agents-best-practices

Portable Tools - Cross-Device Development Methodology

Methodology for building tools that work across different devices, naming schemes, and configurations. Based on lessons from OAuth refresher debugging session (2026-01-23).

Core Principle

Never assume your device is the only device.

Your local setup is just one of many possible configurations. Build for the general case, not the specific instance.

---

The Three Questions (Before Writing Code)

1. "What varies between devices?"

Before writing any code that reads configuration, data, or credentials:

Ask:

• File paths? (macOS vs Linux, different home dirs)
• Account names? (user123 vs default vs oauth)
• Service names? (slight variations in spelling/capitalization)
• Data structure? (different versions, different formats)
• Environment? (different shells, different tools available)

Example from OAuth refresher:

• ❌ Assumed: Account is always "claude"
• ✅ Reality: Could be "claude", "Claude Code", "default", etc.

Action: List variables, make them configurable or auto-discoverable

---

2. "How do I prove this works?"

Before claiming success:

Require:

• Concrete BEFORE state (exact values)
• Concrete AFTER state (exact values)
• Proof they're different (side-by-side comparison)

Example from OAuth refresher:

BEFORE:
- Access Token: POp5z1fi...eSN9VAAA
- Expires: 1769189639000

AFTER:
- Access Token: 01v0RrFG...eOE9QAA ✅ Different
- Expires: 1769190268000 ✅ Extended

Action: Always show data transformation with real values

---

3. "What happens when it breaks?"

Before pushing to production:

Test:

• Wrong configuration (intentionally break config)
• Missing data (remove expected fields)
• Multiple entries (ambiguous case)
• Edge cases (empty values, special characters)

Example from OAuth refresher:

• Test with keychain_account: "wrong-name" → Fallback should work
• Test with incomplete keychain data → Should fail gracefully with helpful error

Action: Test failure modes, not just happy path

---

Mandatory Patterns

Pattern 1: Explicit Over Implicit

❌ Wrong:

# Ambiguous - returns first match
security find-generic-password -s "Service" -w

✅ Correct:

# Explicit - returns specific entry
security find-generic-password -s "Service" -a "account" -w

Rule: If a command can be ambiguous, make it explicit.

---

Pattern 2: Validate Before Use

❌ Wrong:

DATA=$(read_config)
USE_VALUE="$DATA"  # Hope it's valid

✅ Correct:

DATA=$(read_config)
if ! validate_structure "$DATA"; then
    error "Invalid data structure"
fi
USE_VALUE="$DATA"

Rule: Never assume data has expected structure.

---

Pattern 3: Fallback Chains

❌ Wrong:

ACCOUNT="claude"  # Hardcoded

✅ Correct:

# Try configured → Try common → Error with help
ACCOUNT="${CONFIG_ACCOUNT}"
if ! has_data "$ACCOUNT"; then
    for fallback in "claude" "default" "oauth"; do
        if has_data "$fallback"; then
            ACCOUNT="$fallback"
            break
        fi
    done
fi
[[ -z "$ACCOUNT" ]] && error "No account found. Tried: ..."

Rule: Provide automatic fallbacks for common variations.

---

Pattern 4: Helpful Errors

❌ Wrong:

[[ -z "$TOKEN" ]] && error "No token"

✅ Correct:

[[ -z "$TOKEN" ]] && error "No token found

Checked:
- Config: $CONFIG_FILE
- Field: $FIELD_NAME
- Expected: { \"tokens\": { \"refresh\": \"...\" } }

Verify with:
  cat $CONFIG_FILE | jq '.tokens'
"

Rule: Error messages should help user diagnose and fix.

---

Debugging Methodology (Patrick's Approach)

Step 1: Get Exact Data

Don't ask: "Is it broken?"
Ask: "What exact values do you see? How many entries exist? Which one has the data?"

Example:

# Vague
"Check keychain"

# Specific
"Run: security find-generic-password -l 'Service' | grep 'acct'"
"Tell me: 1. How many entries 2. Which has tokens 3. Last modified"

---

Step 2: Prove With Concrete Examples

Don't say: "It should work now"
Show: "Here's the BEFORE token (POp5z...), here's AFTER (01v0R...), they're different"

Template:

BEFORE:
- Field1: <exact_value>
- Field2: <exact_value>

AFTER:
- Field1: <new_value> ✅ Changed
- Field2: <new_value> ✅ Changed

PROOF: Values are different

---

Step 3: Think Cross-Device Immediately

Don't think: "Works on my machine"
Think: "What if their setup differs in [X]?"

Checklist:

• [ ] Different account names?
• [ ] Different file paths?
• [ ] Different tools/versions?
• [ ] Different permissions?
• [ ] Different data formats?

