Hiroshi Tanaka c2950121c9
feat(session): LLM-powered session summary via claude -p (#2388)
Replace mechanical text extraction in session-end.js and pre-compact.js
with LLM-generated summaries using `claude -p`. Summaries now capture
design decisions, resolved bugs, changed files, and carry-over context
rather than just truncated user message snippets.

- Add scripts/lib/llm-summary.js: generateSessionSummary, extractConversationText,
  getContextRemainingPct, getContextThreshold, getLLMModel
- Update scripts/hooks/session-end.js: trigger LLM when context < 20% or
  every 50 messages (env-configurable via ECC_LLM_SUMMARY_*)
- Update scripts/hooks/pre-compact.js: generate LLM summary right before
  compaction and write it to the active session .tmp file
- Add tests/lib/llm-summary.test.js: 18 unit tests
- Update tests/hooks/hooks.test.js: 3 integration tests for new behaviour

Recursion guard: sets ECC_SKIP_LLM_SUMMARY=1 in subprocess env so Stop
hooks fired by the claude -p subprocess do not re-enter summarisation.
Requires no ANTHROPIC_API_KEY — reuses Claude Code's own authentication.

Co-authored-by: Hiroshi Tanaka <hiroshi_tanaka@MBAM3.local>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-06-29 15:55:01 -07:00

177 lines
5.2 KiB
JavaScript

#!/usr/bin/env node
/**
* LLM-powered session summary generator
*
* Uses `claude -p` (Claude Code CLI) to generate rich, contextual session
* summaries from JSONL transcripts. Requires no API key — reuses Claude Code's
* own authentication.
*
* Recursion guard: sets ECC_SKIP_LLM_SUMMARY=1 in subprocess env so any Stop
* hooks fired by the subprocess do NOT re-enter LLM summarization.
*/
'use strict';
const { spawnSync } = require('child_process');
const fs = require('fs');
const MAX_TRANSCRIPT_CHARS = 7000;
const MAX_TURNS = 25;
const LLM_TIMEOUT_MS = 90000;
function getLLMModel() {
return process.env.ECC_LLM_SUMMARY_MODEL || 'haiku';
}
function getContextThreshold() {
const raw = parseInt(process.env.ECC_LLM_SUMMARY_CONTEXT_THRESHOLD || '20', 10);
return Number.isFinite(raw) && raw > 0 && raw <= 100 ? raw : 20;
}
/**
* Extract the last MAX_TURNS user+assistant turns from a JSONL transcript.
* Returns null when the transcript is missing or has no parseable turns.
*/
function extractConversationText(transcriptPath) {
let content;
try {
content = fs.readFileSync(transcriptPath, 'utf8');
} catch {
return null;
}
const lines = content.split('\n').filter(Boolean);
const turns = [];
for (const line of lines) {
try {
const entry = JSON.parse(line);
const isUser = entry.type === 'user' || entry.message?.role === 'user';
const isAssistant = entry.type === 'assistant';
if (isUser) {
const rawContent = entry.message?.content ?? entry.content;
const text =
typeof rawContent === 'string'
? rawContent
: Array.isArray(rawContent)
? rawContent
.filter(c => c?.type === 'text')
.map(c => c.text)
.join(' ')
: '';
const cleaned = text.replace(/\n+/g, ' ').trim();
if (cleaned) {
turns.push({ role: 'User', text: cleaned.slice(0, 400) });
}
}
if (isAssistant && Array.isArray(entry.message?.content)) {
const textParts = entry.message.content
.filter(b => b?.type === 'text')
.map(b => b.text)
.join(' ')
.replace(/\n+/g, ' ')
.trim();
if (textParts) {
turns.push({ role: 'Claude', text: textParts.slice(0, 600) });
}
}
} catch {
// Skip unparseable lines
}
}
if (turns.length === 0) return null;
const recent = turns.slice(-MAX_TURNS);
const formatted = recent.map(t => `**${t.role}:** ${t.text}`).join('\n\n');
return formatted.length > MAX_TRANSCRIPT_CHARS ? '...(前略)\n\n' + formatted.slice(-MAX_TRANSCRIPT_CHARS) : formatted;
}
/**
* Read the context remaining percentage from a transcript's latest usage record.
* Returns null when unavailable.
*/
function getContextRemainingPct(transcriptPath) {
try {
const { readLatestContextTokens, resolveContextWindowTokens } = require('./transcript-context');
const usage = readLatestContextTokens(transcriptPath);
if (!usage) return null;
const windowTokens = resolveContextWindowTokens(usage.tokens, usage.model);
return Math.round((1 - usage.tokens / windowTokens) * 100);
} catch {
return null;
}
}
/**
* Generate a session summary using `claude -p`.
* Returns the summary string, or null on failure or when recursion guard is active.
*/
function generateSessionSummary(transcriptPath) {
if (process.env.ECC_SKIP_LLM_SUMMARY) return null;
const conversation = extractConversationText(transcriptPath);
if (!conversation) return null;
const prompt = [
'Below is a conversation log from a Claude Code coding session.',
'Create a summary to help the next session quickly understand the context.',
'',
'## Prioritize including',
'- Design decisions and technology choices made this session',
'- Bugs and problems solved',
'- Files changed or created, with a brief description of changes',
'- Unfinished tasks and work to continue in the next session',
'- Important context the next session needs to know',
'',
'## Conversation log',
conversation,
'',
'## Output format (Markdown only, no preamble)',
'',
'## Session Summary',
'',
'### Tasks',
'(main tasks worked on this session)',
'',
'### Decisions Made',
'(design decisions and technology choices)',
'',
'### Files Modified',
'(files changed or created)',
'',
'### Unresolved Issues',
'(unfinished tasks and work to continue)',
'',
'### Next Session Context',
'(important context for the next session)'
].join('\n');
try {
const result = spawnSync('claude', ['--model', getLLMModel(), '-p'], {
input: prompt,
encoding: 'utf8',
env: {
...process.env,
CLAUDECODE: '',
ECC_SKIP_LLM_SUMMARY: '1'
},
timeout: LLM_TIMEOUT_MS,
shell: process.platform === 'win32'
});
if (result.error || result.status !== 0) {
return null;
}
const output = (result.stdout || '').trim();
return output || null;
} catch {
return null;
}
}
module.exports = { generateSessionSummary, extractConversationText, getContextRemainingPct, getContextThreshold, getLLMModel };