Skill debug

Skill: AL Debugging & Diagnostics

Purpose

Systematic diagnosis and root cause analysis for AL Business Central issues: runtime errors, logic bugs, intermittent failures, performance bottlenecks, and configuration problems.

When to Load

This skill should be loaded when:

• A runtime error, exception, or unexpected behavior is reported
• A logic bug needs root cause analysis (not just a symptom fix)
• An intermittent issue needs snapshot capture
• A performance bottleneck needs CPU profiling
• Configuration issues (auth, symbols, build, publishing) block development
• An event subscriber is not firing as expected

Core Patterns

Pattern 1: Choose Debugging Strategy

Select the right tool before starting:

Issue Type	Strategy	Tool
Consistent runtime error	Standard debugger	al_debug
Already deployed code	Debug without publish	al_debug
Rapid dev cycle	Incremental publish	al_publish (incremental)
Intermittent / hard-to-reproduce	Snapshot debugging	al_snapshotdebugging
Slow performance	CPU profiling	VS Code command (not an agent tool)
Auth / symbols / build	Configuration troubleshoot	See Workflow Step 2b
Copilot AI feature	Agent session debug	launch.json with clientType: Agent

Pattern 2: Data Flow Tracing

Scenario: "Value is wrong after posting"

1. Set breakpoint at final location (where value is wrong)
2. Work backwards to find where value is set
3. Use `usages` tool to find all assignments
4. Set breakpoints at each assignment point
5. Step through to find which execution path is taken
6. Inspect conditions and variable states at each point

Pattern 3: Event Subscriber Not Firing

Scenario: "My event subscriber doesn't execute"

1. Verify subscriber signature exactly matches publisher (name, parameters, types)
2. Check SkipOnMissingLicense and SkipOnMissingPermission attributes
3. Confirm the extension containing the publisher is active and published
4. Set breakpoint inside subscriber body
5. Verify the publisher event is actually being raised (breakpoint in publisher)
6. Check if IsHandled = true is set before your subscriber runs
7. Inspect subscriber execution order (EventPriority)

Pattern 4: Performance Bottleneck

// ❌ N+1 pattern — common cause of slow pages
repeat
    Item.Get(SalesLine."No.");      // DB call inside loop
    SalesLine.Amount := Item.Price * SalesLine.Quantity;
until SalesLine.Next() = 0;

// ✅ Use SetLoadFields + single pass
SalesLine.SetLoadFields("No.", Quantity, Amount);
if SalesLine.FindSet() then
    repeat
        // ...
    until SalesLine.Next() = 0;

CPU profile analysis focus:

• Top time-consuming procedures (hotspots)
• Frequently called procedures in loops
• Expensive database operations
• FlowField CalcFormula complexity

Pattern 5: Snapshot Debugging for Intermittent Issues

⚠️ HUMAN GATE: Snapshots may capture sensitive runtime data.
Before initializing:
1. Confirm what data will be captured
2. Security review for sensitive information
3. Obtain explicit user approval

Commands (one tool — `al_snapshotdebugging` — covers initialize / finish / view):
  al_snapshotdebugging (initialize)   ← start capture session
  [reproduce scenario 10-20 times]
  al_snapshotdebugging (finish)       ← end capture
  al_snapshotdebugging (view)         ← view and compare

Compare snapshots between success and failure cases:
- Variable values at failure point
- Timing differences
- Execution paths taken
- Data state variations

Workflow

Step 1: Context Gathering (MANDATORY before any debugging)

Read existing plans context first:

.github/plans/memory.md              ← project state and recent decisions
.github/plans/*-diagnosis.md         ← previous debug sessions (similar issues)

Gather issue information:

• What is the expected behavior vs actual behavior?
• When does it happen: always / sometimes / specific conditions?
• Exact error message and stack trace
• Recent code changes (potential regression source)
• User/permission set and environment where it occurs

Stop criterion: can reproduce the issue consistently (≥80% success rate).

