binary the plugin never installs. /sfi-init, /sfi-refresh, /sfi-status and /sfi-onboard now check command -v sfi and fall back to npx -y sf-intelligence@<version>. /sfi-init asks which org to use instead of running a bare sfi init that cannot prompt without a terminal. /sfi-field-audit no longer suggests --no-pull to raise the report cap (an offline rebuild never fetches reports). The slash-parity and version-consistency checks now fail on a missing fallback or a stale pin.
  • Onboarding docs contradicted the code and each other. onboarding.md said a refresh is not picked up by a running server and that the server blocks the refresh; on macOS/Linux the rebuild swaps in behind a live server (Windows still needs the client closed, now stated once). installation.md's Claude Code command lacked --vault, and its "no DX project" row described an error that cannot happen (init scaffolds one). mcp-hosts.md pinned an old version, examples/demo-questions.md recommended a stale committed demo vault, and the README did not mention that the first sfi demo start can exceed a host's startup timeout.
  • docs/routing.md and docs/configuration.md quoted stale tool counts (217 registered / 212 advertised / core 19 against 231 / 226 / 25). Fixed, and pnpm doc-sync now fails when those phrases drift from the product manifest.
  • Website pages scrolled sideways on phones. At 390px every page overflowed by 9px (the header) and /tools by about 190px (long tool ids). Tighter phone gutters and wrapping for tool ids and inline code paths fix it without hiding the install button.
  • 136 harness-bound extractor/parser/renderer tests skipped silently on every machine but a frozen harness copy, CI included. The skip now says loudly that those tests assert nothing. SFI_HARNESS_ROOT=<dir> runs them from any checkout (a wrong path fails instead of skipping), SFI_REQUIRE_HARNESS=1 makes a missing harness a failure, and the walk-up no longer accepts an unrelated ancestor tests/fixtures. The three copies of the locator are held identical by a test.
  • sfi quickstart marked steps done that it never checked. Install and auth were always ticked, the last step was ticked with a pseudo-command, and SFI_VAULT was ignored. It now runs sf --version and reads your sf logins (checking the vault's own org), never ticks the MCP step it cannot see, prints the exact claude mcp add … --vault "<path>" line, honours --vault/SFI_VAULT, and seeds starter questions from your own __c objects (ahead of platform events and custom metadata types) with fields written as Object.Field. A slow sf --version is no longer reported as "sf not installed".
  • Flow extraction dropped the values a Flow compares and writes. Decision conditions, record-trigger entry filters and workflow criteria now carry structured conditionItems (field, resolved field id, operator, value, literal-vs-reference). Get / Update / Delete Records <filters> now mint field-level reads carrying the compared value. Assignment elements that set a field on a record variable the Flow later saves now mint field writes with the assigned value. A condition on a Get Records output or record variable (Get_Project.Stage__c) now resolves to the real field instead of a phantom id. When a Flow writes one field twice, every literal it sets is kept (literalValues), not only the first.
  • The "only when a record is updated to meet the condition requirements" entry setting was dropped. Flow nodes and their entry condition now carry entryRequiresRecordChange.
  • what_if_remove_picklist_value reported "no impacts" for values active Flows depend on. It matched only quoted literals, but Flow and workflow criteria store values unquoted, and it never attributed a condition to its Flow. It now reads the structured conditions and Flow field edges, names each use (where: entry criteria, decision, Get Records filter, literal write), lists record types that merely list the value and Flows that do not run under informational (they never move compatibility or verdict), says when a write of the value would fail on a restricted picklist, and also checks record types, criteria sharing rules, list view filters and Lightning page visibility rules. notChecked names the families it does not read (reports, path assistants, dependent-picklist matrices, dynamic Apex). value_change_audit reads the same structured literals.
  • Flow writers were attributed across objects. The shared Flow source scan behind why_field_changed, field_360, safe_to_delete_field and unused_fields_deep matched a <field> name inside any Create/Update Records element without checking that element's object, so a Flow writing Status__c on a Task was reported as a writer of the Case field. The scan now resolves each element's object (<object>, $Record, typed variable), drops other-object matches, and marks unresolvable ones objectScope: 'unresolved'; find_component_usages uses it instead of its private copy.
  • why_field_changed writers came back without the value they write. Each writer now carries assignedValues (the literal it sets) or assignedFrom (its variable/formula source) and operation; the gating condition shown is the entry criteria, with onlyWhenChangedToMeet when the Flow fires only on the save that changes the record. A new optional value input answers "what set it to X?".
