@event4u/agent-config 9.16.0 → 9.17.0

package/src/scripts/_cli/handoff_generate.ts

@@ -0,0 +1,453 @@
+/**
+ * Handoff generator (road-to-agent-handoff-resume Phase 2). Takes a session
+ * picked by `handoff_sessions.ts`, reads its FULL transcript, and emits the
+ * handoff document with the exact section set from
+ * `src/domains/meta/agent-handoff/command.md` step 2 — deterministically:
+ * no LLM call, reproducible, privacy-floored per line.
+ *
+ *   - User messages are preserved VERBATIM (the template's load-bearing
+ *     rule). A prompt containing any line the low-impact redactor refuses
+ *     is withheld whole (drop, never rewrite) and counted in the
+ *     `[N line(s) withheld by privacy floor]` marker.
+ *   - Assistant/tool content is extracted with the hot-context heuristics:
+ *     last intents, tool-result changes, failure-pattern regex for errors.
+ *   - Output lands atomically in `agents/runtime/state/handoff-context.md`
+ *     with `Generated:` / `Source-Session:` / `Branch:` parse anchors for
+ *     the session_start hook, hard-capped at ~1200 words (priority-trim
+ *     like `build_hot_context`).
+ *   - `--llm` polish is a seam only in v1: requesting it raises a clear
+ *     "not implemented" error. No spend without the flag, and none with it.
+ */
+
+import { spawnSync } from 'node:child_process';
+import * as fs from 'node:fs';
+import * as path from 'node:path';
+import process from 'node:process';
+
+import { hardenedSpawnEnv } from '../_lib/spawn_env.js';
+import { redact_low_impact_entry } from '../ai_council/redact_low_impact_entry.js';
+import { read_entries } from '../chat_history.js';
+import { HANDOFF_CONTEXT_REL } from '../handoff_context_hook.js';
+import type { HandoffSession } from './handoff_sessions.js';
+
+export const WORD_CAP = 1200;
+
+const SNIPPET_CHARS = 200;
+const DONE_SNIPPET_CHARS = 400;
+const MAX_ERRORS = 5;
+const MAX_DECISIONS = 5;
+const MAX_FILES = 15;
+
+const FAILURE_RE = /\b(fail(ing|ed|ure)?|error|exit[= ][1-9]|red\b|broken)\b/i;
+const DECISION_RE = /\b(decided|decision|chose|choosing|locked|we'll use|we will use)\b/i;
+const FILE_PATH_RE = /(?:src|tests|docs|agents|scripts|config)\/[\w./-]+\.\w{1,8}/g;
+const ROADMAP_RE = /agents\/roadmaps\/[\w.-]+\.md/g;
+
+export class LlmPolishNotImplementedError extends Error {
+    constructor() {
+        super(
+            '--llm polish is a v1 seam and not implemented yet — ' +
+                'the deterministic handoff (default) is the supported path.',
+        );
+    }
+}
+
+// ---------------------------------------------------------------------
+// unified transcript model
+// ---------------------------------------------------------------------
+
+export interface TranscriptTurn {
+    kind: 'user' | 'assistant' | 'tool';
+    text: string;
+    ts: string | null;
+    tool?: string;
+}
+
+export interface TranscriptData {
+    turns: TranscriptTurn[];
+    branch: string | null;
+}
+
+function _read_lines(file: string): string[] {
+    try {
+        return fs.readFileSync(file, 'utf-8').split('\n');
+    } catch {
+        return [];
+    }
+}
+
+function _parse_json(line: string): Record<string, unknown> | null {
+    const trimmed = line.trim();
+    if (!trimmed) return null;
+    try {
+        const parsed = JSON.parse(trimmed) as unknown;
+        if (parsed && typeof parsed === 'object' && !Array.isArray(parsed)) {
+            return parsed as Record<string, unknown>;
+        }
+    } catch {
+        // fail-open
+    }
+    return null;
+}
+
+function _from_chat_history(sessionId: string, chatHistoryPath: string | null): TranscriptData {
+    let entries: Array<Record<string, unknown>>;
+    try {
+        entries = read_entries({ session: sessionId, path: chatHistoryPath, last: null });
+    } catch {
+        entries = [];
+    }
+    const turns: TranscriptTurn[] = [];
+    for (const e of entries) {
+        const text = typeof e.text === 'string' ? e.text : '';
+        if (!text) continue;
+        const t = String(e.t ?? '');
+        const kind: TranscriptTurn['kind'] = t === 'user' ? 'user' : t === 'tool' ? 'tool' : 'assistant';
+        turns.push({
+            kind,
+            text,
+            ts: typeof e.ts === 'string' ? e.ts : null,
+            ...(typeof e.tool === 'string' && e.tool ? { tool: e.tool } : {}),
+        });
+    }
+    return { turns, branch: null };
+}
+
+function _cc_text_from_content(content: unknown): string {
+    if (typeof content === 'string') return content;
+    if (Array.isArray(content)) {
+        return content
+            .map((block) => {
+                if (block && typeof block === 'object' && !Array.isArray(block)) {
+                    const b = block as Record<string, unknown>;
+                    if (b.type === 'text' && typeof b.text === 'string') return b.text;
