Developer Apex Refactor

Developer Apex refactor

Usage & discovery (§C3 contract)

For "where is X used / who references X / what depends on X" — for ANY component type — call sfi.run_analysis with { "name": "sfi.find_component_usages", "args": { … } }, or the family specialist (find_code_usages for code, find_field_anywhere for a field, layout_assignments for a layout). Route by VERB: describe questions (what is / list / what values) use describe tools, NOT usage tools. Never improvise a multi-tool fan-out without citing evidence tiers (graph edge confidence + grep text-match). An empty result is "no static evidence in the vault" — NEVER "nothing uses this".

Overview

The developer persona has one headline tool, sfi.find_code_usages, with an Apex-only narrowing MODE.

• sfi.find_code_usages (v1.4 broaden) is the default. Given a targetId, it walks incoming edges of type readsFrom, writesTo, callsApex, or references whose source node is any of ApexClass, ApexTrigger, LightningComponentBundle, AuraDefinitionBundle, VisualforcePage, or VisualforceComponent and returns each referrer alongside the edge's metadata. It answers "which source files anywhere in the code surface touch this component?".
• Apex-only narrowing. Pass nodeTypes: ['ApexClass','ApexTrigger'] (and, if desired, edgeTypes: ['readsFrom','writesTo','callsApex']) to get the Apex-source-only view — the same answer the former sfi.find_apex_usages tool gave (now folded into this tool as a hidden back-compat alias). Use when the user's question is explicitly Apex-scoped ("where is X used in Apex", "refactor Apex method", "Apex callers only") — the smaller result list keeps the surface focused.

sfi.find_code_usages answers the developer-persona question — "which source files touch this component?" — that developers ask before renaming a field, refactoring a method, or removing dead code.

Edge confidence varies by producer and edge type. Apex edges come from two producers. The default-on parser-grade Apex AST pass (source: 'apex-ast', confidence: 'parsed') resolves receivers through a symbol table and emits most of the field reads/writes and cross-class calls on a current vault; it is supplemented by a heuristic recall scanner (source: 'apex-scanner', confidence: 'heuristic') that backfills files the AST could not parse and patterns it does not resolve. Cite each edge's own confidence — do not assume Apex means heuristic. The LWC/Aura/VF scanner emits declared confidence for the unambiguous cases (LWC @salesforce/apex/Class.method imports, VF controller=/extensions= attributes) and heuristic confidence for everything else (LWC field reads, Aura attribute references, VF expression bindings). The result lists from either tool will include false positives and miss real references. A clean find means "the scanner saw these matches; verify before refactoring," not "these are the only places this is referenced." The Boundary disclosure section below names what the scanners cannot see.

v1.4 broadens to LWC/Aura/VF

sfi.find_code_usages is the v1.4 broader tool. Use it for any code-persona question that is not explicitly Apex-scoped — and even in many Apex-scoped questions, the broad (default) result is the "right" answer because real refactors care about LWC and VF callers too. Narrow to nodeTypes: ['ApexClass','ApexTrigger'] only when the user pins the answer to Apex.

Key confidence rules to cite when reporting v1.4 results:

• LWC apex imports are declared. A statement like import processOpportunity from '@salesforce/apex/OpportunityService.process' produces a callsApex edge from LightningComponentBundle:opportunityCard to ApexClass:OpportunityService at confidence: 'declared' — the import path is the declaration; there is no inference. Cite as such.
• LWC field reads are heuristic. Per the v1.4 R3a report, the scanner does not distinguish a schema-import field (e.g., import FIELD from '@salesforce/schema/Account.Industry') from a body-text field reference; both default to heuristic. State the caveat explicitly when reporting.
• Aura field accesses are heuristic. The scanner pattern-matches Object.Field__c shapes in the markup and the controller/helper JS; it does not resolve v.{attr} framework attributes (those are intentionally excluded — see Boundary disclosure).
• VF controller / extension references are declared. A <apex:page controller="OpportunityController" extensions="OppExt1,OppExt2"> produces three declared references edges from the VF page to the three classes. The attribute IS the declaration.
• All VF/Aura field touches and expressions in HTML are heuristic. Field name strings inside JS literals, dynamic binding expressions, and attribute-embedded computed names all default to heuristic — the scanner cannot resolve them symbolically.

