diff --git a/src/services/mcp/client.ts b/src/services/mcp/client.ts
index 7cdc1c17..06e911e8 100644
--- a/src/services/mcp/client.ts
+++ b/src/services/mcp/client.ts
@@ -55,6 +55,7 @@ import { type MCPProgress, MCPTool } from '../../tools/MCPTool/MCPTool.js'
 import { createMcpAuthTool } from '../../tools/McpAuthTool/McpAuthTool.js'
 import { ReadMcpResourceTool } from '../../tools/ReadMcpResourceTool/ReadMcpResourceTool.js'
 import { createAbortController } from '../../utils/abortController.js'
+import { AbortError, isAbortError } from '../../utils/errors.js'
 import { count } from '../../utils/array.js'
 import {
   checkAndRefreshOAuthTokenIfNeeded,
@@ -3283,11 +3284,18 @@ async function callMCPTool({
       }
     }
 
-    // When the users hits esc, avoid logspew
-    if (!(e instanceof Error) || e.name !== 'AbortError') {
-      throw e
+    // When the user hits esc, convert to our AbortError class so the tool
+    // execution framework handles it properly (skips logging, creates
+    // is_error: true result with [Request interrupted by user for tool use]).
+    // Previously this returned { content: undefined }, which masked the
+    // cancellation and caused mapToolResultToToolResultBlockParam to send
+    // empty/undefined content to the API as if it were a successful result.
+    if (isAbortError(e)) {
+      throw new AbortError(
+        e instanceof Error ? e.message : 'Tool execution cancelled',
+      )
     }
-    return { content: undefined }
+    throw e
   } finally {
     // Always clear intervals
     if (progressInterval !== undefined) {
diff --git a/src/tools/BashTool/commandSemantics.test.ts b/src/tools/BashTool/commandSemantics.test.ts
new file mode 100644
index 00000000..68a5331d
--- /dev/null
+++ b/src/tools/BashTool/commandSemantics.test.ts
@@ -0,0 +1,163 @@
+import { describe, expect, test } from 'bun:test'
+import { interpretCommandResult } from './commandSemantics.js'
+
+// =============================================================================
+// interpretCommandResult — exit code semantics per command
+// =============================================================================
+
+describe('interpretCommandResult', () => {
+  // --- Default semantics (most commands) ---
+  describe('default semantics', () => {
+    test('exit code 0 = success, no error', () => {
+      const result = interpretCommandResult('python script.py', 0, '', '')
+      expect(result.isError).toBe(false)
+      expect(result.message).toBeUndefined()
+    })
+
+    test('exit code 1 = error', () => {
+      const result = interpretCommandResult('python script.py', 1, '', '')
+      expect(result.isError).toBe(true)
+      expect(result.message).toContain('exit code 1')
+    })
+
+    test('exit code 127 = command not found', () => {
+      const result = interpretCommandResult('foobar', 127, '', '')
+      expect(result.isError).toBe(true)
+      expect(result.message).toContain('127')
+    })
+
+    test('exit code 126 = permission denied', () => {
+      const result = interpretCommandResult('./script.sh', 126, '', '')
+      expect(result.isError).toBe(true)
+      expect(result.message).toContain('126')
+    })
+
+    test('exit code 130 = SIGINT (but not treated as interrupted here)', () => {
+      const result = interpretCommandResult('long-command', 130, '', '')
+      expect(result.isError).toBe(true)
+    })
+  })
+
+  // --- grep: 0=matches, 1=no matches, 2+=error ---
+  describe('grep', () => {
+    test('exit code 0 = matches found (not error)', () => {
+      const result = interpretCommandResult('grep foo file.txt', 0, 'foo\n', '')
+      expect(result.isError).toBe(false)
+    })
+
+    test('exit code 1 = no matches (not error)', () => {
+      const result = interpretCommandResult('grep foo file.txt', 1, '', '')
+      expect(result.isError).toBe(false)
+      expect(result.message).toContain('No matches found')
+    })
+
+    test('exit code 2 = real error', () => {
+      const result = interpretCommandResult('grep foo file.txt', 2, '', 'No such file')
+      expect(result.isError).toBe(true)
+    })
+  })
+
+  // --- ripgrep: same as grep ---
