Salesforce-Intelligence

Answers Salesforce admin documentation-generation requests: "generate a data dictionary for Account", "create an admin handbook for this org", "write an onboarding doc for a new developer", "generate an architecture overview", "produce a sharing-model summary", "draft a compliance report on PII". Drives the v2.5 documentation-generator tier (`generate_data_dictionary`, `generate_admin_handbook`, `generate_onboarding_doc`, `generate_architecture_overview`, `generate_sharing_summary`, `generate_compliance_report`). Every generator emits a `GeneratedDocument` with frontmatter (`generatedAt`, `sourceTreeHash`, `componentIds`), structured markdown body, per-section confidence labels, and verbatim `boundaries[]` disclosures. Discloses the v2.5 Q125 freshness disclosure on every generated doc, the inherited-confidence rule, the heuristic-tagged sections (Domain Clustering, PII Inventory, Field Label Glossary), and the v1.7-enrichment- missing fallback for Recent Changes / Key Contacts.

Answers admin questions about Salesforce's legacy automation layer: "what workflow rules touch Account", "which approval processes are still active", "what assignment rule routes Cases", "is this workflow being replaced by a flow", "which fields does this duplicate rule check", "what emails go out for this approval", "show me the auto-response rules for Lead". Calls `sfi.get_impact` and `sfi.get_edges` to find workflow rules, approval processes, assignment/auto-response/escalation rules, duplicate/matching rules, email templates, and letterheads — plus the `sendsEmail`, `triggersOn`, and `references` edges that connect them. Discloses the v1.3 boundary honestly: workflow rules and processes are extracted in v1.3, but their actual execution semantics (criteria evaluation, time-dependent actions, approval routing) need record-level data v1.x doesn't have. Recommends Flow migration honestly where applicable.

Saved, repeatable admin routines for the sf-intelligence org that batch the right `sfi.*` tools in one request, then synthesize ONE report. Fire when the user asks for a named routine rather than a single fact — "run my pre-deploy checklist", "I inherited this org, give me the tour", "run a security audit", "what's junk in this org / what can I clean up", "is my org healthy / on fire", "audit sweep", "cleanup sweep", "go-live check". Each playbook is an ordered tool batch; run it end to end, then summarize. Single-fact questions go to `answering-org-questions` instead.

Answers Salesforce admin sharing/visibility questions like "why can't user X see Y", "who can see this record", "what would let this user see Y", "the user says they can't see Z", "visibility check for X on Y", "sharing rules touching Z", "role hierarchy from Foo up", "what grants Edit access to Bar". Calls `sfi.why_cant_user_see_record` (cascade over OWD → permission grants → role hierarchy → owner sharing rules → criteria sharing rules → manual/sets/teams) and explains the verdict honestly: stages return `unknown` rather than fabricating decisions when v1.1's metadata model can't tell. Discloses v1.1's boundary explicitly — manual sharing, sharing sets, account teams, and record-level criteria evaluation all return `unknown` for an admin to verify manually.

Conversational interface for questions about the Salesforce org backed by this repo's `org-kb/` vault. Use for any question about schema, dependencies, permissions, naming conventions, Apex source, Flow metadata, or vault freshness — except refresh requests (those fire `refreshing-the-org-vault` instead) and initialization (which fires `/sfi-init`). Triggers on questions starting with "what", "which", "show me", "find", "list", "who can", "where is", "how many", "does", "is there", or "when was" when the topic is a Salesforce entity (object, field, validation rule, flow, Apex class, trigger, layout, permission set, profile) or the vault itself.

Answers cross-org metadata comparison questions for Salesforce architects — sandbox-vs-prod drift, prod-vs-prod (multi-customer) audits, "what changed since I last refreshed?" diffs, and Lead-vs- Contact (intra-org) field-mapping for the conversion. Triggers on phrases like "compare sandbox vs prod", "compare two orgs", "what's drifted between {vaultA} and {vaultB}?", "is this object different in sandbox?", "does System Administrator have the same permissions in sandbox?", "Lead conversion mapping", "map Lead fields to Contact". Routes to four MCP tools — `sfi.compare_vaults`, `sfi.compare_object_across_vaults`, `sfi.compare_profile_across_vaults`, `sfi.field_mapping_between_objects` — surfacing the verbatim honesty disclosures on every response: volatile-property filter, api-name-match correspondence, profile-edition rollup, and the Q174 heuristic-mapping disclosure.