+                }
+                return '';
+            })
+            .filter((t) => t)
+            .join('\n');
+    }
+    return '';
+}
+
+function _from_claude_transcript(file: string): TranscriptData {
+    const turns: TranscriptTurn[] = [];
+    let branch: string | null = null;
+    for (const line of _read_lines(file)) {
+        const rec = _parse_json(line);
+        if (!rec || rec.isSidechain === true) continue;
+        if (!branch && typeof rec.gitBranch === 'string' && rec.gitBranch) branch = rec.gitBranch;
+        const ts = typeof rec.timestamp === 'string' ? rec.timestamp : null;
+        const msg =
+            rec.message && typeof rec.message === 'object' && !Array.isArray(rec.message)
+                ? (rec.message as Record<string, unknown>)
+                : null;
+        if (!msg) continue;
+        if (rec.type === 'user') {
+            // real prompts have string content; tool_results arrive as lists
+            if (typeof msg.content === 'string') {
+                const text = msg.content.trim();
+                if (text && !text.startsWith('<') && !text.startsWith('Caveat:')) {
+                    turns.push({ kind: 'user', text, ts });
+                }
+            } else {
+                const text = _cc_text_from_content(msg.content).trim();
+                if (text) turns.push({ kind: 'tool', text, ts });
+            }
+        } else if (rec.type === 'assistant') {
+            const text = _cc_text_from_content(msg.content).trim();
+            if (text) turns.push({ kind: 'assistant', text, ts });
+        }
+    }
+    return { turns, branch };
+}
+
+function _from_codex_session(file: string): TranscriptData {
+    const turns: TranscriptTurn[] = [];
+    let branch: string | null = null;
+    for (const line of _read_lines(file)) {
+        const rec = _parse_json(line);
+        if (!rec) continue;
+        const ts = typeof rec.timestamp === 'string' ? rec.timestamp : null;
+        const p =
+            rec.payload && typeof rec.payload === 'object' && !Array.isArray(rec.payload)
+                ? (rec.payload as Record<string, unknown>)
+                : null;
+        if (!p) continue;
+        if (rec.type === 'session_meta') {
+            const git =
+                p.git && typeof p.git === 'object' && !Array.isArray(p.git)
+                    ? (p.git as Record<string, unknown>)
+                    : {};
+            if (typeof git.branch === 'string' && git.branch) branch = git.branch;
+        } else if (rec.type === 'event_msg' && p.type === 'user_message') {
+            const text = typeof p.message === 'string' ? p.message.trim() : '';
+            if (text && !text.startsWith('<')) turns.push({ kind: 'user', text, ts });
+        } else if (rec.type === 'event_msg' && p.type === 'agent_message') {
+            const text = typeof p.message === 'string' ? p.message.trim() : '';
+            if (text) turns.push({ kind: 'assistant', text, ts });
+        }
+    }
+    return { turns, branch };
+}
+
+export function load_transcript(
+    session: Pick<HandoffSession, 'id' | 'source' | 'transcriptPath'>,
+    options: { chatHistoryPath?: string | null } = {},
+): TranscriptData {
+    if (session.source === 'chat-history') {
+        return _from_chat_history(session.id, options.chatHistoryPath ?? null);
+    }
+    if (!session.transcriptPath) return { turns: [], branch: null };
+    return session.source === 'claude-transcript'
+        ? _from_claude_transcript(session.transcriptPath)
+        : _from_codex_session(session.transcriptPath);
+}
+
+// ---------------------------------------------------------------------
+// deterministic section extraction
+// ---------------------------------------------------------------------
+
+function _snippet(text: string, max: number): string {
+    const oneLine = text.replace(/\s+/g, ' ').trim();
+    return oneLine.length <= max ? oneLine : oneLine.slice(0, max - 1) + '…';
+}
+
+function _word_count(text: string): number {
+    return text.split(/\s+/).filter((w) => w.length > 0).length;
+}
+
+/** Privacy floor for NON-verbatim items: drop refused lines, count them. */
+function _redact_items(items: string[]): { kept: string[]; dropped: number } {
+    const kept: string[] = [];
+    let dropped = 0;
+    for (const item of items) {
+        try {
+            if (redact_low_impact_entry(item).ok) kept.push(item);
+            else dropped += 1;
+        } catch {
+            dropped += 1; // fail-closed per line
+        }
+    }
+    return { kept, dropped };
+}
+
+/**
+ * Verbatim items: a prompt is included ONLY when every one of its lines
+ * passes the redactor — otherwise it is withheld whole and its failing
+ * lines are counted (drop, never rewrite).