When the user's question is broad ("what touches this from code?"), default to sfi.find_code_usages and surface the per-tier breakdown. When the user pins to a single tier ("what LWC components touch this?"), call sfi.run_analysis with { "name": "sfi.find_code_usages", "args": { … } } with nodeTypes: ['LightningComponentBundle']. Only narrow to nodeTypes: ['ApexClass','ApexTrigger'] when the user is explicit about Apex-only.

When to fire

Fire on developer-flavored phrasing. Concrete triggers:

• "Where is X used in Apex…" — "where is Account.Industry__c used in Apex?", "where do we read/write this field from code?", "what Apex code touches this object?". → sfi.find_code_usages.
• "Where is X used in code…" — "where is Industry__c touched from the codebase?", "what reads this field anywhere in code?", "every code reference to MyClass". → sfi.find_code_usages.
• "Find LWC / Aura / VF referrers…" — "find LWC components reading this field", "which LWC bundles call this Apex method?", "what Aura components reference this object?", "what VF pages touch this field?", "what Visualforce uses OpportunityController?". → sfi.find_code_usages with nodeTypes narrowed (e.g., ['LightningComponentBundle']).
• "Who reads / writes / calls…" — "who reads Contact.Tier__c?", "what classes write to Industry__c?", "what classes call OpportunityService.process?", "who invokes MyClass.run?". → sfi.find_code_usages (default broad answer) or narrow to nodeTypes: ['ApexClass','ApexTrigger'] if the user pins to Apex.
• "What Apex/code references / touches…" — "find all Apex references to this field", "what code touches this trigger?", "what references MyClass?". → broad vs. narrow split as above.
• "Is it safe to rename…" — "is it safe to rename this Apex method?", "can I rename processOpportunity?", "what breaks in code if I rename this field?". → sfi.find_code_usages is the safer default because a rename breaks LWC @salesforce/apex imports and VF controller attributes too, not just Apex callers.
• "Refactor / trace…" — "refactor Industry__c across Apex", "trace this method through the codebase", "show me the call sites for MyClass.helper". → broad by default; narrow if user pins.

When NOT to fire

Defer to another skill when:

• The user asks for full cross-metadata impact ("what breaks if I change X?", "what depends on this field?", "is it safe to deprecate?"). Those questions need Flow, layout, validation rule, and permission edges too — defer to architect-impact-analysis → sfi.get_impact.
• The user asks a schema lookup question ("what fields does Account have?", "what's the type of Industry__c?"). Defer to answering-org-questions → sfi.list_components or sfi.get_component.
• The user asks about a naming pattern or convention ("what's our convention for status fields?"). Defer to recognizing-naming-conventions → sfi.get_naming_convention_report.
• The user asks where a literal string appears in Apex ("find any class that mentions Database.upsert"). That's a grep question for sfi.search_apex_source directly — no graph lookup needed.
• The user wants Flow callers ("what flows call this Apex class?"). sfi.find_code_usages (in any node-type narrowing) deliberately excludes Flow-emitted callsApex edges; route through architect-impact-analysis for the full callsApex picture.

Steps

Walk these in order. The pattern is "tool gives the file; grep gives the line."

Step 1 — Identify the target component ID

Translate the user's reference into a canonical ID. Both tools accept the same targetId formats — what differs is the referrer type, not the target. Common targets:

• CustomField:{Object}.{Field} — e.g., CustomField:Account.Industry__c.
• ApexClass:{ClassName} — e.g., ApexClass:OpportunityService. The v1.4 broadening means LWC @salesforce/apex callers and VF controller= references will now surface here too.
• ApexTrigger:{TriggerName} — e.g., ApexTrigger:AccountTrigger.
• CustomObject:{ApiName} — e.g., CustomObject:Account (returns code referrers of the object itself; rare).
• LightningComponentBundle:{name}, AuraDefinitionBundle:{name}, VisualforcePage:{Name}, VisualforceComponent:{Name} — when the target IS a frontend component (e.g., "what references this LWC bundle?").

If the user's phrasing is unambiguous, translate directly. If it could match multiple components, run sfi.search_components first and confirm the top match with the user before proceeding.