Step 2a: Isolate the Problem (Runtime / Logic)

1. Narrow down scope with search and usages tools
2. Identify suspect objects (tables, pages, codeunits, event subscribers)
3. Attach debugger with selected strategy (Pattern 1)
4. Set strategic breakpoints:
   • Just before the error occurs
   • At data modification points
   • In event subscribers and validation triggers
5. Inspect: variable values, record filters, call stack, parameter values
6. Step execution: F10 (step over), F11 (step into), Shift+F11 (step out)

Step 2b: Configuration Troubleshooting

Authentication failures (401/403, cannot download symbols):

⚠️ HUMAN GATE: Clearing credentials disconnects active sessions.
Confirm impact and obtain approval before clearing the credentials cache
(a VS Code command, not an agent tool).
Then re-authenticate and verify launch.json authentication method.

Missing symbols (unresolved references, red squiggles):

al_downloadsymbols                      ← first attempt
al_downloadsymbols (globalSourcesOnly)  ← if no BC server connection is available
al_build                                ← verify compilation

Check app.json dependencies version alignment.

Build errors — common codes:

• AL0896 — Recursive FlowField: map dependency graph, break circular chain
• AL0185 — Object ID conflict: check ID range in app.json, no duplicates
• AL0118 — Field length mismatch: align extension field length with base table

Publishing failures: check environment connectivity, extension version increment, dependency resolution via al_packages.

Step 3: Diagnose Root Cause

Identify the exact point of failure and — critically — WHY it occurs:

Common AL root causes by scenario:

Symptom	Root Cause Candidates
Record Not Found	Wrong key values, filters blocking read, wrong company context, permission on read
FlowField wrong value	CalcFields not called, incorrect CalcFormula filters, circular dependency (AL0896)
OnValidate not firing	Direct assignment without Validate(), IsHandled = true in subscriber, disabled trigger
Custom event not firing	Signature mismatch, wrong ObjectType/ObjectId, publisher not raising event, compilation error in subscriber
Slow page	Missing SetLoadFields, N+1 loop, FlowField CalcFormula too complex, missing index/key
Intermittent failure	Race condition, timing dependency, data-state dependency, environment-specific

Step 4: Document Diagnosis (MANDATORY)

Create .github/plans/<issue-kebab-case>-diagnosis.md before proposing any fix:

# Debug Session: <Issue Title>

**Date**: YYYY-MM-DD
**Severity**: Critical / High / Medium / Low
**Status**: Investigating → Diagnosed → Fixed → Verified

## Issue Summary
[Brief description]

## Symptoms
- Expected: [what should happen]
- Actual: [what happens — include exact error text]

## Reproduction Steps
1. [Step]
2. [Step]
**Reproducibility**: Always / Sometimes / Rare

## Root Cause
[Technical explanation with evidence — code snippet, line reference]

## Recommended Fix
[Specific solution. Short-term hotfix + long-term if needed]

## Testing Strategy
- Unit: [what to mock/test]
- Regression: [related functionality to verify]

## Next Steps
1. [Action]
2. [Action]

File naming: sales-posting-error-diagnosis.md, slow-customer-list-diagnosis.md.

Step 5: Propose Fix & Handoff (MANDATORY HITL gate)

1. Design fix addressing root cause (never symptoms only)
2. Consider edge cases and side effects
3. PAUSE — present diagnosis and proposed fix to user for approval
4. After approval: handoff implementation to al-developer (simple) or via al-conductor TDD cycle (complex refactor)
5. After fix: run tests, re-profile if performance issue, verify regression scenarios

Clean up: remove debugging artifacts, breakpoints, and temporary logging from code.

References

• AL Debugger — Microsoft Docs
• Snapshot Debugging
• AL Code Analysis — AL0896
• Performance Profiler

Constraints

• This skill covers diagnosis and root cause analysis only — code changes are implemented by al-developer or via al-conductor
• NEVER guess solutions without evidence from debugging tools
• NEVER skip Step 4 (diagnosis document) for non-trivial issues (>5 min investigation)
• NEVER proceed to fix without user approval (Step 5 HITL gate)
• NEVER leave debugging code, breakpoints, or temporary logging in production code
• Performance deep-dive (SetLoadFields, FlowField patterns, query optimization) → load skill-performance.md
• Event publisher/subscriber design patterns → load skill-events.md
• Permission set generation → load skill-permissions.md