salesforce-agentforce-stdm-observer-skill

salesforce-agentforce-stdm-observer-skill

Production observability for Agentforce agents via STDM and Data Cloud. This skill is a live evidence reader, not a configuration reviewer. It queries session telemetry, quality scores, and action traces to answer: "Is my agent working correctly right now?" It does not modify agents, configurations, or any org data.

Adaptation note: Query mechanics in this skill are adapted from the observing-agentforce skill published by Salesforce in the forcedotcom/sf-skills repository (Apache-2.0). Vanguard-specific additions include the T1 least-privilege contract, structured audit envelope, explicit aggregate-only output policy, and the handoff routing model.

Verify-before-merge notice: All Agentforce, STDM, Data Cloud, and Einstein AI feature names evolve rapidly. Validate all product references, DMO field names, and API structures against current official Salesforce documentation before use in production.

When This Skill Owns the Task

Use salesforce-agentforce-stdm-observer-skill when the goal is live production observability for an Agentforce agent:

• "How is my Order Service agent performing over the last 7 days?"
• "Show me faithfulness and answer relevance scores for the Support agent"
• "Are there action invocation errors for the Case Routing agent?"
• "What is the session abandonment rate for the HR Self-Service agent?"
• "Show me STDM session telemetry for the last 48 hours"
• "Is the agent hallucinating? What do the quality scores say?"

Delegate elsewhere when:

Required Context to Gather First

Before executing any STDM query, confirm all of the following. Ask if missing.

1. Target org alias — the --target-org value recognized by sf org list. Never accept a raw instance URL or session token.
2. Org type — production or sandbox. Flag production orgs; apply stricter output scrutiny and confirm the Connected App allowlist authorizes this alias.
3. Agent name(s) to observe — the agent's display label (MasterLabel) or API name. Resolve against the org before querying (see Step 2).
4. Time window — ISO 8601 start and end timestamps (UTC). Default to last 7 days if not specified.
5. Specific metrics requested — sessions summary, faithfulness, answer relevance, action telemetry, or all. Prefer getAggregatedMetrics as the first call to bound scope.
6. Sensitivity classification — does the org serve Health Cloud, Financial Services Cloud, or other regulated verticals? If yes, extra redaction scrutiny applies; route any anomalies involving session content to salesforce-compliance-privacy-agent before sharing externally.
7. Data Cloud provisioned? — STDM requires Data Cloud and the Agentforce Activity data stream enabled. Confirm before executing (see Step 1).

Recommended Workflow

Step 1 — Verify org connectivity and Data Cloud availability

Confirm the org alias is reachable and Data Cloud is provisioned:

sf org display --target-org <alias>

Then probe the Data Cloud data spaces endpoint to confirm STDM DMOs are available:

sf api request rest "/services/data/v63.0/ssot/data-spaces" \
  --target-org <alias>

Note: sf api request rest is a beta command — do not add --json (that flag is unsupported and causes an error in this command).

Decision logic:

• If the command fails (404 or permission error), assume DATA_SPACE=default and log it as an assumption.
• Filter to only status: "Active" data spaces.
• If exactly one active data space exists, use it automatically.
• If multiple exist, show the list and ask the user which to use.
• Store the selected name as DATA_SPACE for all subsequent steps.

If Data Cloud is unavailable, stop and inform the user:

STDM requires Data Cloud with "Agentforce Activity" data stream active. Navigate to Setup → Data Cloud → Data Streams to verify. This skill cannot proceed without STDM. For local trace analysis without Data Cloud, see the observing-agentforce skill from forcedotcom/sf-skills.

Step 2 — Resolve target agent and confirm AiAgentTag

Resolve the user-provided agent name to the exact MasterLabel used by STDM. Field names and exact object names are drift-prone — run this query and use the returned values, not the user-provided string:

sf data query \
  --query "SELECT Id, MasterLabel, DeveloperName FROM GenAiPlannerDefinition WHERE MasterLabel LIKE '%<user-provided-name>%' OR DeveloperName LIKE '%<user-provided-name>%'" \
  --target-org <alias> \
  --result-format json

Store:

• AGENT_MASTER_LABEL — for STDM findSessions agent filter
• PLANNER_ID — the Salesforce record ID for this agent (redact in output)

If the query returns no results: The agent does not exist in this org. Show the full list of agents and ask the user to identify the target.

Step 3 — Query AiAgentTagAssociation for tagging context

Retrieve tagging metadata to understand what quality evaluation definitions are configured. This confirms the org has quality scoring enabled before querying scores:

sf data query \
  --query "SELECT Id, AiAgentTagId, EntityId, EntityType FROM AiAgentTagAssociation LIMIT 10" \
  --target-org <alias> \
  --result-format json

If AiAgentTagAssociation returns no rows, quality scores may not be configured. Note this in the output and proceed with session-level metrics only.