+  describe('rg', () => {
+    test('exit code 1 = no matches (not error)', () => {
+      const result = interpretCommandResult('rg pattern', 1, '', '')
+      expect(result.isError).toBe(false)
+    })
+
+    test('exit code 2 = error', () => {
+      const result = interpretCommandResult('rg pattern', 2, '', '')
+      expect(result.isError).toBe(true)
+    })
+  })
+
+  // --- find: 0=success, 1=partial, 2+=error ---
+  describe('find', () => {
+    test('exit code 0 = success', () => {
+      const result = interpretCommandResult('find . -name "*.ts"', 0, 'file.ts\n', '')
+      expect(result.isError).toBe(false)
+    })
+
+    test('exit code 1 = partial success (not error)', () => {
+      const result = interpretCommandResult('find . -name "*.ts"', 1, 'file.ts\n', '')
+      expect(result.isError).toBe(false)
+      expect(result.message).toContain('inaccessible')
+    })
+
+    test('exit code 2 = error', () => {
+      const result = interpretCommandResult('find . -name "*.ts"', 2, '', 'Permission denied')
+      expect(result.isError).toBe(true)
+    })
+  })
+
+  // --- diff: 0=same, 1=different, 2+=error ---
+  describe('diff', () => {
+    test('exit code 0 = files identical', () => {
+      const result = interpretCommandResult('diff a.txt b.txt', 0, '', '')
+      expect(result.isError).toBe(false)
+    })
+
+    test('exit code 1 = files differ (not error)', () => {
+      const result = interpretCommandResult('diff a.txt b.txt', 1, '< line1\n> line2', '')
+      expect(result.isError).toBe(false)
+      expect(result.message).toContain('differ')
+    })
+
+    test('exit code 2 = error', () => {
+      const result = interpretCommandResult('diff a.txt b.txt', 2, '', 'No such file')
+      expect(result.isError).toBe(true)
+    })
+  })
+
+  // --- test/[: 0=true, 1=false, 2+=error ---
+  describe('test and [', () => {
+    test('test exit code 0 = condition true', () => {
+      const result = interpretCommandResult('test -f file.txt', 0, '', '')
+      expect(result.isError).toBe(false)
+    })
+
+    test('test exit code 1 = condition false (not error)', () => {
+      const result = interpretCommandResult('test -f file.txt', 1, '', '')
+      expect(result.isError).toBe(false)
+      expect(result.message).toContain('false')
+    })
+
+    test('[ exit code 1 = condition false (not error)', () => {
+      const result = interpretCommandResult('[ -f file.txt ]', 1, '', '')
+      expect(result.isError).toBe(false)
+    })
+  })
+
+  // --- Compound commands ---
+  describe('compound commands', () => {
+    test('last command determines semantics: grep last', () => {
+      const result = interpretCommandResult('cd /tmp && grep foo file.txt', 1, '', '')
+      // grep exit code 1 = no matches, not error
+      expect(result.isError).toBe(false)
+    })
+
+    test('last command determines semantics: python last', () => {
+      const result = interpretCommandResult('cd /tmp && python script.py', 1, '', '')
+      // python exit code 1 = error
+      expect(result.isError).toBe(true)
+    })
+  })
+
+  // --- systemctl, apt, docker (real-world commands) ---
+  describe('system/service commands', () => {
+    test('systemctl failure = error', () => {
+      const result = interpretCommandResult('systemctl start nginx', 1, '', 'Job for nginx.service failed')
+      expect(result.isError).toBe(true)
+      expect(result.message).toContain('exit code 1')
+    })
+
+    test('apt failure = error', () => {
+      const result = interpretCommandResult('apt install foo', 100, '', 'Unable to locate package')
+      expect(result.isError).toBe(true)
+    })
+
+    test('docker failure = error', () => {
+      const result = interpretCommandResult('docker run ubuntu', 1, '', 'Unable to find image')
+      expect(result.isError).toBe(true)
+    })
+  })
+})
diff --git a/src/tools/BashTool/utils.test.ts b/src/tools/BashTool/utils.test.ts
new file mode 100644
index 00000000..89272282
--- /dev/null
+++ b/src/tools/BashTool/utils.test.ts
@@ -0,0 +1,212 @@
+import { describe, expect, test } from 'bun:test'
+import {
+  stripEmptyLines,
+  isImageOutput,
+  parseDataUri,
