* Refactor CLAUDE.md and related files for December 2025 updates - Updated CLAUDE.md in src/services/worker with new entries for December 2025, including changes to Search.ts, GeminiAgent.ts, SDKAgent.ts, and SessionManager.ts. - Revised CLAUDE.md in src/shared to reflect updates and new entries for December 2025, including paths.ts and worker-utils.ts. - Modified hook-constants.ts to clarify exit codes and their behaviors. - Added comprehensive hooks reference documentation for Claude Code, detailing usage, events, and examples. - Created initial CLAUDE.md files in various directories to track recent activity. * fix: Merge user-message-hook output into context-hook hookSpecificOutput - Add footer message to additionalContext in context-hook.ts - Remove user-message-hook from SessionStart hooks array - Fixes issue where stderr+exit(1) approach was silently discarded Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> * Update logs and documentation for recent plugin and worker service changes - Added detailed logs for worker service activities from Dec 10, 2025 to Jan 7, 2026, including initialization patterns, cleanup confirmations, and diagnostic logging. - Updated plugin documentation with recent activities, including plugin synchronization and configuration changes from Dec 3, 2025 to Jan 7, 2026. - Enhanced the context hook and worker service logs to reflect improvements and fixes in the plugin architecture. - Documented the migration and verification processes for the Claude memory system and its integration with the marketplace. * Refactor hooks architecture and remove deprecated user-message-hook - Updated hook configurations in CLAUDE.md and hooks.json to reflect changes in session start behavior. - Removed user-message-hook functionality as it is no longer utilized in Claude Code 2.1.0; context is now injected silently. - Enhanced context-hook to handle session context injection without user-visible messages. - Cleaned up documentation across multiple files to align with the new hook structure and removed references to obsolete hooks. - Adjusted timing and command execution for hooks to improve performance and reliability. * fix: Address PR #610 review issues - Replace USER_MESSAGE_ONLY test with BLOCKING_ERROR test in hook-constants.test.ts - Standardize Claude Code 2.1.0 note wording across all three documentation files - Exclude deprecated user-message-hook.ts from logger-usage-standards test Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> * fix: Remove hardcoded fake token counts from context injection Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> * Address PR #610 review issues by fixing test files, standardizing documentation notes, and verifying code quality improvements. * fix: Add path validation to CLAUDE.md distribution to prevent invalid directory creation - Add isValidPathForClaudeMd() function to reject invalid paths: - Tilde paths (~) that Node.js doesn't expand - URLs (http://, https://) - Paths with spaces (likely command text or PR references) - Paths with # (GitHub issue/PR references) - Relative paths that escape project boundary - Integrate validation in updateFolderClaudeMdFiles loop - Add 6 unit tests for path validation - Update .gitignore to prevent accidental commit of malformed directories - Clean up existing invalid directories (~/, PR #610..., git diff..., https:) Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: Implement path validation in CLAUDE.md generation to prevent invalid directory creation - Added `isValidPathForClaudeMd()` function to validate file paths in `src/utils/claude-md-utils.ts`. - Integrated path validation in `updateFolderClaudeMdFiles` to skip invalid paths. - Added 6 new unit tests in `tests/utils/claude-md-utils.test.ts` to cover various rejection cases. - Updated `.gitignore` to prevent tracking of invalid directories. - Cleaned up existing invalid directories in the repository. * feat: Promote critical WARN logs to ERROR level across codebase Comprehensive log-level audit promoting 38+ WARN messages to ERROR for improved debugging and incident response: - Parser: observation type errors, data contamination - SDK/Agents: empty init responses (Gemini, OpenRouter) - Worker/Queue: session recovery, auto-recovery failures - Chroma: sync failures, search failures (now treated as critical) - SQLite: search failures (primary data store) - Session/Generator: failures, missing context - Infrastructure: shutdown, process management failures - File Operations: CLAUDE.md updates, config reads - Branch Management: recovery checkout failures Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> * fix: Address PR #614 review issues - Remove incorrectly tracked tilde-prefixed files from git - Fix absolute path validation to check projectRoot boundaries - Add test coverage for absolute path validation edge cases Closes review issues: - Issue 1: ~/ prefixed files removed from tracking - Issue 3: Absolute paths now validated against projectRoot - Issue 4: Added 3 new test cases for absolute path scenarios Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> * build assets and context --------- Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
14 KiB
Issue #588: Unexpected API Charges from ANTHROPIC_API_KEY Discovery
Date: January 7, 2026 Status: INVESTIGATION COMPLETE - Critical UX/Financial Issue Priority: HIGH Labels: bug, financial-impact, ux Author: imkane Version Affected: 9.0.0 and earlier
Executive Summary
A user with a Claude Max subscription ($100/month) began receiving unexpected "Auto-recharge credits" invoice emails from Anthropic after installing the claude-mem plugin. The plugin discovered an ANTHROPIC_API_KEY in a .env file in the project root and used it for AI operations (observation compression, summary generation), causing direct API charges that were not anticipated by the user.
