salesforce-flow-debugger-skill

salesforce-flow-debugger-skill

Hybrid T0/T1 skill that diagnoses Salesforce Flow failures. Works in two modes:

• T0 mode (default): Takes a pasted Flow error message or pasted Flow Interview log export. No org connection needed.
• T1 mode (optional): Fetches the Flow definition via sf org list metadata and queries FlowInterviewLog records via sf data query to retrieve live failure context.

Identifies the failing node, root cause, fix recommendation, and fault path design guidance. Outputs a structured diagnosis with an audit envelope.

When This Skill Owns the Task

Use salesforce-flow-debugger-skill when the work is to diagnose why a Flow failed or is not working:

• "My Flow failed with: 'This record failed to save because...'"
• "Here's the Flow Interview log — what went wrong?"
• "Why does my Account Before-Save Flow throw a null pointer on the Decision node?"
• "Flow: Order Notification fails on Send Email Action with UNHANDLED_FAULT"
• "How do I add a fault path to prevent this error?"

Delegate elsewhere when:

Required Context to Gather First

Before diagnosing, gather:

1. Error message text — the exact error string from the flow failure notification, debug log, or Flow Interview log entry.
2. Flow API name — the developer name of the failing Flow, if known.
3. Flow type — Screen Flow, Auto-launched Flow, Record-Triggered Flow, Scheduled Flow, or Subflow.
4. Trigger context — what action or event triggered the Flow? User action, record save, scheduled batch, or REST API call?
5. Failing node name — the element name (if visible in the error or the Flow builder debug view).
6. Org type — sandbox or production. Production errors require stricter redaction.
7. Recent changes — was the Flow recently modified or deployed? What changed?

In T1 mode, additionally confirm:

• Target org alias recognized by sf org list
• Org is sandbox (T1 mode is restricted to sandbox for live log fetch)

Recommended Workflow

Step 1 — Parse the error message

Extract from the error text:

• Fault type — UNHANDLED_FAULT, NullPointerException, DML Exception, Governor Limit, Type Mismatch, Recursive Flow
• Failing element — the element name or API action mentioned
• Variable state context — any variable values or record ID fragments in the error (mask immediately per redaction rules)
• Stack trace — if present, extract the top frame

Apply redaction (see Redaction Rules) before continuing analysis.

Step 2 (T1 mode only) — Fetch Flow metadata

sf org list metadata \
  --metadata-type Flow \
  --target-org <alias>

Filter for the failing Flow by API name. Use the result to confirm the Flow version deployed and retrieve the element count.

Step 3 (T1 mode only) — Query FlowInterviewLog

sf data query \
  --query "SELECT Id, FlowApiName, InterviewLabel, CurrentElement, ErrorCode, ErrorMessage, StartTime, EndTime FROM FlowInterviewLog WHERE FlowApiName = '<FlowApiName>' AND Status = 'Fault' ORDER BY StartTime DESC LIMIT 10" \
  --target-org <alias> \
  --result-format json

Redact all record IDs and variable values before analysis. Note: FlowInterviewLog and FlowInterviewLogEntry are available in orgs with Flow Interview Logging enabled (must be activated in Setup → Process Automation Settings).

Step 4 — Classify the root cause

Map the parsed error to a root cause pattern (see references/flow-error-patterns.md):

Step 5 — Generate specific fix recommendation

For each root cause, generate:

1. Immediate fix — what to change in the Flow builder right now
2. Fault path recommendation — should a fault connector be added? Where should the fault path go? (See references/fault-path-design.md)
3. Null guard — if null variables are involved, where should the null check Decision element be placed before the failing element?
4. Data type fix — if type mismatch, what types need to match and how to add an explicit conversion or intermediate variable

Step 6 — Emit structured diagnosis with audit envelope

Produce the full output block per the Output Format section below.

Quality Scoring Rubric (100-point)

Score every diagnosis before emitting. Threshold: 80+ ship, 60–79 ship with caveat, below 60 reject and request more context.

Scoring penalties:

• Generic recommendation ("check the error") with no specific element: -25
• Missing fault path guidance when UNHANDLED_FAULT is the error type: -20
• Unredacted record ID or user ID in output: -30 (immediate caveat)
• Incorrect fault type classification: -20

T0/T1 Contract

T0 Mode (default, no org connection)

• Takes pasted error message or pasted FlowInterviewLog export as input.
• All analysis is static — no org connection.
• allowed-tools: Read Grep Glob only.
• No sf CLI commands executed.

T1 Mode (optional, read-only runtime)

• Activates when user confirms org alias and explicitly requests live log fetch.
• OAuth scopes: api and refresh_token only.
• Run As permissions: View Setup and Configuration, View Setup.
• Denied: ModifyAllData, ViewAllData, ViewEncryptedData, ModifyMetadata, AuthorApex.
• Restricted to sandbox orgs — do not query FlowInterviewLog on production in T1 mode without explicit user confirmation and production safety review.
• Maximum query: 10 most recent fault records. No bulk log extraction.