Answers "what breaks if I change X?" CHANGE-IMPACT questions for Salesforce architects. Triggers on phrases like "what breaks if I change/delete/rename", "what depends on", "what's the impact of", "is it safe to deprecate", "blast radius of". USAGE phrasings — "where is X used", "what touches X", "which components reference X", "show me everything that uses X" — are NOT this skill: route them to `sfi.find_component_usages` (the §C3 universal usage dispatcher), and "find formulas using X" to `sfi.find_formula_references`. Calls `sfi.get_impact` (BFS over incoming edges, default 2 hops), then reports the dependency subgraph organized by component type with canonical IDs and per-edge confidence. Discloses the current edge-coverage boundary explicitly: dynamic Apex/SOQL blind spots, heuristic-confidence edges, and any metadata families missing from the vault coverage report.

Answers Salesforce architect integration questions: "draw me our integration map", "what subscribes to this platform event", "who calls this REST API", "which classes are queueable / batchable", "what runs async in this org", "where does this external data source connect", "what is our external API surface", "which external services use this named credential", "audit our auth providers". Calls `sfi.integration_map` (organized map of auth + endpoints + APIs + access lists) and `sfi.event_subscribers` (given a platform event, return all Apex/Flow subscribers). Plus `sfi.list_components` with the new v1.5 property filters (isQueueable, isSchedulable, isBatchable, hasFutureMethod, hasInvocableMethod, isRestResource, hasAuraEnabledMethod) for async/job and API surface questions. Discloses v1.5 honesty axis: some patterns are heuristic (dispatchesAsync via reflection, LWC fetch URL correlation, CDC subscription detection) and require manual verification.

Answers business-user-flavored Salesforce questions like "what does this field mean in English", "what happens when I close this opportunity", "why can't I do X", "how do I find report Y", "what's the SLA on this case", "how are cases routed to agents", "is this field required", "what's this object for", "how is this calculated". Classifies the intent and routes to the right `sfi.*` tool, OR honestly refuses when v1.x doesn't yet model the answer (record-level data, reports, dashboards, live entitlement-milestone status/timers, flow-to-narrative). R6-18 extracts EntitlementProcess/MilestoneType (which milestones apply, business hours, active state) and ServiceChannel/QueueRoutingConfig (Omni-Channel routing model, capacity, overflow queue) — SLA/ entitlement and case-routing questions get a real partial answer now, not a blanket refusal. Bridges the vocabulary gap between business-user phrasing and the architect/admin tools the product already ships. v1.6 broadens to record-value lookups for Custom Metadata T

Answers code-source-focused developer questions across the full org code surface: Apex classes, Apex triggers, and the v1.4 frontend tier (LWC, Aura, Visualforce). Triggers: "where is field X used in Apex", "find LWC components reading this field", "what calls this Apex from VF", "Aura references to this object", "who reads/writes this field from code", "what classes call MyClass.someMethod", "is it safe to rename this method", "find all Apex references to this", "what Apex/LWC/VF touches this object". Calls `sfi.find_code_usages` (incoming readsFrom/writesTo/callsApex/references edges from ApexClass, ApexTrigger, LightningComponentBundle, AuraDefinitionBundle, VisualforcePage, or VisualforceComponent nodes) — the single code-usage tool. For strictly Apex-only questions, narrow it with `nodeTypes: ['ApexClass','ApexTrigger']` (the former `find_apex_usages` Apex-only view, now folded in). Optionally follows up with `sfi.search_apex_source` to surface the actual lines. Discloses the heuristic-confidence boundar

Answers Salesforce developer code-quality questions: "audit my Apex for quality issues", "find governor-limit risks", "where do we have hardcoded IDs / emails / usernames", "check CRUD / FLS enforcement", "what tests are missing / fake-covered", "find dead code". Drives the v2.1 quality-recognizer cascade (`code_quality_audit`, `governor_limit_risks`, `find_hardcoded_values`, `crud_fls_audit`, `test_coverage_gaps`) and the v2.4 hygiene tool `find_dead_code`. Every finding is `confidence: 'heuristic'` — pattern recognition is recognition, not declaration. Discloses the v2.1 boundary verbatim: the recognizer pattern-matches on tokenized Apex source (not a compiler AST), so cross-class blindness, custom security utility helpers invisible, dynamic SOQL invisible, reflective field access invisible, custom test-assertion frameworks invisible. (Note: Apex extraction itself uses a parser-grade AST by default for confidence: parsed edges; the recognizer's scope is narrower.)