+  formatOutput,
+  createContentSummary,
+} from './utils.js'
+
+// =============================================================================
+// stripEmptyLines — removes leading/trailing blank lines
+// =============================================================================
+
+describe('stripEmptyLines', () => {
+  test('strips leading empty lines', () => {
+    expect(stripEmptyLines('\n\n\nhello')).toBe('hello')
+  })
+
+  test('strips trailing empty lines', () => {
+    expect(stripEmptyLines('hello\n\n\n')).toBe('hello')
+  })
+
+  test('strips both ends', () => {
+    expect(stripEmptyLines('\n\nhello\n\n')).toBe('hello')
+  })
+
+  test('preserves internal empty lines', () => {
+    expect(stripEmptyLines('a\n\nb')).toBe('a\n\nb')
+  })
+
+  test('all empty lines returns empty string', () => {
+    expect(stripEmptyLines('\n\n\n')).toBe('')
+  })
+
+  test('empty string returns empty string', () => {
+    expect(stripEmptyLines('')).toBe('')
+  })
+
+  test('preserves whitespace-only lines in the middle', () => {
+    expect(stripEmptyLines('a\n   \nb')).toBe('a\n   \nb')
+  })
+
+  test('single line no change', () => {
+    expect(stripEmptyLines('hello')).toBe('hello')
+  })
+})
+
+// =============================================================================
+// isImageOutput — detects base64 data URIs
+// =============================================================================
+
+describe('isImageOutput', () => {
+  test('detects PNG data URI', () => {
+    expect(isImageOutput('data:image/png;base64,iVBORw0KGgo=')).toBe(true)
+  })
+
+  test('detects JPEG data URI', () => {
+    expect(isImageOutput('data:image/jpeg;base64,/9j/4AAQ')).toBe(true)
+  })
+
+  test('detects GIF data URI', () => {
+    expect(isImageOutput('data:image/gif;base64,R0lGODlhAQABAIAAAP')).toBe(true)
+  })
+
+  test('detects SVG data URI', () => {
+    expect(isImageOutput('data:image/svg+xml;base64,PHN2Zz4=')).toBe(true)
+  })
+
+  test('rejects plain text', () => {
+    expect(isImageOutput('hello world')).toBe(false)
+  })
+
+  test('rejects empty string', () => {
+    expect(isImageOutput('')).toBe(false)
+  })
+
+  test('rejects non-image data URI', () => {
+    expect(isImageOutput('data:text/plain;base64,aGVsbG8=')).toBe(false)
+  })
+
+  test('rejects partial data URI', () => {
+    expect(isImageOutput('data:image/png,')).toBe(false)
+  })
+})
+
+// =============================================================================
+// parseDataUri — extracts media type and base64 payload
+// =============================================================================
+
+describe('parseDataUri', () => {
+  test('parses valid PNG data URI', () => {
+    const result = parseDataUri('data:image/png;base64,iVBORw0KGgo=')
+    expect(result).toEqual({
+      mediaType: 'image/png',
+      data: 'iVBORw0KGgo=',
+    })
+  })
+
+  test('parses valid JPEG data URI', () => {
+    const result = parseDataUri('data:image/jpeg;base64,/9j/4AAQ')
+    expect(result).toEqual({
+      mediaType: 'image/jpeg',
+      data: '/9j/4AAQ',
+    })
+  })
+
+  test('handles whitespace around URI', () => {
+    const result = parseDataUri('  data:image/png;base64,abc123  ')
+    expect(result).toEqual({
+      mediaType: 'image/png',
+      data: 'abc123',
+    })
+  })
+
+  test('returns null for non-data URI', () => {
+    expect(parseDataUri('https://example.com/image.png')).toBeNull()
+  })
+
+  test('returns null for empty string', () => {
+    expect(parseDataUri('')).toBeNull()
+  })
+
+  test('returns null for incomplete data URI', () => {
+    expect(parseDataUri('data:image/png')).toBeNull()
+  })
+
+  test('returns null for non-base64 data URI', () => {
+    expect(parseDataUri('data:text/plain,hello')).toBeNull()
+  })
+})
+
+// =============================================================================
+// formatOutput — truncates long output with line count
+// =============================================================================
+
+describe('formatOutput', () => {
+  test('short output passes through unchanged', () => {
+    const result = formatOutput('line1\nline2\nline3')