+ */
+function _redact_verbatim(prompts: string[]): { kept: string[]; withheldLines: number } {
+    const kept: string[] = [];
+    let withheldLines = 0;
+    for (const prompt of prompts) {
+        const lines = prompt.split('\n');
+        let failing = 0;
+        for (const line of lines) {
+            try {
+                if (!redact_low_impact_entry(line).ok) failing += 1;
+            } catch {
+                failing += 1;
+            }
+        }
+        if (failing === 0) kept.push(prompt);
+        else withheldLines += failing;
+    }
+    return { kept, withheldLines };
+}
+
+function _current_head(root: string): string | null {
+    try {
+        const proc = spawnSync('git', ['log', '--oneline', '-1'], {
+            cwd: root,
+            encoding: 'utf-8',
+            timeout: 5000,
+            env: hardenedSpawnEnv(),
+        });
+        if (proc.status === 0) return (proc.stdout || '').trim() || null;
+    } catch {
+        // fall through
+    }
+    return null;
+}
+
+function _current_branch(root: string): string {
+    try {
+        const proc = spawnSync('git', ['rev-parse', '--abbrev-ref', 'HEAD'], {
+            cwd: root,
+            encoding: 'utf-8',
+            timeout: 5000,
+            env: hardenedSpawnEnv(),
+        });
+        if (proc.status === 0) return (proc.stdout || '').trim() || 'unknown';
+    } catch {
+        // fall through
+    }
+    return 'unknown';
+}
+
+export interface BuildOptions {
+    cwd?: string;
+    chatHistoryPath?: string | null;
+    now?: Date;
+    wordCap?: number;
+    /** v1 seam — requesting it throws LlmPolishNotImplementedError. */
+    llm?: boolean;
+}
+
+export function build_handoff(
+    session: Pick<HandoffSession, 'id' | 'source' | 'transcriptPath' | 'branch'>,
+    options: BuildOptions = {},
+): string {
+    if (options.llm) throw new LlmPolishNotImplementedError();
+
+    const root = options.cwd ?? process.cwd();
+    const now = options.now ?? new Date();
+    const wordCap = options.wordCap ?? WORD_CAP;
+    const transcript = load_transcript(session, { chatHistoryPath: options.chatHistoryPath ?? null });
+
+    const users = transcript.turns.filter((t) => t.kind === 'user');
+    const assistants = transcript.turns.filter((t) => t.kind === 'assistant');
+    const toolish = transcript.turns.filter((t) => t.kind !== 'user');
+
+    // User instructions — verbatim or withheld, never rewritten.
+    const verbatim = _redact_verbatim(users.map((t) => t.text));
+
+    // Done — the session's final assistant text.
+    const lastAssistant = assistants[assistants.length - 1]?.text ?? '';
+    const done = _redact_items(lastAssistant ? [_snippet(lastAssistant, DONE_SNIPPET_CHARS)] : []);
+
+    // Open + Resume pointer — the last user ask is what was in flight.
+    const lastUser = users[users.length - 1]?.text ?? '';
+    const open = _redact_items(lastUser ? [`in-flight ask: ${_snippet(lastUser, SNIPPET_CHARS)}`] : []);
+    const resume = _redact_items(
+        lastUser ? [`continue from the last user ask: ${_snippet(lastUser, SNIPPET_CHARS)}`] : [],
+    );
+
+    // Errors + fixes — failure-pattern turns, oldest→newest, last 5.
+    const errors = _redact_items(
+        toolish
+            .filter((t) => FAILURE_RE.test(t.text))
+            .slice(-MAX_ERRORS)
+            .map((t) => _snippet(`${t.tool ? `${t.tool}: ` : ''}${t.text}`, SNIPPET_CHARS)),
+    );
+
+    // Key decisions — decision-pattern assistant turns, last 5.