Financial Impact: The user expected all AI costs to be covered by their Claude Max subscription. Instead, the plugin consumed their Anthropic API credits separately, triggering auto-recharge billing.
Root Cause: The Claude Agent SDK (@anthropic-ai/claude-agent-sdk) automatically discovers and uses ANTHROPIC_API_KEY from environment variables or .env files. Claude-mem's worker service runs AI operations (observation compression, summaries) through this SDK, which consumes API credits independently of the user's Claude Max subscription.
Problem Analysis
User Expectations vs. Reality
| Expectation | Reality |
|---|---|
| Claude Max ($100/mo) covers all Claude usage | Claude Max covers Claude Code IDE usage only |
| Plugin enhances Claude Code without extra cost | Plugin uses separate API calls via SDK |
| No API key needed since using Claude Max | SDK auto-discovers .env API keys |
| Billing would be transparent | Silent API key discovery leads to surprise charges |
The Discovery Flow
- User installs claude-mem plugin via marketplace
- User has an
ANTHROPIC_API_KEYin project.envfile (for other purposes) - Plugin worker starts on first Claude Code session
- Worker spawns Claude Agent SDK for observation processing
- SDK auto-discovers
ANTHROPIC_API_KEYfrom environment - Every observation compression and session summary uses API credits
- User receives unexpected invoice for API usage
Technical Details
How claude-mem Uses the Claude Agent SDK
Claude-mem uses @anthropic-ai/claude-agent-sdk (version ^0.1.76) for AI-powered operations:
File: /Users/alexnewman/conductor/workspaces/claude-mem/budapest/src/services/worker/SDKAgent.ts
// Line 26
import { query } from '@anthropic-ai/claude-agent-sdk';
// Lines 100-114 - SDK query execution
const queryResult = query({
prompt: messageGenerator,
options: {
model: modelId,
...(hasRealMemorySessionId && session.lastPromptNumber > 1 && { resume: session.memorySessionId }),
disallowedTools,
abortController: session.abortController,
pathToClaudeCodeExecutable: claudePath
}
});
API Key Discovery Chain
The Claude Agent SDK uses a standard discovery mechanism:
- Environment Variable:
process.env.ANTHROPIC_API_KEY - File Discovery:
~/.anthropic/api_keyor project.envfiles - Inherited Environment: Claude Code passes its environment to spawned processes
From hooks architecture documentation (line 826-828):
### API Key Protection
**Configuration:**
- Anthropic API key in `~/.anthropic/api_key` or `ANTHROPIC_API_KEY` env var
- Worker inherits environment from Claude Code
- Never logged or stored in database
Worker Service Environment Inheritance
File: /Users/alexnewman/conductor/workspaces/claude-mem/budapest/src/services/worker-service.ts
// Line 263 - Worker spawns with full environment
env: process.env
File: /Users/alexnewman/conductor/workspaces/claude-mem/budapest/src/services/infrastructure/ProcessManager.ts
// Line 273 - Process inherits environment
...process.env,
This means any ANTHROPIC_API_KEY present in the parent process environment or discovered from .env files will be used by the worker.
What Operations Consume API Credits
| Operation | Trigger | API Usage |
|---|---|---|
| Observation Compression | PostToolUse hook | ~0.5-2K tokens per observation |
| Session Summary | Summary hook | ~2-5K tokens per session |
| Follow-up Queries | Multi-turn processing | Variable |
Estimated Usage Per Session:
- Active coding session: 20-50 tool uses
- At ~1.5K tokens per observation: 30-75K tokens
- Session summary: ~3K tokens
- Total per session: 33-78K tokens
Alternative Providers (Not Using Anthropic API)
Claude-mem supports alternative AI providers that DO NOT use the Anthropic API:
File: /Users/alexnewman/conductor/workspaces/claude-mem/budapest/src/services/worker/GeminiAgent.ts
// Line 376
const apiKey = settings.CLAUDE_MEM_GEMINI_API_KEY || process.env.GEMINI_API_KEY || '';
File: /Users/alexnewman/conductor/workspaces/claude-mem/budapest/src/services/worker/OpenRouterAgent.ts
// Line 418
const apiKey = settings.CLAUDE_MEM_OPENROUTER_API_KEY || process.env.OPENROUTER_API_KEY || '';
These providers require explicit configuration and do not auto-discover.