+    expect(result.truncatedContent).toBe('line1\nline2\nline3')
+    expect(result.totalLines).toBe(3)
+    expect(result.isImage).toBe(false)
+  })
+
+  test('empty output', () => {
+    const result = formatOutput('')
+    expect(result.truncatedContent).toBe('')
+    expect(result.totalLines).toBe(1)
+  })
+
+  test('image output is passed through', () => {
+    const img = 'data:image/png;base64,iVBORw0KGgo='
+    const result = formatOutput(img)
+    expect(result.truncatedContent).toBe(img)
+    expect(result.totalLines).toBe(1)
+    expect(result.isImage).toBe(true)
+  })
+
+  test('single line no trailing newline', () => {
+    const result = formatOutput('hello')
+    expect(result.totalLines).toBe(1)
+  })
+})
+
+// =============================================================================
+// createContentSummary — MCP content block summaries
+// =============================================================================
+
+describe('createContentSummary', () => {
+  test('summarizes text blocks', () => {
+    const content = [
+      { type: 'text' as const, text: 'Hello world' },
+    ]
+    const result = createContentSummary(content)
+    expect(result).toContain('MCP Result')
+    expect(result).toContain('1 text block')
+    expect(result).toContain('Hello world')
+  })
+
+  test('summarizes image blocks', () => {
+    const content = [
+      { type: 'image' as const, data: 'base64data', mimeType: 'image/png' },
+    ]
+    const result = createContentSummary(content)
+    expect(result).toContain('1 image')
+  })
+
+  test('summarizes mixed content', () => {
+    const content = [
+      { type: 'text' as const, text: 'Description' },
+      { type: 'image' as const, data: 'base64data', mimeType: 'image/png' },
+      { type: 'text' as const, text: 'More text' },
+    ]
+    const result = createContentSummary(content)
+    expect(result).toContain('1 image')
+    expect(result).toContain('2 text blocks')
+  })
+
+  test('truncates long text preview at 200 chars', () => {
+    const longText = 'x'.repeat(300)
+    const content = [
+      { type: 'text' as const, text: longText },
+    ]
+    const result = createContentSummary(content)
+    expect(result).toContain('...')
+    expect(result).toContain('x'.repeat(200))
+  })
+
+  test('empty content array', () => {
+    const result = createContentSummary([])
+    expect(result).toContain('MCP Result')
+  })
+})
diff --git a/src/tools/MCPTool/MCPTool.test.ts b/src/tools/MCPTool/MCPTool.test.ts
new file mode 100644
index 00000000..9becdf92
--- /dev/null
+++ b/src/tools/MCPTool/MCPTool.test.ts
@@ -0,0 +1,132 @@
+import { describe, expect, test, beforeEach } from 'bun:test'
+import { MCPTool } from './MCPTool.js'
+
+// =============================================================================
+// MCPTool.validateInput — AJV schema validation
+// =============================================================================
+
+describe('MCPTool.validateInput', () => {
+  test('passes when no inputJSONSchema is set', async () => {
+    const tool = { ...MCPTool, inputJSONSchema: undefined }
+    const result = await tool.validateInput({ anything: 'goes' }, {} as never)
+    expect(result.result).toBe(true)
+  })
+
+  test('validates against inputJSONSchema when set', async () => {
+    const schema = {
+      type: 'object' as const,
+      properties: {
+        name: { type: 'string' },
+      },
+      required: ['name'],
+      additionalProperties: false,
+    }
+    const tool = { ...MCPTool, inputJSONSchema: schema }
+
+    // Valid input
+    const valid = await tool.validateInput({ name: 'test' }, {} as never)
+    expect(valid.result).toBe(true)
+
+    // Missing required field
+    const invalid = await tool.validateInput({}, {} as never)
+    expect(invalid.result).toBe(false)
+    expect(invalid.result === false && invalid.message).toContain('name')
+  })
+
+  test('rejects extra properties when additionalProperties is false', async () => {
+    const schema = {
+      type: 'object' as const,
+      properties: {
+        x: { type: 'number' },
+      },
+      additionalProperties: false,
+    }