+    const decisions = _redact_items(
+        assistants
+            .filter((t) => DECISION_RE.test(t.text))
+            .slice(-MAX_DECISIONS)
+            .map((t) => _snippet(t.text, SNIPPET_CHARS)),
+    );
+
+    // Relevant files — path mentions across the whole transcript, deduped.
+    const fileSet = new Set<string>();
+    for (const t of transcript.turns) {
+        for (const m of t.text.match(FILE_PATH_RE) ?? []) fileSet.add(m);
+    }
+    const files = _redact_items([...fileSet].slice(-MAX_FILES));
+
+    // Roadmap — last roadmap-path mention, when present.
+    let roadmap: string | null = null;
+    for (const t of transcript.turns) {
+        for (const m of t.text.match(ROADMAP_RE) ?? []) roadmap = m;
+    }
+
+    const branch = session.branch ?? transcript.branch ?? 'unknown';
+    const droppedTotal =
+        done.dropped + open.dropped + resume.dropped + errors.dropped + decisions.dropped + files.dropped;
+
+    const sections: Array<{ title: string; items: string[]; verbatim?: boolean }> = [
+        { title: 'User instructions (VERBATIM)', items: verbatim.kept, verbatim: true },
+        { title: 'Done', items: done.kept },
+        { title: 'Open', items: open.kept },
+        { title: 'Resume pointer', items: resume.kept },
+        { title: 'Errors + fixes', items: errors.kept },
+        { title: 'Key decisions', items: decisions.kept },
+        { title: 'Relevant files', items: files.kept },
+    ];
+    // trim lowest-priority first; the verbatim record survives longest
+    const trimOrder = [
+        'Relevant files',
+        'Key decisions',
+        'Errors + fixes',
+        'Done',
+        'Open',
+        'Resume pointer',
+        'User instructions (VERBATIM)',
+    ];
+
+    const head = branch !== 'unknown' && branch === _current_branch(root) ? _current_head(root) : null;
+
+    const render = (): string => {
+        const lines: string[] = [
+            '# Handoff',
+            '',
+            `Generated: ${now.toISOString()}`,
+            `Source-Session: ${session.id}`,
+            `Branch: ${branch}`,
+        ];
+        if (head) lines.push(`Last commit: ${head}`);
+        if (roadmap) lines.push(`Roadmap: ${roadmap}`);
+        if (verbatim.withheldLines > 0 || droppedTotal > 0) {
+            lines.push(
+                `Privacy floor: [${verbatim.withheldLines + droppedTotal} line(s) withheld by privacy floor]`,
+            );
+        }
+        for (const s of sections) {
+            lines.push('', `## ${s.title}`, '');
+            if (s.items.length === 0) {
+                if (s.verbatim && verbatim.withheldLines > 0) {
+                    lines.push(`- [${verbatim.withheldLines} line(s) withheld by privacy floor]`);
+                } else {
+                    lines.push('- (none extracted)');
+                }
+                continue;
+            }
+            for (const item of s.items) {
+                if (s.verbatim) {
+                    // verbatim: keep internal newlines, indent continuations
+                    const [first, ...rest] = item.split('\n');
+                    lines.push(`- ${first ?? ''}`);
+                    for (const cont of rest) lines.push(`  ${cont}`);
+                } else {
+                    lines.push(`- ${item}`);
+                }
+            }
+        }
+        lines.push('');
+        return lines.join('\n');
+    };
+
+    let text = render();
+    for (const title of trimOrder) {
+        if (_word_count(text) <= wordCap) break;
+        const section = sections.find((s) => s.title === title);
+        if (!section) continue;
+        while (section.items.length > 0 && _word_count(text) > wordCap) {
+            // verbatim: drop the OLDEST prompt first (shift), keep the latest asks
+            if (section.verbatim) section.items.shift();
+            else section.items.pop();
+            text = render();
+        }
+    }
+    return text;
+}
+
+// ---------------------------------------------------------------------
+// atomic write
+// ---------------------------------------------------------------------
+
+export function handoff_target_path(root: string): string {
+    return process.env.AGENT_HANDOFF_CONTEXT_FILE || path.join(root, HANDOFF_CONTEXT_REL);
+}
+
+export function write_handoff(text: string, options: { cwd?: string } = {}): string {
+    const root = options.cwd ?? process.cwd();
+    const target = handoff_target_path(root);
+    fs.mkdirSync(path.dirname(target), { recursive: true });
+    const tmp = `${target}.tmp-${process.pid}`;
+    fs.writeFileSync(tmp, text, { encoding: 'utf-8' });
+    fs.renameSync(tmp, target);
+    return target;
+}