Files
claude-mem/docs/reports/2026-01-05--issue-545-formattool-json-parse-crash.md
T
Alex Newman f38b5b85bc fix: resolve issues #543, #544, #545, #557 (#558)
* docs: add investigation reports for 5 open GitHub issues

Comprehensive analysis of issues #543, #544, #545, #555, and #557:

- #557: settings.json not generated, module loader error (node/bun mismatch)
- #555: Windows hooks not executing, hasIpc always false
- #545: formatTool crashes on non-JSON tool_input strings
- #544: mem-search skill hint shown incorrectly to Claude Code users
- #543: /claude-mem slash command unavailable despite installation

Each report includes root cause analysis, affected files, and proposed fixes.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* fix(logger): handle non-JSON tool_input in formatTool (#545)

Wrap JSON.parse in try-catch to handle raw string inputs (e.g., Bash
commands) that aren't valid JSON. Falls back to using the string as-is.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* fix(context): update mem-search hint to reference MCP tools (#544)

Update hint messages to reference MCP tools (search, get_observations)
instead of the deprecated "mem-search skill" terminology.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* fix(settings): auto-create settings.json on first load (#557, #543)

When settings.json doesn't exist, create it with defaults instead of
returning in-memory defaults. Creates parent directory if needed.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* fix(hooks): use bun runtime for hooks except smart-install (#557)

Change hook commands from node to bun since hooks use bun:sqlite.
Keep smart-install.js on node since it bootstraps bun installation.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* chore: rebuild plugin scripts

* docs: clarify that build artifacts must be committed

* fix(docs): update build artifacts directory reference in CLAUDE.md

* test: add test coverage for PR #558 fixes

- Fix 2 failing tests: update "mem-search skill" → "MCP tools" expectations
- Add 56 tests for formatTool() JSON.parse crash fix (Issue #545)
- Add 27 tests for settings.json auto-creation (Issue #543)

Test coverage includes:
- formatTool: JSON parsing, raw strings, objects, null/undefined, all tool types
- Settings: file creation, directory creation, schema migration, edge cases

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* fix(tests): clean up flaky tests and fix circular dependency

Phase 1 of test quality improvements:

- Delete 6 harmful/worthless test files that used problematic mock.module()
  patterns or tested implementation details rather than behavior:
  - context-builder.test.ts (tested internal implementation)
  - export-types.test.ts (fragile mock patterns)
  - smart-install.test.ts (shell script testing antipattern)
  - session_id_refactor.test.ts (outdated, tested refactoring itself)
  - validate_sql_update.test.ts (one-time migration validation)
  - observation-broadcaster.test.ts (excessive mocking)

- Fix circular dependency between logger.ts and SettingsDefaultsManager.ts
  by using late binding pattern - logger now lazily loads settings

- Refactor mock.module() to spyOn() in several test files for more
  maintainable and less brittle tests:
  - observation-compiler.test.ts
  - gemini_agent.test.ts
  - error-handler.test.ts
  - server.test.ts
  - response-processor.test.ts

All 649 tests pass.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* refactor(tests): phase 2 - reduce mock-heavy tests and improve focus

- Remove mock-heavy query tests from observation-compiler.test.ts, keep real buildTimeline tests
- Convert session_id_usage_validation.test.ts from 477 to 178 lines of focused smoke tests
- Remove tests for language built-ins from worker-spawn.test.ts (JSON.parse, array indexing)
- Rename logger-coverage.test.ts to logger-usage-standards.test.ts for clarity

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* docs(tests): phase 3 - add JSDoc mock justification to test files

Document mock usage rationale in 5 test files to improve maintainability:
- error-handler.test.ts: Express req/res mocks, logger spies (~11%)
- fallback-error-handler.test.ts: Zero mocks, pure function tests
- session-cleanup-helper.test.ts: Session fixtures, worker mocks (~19%)
- hook-constants.test.ts: process.platform mock for Windows tests (~12%)
- session_store.test.ts: Zero mocks, real SQLite :memory: database