+    const tool = { ...MCPTool, inputJSONSchema: schema }
+
+    const result = await tool.validateInput({ x: 1, extra: 'bad' }, {} as never)
+    expect(result.result).toBe(false)
+  })
+
+  test('handles invalid schema gracefully', async () => {
+    // Schema that will cause ajv.compile to throw
+    const schema = { type: 'invalid_type' } as any
+    const tool = { ...MCPTool, inputJSONSchema: schema }
+
+    const result = await tool.validateInput({}, {} as never)
+    expect(result.result).toBe(false)
+    expect(result.result === false && result.errorCode).toBe(500)
+    expect(result.result === false && result.message).toContain('Failed to compile')
+  })
+
+  test('error message is readable (not [object Object])', async () => {
+    const schema = { type: 'invalid_type' } as any
+    const tool = { ...MCPTool, inputJSONSchema: schema }
+
+    const result = await tool.validateInput({}, {} as never)
+    expect(result.result).toBe(false)
+    // Should NOT contain [object Object]
+    expect(result.result === false && result.message).not.toContain('[object Object]')
+  })
+})
+
+// =============================================================================
+// MCPTool.mapToolResultToToolResultBlockParam — null safety
+// =============================================================================
+
+describe('MCPTool.mapToolResultToToolResultBlockParam', () => {
+  test('handles string content', () => {
+    const result = MCPTool.mapToolResultToToolResultBlockParam('hello', 'tool-1')
+    expect(result.content).toBe('hello')
+    expect(result.tool_use_id).toBe('tool-1')
+    expect(result.type).toBe('tool_result')
+  })
+
+  test('handles array content', () => {
+    const blocks = [{ type: 'text', text: 'hello' }]
+    const result = MCPTool.mapToolResultToToolResultBlockParam(blocks as any, 'tool-2')
+    expect(result.content).toEqual(blocks)
+  })
+
+  test('handles undefined content gracefully', () => {
+    const result = MCPTool.mapToolResultToToolResultBlockParam(undefined as any, 'tool-3')
+    expect(result.content).toBe('[No content returned from MCP tool]')
+    expect(result.tool_use_id).toBe('tool-3')
+  })
+
+  test('handles null content gracefully', () => {
+    const result = MCPTool.mapToolResultToToolResultBlockParam(null as any, 'tool-4')
+    expect(result.content).toBe('[No content returned from MCP tool]')
+    expect(result.tool_use_id).toBe('tool-4')
+  })
+})
+
+// =============================================================================
+// MCPTool.isResultTruncated
+// =============================================================================
+
+describe('MCPTool.isResultTruncated', () => {
+  test('returns false for short string', () => {
+    expect(MCPTool.isResultTruncated('short')).toBe(false)
+  })
+
+  test('returns false for empty array', () => {
+    expect(MCPTool.isResultTruncated([])).toBe(false)
+  })
+
+  test('returns false for array with short text blocks', () => {
+    expect(MCPTool.isResultTruncated([{ type: 'text', text: 'short' }])).toBe(false)
+  })
+
+  test('handles null blocks in array', () => {
+    expect(MCPTool.isResultTruncated([null as any, { type: 'text', text: 'ok' }])).toBe(false)
+  })
+
+  test('handles undefined blocks in array', () => {
+    expect(MCPTool.isResultTruncated([undefined as any])).toBe(false)
+  })
+
+  test('returns false for non-string non-array', () => {
+    expect(MCPTool.isResultTruncated(42 as any)).toBe(false)
+    expect(MCPTool.isResultTruncated({} as any)).toBe(false)
+  })
+})
diff --git a/src/tools/MCPTool/MCPTool.ts b/src/tools/MCPTool/MCPTool.ts
index d7aa68a1..c2d3dad8 100644
--- a/src/tools/MCPTool/MCPTool.ts
+++ b/src/tools/MCPTool/MCPTool.ts
@@ -40,6 +40,21 @@ export type { MCPProgress } from '../../types/tools.js'
 
 const ajv = new Ajv({ strict: false })
 
+// Cache compiled validators to avoid recompiling on every validateInput call.
+// AJV compilation is expensive — schemas don't change between calls.
+// Uses WeakMap to allow garbage collection of schemas from disconnected/refreshed
+// MCP tools, preventing memory leaks from accumulating strong references.