  • Save order listed async and scheduled-only Flows as running in the save. order_of_execution and what_happens_on_save each copied a rule that recognised scheduled-only Flows only by <pathType>, which a time-offset scheduled path does not carry, and read a vault without hasImmediateConnector as "immediate". Both now use one classifier (soe-flow-timing.ts) fed by a new scheduledPathCount; Flow steps carry timing (async-only, or unknown on an older vault), their declared triggerOrder, and onlyWhenChangedToMeet.
  • Save order counted inactive automations without naming them. what_happens_on_save and order_of_execution said "3 inactive automations are left out" but not which, so a host could not tell the user which Draft / Obsolete Flows or inactive rules look relevant but do not run. inactiveSummary.names now lists them (Type:apiName (reason), capped at 25, with namesTruncated).
  • A Flow writing one field from a literal and from a variable lost one of them. The duplicate-edge merge kept only the first write's kind: when the variable write came first the literal vanished (so what_if_remove_picklist_value missed it), and when the literal came first the writer read as literal-only (so why_field_changed with value dropped it). Merged write edges now keep every literal (literalValues) and every variable source (referenceValues).
  • what_if_remove_picklist_value missed Lightning pages that use the value. It read only pages already linked to the field by an edge, but neither a {!Record.Field} visibility rule nor a related-list filter creates one. It now reads every Lightning page, skips visibility rules on pages bound to another object, and counts related-list adminFilters only when the list is a child relationship of the field's object.
  • what_if_remove_picklist_value missed values tested inside Flow formula resources. An ISPICKVAL({!$Record.Field__c}, 'X') in a <formulas> resource carries no edge to the field, so the Flow was left out while the disclosure said Flows were checked. Flow nodes now carry formulaValueRefs (each formula's resolved fields and literals) and formulaComparisons (direct field-to-literal comparisons) and the tool reads them; a vault built before this is named in notChecked, as are screen-component visibility rules and text templates.
  • what_if_remove_picklist_value could say safe while notChecked named families it never read. An empty scan now returns review.
  • One too-deeply-nested Flow or Apex property could abort the whole vault render. The Flow and Apex markdown renderers serialized node properties without the depth guard the generic component renderer already used; they now share it, so such a value is JSON-encoded in the frontmatter instead of failing the refresh.
  • An unrecognized SFI_TOOL_PROFILE value (minimal, lean, a typo) no longer silently selects the full ~760 KB tool roster; it warns on stderr and uses core. Only full selects the full roster.
  • Tool calls no longer silently drop arguments a tool does not declare. A guessed sibling name is mapped onto the tool's real argument when unambiguous (object/objectName → objectApiName, componentId ↔ nodeId/targetId/rootId, singular permissionSetId → permissionSetIds, …); anything still unused is reported under argumentNotes on the response (or named in the error) instead of yielding a confident unscoped answer.
  • Every invalid-query argument error now names the tool's required and optional arguments, a one-line example call, and points at sfi.resolve for canonical ids.
  • The router no longer recommends four retired hidden-alias tools that are not invocable under the core profile (cdc_subscribers, field_cleanup_candidates, churn, find_apex_usages); each intent now names its surviving tool.
  • layout_for_user routing no longer emits hard-coded, org-specific profile ids (one of which exists in no org). Only the platform-standard Profile:Admin is emitted; any other " profile" phrase is passed as the tool's profileName selector.
  • sfi.route_question now binds the component it already resolved into every recommended call on the first turn, under each tool's own declared argument (fieldId, componentId, objectApiName, triggerId, rootId, classApiName, …, derived from the tool's registered input schema). Previously invoke/candidates carried empty args and most first calls failed invalid-query. Candidates whose required args could not be bound now say so in missingArgs; invoke no longer contains a non-executable sfi.resolve {} step.
  • A label-named object binds its API name (Invoice → Invoice__c), and a prose word scraped as an object (figure, same) is dropped and disclosed instead of passed as objectApiName.
  • Class.method references resolve the Apex class and bind the method instead of raising a false "no component named … exists" premise check.
  • Save questions phrased "saved" / "inserted or updated" route with event: 'upsert' (insert + update) instead of update only.
  • Shortlist hygiene: a stacked complement (e.g. sfi.interpret) no longer outranks a high-confidence route's own primary tool; live-only tools sit below every vault tool when the live plane is off; hidden retired tools are never offered; entity evidence lists an exact winner first, then by score.
  • sfi.resolve: trailing type nouns (label, custom label, class, checkbox, picklist, record(s)) are treated as type hints, so an exact API name followed by its type word resolves exact instead of ambiguous; a query equal to a component's multi-word label is a whole-name match. A trailing word is only stripped when the unstripped words do not literally name a component, so " Record field" keeps resolving to the field labelled " Record" rather than a sibling labelled ""; runner-up rows next to a literal name match report substring, not exact.