Step 2 — Call sfi.run_analysis with { "name": "sfi.find_code_usages", "args": { … } }

Default invocation (broad — the v1.4 tool):

{ "targetId": "CustomField:Account.Industry__c" }

LWC-only narrowing:

{
  "targetId": "CustomField:Account.Industry__c",
  "nodeTypes": ["LightningComponentBundle"]
}

Apex-only question (narrow nodeTypes for the smaller, Apex-scoped surface — same parsed-plus-heuristic edges, just filtered to ApexClass/ApexTrigger referrers; this is the folded-in find_apex_usages view):

{
  "targetId": "CustomField:Account.Industry__c",
  "nodeTypes": ["ApexClass", "ApexTrigger"]
}

Optional parameters (sfi.find_code_usages):

• limit — defaults to 50, max 500. Real components rarely have more than a handful of code referrers; the default is almost always right. Raise it only when the user asks for an exhaustive sweep.
• edgeTypes — defaults to all four (readsFrom, writesTo, callsApex, references). Narrow only when the user pins the question ("which classes write to this field?" → ['writesTo']). An empty array is allowed and returns an empty result.
• nodeTypes — defaults to all six code node types. Narrow to a single tier when the user pins it (['LightningComponentBundle'] for LWC-only, ['VisualforcePage', 'VisualforceComponent'] for VF, ['ApexClass', 'ApexTrigger'] for the Apex-only view — the folded-in find_apex_usages behavior). An empty array is allowed and returns an empty result.

The hidden sfi.find_apex_usages back-compat alias still resolves by exact name (it delegates here with nodeTypes fixed to ['ApexClass','ApexTrigger'] and edgeTypes restricted to the Apex triad), but it is un-advertised — prefer sfi.find_code_usages with the narrowing above.

The response shape:

{
  "data": {
    "usages": [
      {
        "id": "ApexClass:OpportunityService",
        "type": "ApexClass",
        "apiName": "OpportunityService",
        "edgeType": "readsFrom",
        "source": "apex-ast",
        "confidence": "parsed",
        "properties": { "offset": 412, "length": 18 }
      },
      {
        "id": "LightningComponentBundle:accountTile",
        "type": "LightningComponentBundle",
        "apiName": "accountTile",
        "edgeType": "readsFrom",
        "source": "lwc-aura-vf-scanner",
        "confidence": "heuristic",
        "properties": { "wirePath": "Account.Industry" }
      }
    ]
  }
}

An empty usages list means "the scanner saw no matching references." It does not mean "no code references this component" — see the boundary disclosure below.

Step 3 — Group by edgeType and present

Bucket usages by edgeType so the developer sees reads, writes, and calls separately. Within each bucket, sort by id (the API already returns the list sorted by (id, edgeType) ASC; preserve that order).

Every referrer gets its canonical id in backticks. Every edge gets its own confidence cited explicitly, every time — parsed (with source: apex-ast) for AST-resolved Apex edges, heuristic (with source: apex-scanner) for the recall-scanner backfill and for LWC/Aura/VF field reads, declared for LWC apex imports and VF controller/extension attributes. Never blanket-label the bucket heuristic.

Step 4 — Optionally call sfi.run_analysis with { "name": "sfi.search_apex_source", "args": { … } } to surface lines

The scanner gives you the file that referenced the target. To surface the actual line, follow up with sfi.search_apex_source using the field or method name as the query:

{ "query": "Industry__c", "limit": 25 }

This is the "tool gives the file; grep gives the line" pattern. Run the second call when the result list is short and the developer is likely to act on it (rename, refactor, deprecate). Skip it when there are many matches and the developer is just scanning the impact surface.

Reporting format

Worked example. User asks: "Where do we write to Account.Industry__c from Apex?"

Apex usages of CustomField:Account.Industry__c (per-edge confidence — most parsed from the default-on Apex AST, one heuristic scanner-backfill):

Writes (writesTo) — 2 referrers

• ApexClass:AccountTriggerHandler — source: apex-ast, confidence: parsed, properties: { offset: 412, length: 23 }. The AST resolved the receiver to Account.
• ApexClass:OpportunityService — source: apex-ast, confidence: parsed, properties: { offset: 1287, length: 19 }.

Reads (readsFrom) — 1 referrer

• ApexClass:AccountReportBuilder — source: apex-scanner, confidence: heuristic, properties: { offset: 215, length: 18 }. This file fell back to the recall scanner (AST parse failure), so the read is a heuristic match on acc.Industry__c with acc unresolved — spot-check with sfi.search_apex_source({ query: 'Industry__c' }).

The two parsed writes are high-confidence; the one heuristic read needs verification.

Boundary: the default-on Apex AST pass resolves inline static SOQL ([SELECT ... WHERE Industry__c ...]) and CONSTANT-string Database.query('SELECT Industry__c FROM Account') field references at confidence: 'parsed' — those appear as real edges above. Still invisible: string-BUILT dynamic SOQL (Database.query('SELECT ' + f + ...)), dynamic field access (record.get('Industry__c')), and reflective writes via SObject.put. Real references may exist that this list misses; files the AST failed to parse fall back to the regex scanner only (the manifest's apexAst block counts them).

