codeburn/tests/parser-streamed-tool-blocks.test.ts
Resham Joshi 75d3f0a9b6
fix(claude): read thinking tokens into the reasoning breakdown (#1848)
* fix(claude): read thinking tokens into the reasoning breakdown

Claude rows carry output_tokens_details.thinking_tokens, but reasoningTokens was always 0. Read it so the token breakdown shows thinking. Claude is in REASONING_INCLUDED_IN_OUTPUT, so billed output and cost are unchanged. Bumps the Claude parse version so cached calls re-parse.

* docs: note the thinking-tokens parse suffix
2026-11-04 19:08:10 +01:00

94 lines
4.3 KiB
TypeScript

import { describe, expect, it } from 'vitest'
import { dedupeStreamingMessageIds, parseApiCall } from '../src/parser.js'
import type { ContentBlock, JournalEntry } from '../src/types.js'
// Claude Code writes one JSONL line per content block of an assistant message:
// every line repeats the message id and usage, and each carries only its own
// block. Parallel tool calls therefore arrive as several lines with the same id.
// Older streams (#110) instead restate the message and put every block in the
// final write. Both shapes must yield each tool_use exactly once.
const usage = { input_tokens: 10, output_tokens: 200, cache_read_input_tokens: 5_000, cache_creation_input_tokens: 0 }
function line(id: string, content: ContentBlock[], timestamp: string): JournalEntry {
return {
type: 'assistant',
sessionId: 's1',
timestamp,
message: { id, type: 'message', role: 'assistant', model: 'claude-sonnet-4-5', content, usage },
}
}
const read = (id: string): ContentBlock => ({ type: 'tool_use', id, name: 'Read', input: { file_path: `/work/${id}.ts` } })
const bash = (id: string): ContentBlock => ({ type: 'tool_use', id, name: 'Bash', input: { command: 'npm test' } })
const mcp = (id: string): ContentBlock => ({ type: 'tool_use', id, name: 'mcp__playwright__browser_navigate', input: { url: 'http://localhost' } })
function toolsOf(entries: JournalEntry[]): string[][] {
return dedupeStreamingMessageIds(entries).map(e => parseApiCall(e)?.tools ?? [])
}
describe('streamed assistant messages with one content block per line', () => {
it('keeps every parallel tool_use, not only the last line', () => {
const entries = [
line('msg-1', [{ type: 'thinking', thinking: 'plan' }], '2026-09-11T10:00:00.000Z'),
line('msg-1', [read('t1')], '2026-09-11T10:00:01.000Z'),
line('msg-1', [read('t2')], '2026-09-11T10:00:02.000Z'),
line('msg-1', [bash('t3')], '2026-09-11T10:00:03.000Z'),
]
const deduped = dedupeStreamingMessageIds(entries)
expect(deduped).toHaveLength(1)
expect(deduped[0]!.timestamp).toBe('2026-09-11T10:00:00.000Z')
expect(toolsOf(entries)).toEqual([['Read', 'Read', 'Bash']])
})
it('keeps an MCP call made in parallel with a built-in tool', () => {
const entries = [
line('msg-1', [mcp('t1')], '2026-09-11T10:00:01.000Z'),
line('msg-1', [read('t2')], '2026-09-11T10:00:02.000Z'),
]
const call = parseApiCall(dedupeStreamingMessageIds(entries)[0]!)
expect(call?.mcpTools).toEqual(['mcp__playwright__browser_navigate'])
expect(call?.tools).toContain('Read')
})
it('still counts usage once per message', () => {
const entries = [
line('msg-1', [read('t1')], '2026-09-11T10:00:01.000Z'),
line('msg-1', [bash('t2')], '2026-09-11T10:00:02.000Z'),
]
const calls = dedupeStreamingMessageIds(entries).map(e => parseApiCall(e)!)
expect(calls).toHaveLength(1)
expect(calls[0]!.usage.outputTokens).toBe(200)
})
})
describe('streamed assistant messages restated with every block in the final write (#110)', () => {
it('does not double count a tool_use repeated in a later restatement', () => {
const entries = [
line('msg-1', [], '2026-09-11T10:00:00.000Z'),
line('msg-1', [read('t1')], '2026-09-11T10:00:01.000Z'),
line('msg-1', [read('t1'), bash('t2')], '2026-09-11T10:00:02.000Z'),
]
expect(toolsOf(entries)).toEqual([['Read', 'Bash']])
})
it('leaves a message whose last line already holds every block unchanged', () => {
const last = line('msg-1', [read('t1'), bash('t2')], '2026-09-11T10:00:02.000Z')
const entries = [line('msg-1', [], '2026-09-11T10:00:00.000Z'), last]
expect(dedupeStreamingMessageIds(entries)[0]!.message).toBe(last.message)
})
})
describe('thinking tokens', () => {
it('fills reasoningTokens without changing output tokens or cost', () => {
const base = line('msg-t', [{ type: 'text', text: 'ok' }], '2026-09-11T10:00:00.000Z')
const withThinking = line('msg-t', [{ type: 'text', text: 'ok' }], '2026-09-11T10:00:00.000Z')
withThinking.message!.usage = { ...usage, output_tokens_details: { thinking_tokens: 120 } }
const a = parseApiCall(base)!
const b = parseApiCall(withThinking)!
expect(b.usage.reasoningTokens).toBe(120)
expect(b.usage.outputTokens).toBe(a.usage.outputTokens)
expect(b.costUSD).toBe(a.costUSD)
})
})