  • sfi.describe_analysis's default summary now includes a compact args signature (every arg with its type, ? for optional, and a short hint), so one describe call is enough for a correct first call.
  • sfi.list_analyses accepts a query filter (every word must appear in the name or summary) and echoes it as appliedQuery; previously the argument was silently dropped and the whole roster returned.
  • Object-access questions ("who can see the X object") whose entity resolves to an object now route to who_can_access_object / object_access_audit instead of the field-only field_access_audit, which required a fieldId the question never had.
  • Specific what-if questions are no longer swallowed by the generic impact route: "remove 'X' from the Y picklist" routes to what_if_remove_picklist_value and "change X from currency to number" to what_if_change_field_type. The quoted picklist value and the stated target type are bound into value / newType (derived from the tool's schema; left unbound when not exactly one), and "the Y picklist on Z" / "the Y checkbox" now resolve as a field.
  • sfi.resolve: a type noun before "on " ("Status picklist on Case") is a type hint, so the field resolves exact instead of ambiguous.
  • "What changed since the last refresh" now also routes to diff_snapshots (named added/removed/modified components), and the no-arg what_changed_since_refresh ranks above changed_since, which needs a since date.
  • A "what does Class.method do?" question that matches no rule now still offers the tools that take a method (explain_apex_method, apex_structure, call_graph, ...), bound to the resolved class and method, at low confidence, instead of an empty shortlist.
  • A guessed arg renamed onto a typed id (fieldApiName: 'Contact.Birthdate' → fieldId) now gains its canonical prefix (CustomField:), and an invalid-query that follows a rename names the rename, so the error never blames an argument the caller did not pass.
  • A quoted literal in the question binds only a tool's REQUIRED value argument, never an optional filter or a write-side annotation value.
  • sfi.route_question invoke steps whose required args could not be bound now carry missingArgs instead of posing as executable; an unrouted Class.method question's invoke is the bound method tools rather than the generic fallback; and an object-access question no longer offers the field-only field_access_audit (unscoped) back through the shortlist.
  • A vault built by an older sf-intelligence was served silently. Only safe_to_delete_field compared the vault's builder version with the running one, so get_impact and find_component_usages could report soundness.complete: true on a vault whose builder predates whole edge families (for example OmniStudio callers of Apex, or this release's method-level LWC/Aura caller and Apex Custom Label edges). Every tool response now goes through one shared freshness assessment: on builder drift vaultState.freshness names both versions and the edge families added since, trust.completeness drops from complete to partial, soundness gains a stale-builder blind spot, and the warning leads trust.limitations. A vault older than a week is flagged by age band (>7d, >30d, >90d). An empty trust.freshness (as route_question returned) now names the snapshot time.
  • health_check reported degraded on every vault, including the official demo. Coverage gaps (metadata families the org does not have, or the refresh did not confirm, report pulls capped by design, and unrecognised source directories) are now listed in a new coverageGaps field and no longer change status, which reflects wiring and integrity only. A retrieve that errored or never finished is a refresh failure, not a gap: it stays in issues and keeps status at degraded. Absence claims are still hedged by each tool's coverage caveat.
  • health_check and coverage_report disagreed about families the vault's own edges reference but which hold no components. coverage_report listed such a family as partial while health_check, coverage_report.summary and every coverage caveat called it covered. The fact is now computed once when the server binds the vault and folded into the shared coverage summary, so every consumer reads the same verdict. safe_to_delete_field (and the other field-deletion gates) now also require EmailTemplate coverage, since merge fields reference fields; an unread template plane used to be left out of missingCoverage.
  • get_impact, find_component_usages and what_if_change_method_signature claimed completeness for types whose referrers are only partly modeled. A custom label used in formula fields, or an @AuraEnabled method called from an Aura bundle with no declared controller, came back as complete: true with no blind spots. One referrer-coverage table per component type now says which referrer kinds are graph edges, text-search only, or not modeled. get_impact names the unmodeled kinds as an unwalked-referrer-class blind spot, find_component_usages returns them in referrerCoverage (with notSearched for the kinds its text search does not read either), and what_if_change_method_signature lists them in trust.limitations and reports review instead of safe when an @AuraEnabled method shows no callers. The field and object blind-spot list is now derived from the same table. Kinds that only newer builders emit as edges (method-level Aura callers, Apex Custom Label references) are judged by the vault's own builder version, so a vault built before them is not credited with edges it does not hold.