Answers Salesforce developer "show me EVERYTHING about this field" and "trace this field's lineage" deep-forensics questions: "what does Account.Industry__c touch", "give me the full field profile for X", "where does this field's data come from", "what fires when this field changes", "show me the upstream + downstream walk for this field". Cascades the v3.0 synthesis tools (`field_360`, `field_lineage`) over the v2.0a conditional-context tier (`firesWhen` edges, `ConditionalContext` nodes). Discloses the Q165 honesty anchor verbatim (`dataNotAvailable[]` array of list-view filters, reports, dashboards) and the v2.0a synthetic ConditionalContext id-prefix classification brittleness.

Answers Salesforce developer "where is X used everywhere" and "what does this field mean in our org" discovery questions: "find every reference to `Industry__c`", "do we already have a field for customer health", "what does this field actually mean", "is `Status` the same as `Stage` here", "is this field manually entered or automated", "what does this acronym mean", "find this literal anywhere in the org". Cascades over the v2.2 universal- search tools (`find_field_anywhere`, `find_semantic_field`, `find_hardcoded_values_anywhere`) and the v2.9 vocabulary + semantic-disambiguation tools (`field_meaning`, `disambiguate_concepts`, `field_provenance`). Discloses the v2.2 Q95 honesty anchor verbatim (lexical-only, no synonyms) and the v2.9 classifier-not-run boundary (`sourceOfTruth` / `semanticCategory` absent on pre-v2.9 vaults).

Answers Salesforce developer "what breaks if I change X / where is M reachable from / what does this method touch / is this dead code" questions. Cascades the v2.3 what-if composer family (`what_if_deactivate_flow`, `what_if_disable_trigger`, `what_if_change_method_signature`, `what_if_change_field_type`, `what_if_remove_picklist_value`, `what_if_make_field_required`, `what_if_merge_profiles`, `what_if_split_profile`) and the v2.7 deep code-understanding tools (`call_graph`, `downstream_effects`, `test_coverage_for_method`, `method_reachability`, `meaningful_test_audit`). Discloses the v2.3 / v2.7 boundary axes verbatim: class-granularity (not method); dynamic-dispatch and reflective-invocation invisibility; method-name echo (`method_name` surfaced verbatim, not subsetted); cross-class transitive analysis not available.

Answers Salesforce admin/architect freshness questions like "when was X modified?", "who built this?", "what changed since last week?", "has this been touched since go-live?", "what changed in the last sprint?". Calls `sfi.last_modified` and `sfi.changed_since`. Discloses v1.7 honesty axis: when Tooling API enrichment hasn't run, returns `enriched: false` with explicit guidance to run `sfi refresh --with-tooling-api`; never fabricates dates.

Refreshes the SfIntelligence vault from a live Salesforce org. Use when the user invokes `/sfi-refresh`, or says "refresh", "refresh the vault", "pull latest", "pull the latest from my org", "rebuild the vault", "rebuild the index", "my data is stale", "the vault is stale", "I just deployed changes", or any phrase signaling the vault is out of date relative to the org. Drives `sfi refresh`, which shells out to `sf project retrieve`, runs the full extractor graph, imports into DuckDB, renders Markdown, and writes a new manifest (including optional **coverage** metadata).

Entry point for the SfIntelligence plugin. Use whenever the conversation touches a Salesforce org in this repo — schema, dependencies, permissions, naming conventions, automation, Apex, Flow, integrations, sharing, or any question about a component the org contains. Loads first; teaches Claude the `sfi.*` MCP tool cascade, the resolve-first / ask-a-clarifying-question pattern, the canonical component-ID format, and how to route to the other sf-intelligence skills.

Answers Salesforce admin questions about which page layout a user sees and why: "what layout does user X see for Account", "which page layout for Sales rep on Opportunity", "why is this user seeing this layout", "which record type does this profile default to", "what's the layout for record type Y", "user reports wrong layout", "page layout assignments for profile X". Calls `sfi.layout_for_user` (cascade over profile lookup, layoutAssignment matching, recordType resolution) and explains the routing decision honestly: stages return `unknown` rather than fabricating when v1.2's metadata model can't tell (e.g., org-default layouts, permission-set layout assignments). Discloses v1.2's boundary: only profile-based layout assignments are resolved; permission-set layouts, app-default routing, and Lightning page assignments are noted for manual verification.