Group by edgeType with counts. Cite every component by canonical ID. State each edge's own confidence and source (parsed / apex-ast for AST-resolved edges, heuristic / apex-scanner for scanner-backfill edges) — never blanket-label the bucket. Name the boundary; point the developer at the second-step grep.

Boundary disclosure

The Apex tier runs a parser-grade AST pass by default (source: apex-ast, confidence: parsed) with a heuristic recall scanner (source: apex-scanner, confidence: heuristic) backfilling files the AST cannot parse and patterns it does not resolve. The frontend LWC/Aura/VF tier is a regex scanner — intentionally heuristic for the common cases, with only the unambiguous declared cases (LWC @salesforce/apex imports, VF controller/extension attributes) at confidence: 'declared'. The AST resolves dot-access, cross-class calls, and inline static SOQL, but it is a parser, not a full dataflow engine — the residual limits below stay invisible. Surface this list whenever the developer is about to act on the result.

Apex tier limits (residual, after the default-on AST pass)

• Dynamic field access is invisible. record.get('Industry__c') and SObject.put('Industry__c', value) are string arguments; the scanner blanks strings before pattern-matching. Zero edges.
• Only string-BUILT SOQL is invisible. The default-on AST pass (source: apex-ast, confidence: 'parsed') fully parses inline static SOQL in [ ... ] brackets — SELECT, WHERE, ORDER BY, and GROUP BY fields all yield field-level edges — and also parses CONSTANT-string Database.query('SELECT ...') literals. What remains invisible is SOQL assembled from string concatenation or variables (Database.query('SELECT ' + f + ' FROM ...')). On a file the AST failed to parse (counted in the manifest apexAst block), only the regex scanner ran and SOQL field references for that file are unreliable.
• Reflective method dispatch is invisible. A helper like getFieldValue(record, 'Industry__c') produces no edges; the field name is a string argument.
• Intra-class self-calls produce no cross-class edge. The AST recognizes a bare doStuff(x) as a self-call, but a class calling its own method is not a callsApex graph edge (there's no second class to point at), so it won't surface as a referrer of another component. Cross-class calls (ClassName.method() and instance calls on a typed receiver ARE resolved and emit parsed edges.
• Receiver resolution is a symbol-table pass, not full type inference. The default-on AST resolves acc.Industry__c to CustomField:Account.Industry__c (a parsed edge) when acc is declared, parameter-typed, or for-each-bound as Account. A receiver whose type the local symbol table cannot establish (deep generics, chained returns, untyped sObject) falls to the recall scanner, which emits a heuristic edge keyed on the variable name (CustomField:acc.Industry__c) — treat those as approximate.

LWC scanner (v1.4) limits

• LWC record[fieldName] dynamic field access is invisible. record[this.fieldName] and record[someVar] are computed property accesses; the scanner cannot resolve the bracket expression. Zero edges.
• LWC @wire(getRecord, ...) static-field detection works for fixed lists. @wire(getRecord, { recordId: '$recordId', fields: [INDUSTRY_FIELD] }) resolves INDUSTRY_FIELD when the symbol comes from an @salesforce/schema/Account.Industry import. If the fields list is built dynamically (fields: this.fieldsToFetch), the scanner sees no field reference.
• LWC field-import vs. body-text reads are not distinguished. Per the v1.4 R3a report, both default to confidence: 'heuristic'. A schema-import field that is only ever imported (not used in the body) still produces a heuristic edge; consumers cannot use confidence alone to separate "imported but unused" from "imported and read."

Aura scanner (v1.4) limits

• Aura framework-internal v.{attr} is intentionally NOT extracted. v.recordId, v.records, and other component-local attribute references are framework-level state; they do not correspond to a CustomField or CustomObject and emitting edges for them would produce massive false-positive noise. The scanner silently skips them.
• Aura dynamic field expressions are invisible. v.record[fieldName] and {!v.record[expr]} are runtime-evaluated; the scanner does not resolve them.