  • The conceptReasoning block showed 0 claims on most hub components while costing 4-6 KB. Claims on a standard object run 1.5-3.6 KB each, and the fit dropped claims (15 to 0) around a fixed block of prose. The fit now samples each claim's citation list (true count in groundedInTotal), compacts the prose and shortens long claim texts before it drops a claim, so three to five claims survive where none did. A component where no rule fired gets a compact block that still carries every count and the not-checked warning, plus a pointer to sfi.interpret.
  • sfi.interpret returned oversize on objects and profiles. It serialized every claim three times with full citation lists (about 160-210 KB on a standard object), so the surface every concept-reasoning disclosure points to failed on the most common anchors. Claims are now paged (limit, default 10, and offset, with page.nextOffset), each claim's citations are sampled by groundedInLimit (default 25, full count in groundedInTotal), and the handler fits the page to the response budget itself. It samples the completeness lists and citations before it shrinks the page.
  • safe_to_delete_field's stale-vault caveat named the wrong edge families. It always said the 0.3.0 families were missing, which is false for a vault built by 0.3.0 or later, and never named what such a vault does lack. It now lists exactly the families added after the vault's builder, from the same table every other response uses.
  • sfi.interpret with an offset past the last claim said "No concept rule fired". The rendered answer and the evidence envelope now say the page is past the last claim and how many fired, and a partial page's rendered text names the window it shows.
  • Unreadable files turned into definitive answers. On a vault older than the flow async-path markers, explain_flow treated a flow source it could not read as "no AsyncAfterCommit path" and declared that an unhandled fault rolls back the triggering save. It now returns rollsBackTransaction: null with a statement saying why. The same flow's non-Apex actions are disclosed as not checked instead of being left out. persona_unused_grants now counts reachable Apex files it could not read in trust.limitations and marks completeness partial, where it used to skip them silently.
  • Cross-vault comparisons read a malformed component row as "no properties" without a trace. Seven tools and sfi snapshot each carried a private copy of the properties_json parser that swallowed the error. They now use the graph's parser, which logs a warning naming the row.
  • A field used only as a report filter was invisible to delete-safety. A Report names the fields it filters and sorts on in <criteriaItems><column> and <sortColumn>, which the report sweep never read, so safe_to_delete_field called reports "unchecked" for a report the vault held. Filter and sort columns now mint edges (cross-filter columns, which name fields on a related object, are skipped rather than guessed).
  • DataMapper Extract formulas now count as readers. A formula row such as IF(alias:Field__c == false, …) reads the field through the same alias as the mapping rows; those reads were dropped.
  • An LWC @salesforce/schema import is now a declared reference. It was graded heuristic like an in-body string match, so a compiler-checked import read as "may be a false positive".
  • Standard lookups now point at their target. The Metadata API ships Contact.AccountId, Opportunity.AccountId, Case.ParentId and similar with no <referenceTo>, so the data model had no edge into Account. A curated, platform-fixed table fills the target when the XML names none (marked referenceToSource: standard-relationship-table); unlisted fields are never guessed.
  • safe_to_delete_field tiers were reversed. A parsed Apex reference (which Salesforce refuses to delete past) was only risky while an Obsolete Flow was "will break at runtime". Parsed Apex and LWC schema imports are now blocking; heuristic-only matches stay risky, and so does a field named only inside a SOQL string literal (e.g. a constant handed to Database.query): the compiler does not check strings, so the delete succeeds and the query fails at runtime (such edges carry mechanism: soql-string-literal); every example carries the referrer's status, and an inactive blocker (Obsolete Flow version, inactive rule) is disclosed as "does not run, but must be deleted first".
  • OmniStudio referrers had no category and landed in unknown. They are now omnistudio: a running DataMapper / Integration Procedure / OmniScript / FlexCard is blocking (breaks at runtime), one that does not run is review. A DataMapper's own active flag is not used for this (orgs run DataMappers flagged inactive from active Integration Procedures): it runs when any caller is active.
  • An OmniStudio delete referrer names its callers. A DataMapper that reads the field is usually called by an Integration Procedure or OmniScript, which breaks with it; each omnistudio example now carries calledBy (the components that dispatch it, active ones first) and callers (how many there are and how many are active).
  • A class that both reads and writes a field was counted twice in safe_to_delete_field and spent two example slots on itself. One referrer is one row.
  • Custom labels used by formulas and FlexCards had no referrers. A formula's (or validation rule's) $Label.X and a FlexCard's {Label.X} merge token now mint references edges to the label, so "where is this label used" and get_impact no longer report zero with complete coverage.