Part of ongoing effort to document mock justifications per TESTING.md guidelines.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* test(integration): phase 5 - add 72 tests for critical coverage gaps

Add comprehensive test coverage for previously untested areas:

- tests/integration/hook-execution-e2e.test.ts (10 tests)
  Tests lifecycle hooks execution flow and context propagation

- tests/integration/worker-api-endpoints.test.ts (19 tests)
  Tests all worker service HTTP endpoints without heavy mocking

- tests/integration/chroma-vector-sync.test.ts (16 tests)
  Tests vector embedding synchronization with ChromaDB

- tests/utils/tag-stripping.test.ts (27 tests)
  Tests privacy tag stripping utilities for both <private> and
  <meta-observation> tags

All tests use real implementations where feasible, following the
project's testing philosophy of preferring integration-style tests
over unit tests with extensive mocking.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* context update

* docs: add comment linking DEFAULT_DATA_DIR locations

Added NOTE comment in logger.ts pointing to the canonical DEFAULT_DATA_DIR
in SettingsDefaultsManager.ts. This addresses PR reviewer feedback about
the fragility of having the default defined in two places to avoid
circular dependencies.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
2026-01-05 19:45:09 -05:00

6.9 KiB

Issue #545: formatTool Crashes on Non-JSON Tool Input Strings

Summary

Issue: formatTool method in src/utils/logger.ts crashes when toolInput is a string that is not valid JSON Type: Bug (Critical - Silent Data Loss) Status: Open Author: @Rob-van-B Created: January 4, 2026

The formatTool method unconditionally calls JSON.parse() on string inputs without error handling. When tool inputs are raw strings (not JSON), this throws an exception that propagates up the call stack, causing 400 errors for valid observation requests and silently stopping claude-mem from recording tool usage.

Root Cause Analysis

Verified Issue Location

File: /Users/alexnewman/Scripts/claude-mem/src/utils/logger.ts Line: 139 Method: formatTool

formatTool(toolName: string, toolInput?: any): string {
  if (!toolInput) return toolName;

  const input = typeof toolInput === 'string' ? JSON.parse(toolInput) : toolInput;
  // ... rest of method
}

The Problem

The code assumes that if toolInput is a string, it must be valid JSON. This assumption is incorrect. Tool inputs can be:

  1. Already-parsed objects (no parsing needed)
  2. JSON strings (need parsing)
  3. Raw strings that are not JSON (will crash on parse)

When a raw string is passed (e.g., a Bash command like ls -la), JSON.parse("ls -la") throws:

SyntaxError: Unexpected token 'l', "ls -la" is not valid JSON

Existing Correct Pattern in Codebase

The issue is notable because the correct pattern already exists in src/sdk/prompts.ts (lines 96-102):

try {
  toolInput = typeof obs.tool_input === 'string' ? JSON.parse(obs.tool_input) : obs.tool_input;
} catch (error) {
  logger.debug('SDK', 'Tool input is plain string, using as-is', {
    toolName: obs.tool_name
  }, error as Error);
  toolInput = obs.tool_input;
}

This demonstrates the correct defensive approach was implemented elsewhere but missed in logger.ts.

Call Sites Affected

The formatTool method is called from 4 locations:

File Line Context Impact
src/hooks/save-hook.ts 38 PostToolUse hook logging Hook crashes, observation lost
src/services/worker/http/middleware.ts 110 HTTP request logging 400 error returned to client
src/services/worker/SessionManager.ts 220 Observation queue logging Observation not queued

All call sites pass tool_input directly from Claude Code's PostToolUse hook, which can be any type including raw strings.