Note: AiAgentTag, AiAgentTagDefinition, and AiAgentTagAssociation are Tooling API objects. Use --use-tooling-api if the standard SOQL path returns an "object not found" error:

sf data query \
  --query "SELECT Id, AiAgentTagId, EntityId, EntityType FROM AiAgentTagAssociation LIMIT 10" \
  --target-org <alias> \
  --use-tooling-api \
  --result-format json

Step 4 — Query STDM sessions via aggregated metrics

Start with getAggregatedMetrics via the AgentforceOptimizeService Apex helper class to get the health dashboard before drilling into individual sessions. This is the most efficient first call and avoids fetching session content.

Full Apex service deployment steps and invocation patterns are documented in references/stdm-queries.md (adapted from forcedotcom/sf-skills observing-agentforce).

String result = AgentforceOptimizeService.getAggregatedMetrics(
    '<DATA_SPACE>',
    '<START_ISO>',
    '<END_ISO>',
    50,
    '<AGENT_MASTER_LABEL>'
);
System.debug('STDM_RESULT:' + result);

sf apex run --json --file /tmp/stdm_metrics.apex --target-org <alias>

Parse the result using the DEBUG|STDM_RESULT: pattern (see references/stdm-queries.md). The aggregated metrics return:

• total_sessions, total_turns, avg_quality_score
• avg_faithfulness, avg_answer_relevance, avg_context_precision
• abandonment_rate, deflection_rate, escalation_rate
• end_type_counts, quality_distribution, top_intents
• unavailable_dmos — list of DMOs that could not be queried

If findSessions returns empty: No production sessions exist in this date window. Check that the date range is correct and that the agent is actively receiving traffic. Consider widening the window.

Step 5 — Drill into faithfulness and answer relevance (if warranted)

If avg_faithfulness or avg_answer_relevance falls below the thresholds defined in references/observability-rubric.md, run targeted observability queries:

AgentforceOptimizeService.ObservabilityInput inp = new AgentforceOptimizeService.ObservabilityInput;
inp.queryType = 'Hallucination';
inp.agentApiName = '<AGENT_MASTER_LABEL>';
inp.lookbackDays = 7;

List<AgentforceOptimizeService.ObservabilityOutput> results =
    AgentforceOptimizeService.runObservabilityQuery(
        new List<AgentforceOptimizeService.ObservabilityInput>{ inp }
    );
System.debug('STDM_RESULT:' + results[0].resultJson);

Available query types: KnowledgeGap, Hallucination, RetrievalQuality, AnswerRelevancy, Leaderboard — see references/stdm-queries.md for the full table.

Do NOT use getMultipleConversationDetails or getLlmStepDetails in this skill. Those methods return raw session content (user messages, agent responses) which may contain PII. This skill operates on aggregate metrics only. See Redaction Rules below and references/privacy-redaction.md.

Step 6 — Sanitize output

Before emitting any result, apply all redaction rules from the Redaction Rules section and references/privacy-redaction.md. Specifically:

• Never include session text content (user messages, agent responses). Aggregate metrics and structured scores only.
• Replace all Salesforce record IDs with <record_id_placeholder>.
• Replace all user IDs (session participants) with <user_id_placeholder>.
• Mask raw record IDs in action invocation steps — hash them to detect duplicates but never echo raw.
• If the org is a regulated vertical (Health Cloud, Financial Services Cloud ), apply the compliance-vertical flag and route any anomalies through salesforce-compliance-privacy-agent.

Step 7 — Emit audit envelope

Every execution must produce a complete audit envelope (see Audit Envelope Schema). Emit it unconditionally — even if the result set is empty or an error occurred.

Step 8 — Hand off if anomalies warrant

Compare results against the rubric in references/observability-rubric.md:

• Faithfulness below threshold → hand off to salesforce-agentforce-risk-review-skill
• Action error rate above threshold → hand off to salesforce-agentforce-risk-review-skill
• A configuration change is required → route through salesforce-live-guard-agent
• Regulated-vertical session anomalies → salesforce-compliance-privacy-agent

See Handoff Rules for the full escalation matrix.

Quality Scoring Rubric (100-point)

Score the observability execution quality before emitting results. Threshold: 80+ acceptable, 60–79 emit with caveat, below 60 reject and request revision.