Visualforce scanner (v1.4) limits

• VF expressions inside HTML attribute strings work. A binding like <apex:inputField value="{!account.Industry__c}"/> produces a heuristic readsFrom edge — the {! } shape is recognised inside attribute strings.
• VF expressions inside JS string literals don't. A block like <script>var x = '{!account.Industry__c}';</script> is a JS string literal; the scanner blanks JS strings the same way it blanks Apex strings. Zero edges for that line.
• VF controller and extension references are declared. The controller= and extensions= attributes ARE the declaration; the edge confidence is declared, not heuristic.

Treat missing edges as "may exist; the static passes couldn't see them." For present edges, read the confidence: a parsed (apex-ast) edge is a high-confidence, symbol-resolved match; a heuristic (apex-scanner or frontend-scanner) edge is a best-effort pattern match that needs a human pass before any irreversible change. Heuristic does not mean "approximately correct." Declared edges (LWC apex imports, VF controller/extension attributes) still need spot-checks when the referenced class is renamed, but they will not produce false positives — the import or attribute name IS the contract.

Anti-patterns

Example interaction

User: "Is it safe to rename processOpportunity on OpportunityService? I want to make sure no Apex calls it."

Claude's flow:

1. Translate → targetId: 'ApexClass:OpportunityService', edgeTypes: ['callsApex'] (the question pins to call sites), nodeTypes: ['ApexClass','ApexTrigger'] (pinned to Apex callers).
2. Call sfi.run_analysis with { "name": "sfi.find_code_usages", "args": { … } } with that input.
3. Receive (hypothetical): two callsApex referrers.
4. Call sfi.run_analysis with { "name": "sfi.search_apex_source", "args": { … } } with query: 'OpportunityService.processOpportunity' to surface the lines.
5. Respond:

Apex call sites for ApexClass:OpportunityService (filtered to callsApex; both AST-resolved at confidence: parsed):

Calls (callsApex) — 2 referrers

• ApexClass:OpportunityController — source: apex-ast, confidence: parsed, properties.methodName: 'processOpportunity'.
• ApexTrigger:OpportunityTrigger — source: apex-ast, confidence: parsed, properties.methodName: 'processOpportunity'.

Grep follow-up (sfi.search_apex_source, query OpportunityService.processOpportunity):

• source/classes/OpportunityController.cls:42 — OpportunityService.processOpportunity(opp);
• source/triggers/OpportunityTrigger.trigger:8 — OpportunityService.processOpportunity(Trigger.new);

Both edges are parsed: the AST resolved the OpportunityService receiver and the processOpportunity call-site. It does not, however, resolve dynamic dispatch — before renaming, check for indirect invocation (Type.forName(...).newInstance(), reflective dispatch), which no static pass sees.

Example interaction — code-broad field rename (v1.4)

User: "I want to rename Account.Industry__c to Account.IndustrySegment__c. What touches this field anywhere in code?"

Claude's flow:

1. Translate → targetId: 'CustomField:Account.Industry__c'. The question is code-broad ("anywhere in code"), so call the broader tool with default edges.
2. Call sfi.run_analysis with { "name": "sfi.find_code_usages", "args": { … } } with { targetId: 'CustomField:Account.Industry__c' }.
3. Receive (hypothetical): one Apex read, one LWC read via @wire(getRecord), one VF page binding.
4. Call sfi.run_analysis with { "name": "sfi.search_apex_source", "args": { … } } with query: 'Industry__c', limit: 25 to surface the Apex lines.
5. Respond:

Code usages of CustomField:Account.Industry__c across all six code surfaces (default broad query, no nodeTypes filter):

Reads (readsFrom) — 3 referrers across 3 tiers

• ApexClass:AccountReportBuilder — source: apex-ast, confidence: parsed, properties: { line: 215 }. The AST resolved acc to Account. High-confidence.
• LightningComponentBundle:accountTile — source: lwc-aura-vf-scanner, confidence: heuristic, properties: { wirePath: 'Account.Industry' }. The scanner saw the schema-import field in the @wire(getRecord, ...) fields list.
• VisualforcePage:AccountReport — source: lwc-aura-vf-scanner, confidence: heuristic, properties: { line: 22 }. The scanner saw {!account.Industry__c} in an attribute string.