+const compiledValidatorCache = new WeakMap<object, ReturnType<typeof ajv.compile>>()
+
+function getCompiledValidator(schema: object) {
+  let validator = compiledValidatorCache.get(schema)
+  if (!validator) {
+    validator = ajv.compile(schema)
+    compiledValidatorCache.set(schema, validator)
+  }
+  return validator
+}
+
 export const MCPTool = buildTool({
   isMcp: true,
   // Overridden in mcpClient.ts with the real MCP tool name + args
@@ -78,7 +93,7 @@ export const MCPTool = buildTool({
   async validateInput(input, context): Promise<ValidationResult> {
     if (this.inputJSONSchema) {
       try {
-        const validate = ajv.compile(this.inputJSONSchema)
+        const validate = getCompiledValidator(this.inputJSONSchema)
         if (!validate(input)) {
           return {
             result: false,
@@ -87,9 +102,10 @@ export const MCPTool = buildTool({
           }
         }
       } catch (error) {
+        const errMsg = error instanceof Error ? error.message : String(error)
         return {
           result: false,
-          message: `Failed to compile JSON schema for validation: ${error}`,
+          message: `Failed to compile JSON schema for validation: ${errMsg}`,
           errorCode: 500,
         }
       }
@@ -109,7 +125,8 @@ export const MCPTool = buildTool({
     if (Array.isArray(output)) {
       return output.some(
         block =>
-          block?.type === 'text' &&
+          block != null &&
+          block.type === 'text' &&
           typeof block.text === 'string' &&
           isOutputLineTruncated(block.text),
       )
@@ -117,6 +134,16 @@ export const MCPTool = buildTool({
     return false
   },
   mapToolResultToToolResultBlockParam(content, toolUseID) {
+    // Defensive guard: if content is undefined/null (shouldn't happen after
+    // the abort path fix in client.ts), return a clear indicator rather than
+    // sending undefined to the API which would cause an error.
+    if (content === undefined || content === null) {
+      return {
+        tool_use_id: toolUseID,
+        type: 'tool_result',
+        content: '[No content returned from MCP tool]',
+      }
+    }
     return {
       tool_use_id: toolUseID,
       type: 'tool_result',
diff --git a/src/tools/ReadMcpResourceTool/ReadMcpResourceTool.ts b/src/tools/ReadMcpResourceTool/ReadMcpResourceTool.ts
index 593131e7..226b5e56 100644
--- a/src/tools/ReadMcpResourceTool/ReadMcpResourceTool.ts
+++ b/src/tools/ReadMcpResourceTool/ReadMcpResourceTool.ts
@@ -149,6 +149,15 @@ export const ReadMcpResourceTool = buildTool({
     return isOutputLineTruncated(jsonStringify(output))
   },
   mapToolResultToToolResultBlockParam(content, toolUseID) {
+    // Defensive guard: if content is undefined/null, return a clear indicator
+    // rather than sending undefined to jsonStringify which would cause an error.
+    if (content === undefined || content === null) {
+      return {
+        tool_use_id: toolUseID,
+        type: 'tool_result',
+        content: '[No content returned from MCP resource]',
+      }
+    }
     return {
       tool_use_id: toolUseID,
       type: 'tool_result',
diff --git a/src/utils/toolErrors.test.ts b/src/utils/toolErrors.test.ts
new file mode 100644
index 00000000..6baff01c
--- /dev/null
+++ b/src/utils/toolErrors.test.ts
@@ -0,0 +1,113 @@
+import { describe, expect, test } from 'bun:test'
+import { ShellError, AbortError } from './errors.js'
+import { formatError, getErrorParts } from './toolErrors.js'
+
+// =============================================================================
+// getErrorParts — what the model sees when a tool fails
+// =============================================================================
+
+describe('getErrorParts', () => {
+  test('ShellError: exit code + stderr + stdout', () => {
+    const err = new ShellError('output here', 'error here', 1, false)
+    const parts = getErrorParts(err)
+    expect(parts).toEqual([
+      'Exit code 1',
+      '',
+      'error here',
+      'output here',
+    ])
+  })
+
+  test('ShellError: interrupted flag adds interrupt message', () => {
+    const err = new ShellError('', 'partial output', 130, true)
+    const parts = getErrorParts(err)
+    expect(parts[0]).toBe('Exit code 130')