Scoring penalties:

• Session text content emitted in output: immediate reject (score voided, refusal issued)
• Missing audit envelope: -20 (automatic caveat regardless of total score)
• No time-window filter on STDM query: -15
• Encrypted or PII fields echoed: immediate reject
• Configuration mutation attempted: immediate stop, route to Live Guard

T1 Least-Privilege Contract

This skill operates exclusively at T1 — read-only runtime. The contract is:

• OAuth scopes used: api, refresh_token, and cdp_query_api only. The cdp_query_api scope is required for Data Cloud SQL queries via the ConnectApi.CdpQuery namespace. No full, web, sfap_api, or any other scope.
• Run As account profile: System permissions: View Setup and Configuration and Data Cloud Query API access. Object permissions: Read only on objects in scope. FLS restricted to non-PII, non-encrypted fields by default.
• Denied permissions (enforced at Connected App and profile level):
  • Modify All Data
  • View All Data (system-level bypass)
  • View Encrypted Data
  • Modify Metadata Through Metadata API Functions
  • Author Apex
  • Customize Application
  • Manage Connected Apps
  • Manage Agentforce ( — permission API name subject to change)
• Org allowlist: Enforced by Connected App IP restrictions and explicit org alias allowlist. Skill verifies via sf org display that the target alias is in the authorized set before any query.
• No DML under any circumstances: This skill will not construct or execute any statement containing INSERT, UPDATE, DELETE, MERGE, or UPSERT.
• No agent mutation: This skill does not call sf agent publish, sf agent activate, sf project deploy start, or any command that modifies agent configuration or org state.
• Aggregate metrics only: This skill does not retrieve or emit raw session content, individual message texts, or LLM prompt/response pairs. Those are available in the observing-agentforce sf-skills pattern for use by human operators with appropriate data handling controls.
• Revocation: The least-privilege Run As account's refresh token can be rotated to instantly revoke all access without affecting other integrations.

Refusal Triggers

Stop immediately and do not execute if any of the following apply:

• The target org appears to be a production org but the Connected App allowlist does not include it.
• The user requests session text content (user messages, agent responses, LLM prompt/response pairs). This skill emits aggregate metrics only; raw content requires explicit human-in-the-loop confirmation via salesforce-live-guard-agent.
• The user requests cross-tenant queries (sessions from other orgs or Data Cloud tenants not owned by the connected org).
• The Run As account has Manage Agentforce permission present — this skill requires that permission to be explicitly denied.
• The Run As account is missing View Setup and Configuration — stop and escalate to org administrator.
• The org is a regulated-vertical production org (Health Cloud, Financial Services Cloud ) and jurisdiction or data classification is unknown.
• The audit envelope cannot be completed (missing matter_id, unresolvable org alias, or run_as_user_id unavailable).
• The user requests that redaction be skipped or disabled.
• Any mutation command is requested: sf project deploy start, sf agent publish, Apex DML, or equivalent.

Audit Envelope Schema

Every execution emits an audit envelope. The envelope travels with the sanitized output to any downstream skill.

audit_envelope:
  matter_id: "<caller-provided-or-generated-uuid>"
  skill_id: "salesforce-agentforce-stdm-observer-skill"
  skill_version: "0.1.0"
  target_org_alias: "<alias>"              # never the raw org ID
  run_as_user_id: "<user_id_placeholder>"  # placeholder; never real ID in output
  agent_master_label: "<label>"            # display name used for STDM filter
  data_space: "<data_space_name>"          # resolved Data Cloud data space
  query_types_executed: ["getAggregatedMetrics", "runObservabilityQuery"]
  time_window_start: "<ISO-8601-UTC>"
  time_window_end: "<ISO-8601-UTC>"
  redactions_applied:
    - type: "<session_content|user_id|record_id|pii>"
      reason: "<aggregate-only-policy|pii-risk|encrypted>"
  timestamp: "<ISO-8601-UTC>"
  org_type_verified: "sandbox | production"
  regulated_vertical_flag: true | false
  downstream_skill_recommended: "<skill-id or null>"

Output Format

All output is in YAML. Emit this structure for every execution.