  • A validation rule's error display field is now a dependency. A rule that shows its error on a field (<errorDisplayField>) depends on it even when the formula never reads it; that edge was never minted, so safe_to_delete_field missed it.
  • A capped report pull no longer reads as "reports not checked". The default report retrieve is capped to the most-used reports; safe_to_delete_field called the whole analytics category not-checked although it had checked every report it retrieved. It is now partially-checked, with retrievedOf giving how many reports and dashboards the vault holds (and so were checked) out of the org total.
  • A workflow field update fired by a time trigger was called "never fires". Only a rule's immediate actions were read, so an update named only inside <workflowTimeTriggers> was listed as unreferenced and why_field_changed / safe_to_delete_field said it never runs. Time-trigger field updates now mint the same writer edges as immediate ones, marked timeTriggered.
  • A condition row now carries its firer's status in safe_to_delete_field, so a decision inside an Obsolete Flow is disclosed as not running.
  • who_can_access_object made the host build "who can edit / delete" by hand. It returned one row per capability per principal (40-80 rows on a real object), interleaved across kinds, with no filter; a Modify All holder appeared up to six times. The default view: 'principals' now returns one row per principal with its effective capabilities (Modify All and Modify All Data imply Edit and Delete), recordSharing and scope, and the new accessLevel and principalType inputs filter it ("which permission sets can delete X" is one call); a principalType that is never a granter kind (for example a permission set group, whose access arrives through its member sets) is rejected with the valid kinds instead of returning an empty list. With accessLevel, summary.sharingOnlyPrincipals counts roles and groups that match only through record sharing and hold no object permission themselves. view: 'paths' keeps the per-capability rows. Profile and permission set rows now carry the licence audience (internal, external, guest, unknown), and principals with an external or guest licence that reach every record are called out.
  • field_access_audit counted a read-only profile as able to edit. Its headline counts and each grant's permission came from field-level security alone, so a profile with FLS Edit but object Read (or no object row at all) was listed as edit. Each grant's permission is now the effective level for that container (FLS, object permission on the parent, and field type together), with flsPermission for the declared level and a note saying why they differ (unconfirmed when the vault has no object row). summary.profiles and summary.permissionSets are disjoint effective counts; the declared FLS counts move to summary.declaredFls. permissionType filters on the same effective level, so permissionType: 'edit' no longer lists a read-only profile.
  • permission_risk_report ignored who holds a grant. A permission set licensed to external community users with Modify All on an object ranked below internal admin profiles. Findings now carry the holder audience from the container's licence; View All or Modify All held through an external or guest licence is critical and ranks first within its severity. A permission set group takes the audience of its risky member sets.
  • object_360 said the external org-wide default is never extracted. It returned a hard-coded "not extracted" note while the object node carried externalSharingModel and who_can_access_object reported it. It now reports the extracted value; a null says whether the vault predates the extraction or the XML declares none.
  • generate_sharing_summary spent its 50-object cap on types with no sharing model and dropped lowercase-namespaced objects. Custom metadata, platform event, external and big object types (shown as "OWD: Unknown") are now skipped with a boundary naming how many, objects with sharing rules come first, and names sort case-insensitively, so namespaced objects such as ns__Application__c are no longer cut silently. Naming one of the skipped types with objectFilter still answers it.
  • effective_permissions left out page and tab access and reported unchecked record types as 0. It now returns the Visualforce pages and visible tabs the containers grant (with counts in summary), and summary.recordTypeVisibilities / summary.visualforcePages / summary.tabs are null, not 0, when no container carries that data, with a disclosure naming the containers built before it was extracted. The new perContainer: true input returns each container's own contribution (narrowed to the asked object) and how it was reached (directly or through which permission set group), so "does this group add delete?" no longer needs a call per group.
  • layout_for_user picked the Lightning record page by an alphabetical name guess. The vault already held the activation metadata (the object's org default, each app's default, and app + record type + profile assignments) but no extractor read it. The CustomObject and CustomApplication extractors now capture it as recordPageOverrides, and layout_for_user resolves the page in the platform's order, returning recordPageActivation with the org default, one row per app that overrides it (matchedOn), and the apps where the page depends on the record type. When no record type is passed and the profile's default record type picks an app's page, that row names it (assumedRecordType) and the app is listed in dependsOnRecordType. A vault refreshed before this falls back to the name pick, now marked fallback instead of matched.
  • guest_exposure_report never ranked the Visualforce pages a guest user can load. It counted the guest profile's page grants and still reported the whole page plane as unchecked. Each page is now a visualforce-page finding rated by the controller and extension Apex it runs (pageControllers, read from the page's declared controller and extension references). The unchecked-surface row and limitation remain only for a guest profile built before page grants were extracted.