Refusal Triggers

Stop and decline if:

• The request is to write a new Flow from scratch (out of scope).
• The request is to deploy the fixed Flow to a production org (route to salesforce-deployment-validator-skill).
• In T1 mode: target org appears to be production and user has not explicitly confirmed the production safety review step.
• The audit envelope cannot be populated in T1 mode (org alias unresolvable, flow name missing).
• The user requests that redaction be skipped.

Audit Envelope Schema

Every T1 execution emits an audit envelope. T0 mode emits a reduced envelope (no org fields).

audit_envelope:
  matter_id: "<caller-provided-or-generated-uuid>"
  skill_id: "salesforce-flow-debugger-skill"
  skill_version: "0.1.0"
  mode: "<T0-static | T1-live>"
  target_org_alias: "<alias or 'N/A for T0'>"
  flow_api_name: "<FlowApiName>"
  flow_type: "<RecordTriggered|Screen|AutoLaunched|Scheduled>"
  run_as_user_id: "<user_id_placeholder>"
  log_records_queried: <integer or 0 for T0>
  redactions_applied:
    - field: "<FieldOrVariableName>"
      reason: "<record_id|user_id|variable_value|pii>"
  timestamp: "<ISO-8601-UTC>"
  org_type_verified: "<sandbox | production | N/A for T0>"

Output Format

verdict: "diagnosis-complete | needs-more-context | reject"
quality_score: <0-100>
quality_notes: "<what drove the score>"

diagnosis:
  flow_api_name: "<FlowApiName or 'unknown'>"
  flow_type: "<type>"
  error_type: "<UNHANDLED_FAULT|NullPointerException|DML Exception|Governor Limit|Type Mismatch|Recursive|Other>"
  failing_element_name: "<ElementName or 'unknown'>"
  failing_element_type: "<Action|Decision|Assignment|Loop|GetRecords|UpdateRecords|Screen|Subflow>"
  root_cause_summary: "<1-2 sentence plain-language explanation>"
  root_cause_detail: "<technical detail with element names and variable context>"

fix_recommendation:
  immediate_fix: "<specific steps in Flow Builder>"
  fault_path_required: <true|false>
  fault_path_design: "<where to add it and where it should terminate>"
  null_guard_required: <true|false>
  null_guard_placement: "<before which element and what Decision logic>"
  data_type_fix: "<if type mismatch: what types, where to add intermediate variable>"
  governor_limit_fix: "<if governor limit: what to move outside the loop>"

fault_path_template:
  trigger: "<which element needs the fault connector>"
  fault_variable_capture: "<Fault Message variable assignment>"
  fault_path_ends_at: "<Screen|Custom Notification|Log to Custom Object|Email>"
  sample_fault_message_variable: "{!$Flow.FaultMessage}"

redaction_log:
  - "<description of what was masked>"

audit_envelope:
  <see Audit Envelope Schema>

escalation_triggers_fired:
  - "<trigger name or 'none'>"

missing_evidence:
  - "<what additional context would improve the diagnosis>"

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

Redaction Rules

Apply in order. Do not bypass for any reason.

1. Salesforce Record IDs (15/18-char): Replace with <record_id_placeholder> anywhere they appear in error messages or log entries.
2. User IDs (OwnerId, CreatedById, RunningUserId in logs): Replace with <user_id_placeholder>.
3. Flow variable values that may contain PII (email, phone, name, address): Replace with <variable_value_redacted>.
4. Org IDs (18-char starting 00D): Replace with <org_id_placeholder>.
5. OAuth tokens, session IDs: Strip entirely — never include in output.
6. Instance URLs: Replace with <org_instance_placeholder>.
7. Stack trace class paths containing customer namespace: Preserve class names but redact any embedded record IDs or data values.

Document each redaction in redaction_log.

Handoff Rules

Stop Conditions

• Target org is production and T1 mode was requested without explicit production safety acknowledgment — stop and require acknowledgment.
• FlowInterviewLog object is not queryable (Flow Interview Logging not enabled in org) — stop T1 fetch, fall back to T0 mode with error message.
• Error message contains no element name and no variable context — request additional pasted context (full error text, Flow debug log) before proceeding.
• User requests to run a DML fix directly in T1 mode — refuse; route to human-approval path via salesforce-live-guard-agent.

Security Notes

• T1 read-only operational: Queries FlowInterviewLog only. No DML, no metadata mutation, no Apex execution.
• Sanitized output only: All record IDs, user IDs, and variable values redacted before emission.
• Sandbox-preferred for T1: Production FlowInterviewLog access requires explicit user confirmation and applies stricter redaction.
• Structured audit emitted: Every T1 execution produces a complete audit envelope.
• Least-privilege Run As account: No Modify All Data, View All Data, View Encrypted Data, or Modify Metadata.

Reference File Index