Impact Assessment
Financial Impact
| Scenario | Estimated Monthly Cost |
|---|---|
| Light usage (5 sessions/day) | $10-30 |
| Moderate usage (15 sessions/day) | $30-90 |
| Heavy usage (30+ sessions/day) | $90-200+ |
Compounding Factors:
- Auto-recharge enabled by default on Anthropic accounts
- No notification before charges occur
- User may not realize plugin is source of usage
User Experience Impact
- Trust Violation: Users expect plugins to be transparent about costs
- Subscription Confusion: Claude Max subscription doesn't cover SDK API usage
- No Consent: API key used without explicit opt-in
- Discovery Difficulty: Source of charges not immediately obvious
Affected User Base
- Users with
ANTHROPIC_API_KEYin project.envfiles - Users with API key in
~/.anthropic/api_key - Users who exported
ANTHROPIC_API_KEYfor other tools - Users who don't know they have an API key configured
Root Cause Analysis
Primary Root Cause
Silent API key auto-discovery by the Claude Agent SDK without user consent or notification.
The SDK is designed for developer use cases where explicit API key configuration is expected. When used within a plugin context, the automatic discovery behavior creates a mismatch between user expectations and system behavior.
Contributing Factors
- No Pre-Flight Check: Plugin doesn't warn users that it will use API credits
- No Opt-In Flow: API key usage happens automatically without consent
- No Usage Visibility: No way to see API consumption before it happens
- Documentation Gap: Not clearly documented that separate API credits are used
- Provider Default: Default provider is 'claude' which uses Anthropic API
Why This Wasn't Caught Earlier
- Developer testing uses API keys intentionally
- Claude Max subscription model is newer
- Auto-discovery is a feature for SDK users, not plugin users
- No telemetry on API key discovery
Recommended Solutions
Immediate Fixes (v9.0.1)
1. Add First-Run Warning
Display a prominent warning on first plugin activation:
[claude-mem] IMPORTANT: This plugin uses the Claude Agent SDK for AI operations.
If you have an ANTHROPIC_API_KEY configured, it will be used for:
- Observation compression
- Session summaries
This may incur separate API charges beyond your Claude Max subscription.
To avoid charges, configure an alternative provider in ~/.claude-mem/settings.json:
- Set CLAUDE_MEM_PROVIDER to "gemini" or "openrouter"
- Or ensure no ANTHROPIC_API_KEY is accessible to the plugin
Continue? [Y/n]
2. Detect and Warn About API Key
Add a check during worker initialization:
// Pseudo-code for worker-service.ts
const hasAnthropicKey = !!(
process.env.ANTHROPIC_API_KEY ||
existsSync(join(homedir(), '.anthropic', 'api_key'))
);
if (hasAnthropicKey && provider === 'claude') {
logger.warn('SYSTEM',
'ANTHROPIC_API_KEY detected. Plugin AI operations will consume API credits. ' +
'Configure CLAUDE_MEM_PROVIDER in settings.json to use a free alternative.'