---

Pre-Flight Checklist (Before Publishing)

Discovery Phase

• [ ] List all external dependencies (files, commands, services)
• [ ] Document what each dependency provides
• [ ] Identify which parts could vary between devices

Implementation Phase

• [ ] Make variations configurable (with sensible defaults)
• [ ] Add validation for each input
• [ ] Build fallback chains for common variations
• [ ] Add --dry-run or --test mode

Testing Phase

• [ ] Test with correct config → Should work
• [ ] Test with wrong config → Should fallback or fail gracefully
• [ ] Test with missing data → Should give helpful error
• [ ] Test with multiple entries → Should handle ambiguity

Documentation Phase

• [ ] Document default assumptions
• [ ] Document how to verify local setup
• [ ] Document common variations and how to handle them
• [ ] Include data flow diagram
• [ ] Add troubleshooting section

---

Real-World Example: OAuth Refresher

Original (Broken)

# Assumes single entry, no validation, no fallback
KEYCHAIN_DATA=$(security find-generic-password -s "Service" -w)
REFRESH_TOKEN=$(echo "$KEYCHAIN_DATA" | jq -r '.refreshToken')
# Use token (hope it's valid)

Problems:

• Returns first alphabetical match (wrong entry)
• No validation (could be empty/malformed)
• No fallback (fails if account name differs)

---

Fixed (Portable)

# Explicit account with validation and fallback
validate_data() {
    echo "$1" | jq -e '.claudeAiOauth.refreshToken' > /dev/null 2>&1
}

# Try configured account
DATA=$(security find-generic-password -s "$SERVICE" -a "$ACCOUNT" -w 2>&1)
if validate_data "$DATA"; then
    log "✓ Using account: $ACCOUNT"
else
    log "⚠ Trying fallback accounts..."
    for fallback in "claude" "Claude Code" "default"; do
        DATA=$(security find-generic-password -s "$SERVICE" -a "$fallback" -w 2>&1)
        if validate_data "$DATA"; then
            ACCOUNT="$fallback"
            log "✓ Found data in: $fallback"
            break
        fi
    done
fi

[[ -z "$DATA" ]] || ! validate_data "$DATA" && error "No valid data found
Tried accounts: $ACCOUNT, claude, Claude Code, default
Verify with: security find-generic-password -l '$SERVICE'"

REFRESH_TOKEN=$(echo "$DATA" | jq -r '.claudeAiOauth.refreshToken')

Improvements:

• ✅ Explicit account parameter
• ✅ Validates data structure
• ✅ Automatic fallback to common names
• ✅ Helpful error with verification command

---

Common Anti-Patterns

Anti-Pattern 1: "Works On My Machine"

FILE="/Users/patrick/.config/app.json"  # Hardcoded path

Fix: Use $HOME, detect OS, or make configurable

---

Anti-Pattern 2: "Hope It's There"

TOKEN=$(cat config.json | jq -r '.token')
# What if .token doesn't exist? Script continues with empty value

Fix: Validate before using

TOKEN=$(cat config.json | jq -r '.token // empty')
[[ -z "$TOKEN" ]] && error "No token in config"

---

Anti-Pattern 3: "First Match Is Right"

# If multiple entries exist, which one?
ENTRY=$(find_entry "service")

Fix: Be explicit or enumerate all

ENTRY=$(find_entry "service" "account")  # Specific
# OR
ALL=$(find_all_entries "service")
for entry in $ALL; do
    validate_and_use "$entry"
done

---

Anti-Pattern 4: "Silent Failures"

process_data || true  # Ignore errors

Fix: Fail loudly with context

process_data || error "Failed to process
Data: $DATA
Expected: { ... }
Check: command_to_verify"

---

Integration With Existing Workflows

With sprint-plan.md

Add to testing section:

## Cross-Device Testing
- [ ] Test with different account names
- [ ] Test with wrong config values
- [ ] Test with missing data
- [ ] Document fallback behavior

With PRIVACY-CHECKLIST.md

Add before publishing:

## Portability Check
- [ ] No hardcoded paths (use $HOME, detect OS)
- [ ] No hardcoded names (use config or fallback)
- [ ] Validation on all inputs
- [ ] Helpful errors for common issues

With skill-creator

When building new skills:

1. List what varies between devices
2. Make it configurable or auto-discoverable
3. Test with wrong config
4. Document troubleshooting

---

Quick Reference Card

Before writing code:

1. What varies between devices?
2. How do I prove this works?
3. What happens when it breaks?

Mandatory patterns:

• Explicit over implicit
• Validate before use
• Fallback chains
• Helpful errors

Testing:

• Correct config → Works
• Wrong config → Fallback or helpful error
• Missing data → Clear diagnostic

Documentation:

• Data flow diagram
• Common variations
• Troubleshooting guide

---

Success Criteria

A tool is portable when:

1. ✅ Works on different devices without modification
2. ✅ Auto-discovers common variations in setup
3. ✅ Fails gracefully with actionable error messages
4. ✅ Can be debugged by reading the error output
5. ✅ Documentation covers "what if my setup differs"

Test: Give it to someone with a different setup. If they need to ask you questions, the tool isn't portable yet.

---

Origin Story

This methodology emerged from debugging the OAuth refresher (2026-01-23):

• Script read wrong keychain entry (didn't specify account)
• Assumed single entry existed (multiple did)
• No validation (used empty data)
• No fallback (failed on different account names)

Patrick's approach:

1. Asked for exact data (how many entries, which has tokens)
2. Demanded proof (show BEFORE/AFTER tokens)
3. Thought cross-device (what if naming differs?)

Result: Tool went from single-device/broken to universal/production-ready.

Key insight: The bugs weren't in the logic - they were in the assumptions.

---

When To Use This Skill

Use when:

• Building tools that read system configuration
• Working with keychains, credentials, environment variables
• Creating scripts that run on multiple machines
• Publishing skills to ClawdHub (others will use them)

Apply:

1. Before implementing: Answer the three questions
2. During implementation: Use mandatory patterns
3. Before testing: Run pre-flight checklist
4. After testing: Document variations and troubleshooting

Remember: Your device is just one case. Build for the general case.