+    expect(parts[1]).toContain('interrupted')
+    expect(parts[2]).toBe('partial output')
+  })
+
+  test('ShellError: empty stderr and stdout', () => {
+    const err = new ShellError('', '', 1, false)
+    const parts = getErrorParts(err)
+    expect(parts[0]).toBe('Exit code 1')
+    expect(parts[2]).toBe('')
+    expect(parts[3]).toBe('')
+  })
+
+  test('non-ShellError: returns message + stderr + stdout if present', () => {
+    const err = new Error('something broke')
+    ;(err as any).stderr = 'stderr data'
+    ;(err as any).stdout = 'stdout data'
+    const parts = getErrorParts(err)
+    expect(parts[0]).toBe('something broke')
+    expect(parts[1]).toBe('stderr data')
+    expect(parts[2]).toBe('stdout data')
+  })
+
+  test('non-ShellError: message only when no stderr/stdout', () => {
+    const err = new Error('just a message')
+    const parts = getErrorParts(err)
+    expect(parts).toEqual(['just a message'])
+  })
+})
+
+// =============================================================================
+// formatError — final string sent to the model
+// =============================================================================
+
+describe('formatError', () => {
+  test('AbortError returns message or interrupt default', () => {
+    const err = new AbortError('user cancelled')
+    expect(formatError(err)).toBe('user cancelled')
+  })
+
+  test('AbortError with empty message returns default interrupt', () => {
+    const err = new AbortError('')
+    const result = formatError(err)
+    expect(result.length).toBeGreaterThan(0)
+  })
+
+  test('non-Error value stringified', () => {
+    expect(formatError('raw string')).toBe('raw string')
+    expect(formatError(42)).toBe('42')
+  })
+
+  test('ShellError: combines exit code + stderr + stdout', () => {
+    const err = new ShellError('stdout content', 'stderr content', 1, false)
+    const result = formatError(err)
+    expect(result).toContain('Exit code 1')
+    expect(result).toContain('stderr content')
+    expect(result).toContain('stdout content')
+  })
+
+  test('ShellError: empty output falls back to default message', () => {
+    const err = new ShellError('', '', 1, false)
+    const result = formatError(err)
+    expect(result).toBe('Exit code 1')
+  })
+
+  test('non-ShellError: message only', () => {
+    const err = new Error('something failed')
+    expect(formatError(err)).toBe('something failed')
+  })
+
+  test('truncates at 40KB (not 10KB)', () => {
+    // 50KB of output — should be truncated at 40KB limit
+    const longOutput = 'x'.repeat(50_000)
+    const err = new ShellError('', longOutput, 1, false)
+    const result = formatError(err)
+    expect(result.length).toBeLessThan(50_000)
+    expect(result).toContain('truncated')
+    // Should keep first 20KB + last 20KB
+    expect(result).toContain('x'.repeat(100))
+  })
+
+  test('does NOT truncate under 40KB', () => {
+    // 30KB of output — should NOT be truncated
+    const output = 'y'.repeat(30_000)
+    const err = new ShellError('', output, 1, false)
+    const result = formatError(err)
+    expect(result).not.toContain('truncated')
+    expect(result).toContain(output)
+  })
+})
diff --git a/src/utils/toolErrors.ts b/src/utils/toolErrors.ts
index 7520e513..6fe09612 100644
--- a/src/utils/toolErrors.ts
+++ b/src/utils/toolErrors.ts
@@ -12,13 +12,15 @@ export function formatError(error: unknown): string {
   const parts = getErrorParts(error)
   const fullMessage =
     parts.filter(Boolean).join('\n').trim() || 'Command failed with no output'
-  if (fullMessage.length <= 10000) {
+  // 40KB limit — enough for most command error logs (systemctl, apt, python, etc.)
+  const maxErrorLength = 40000
+  if (fullMessage.length <= maxErrorLength) {
     return fullMessage
   }
-  const halfLength = 5000
+  const halfLength = Math.floor(maxErrorLength / 2)
   const start = fullMessage.slice(0, halfLength)
   const end = fullMessage.slice(-halfLength)
-  return `${start}\n\n... [${fullMessage.length - 10000} characters truncated] ...\n\n${end}`
+  return `${start}\n\n... [${fullMessage.length - maxErrorLength} characters truncated] ...\n\n${end}`
 }
 
 export function getErrorParts(error: Error): string[] {