);
}
3. Default to Free Provider
Change default provider from 'claude' to 'gemini' (free tier available):
File: src/shared/SettingsDefaultsManager.ts
// Line 66 - Change default
CLAUDE_MEM_PROVIDER: 'gemini', // Changed from 'claude' - free tier by default
Medium-Term Solutions (v9.1.0)
4. Opt-In API Key Usage
Require explicit configuration to use Anthropic API:
// ~/.claude-mem/settings.json
{
"CLAUDE_MEM_PROVIDER": "claude",
"CLAUDE_MEM_ANTHROPIC_API_KEY_CONSENT": true // New required field
}
5. Usage Estimation Before Processing
Show estimated token usage before processing:
[claude-mem] Processing 25 observations
Estimated API usage: ~37,500 tokens (~$0.15)
Provider: claude (ANTHROPIC_API_KEY)
6. Environment Isolation
Prevent automatic API key inheritance:
// In worker spawn
env: {
...process.env,
ANTHROPIC_API_KEY: undefined, // Explicitly unset unless opted-in
}
Long-Term Solutions (v10.0.0)
7. Built-In Usage Dashboard
Add API usage tracking to the viewer UI at http://localhost:37777:
- Total tokens consumed this session/day/month
- Estimated costs by provider
- Warning thresholds
8. Provider Configuration Wizard
First-run wizard in viewer UI:
- "Choose your AI provider for memory operations"
- Options: Free (Gemini), Pay-per-use (Claude/OpenRouter), Self-hosted
- Configure API keys through UI, not discovery
Priority/Severity Assessment
Severity: HIGH
Rationale:
- Direct financial impact on users
- Trust violation in plugin ecosystem
- No user consent for charges
- Difficult to discover source of charges
- Affects users who believed Claude Max covered all costs
Priority: P1 - Critical
Rationale:
- Active financial harm to users
- Reputation risk for plugin
- Simple fixes available
- User trust requires immediate action
Recommended Timeline
| Milestone | Target | Description |
|---|---|---|
| Hotfix | 48 hours | Add warning message, update docs |
| v9.0.1 | 1 week | Detection, warning, default provider change |
| v9.1.0 | 2 weeks | Opt-in flow, usage estimation |
| v10.0.0 | 1 month | Full usage dashboard, configuration wizard |
Files to Modify
| File | Change |
|---|---|
src/services/worker-service.ts |
Add API key detection and warning |
src/shared/SettingsDefaultsManager.ts |
Change default provider to 'gemini' |
plugin/scripts/context-hook.js |
Add first-run warning |
docs/public/installation.mdx |
Document API key usage clearly |
docs/public/configuration.mdx |
Add provider selection guidance |
CHANGELOG.md |
Document the change |
Testing Recommendations
Test Cases to Add
- API Key Detection Test: Verify warning appears when ANTHROPIC_API_KEY present
- Default Provider Test: Ensure new installs default to gemini
- Opt-In Test: Verify claude provider requires explicit consent
- Environment Isolation Test: Confirm API key not inherited without consent
Manual Testing
# Test 1: Clean environment (should default to gemini)
unset ANTHROPIC_API_KEY
claude # Start Claude Code with plugin
# Test 2: With API key (should show warning)
export ANTHROPIC_API_KEY="sk-test-key"
claude # Should display warning
# Test 3: Explicit opt-in
# Configure settings.json with consent flag
claude # Should use claude provider without warning
Conclusion
Issue #588 represents a critical UX and financial issue where the plugin's use of the Claude Agent SDK results in unexpected API charges for users who have an ANTHROPIC_API_KEY configured. The auto-discovery behavior, while useful for developers, creates a poor user experience for plugin users who expect their Claude Max subscription to cover all costs.
Immediate Action Required:
- Release hotfix with warning message
- Update documentation to clearly state API usage
- Change default provider to free tier (gemini)
- Implement opt-in consent for Anthropic API usage
The fix is straightforward, but the impact on user trust requires prompt action.
Appendix: Related Issues and Documentation
| Resource | Description |
|---|---|
| Claude Agent SDK Docs | SDK documentation |
docs/public/hooks-architecture.mdx |
Hooks and API key documentation |
docs/public/configuration.mdx |
Settings configuration reference |
| Issue #511 | Related: Gemini model support |
| Issue #527 | Related: Provider detection issues |
Appendix: User Communication Template
Suggested Announcement/Changelog Entry:
## Important Notice for v9.0.1
### API Key Usage Disclosure
Claude-mem uses AI for observation compression and session summaries.
If you have an `ANTHROPIC_API_KEY` configured (in ~/.anthropic/api_key,
environment variables, or project .env files), the plugin will use
Anthropic API credits for these operations.
**This is separate from your Claude Max subscription.**
### Changes in v9.0.1
- **Default provider changed to Gemini** (free tier available)
- **Warning displayed** when ANTHROPIC_API_KEY is detected
- **Opt-in required** to use Anthropic API for plugin operations
### For existing users
If you experienced unexpected charges:
1. Check your provider setting: `~/.claude-mem/settings.json`
2. Set `CLAUDE_MEM_PROVIDER` to `"gemini"` for free operation
3. Or remove/unset `ANTHROPIC_API_KEY` if not needed for other tools
We apologize for any confusion or unexpected charges caused by this behavior.