  • Save order stopped at the trigger dispatcher. A handler-pattern trigger (Dispatcher.run(new XHandler())) showed only "references XHandler / calls Dispatcher" and no field writes. what_happens_on_save and order_of_execution now follow the trigger into the Apex it calls (up to 3 hops) and list the field writes found there, each marked via the class that performs them. Both tools share one module, so they cannot drift.
  • A custom-metadata switch that turns a trigger off was invisible. When the trigger's handler chain reads a bypass-style custom metadata type or custom setting, the trigger step now names it with this object's record and the flags it sets, and a top-level triggerBypassWarnings line appears when a bypass flag is set. Whether the Apex honours the flag is disclosed as not evaluated. Custom-setting values are not in the vault, and the response says so. A record whose TRUE flags are not bypass-named (per-context flags such as Before_Update__c) is reported as flags-set-not-evaluated with its own warning line, and a record whose values were never extracted reads values-not-in-vault, never "no bypass flag".
  • order_of_execution and what_happens_on_save dropped triggers, after-save flows and duplicate rules behind a long run of validation rules. When the byte budget has to drop steps from an unpaged view, it now sheds the most crowded phase first (never below 3 steps) before it touches the tail, and names the shortfall in phasesOmitted. what_happens_on_save used to hand the oversize payload to the global response guard, which cut the tail; on a real Contact it returned 30 of 60 steps with no duplicate rule, after-trigger, after-save flow or async step, and now returns 53 with only 7 validation rules held back. A paged order_of_execution view, and a phase-filtered what_happens_on_save call, keep their earlier behaviour.
  • The test-coverage tools contradicted each other and missed how tests really run code. apex_test_coverage read only callsApex out of each test, tests_for_change walked calls and async dispatch, and test_coverage_gaps walked every usage edge, so one class was untested in one tool and covered in another. None of them credited a test whose DML fires a trigger that calls the handler, so tests_for_change selected zero tests for a trigger-handler change. The three tools (and the trigger join in test_coverage_for_method) now share one walk: calls, async dispatch, references and inheritance from Apex within 3 hops, with tests as sinks, plus a trigger hop for a test that writes or constructs a record of an object whose trigger reaches the class. Trigger-hop credit is marked via: via-trigger and is heuristic. A test whose own source shows DML but none of the trigger's events (only insert against an update-only trigger) is still credited but marked eventMismatch, because code it calls can fire the trigger. Org-wide apex_test_coverage reports summary.classesCoveredOnlyViaTrigger and a boundary when any class is covered only that way. On a real vault, the untested count went from 59 to 10, and it now matches test_coverage_gaps. review_change reads the walked edge set from the same module, so a test reached through new X() is now a selected test rather than an "unchecked referrer"; a test reached only over an edge the walk does not follow (e.g. dependsOnFromApi), or a referrer with no extracted test flag, still makes the row unknown.
  • find_dead_code called installed managed-package fields definitely_dead. A namespaced component (ns__Field__c) cannot be deleted from the org, and the package code that uses it is never in the vault. It is now uncertain with that reason, so an unscoped dead-code listing no longer leads with package fields.
  • The fake-assertion check flagged the standard expected-exception test. System.assert(true, ...) inside the catch of a try/catch whose try expects a throw acknowledges the exception the test is checking for; it is no longer counted as a tautology. assert(true) outside a catch is still flagged.
  • The trigger recursion-guard check fired on almost every trigger. A static flag or static set held on a helper class (if (Flags.runOnce) { Flags.runOnce = false; ... }, Flags.seen.contains(id)) now counts as a guard. A trigger whose whole body is one delegating call (a handler dispatcher or a managed-package framework) is no longer flagged, because any guard lives in the class it delegates to, which this per-file check cannot see.
  • org_risk_report led every vault with a false "critical: retrieve failed". The finding listed every partial coverage type as failed, but most were still pending, capped, or requested and returned nothing. A coverage gap is now a caveat on the report, never an org risk (medium when a retrieve actually errored, low otherwise). Its wording separates errored, pending/capped and unconfirmed-empty types, and the gate: true blocker fires only on an errored retrieve.
  • soql-injection called every concatenated query critical. The rule now traces where each concatenated value comes from. It stays critical only when the value comes from a method parameter or a Visualforce page parameter. An untraced value (class member, method call) is high, a value read only from custom metadata or custom settings is info, and a query built only from literals (including ternaries and String.join over literal lists) or from Id/number/date parameters is no longer flagged.