Answers Salesforce architect async + event-driven topology questions that v2.8 surfaces beyond the v1.5 integration tier. Triggers: "who subscribes to CDC events on Account", "what change-data-capture subscribers do we have", "what's the async chain depth from this Queueable", "show me every scheduled job", "what outbound messages does this org send", "give me every endpoint URL in one place". Drives the v2.8 cascade: `cdc_subscribers` (Change Data Capture subscribers via name-pattern detection), `async_chain_depth` (transitive `dispatchesAsync` walk; chainAsync NOT persisted), `scheduled_job_catalog` (cron-driven jobs from `System.schedule` detection), `outbound_message_catalog` (the new v2.8 OutboundMessage ComponentType promoted from WorkflowRule dangling refs), `endpoint_catalog` (unified URL view across REST exposes / OutboundMessage / ExternalDataSource / NamedCredential). Discloses v2.8 boundaries verbatim: URL-not-validated, name-pattern detection for CDC, depth-cap of 10, Tooling API gap for actual c

Conversation memory and follow-up handling for the sf-intelligence org. Fire on follow-up turns that depend on what was just discussed — pronouns and ellipsis ("what about the other one?", "and on Contact?", "is that one safe to delete?", "show me its dependencies", "the second one"), refinements of a prior answer, or "what should I look at next?". Keeps the resolved component, the last tool/plane, and the disambiguation options from the previous turns so the user does not have to repeat themselves. Pairs with answering-org-questions (which does the actual lookups) and using-sf-intelligence (the cascade).

Diagnostic skill for the SfIntelligence vault. Use when the user invokes `/sfi-status`, says "is something wrong?", "the vault feels off", "something looks stale", "are we connected?", "is the index current?", or when a previous `sfi.*` MCP call returned a `degraded` or `unhealthy` status, an unexpected error envelope, or two consecutive errors. Runs three probes — `sfi.health_check`, `sfi.get_manifest`, and `sf org list --json` — and reports findings as a numbered checklist. Reports only; suggests but does not run `/sfi-refresh` or `/sfi-init`.

Surfaces observed naming-convention patterns from the SfIntelligence vault. Use when the user asks "what's our convention", "what's the naming standard", "do we always …", "is this consistent with the rest of the org", or proposes a new component name and wants it sanity-checked ("can I name this `CustomerOrder__c`?", "should I call this `Foo__c` or `Bar__c`?"). Calls `sfi.get_naming_convention_report` and reports its observations as **heuristic** — observed patterns, not enforced rules.

Answers Salesforce Industries / OmniStudio questions: "walk this OmniScript step by step", "what endpoints does this Integration Procedure hit", "what does this DataRaptor map", "what's inside this FlexCard", "what inputs does this DecisionTable take", "what OmniStudio version is this org on". Drives the v3.2 cascade: `omniscript_flow` (OmniScript step sequence + downstream IP / DataRaptor dispatches), `integration_procedure_chain` (IP action chain + REST endpoints + downstream IP / DataRaptor), `datatransform_field_map` (DataRaptor source→target field map with per-row confidence), `omniuicard_widget_breakdown` (FlexCard widget tree + states + dispatched actions), `decision_table_browse` (DecisionTable parameter shape; rows null — Q179 refusal). Triggers on: "OmniScript", "FlexCard", "DataRaptor", "Integration Procedure", "DecisionTable", "OmniStudio", "OmniUiCard", "Industries", "vlocity", "omnistudio__". Discloses the v3.2 honesty axes verbatim: Native-vs-Vlocity-Legacy detection is heuristic (Q180); the Om

Silent pre-loader for the SfIntelligence vault. Fires automatically when the user's message mentions any Salesforce concept — words like "object", "field", "trigger", "flow", "validation rule", "permission set", "profile", "layout", "Apex", "picklist", "lookup", "record type", "metadata", "org", "sandbox"; common object names like Account, Contact, Opportunity, Lead, Case, User; or the custom-suffix patterns `__c`, `__mdt`, `__e`, `__b`, `__r`, `__x`. Calls `sfi.org_card` (the refresh-time orientation card; falls back to `sfi.get_manifest` + `sfi.list_components(type: 'CustomObject')` on a vault without a card) to warm the cache so the next question answers faster. Does not print to the user; produces no visible output. Skips when the user invokes `/sfi-init`, `/sfi-refresh`, or `/sfi-status`, when the message is off-topic, or when context is already loaded this session.

Routes coverage and destructive-verdict questions. Call sfi.coverage_report when the user asks whether the vault is complete, why something was not found, or before narrating safe-to-delete / what-if answers. Render coverageCaveat BEFORE the verdict — never footnote it. Refuse absence-based answers for not-modeled families; distinguish "not in vault" from "not in org".