verdict: "acceptable | caveat | reject"
quality_score: <0-100>
quality_notes: "<what drove the score>"

aggregate_metrics:
  sessions_count: <integer>
  total_turns: <integer>
  avg_quality_score: <float>           # 1.0-5.0 scale
  avg_faithfulness: <float>            # 0.0-1.0; null if unavailable
  avg_answer_relevance: <float>        # 0.0-1.0; null if unavailable
  avg_context_precision: <float>       # 0.0-1.0; null if unavailable
  action_invocation_count: <integer>   # total across all sessions
  action_error_count: <integer>
  error_rate: <float>                  # action_error_count / action_invocation_count
  abandonment_rate: <float>
  deflection_rate: <float>
  escalation_rate: <float>
  end_type_counts:
    USER_ENDED: <integer>
    AGENT_ENDED: <integer>
    UNKNOWN: <integer>
  quality_distribution:
    "5": <integer>
    "4": <integer>
    "3": <integer>
    "2": <integer>
    "1": <integer>
  top_intents:
    "<intent summary>": <count>
  unavailable_dmos: []

anomalies_detected:
  - dimension: "<faithfulness|relevance|error_rate|abandonment>"
    observed_value: <float>
    threshold: <float>
    severity: "low | medium | high | critical"
    interpretation: "<human-readable explanation>"

sanitized_sample_sessions: null
# Always null in this skill. Session content is never emitted.
# If per-session debugging is genuinely required, route through
# salesforce-live-guard-agent for human-in-the-loop confirmation.

audit_envelope:
  # See Audit Envelope Schema above

downstream_skill_recommendation: "<skill-id or null>"
downstream_routing_reason: "<why this skill was chosen>"

missing_evidence:
  - "<what additional data would improve confidence>"

assumptions:
  - "<explicit list of assumptions made>"

Redaction Rules

Apply in order. Do not bypass for any reason.

1. Session text content (user messages, agent responses, LLM prompts): Never include in any output. This is not a limitation — it is the intentional aggregate-only policy of this skill. Route content access through salesforce-live-guard-agent.
2. OAuth tokens, refresh tokens, session IDs: Never include in any output, log, or audit envelope field.
3. Salesforce Org IDs (18-char starting with 00D): Replace with <org_id_placeholder>.
4. Salesforce Record IDs (15/18-char) in action invocations: Hash to a deterministic short token (e.g., rec_a3f2) to detect duplicates without echoing the raw ID. Never emit raw record IDs.
5. User IDs (session participants, OwnerId, UserId): Replace with <user_id_placeholder>.
6. Customer names, email addresses, phone numbers in any field: Replace with <pii_redacted>.
7. Instance URLs and API endpoints: Omit from output; reference only the org alias in the audit envelope.
8. Agent version suffix on DeveloperName (e.g., _v9): Omit from output — it reveals internal versioning structure.
9. Encrypted fields (Shield PE / PMLE): Skip entirely. Do not emit the field name or any placeholder that implies a value was retrieved.

Handoff Rules

When metrics cross the thresholds in references/observability-rubric.md, hand off to the appropriate skill with the sanitized output and audit envelope as the payload.

Required handoff fields: matter_id, audit_envelope, aggregate_metrics (summary — not raw session data), anomalies_detected, assumptions.

Stop Conditions

Stop and do not continue if:

• Target org appears to be production but the Connected App allowlist excludes it — stop, emit a refusal with reason, do not execute any query.
• Data Cloud is unavailable in the target org — stop, inform the user, reference Setup → Data Cloud → Data Streams.
• The Run As account has Manage Agentforce permission granted — this skill requires it to be denied; stop and escalate to the org administrator.
• Session text content is requested by the user — stop, explain the aggregate-only policy, offer to route through salesforce-live-guard-agent.
• Cross-tenant queries are requested — stop and refuse.
• The audit envelope cannot be completed — stop until resolved.
• The user requests that redaction be disabled — stop and explain the policy.
• A mutation command is requested (deploy, publish, DML, configuration change) — stop immediately and route to salesforce-live-guard-agent for human approval.

Security Notes

• T1 read-only operational: No DML, no metadata mutation, no Apex authoring, no agent publishing or activation.
• Aggregate-only output: Session text content is never emitted under any circumstance. This is a structural policy, not a configurable option.
• Additional OAuth scope (cdp_query_api): Required for Data Cloud SQL queries. This scope does not grant write access to Data Cloud; it permits read-only queries against the cdp_query_api endpoint only.
• Sanitized output only: All Salesforce IDs, user IDs, and PII fields redacted before emission. Record IDs in action invocations are hashed.
• Org allowlist enforced: Connected App restricts which orgs can be targeted; skill verifies before executing.
• Structured audit emitted: Every execution produces a complete audit envelope regardless of result count.
• Revocable: Rotating the Run As account's refresh token immediately revokes all access without affecting other integrations.
• No credential echo: OAuth tokens, refresh tokens, and session IDs are never included in output or audit envelopes.
• Regulated-vertical escalation: Health Cloud and Financial Services Cloud orgs trigger mandatory escalation to salesforce-compliance-privacy-agent before results are shared externally.
• Manage Agentforce permission must be denied: If the Run As account has this permission, this skill refuses to operate. Read-only telemetry access does not require management permissions.

Reference File Index