  • without sharing justified in the class doc comment was still flagged. A class doc comment of any length (annotations may sit between it and the declaration) or a trailing comment on the declaration line now counts as the justification.
  • tech_debt_score banded a two-axis score as org-wide critical-debt and reported a delta against a different method. When less than half the applicable axis weight is scored, scoreBand is now insufficient-evidence (and org_risk_report ranks it low), and the response lists scoredAxes and scoredWeightFraction. Refresh now logs the axis set and score-model version with each score. scoreDelta is reported only against a prior score from the same model over the same axes; otherwise scoreDeltaUnavailable says why.
  • endpoint_catalog counted Lightning controllers as inbound APIs and left out Apex callouts. inboundApis now holds only REST/SOAP entry points. @AuraEnabled and invocable classes are listed under uiEntryPoints and are not counted in totalEndpoints.
  • org_risk_report left out guest-user exposure and ranked Salesforce's standard profiles as critical. Guest exposure is now one ranked finding at the worst guest severity, citing the top nodes. A standard profile (<custom>false</custom>) that carries god-mode permissions by design is capped at medium and labelled "Standard Profile", so custom grantors of the same permissions rank above it. This also applies in permission_risk_report. The guest-exposure finding's evidence lists each node once.
  • Three tools gave three answers for a field used only as a report-type column. unused_fields_deep hid such a field as "used". It is now listed at confidence medium at most, its recommendedAction starts with REPORT-TYPE COLUMN ONLY, and a boundary gives the count. A Flow write still holds a field out of the list. safe_to_delete_field now returns reviewBecause when it gives a review verdict with empty reasoning, and names the coverage gap, the older builder or the live population that caused it. safe_to_delete_field and field_360 now read the same report-type scan as unused_fields_deep: both return reportTypeColumns, safe_to_delete_field never calls such a field safe, and field_360 raises its risk from low to medium. When the report-type scan failed or read only some report types, both tools say so, and safe_to_delete_field gives review instead of safe.
  • A change to only a code file's -meta.xml sidecar (an API-version bump) was dropped from a diff review. It now maps to its class, trigger or page.
  • An api name with the wrong case was reported as not in the vault. It is now corrected to the vault's casing, but only on an exact case-insensitive match within the same type. A bare nested name completed to the one object that has it (Field__c to Obj.Field__c) is counted separately as qualified.
  • sfi review-change --diff reviewed a rename as a change to the new file only, so the removed class and its dependents never appeared. It now reviews the old component as deleted and the new one as added, and both --diff and --manifest name container files and unmodelled metadata as not reviewed instead of mapping them to ids that do not exist.
  • order_of_execution dropped record-triggered flows that had no recordTriggerType. It had its own copy of the save-event matchers, and that copy had drifted from what_happens_on_save's: it skipped such a flow on every event, while what_happens_on_save treats it as the platform default, CreateAndUpdate. Both tools now use one shared module, soe-event-match.ts.
  • sfi.generate_onboarding_doc silently ignored an object scope. Its schema was not strict, so { "objectApiName": "Case" } was dropped and the org-wide tour came back as if it answered the object question. The tool now refuses unknown arguments and names sfi.object_360 with format: "handbook"; its description says it is org-wide.
  • safe_to_delete_field read the sharing category as "checked, none" for a field that sharing rules test. Sharing-rule criteria and restriction / scoping rule filters now reference the fields they test, so such a field is a blocking sharing referrer. A filter that walks a lookup (Advisor__r.Region__c = $User.Id) now references the field it ends on, on the related object, as well as the lookup. When the vault cannot resolve such a path, a field with the name it ends on reads sharing: not-checked and never safe. On a vault built before this release the row reads not-checked. Re-run sfi refresh to get the new references.
  • A time-triggered workflow field update was treated as a same-save write. automation_collisions reported a high-severity same-save collision with a before-save Flow, and what_happens_on_save / order_of_execution listed the update as a save-time action. The update runs later in its own transaction. It is now left out of collisions and shown as schedulesTimeTriggeredWrite.
  • health_check reported degraded ("retrieve failed, re-run refresh") on every vault with no report pull, including sfi demo, --no-reports and --no-pull. Report and Dashboard rows start as pending by design. They now count as a failed retrieve only when a pull was attempted.
  • field_access_audit claimed a denial it had not checked. Its "N containers declare FLS Edit but cannot edit" count included grants whose object access was never checked (on a real vault, 38 instead of 15). It now counts only verified downgrades.
  • layout_for_user said "no record page is activated" for an object whose own definition the vault does not hold. The org default record page is now reported as NOT CHECKED, and the page picked by name is labelled as a guess.