Impact Assessment

Severity: High

  1. Silent Data Loss: Observations fail to save without user notification
  2. No Error Visibility: Worker runs as background process - errors go unnoticed
  3. Intermittent Failures: Only affects certain tool types with string inputs
  4. Cascading Effect: One failed observation can disrupt session tracking

Affected Tool Types

Tools most likely to trigger this bug:

  • Bash: Command strings like git status, npm install
  • Grep: Search patterns
  • Glob: File patterns like **/*.ts
  • Custom MCP tools: May pass raw strings

Data Flow Path

Claude Code
    |
    v
PostToolUse Hook (save-hook.ts:38)
    |-- logger.formatTool() <-- CRASH HERE
    |
    v [if crash, never reaches]
Worker HTTP Endpoint
    |-- middleware.ts:110 logger.formatTool() <-- CRASH HERE TOO
    |
    v [if crash, 400 returned]
SessionManager
    |-- SessionManager.ts:220 logger.formatTool() <-- CRASH HERE TOO
    |
    v [if crash, not queued]
Database

Option 1: User's Proposed Fix (Minimal)

let input = toolInput;
if (typeof toolInput === 'string') {
  try {
    input = JSON.parse(toolInput);
  } catch {
    input = { raw: toolInput };
  }
}

Pros: Simple, encapsulates raw strings in an object Cons: Changes the shape of input for raw strings (may affect downstream logic)

formatTool(toolName: string, toolInput?: any): string {
  if (!toolInput) return toolName;

  let input = toolInput;
  if (typeof toolInput === 'string') {
    try {
      input = JSON.parse(toolInput);
    } catch {
      // Input is a raw string, not JSON - use as-is
      input = toolInput;
    }
  }

  // Bash: show full command
  if (toolName === 'Bash' && input.command) {
    return `${toolName}(${input.command})`;
  }

  // Handle raw string inputs (e.g., from Bash commands passed as strings)
  if (typeof input === 'string') {
    return `${toolName}(${input.length > 50 ? input.slice(0, 50) + '...' : input})`;
  }

  // ... rest of existing logic
}

Pros: Consistent with existing pattern, handles raw strings gracefully Cons: Requires additional check for string display formatting

Option 3: Extract Shared Utility (Best Long-term)

Create a shared utility in src/shared/json-utils.ts:

/**
 * Safely parse JSON that might be a raw string
 * Returns the parsed object if valid JSON, otherwise the original value
 */
export function safeJsonParse<T>(value: T): T | object {
  if (typeof value !== 'string') return value;
  try {
    return JSON.parse(value);
  } catch {
    return value;
  }
}

Then use in both logger.ts and prompts.ts for consistency.

Similar Patterns to Review

Other JSON.parse calls that may need similar protection:

File Line Current Protection
src/sdk/prompts.ts 97, 106 Has try-catch
src/services/sqlite/PendingMessageStore.ts 373-374 No try-catch (lower risk - DB data should be valid)
src/utils/logger.ts 139 No try-catch (BUG)

Testing Considerations

Unit Tests Needed

  1. formatTool with valid JSON string input
  2. formatTool with object input (already parsed)
  3. formatTool with raw string input (the bug case)
  4. formatTool with null/undefined input
  5. formatTool with empty string input

Integration Tests Needed

  1. PostToolUse hook with Bash command string
  2. Observation storage with raw string tool input
  3. Full pipeline from hook through worker to database

Test Cases

// Should handle raw string input without crashing
expect(logger.formatTool('Bash', 'ls -la')).toBe('Bash(ls -la)');

// Should handle JSON string input
expect(logger.formatTool('Read', '{"file_path": "/foo"}'))
  .toBe('Read(/foo)');

// Should handle object input
expect(logger.formatTool('Read', { file_path: '/foo' }))
  .toBe('Read(/foo)');

// Should handle empty/null input
expect(logger.formatTool('Bash')).toBe('Bash');
expect(logger.formatTool('Bash', null)).toBe('Bash');

Complexity

Low - 30 minutes to 1 hour

  • Single file change (src/utils/logger.ts)
  • Clear fix pattern exists in codebase
  • No breaking API changes
  • Unit tests straightforward

References

  • GitHub Issue: #545
  • Related file with correct pattern: src/sdk/prompts.ts (lines 96-102)
  • Logger source: src/utils/logger.ts (lines 136-197)