Grep follow-up (sfi.search_apex_source, query Industry__c):

• source/classes/AccountReportBuilder.cls:215 — report.industry = acc.Industry__c;

The Apex read is parsed (high-confidence); the two frontend edges are heuristic. Before renaming:

• The LWC @wire(getRecord) schema import must be updated; the v1.4 scanner could not distinguish schema-import from body-text reads, so check accountTile/accountTile.js for an @salesforce/schema/Account.Industry import line.
• The VF binding {!account.Industry__c} in AccountReport.page:22 must be updated by hand.
• Inline static SOQL and constant-string Database.query references ARE covered (parsed-confidence apex-ast edges). Still invisible: string-BUILT dynamic SOQL, LWC record[fieldName] dynamic access, Aura v.{attr} framework attributes, and VF <script> JS string literals. Run sfi.search_apex_source for Industry__c to catch concatenated SOQL strings, and do a manual grep across LWC/Aura/VF source for bracket and template-string references.

Example interaction — LWC-only narrowing (v1.4)

User: "Which LWC components read Account.Industry__c?"

Claude's flow:

1. Translate → targetId: 'CustomField:Account.Industry__c', nodeTypes: ['LightningComponentBundle'], edgeTypes: ['readsFrom'] (the question pins both axes).
2. Call sfi.run_analysis with { "name": "sfi.find_code_usages", "args": { … } } with that input.
3. Receive (hypothetical): one LWC bundle.
4. Respond:

LWC bundles reading CustomField:Account.Industry__c (v1.4 scanner, confidence: heuristic, filtered to nodeTypes: ['LightningComponentBundle'] and edgeTypes: ['readsFrom']):

Reads (readsFrom) — 1 referrer

• LightningComponentBundle:accountTile — source: lwc-aura-vf-scanner, confidence: heuristic, properties: { wirePath: 'Account.Industry' }.

Heuristic boundary: the v1.4 LWC scanner cannot distinguish a schema-import field that is imported-but-unused from one that is imported-and-read in the template/JS body. The edge may be a false positive if the bundle imports @salesforce/schema/Account.Industry but never references the value. Also invisible: any record[fieldName] dynamic access in the bundle's controller JS.

Verification

Before sending a response, confirm:

• I translated the user's reference into a canonical targetId (CustomField:Object.Field, ApexClass:Name, ApexTrigger:Name, LightningComponentBundle:name, AuraDefinitionBundle:name, VisualforcePage:Name, or VisualforceComponent:Name), confirming via sfi.search_components when ambiguous.
• I picked the right SCOPE by user intent: broad sfi.find_code_usages (all node types) for any code-broad question (default), narrowed to nodeTypes: ['ApexClass','ApexTrigger'] only when the user explicitly pins to Apex.
• When the user pinned to a single code tier (e.g., "what LWC bundles touch this?"), I narrowed with nodeTypes: ['LightningComponentBundle'] (or the appropriate subset) rather than calling the broad tool and post-filtering by hand.
• I called the tool and grouped results by edgeType (and by code tier when the result spans multiple tiers).
• For non-trivial result lists, I followed up with sfi.search_apex_source to surface the lines.
• I cited every referrer by canonical ID and labeled the edges with their actual per-edge confidence: parsed for AST-resolved Apex edges (source: apex-ast — the default and dominant case); heuristic for the Apex recall-scanner backfill and for LWC/Aura/VF field reads and dynamic expressions; declared for LWC @salesforce/apex callsApex edges and VF controller/ extension references. I did NOT blanket-label Apex edges heuristic.
• I appended the boundary disclosure relevant to the result tier — Apex-scanner invisible patterns when Apex referrers are present; LWC scanner limits (record[fieldName], schema-import vs. body-read indistinguishability) when LWC referrers are present; Aura v.{attr} exclusion when Aura referrers are present; VF JS-string blanking when VF referrers are present.
• I did not present a heuristic edge as ground truth, and I did not claim an empty result means "no references exist."
• I did not present a declared edge with a heuristic disclaimer (LWC apex imports and VF controller/extension attributes do not need the "scanner may be wrong" caveat).

Grounding & routing (shared contract). For a vague or broad ask, call sfi.route_question first — in the default hybrid mode it returns a meaning-ranked toolCandidates shortlist (which YOU pick from) plus a suggested plane and a route hint (and whether to sfi.resolve a name first). Default tool profile is core: only the core spine (including sfi.live_consent) is directly invokable. For every other sfi.* analysis, call sfi.run_analysis with { "name": "sfi.<tool>", "args": { … } } (or follow route_question.invoke, which already wraps non-core steps). Optional: sfi.describe_analysis first when args are unclear. Every org fact must come from an sfi.* tool call, cited by its canonical id — never from memory. Build the answer only from what the tools returned, then pass it through sfi.synthesize_answer, which flags any hallucinatedIds (canonical ids no tool produced). Full cascade: using-sf-intelligence.