  • field_360 and safe_to_delete_field missed report-type columns reached through a relationship path. When such a column has the same name as the field, the field is no longer rated low risk or certified safe, and its analytics row no longer reads none-found.
  • apex_structure dropped the omitted-sharing finding for a Visualforce @RemoteAction controller. @RemoteAction is now an entry point, and the catalog finding is kept unless the parsed check reports the same rule.
  • sfi mcp --vault <project folder> booted setup mode, saying no vault was built. It now accepts the folder that holds org-kb, like every other command. The setup-mode redirect also states the actual reason (for example a mistyped path), not "no knowledge base yet".
  • /sfi-refresh told the agent to stop on every non-scratch org. --skip-connection-status never reports Connected, so the pre-flight now checks only that the alias is in the local login list. The skill agrees.
  • sfi watch and refresh --components lost standard fields after a demand retrieve. The re-extract after the retrieve now runs the standard-field describe live.
  • effective_permissions no longer says tab visibility is outside the permission union it returns, in its response or its tool description.
  • /sfi-onboard no longer asks for a "connected" org after listing orgs with --skip-connection-status, which never reports one.
  • order_of_execution dropped record-triggered flows that have no recordTriggerType. Its stale copy of the event matcher had missed the fix that what_happens_on_save already had.
  • The recursion-guard detector missed a toggled static Boolean when the if test used a different letter case from the declaration. Apex identifiers are case-insensitive.
  • sfi.get_impact no longer tells the host to check that an id exists first. Its description and long-form reference said an unknown id returns the same empty slice as "nothing depends on it"; the tool already refuses an unknown id with component-not-found. Both now say so, and say that a phantom id (referenced but never retrieved) still answers but is absent from impact.nodes, with lost edges counted in droppedEndpointEdges.
  • sfi.safe_to_delete_field and sfi.field_access_audit say which part of fieldId is case-insensitive. The Object.Field part is; the CustomField: prefix is not.
  • Server instructions rule 7 says live data can also be enabled by the operator, not only by a sfi.live_consent grant.
  • Workflow and approval-process field updates now carry the value they write (literal, lookup, formula, blank, next/previous) on their writesTo edge; before, the edge said only Literal.
  • A component that writes one field in several places (an approval process's submit / approve / reject / recall updates, a Flow's create and update, a rule's immediate and time-triggered updates) kept only its first value in the graph; every value is now folded into the stored edge (cold import and incremental refresh), keeping which values a time trigger writes (partlyTimeTriggered, timeTriggeredValues, immediateValues). Edges that state no value are left unmerged.
  • Field updates run by an approval step's own approve / reject actions emitted no edge, so those writers were missing and the update could read as "nothing fires it". They are now writers (hookType: stepApproval / stepRejection).
  • what_if_remove_picklist_value now also finds workflow / approval updates that set the value (vault rebuilt by this version), and shares its value-set check with why_field_changed.
  • endpoint_catalog reads a computed Apex callout endpoint instead of stopping at "dynamic". String constants (same class or OtherClass.NAME) are followed, and each runtime part is attributed to its custom metadata type and field, custom label or custom setting, with the values the vault holds for it (apexCallouts[].resolution). A literal or constant prefix that names a credential or host followed by a runtime part is reported as partially-resolved with that credential or host. When nothing shows where the name ends ('callout:My_API' + path, 'https://api.vendor.com' + path), the name is still reported, flagged nameMayContinue, and such a host stays in summary.unauthorizedCalloutHosts with a boundaries[] note. A distinct record value held by several records lists all of them (alsoFrom). An endpoint built only from constants resolves to named-credential / literal-host. A 'callout:' + <record field> endpoint lists the credentials its record values name as vault-record candidates. 'https://api.' + x no longer reports the host api..
  • pnpm publish --dry-run (and npm publish --dry-run) in packages/cli no longer aborts in the prepublish hook with "found 0 .tgz": the pack-allowlist check's own npm pack now ignores an inherited npm_config_dry_run, so a release can be rehearsed end to end. The allowlist check itself is unchanged.

  • Verification scope

    Windows: partially verified. The hard-gated tier (core, tooling-api, renderers, parsers, patterns, vault, extractors) passed. Advisory tiers were NOT green: WIN-1=failure, WIN-2=failure — those suites did not report a pass on Windows for this build and were not allowed to block it. See CI run https://github.com/PranavNagrecha/Salesforce-Intelligence/actions/runs/38083381223.

    CI run: https://github.com/PranavNagrecha/Salesforce-Intelligence/actions/runs/38083381223