@krtclcdy/tersio 2.25.5 → 2.25.7

package/dist/extensions/shared/usage-store.js

@@ -1,8 +1,8 @@
-// extensions/shared/usage-store.ts — tersio-owned usage.db. A cache of the live parse, never a fork; missing sqlite3 → sync false / read null.
-import { execFileSync } from 'node:child_process';
+// usage-store.ts — tersio-owned usage.db. A cache of the live parse, not a fork; no SQLite → sync false / read null.
 import fs from 'node:fs';
 import path from 'node:path';
-import { RECENT_LIMIT, SQLITE_READ_BUFFER, canonicalModelId, classifySessionLine, classifyOpencodeMessage, costOf, durOf, agentOfSessionText, hostOfSessionFile, ingestSessionRow, messageRowId, OPENCODE_DB_OVERLAP_MS, opencodeDbPath, opencodeSessionsDir, newSessionAccum, readOpencodeDbRows, sessionsDirs, walkJsonl, } from './usage-ledger.js';
+import { sqliteExec, sqliteRead } from './sqlite.js';
+import { RECENT_LIMIT, canonicalModelId, classifySessionLine, classifyOpencodeMessage, costOf, durOf, agentOfSessionText, hostOfSessionFile, ingestSessionRow, messageRowId, OPENCODE_DB_OVERLAP_MS, opencodeDbPath, opencodeSessionsDir, newSessionAccum, opencodeToolNames, readOpencodeDbRows, sessionsDirs, walkJsonl, } from './usage-ledger.js';
 import { tersioDataPath } from './utils.js';
 import { MODEL_ALIASES } from './pricing.js';
 import { tersioSettingsFile } from './plugin-settings.js';
@@ -24,25 +24,21 @@ function nullNum(v) {
 function nullStr(v) {
     return v === undefined ? 'NULL' : esc(v);
 }
+// A failed INSERT stops before COMMIT, so it leaves no DELETEs behind.
 function run(db, sql) {
-    // -bail stops the script before COMMIT, so a failed INSERT never leaves DELETEs behind.
-    execFileSync('sqlite3', ['-bail', db, sql], { stdio: ['ignore', 'ignore', 'ignore'], timeout: 30000 });
+    sqliteExec(db, sql);
 }
 function query(db, sql) {
-    const out = execFileSync('sqlite3', ['-separator', '\t', db, sql], {
-        encoding: 'utf8',
-        timeout: 30000,
-        stdio: ['ignore', 'pipe', 'ignore'],
-        maxBuffer: SQLITE_READ_BUFFER,
-    });
-    return out
+    return sqliteRead(db, sql)
         .split('\n')
         .filter((line) => line.trim())
         .map((line) => line.split('\t'));
 }
 // Bump on a parse change: unchanged transcripts are never re-read.
-const PARSER_VERSION = '12';
-// A re-parse that keeps less than half the rows means transcripts vanished mid-migration; the backup is restored instead of publishing the loss.
+const PARSER_VERSION = '13';
+/** OpenCode rows gained tool names; re-read by id, since a PARSER_VERSION bump drops rotated history. */
+const OPENCODE_TOOLS_VERSION = '1';
+// A re-parse keeping under half the rows means transcripts vanished mid-migration; the backup is restored instead.
 const REPARSE_GUARD_RATIO = 0.5;
 let guardReport = null;
 /** Why the last sync refused to publish a re-parse, if it did. */
@@ -52,18 +48,14 @@ export function reparseGuardReport() {
 const SCHEMA_SQL = `CREATE TABLE IF NOT EXISTS meta (k TEXT PRIMARY KEY, v TEXT NOT NULL);` +
     `CREATE TABLE IF NOT EXISTS files (id INTEGER PRIMARY KEY AUTOINCREMENT, path TEXT UNIQUE NOT NULL, mtime REAL NOT NULL, size INTEGER NOT NULL);` +
     `CREATE TABLE IF NOT EXISTS messages (file_id INTEGER NOT NULL REFERENCES files(id), t REAL, model TEXT NOT NULL,` +
-    ` i INTEGER NOT NULL, o INTEGER NOT NULL, d REAL, cr INTEGER NOT NULL, cw INTEGER NOT NULL,` +
+    ` i INTEGER NOT NULL, o INTEGER NOT NULL, d REAL, tf REAL, cr INTEGER NOT NULL, cw INTEGER NOT NULL,` +
     ` usd REAL, st TEXT NOT NULL, code REAL, note TEXT, tools TEXT NOT NULL DEFAULT '[]', h TEXT, a TEXT, id TEXT);` +
-    // Every read filters or orders by one of these three. Without them each sync and each read
-    // walks the whole table: the watermark scan, the day buckets, and the recent list all scan.
+    // Every read filters or orders by one of these three; without them each sync and read scans the table.
     `CREATE INDEX IF NOT EXISTS idx_messages_file ON messages(file_id);` +
     `CREATE INDEX IF NOT EXISTS idx_messages_t ON messages(t);` +
     `CREATE INDEX IF NOT EXISTS idx_messages_h_t ON messages(h, t);` +
     `CREATE INDEX IF NOT EXISTS idx_messages_model ON messages(model);`;
-/**
- * WAL plus a checkpoint at the end of a sync: the write path stops paying an fsync per commit,
- * and the main file still holds every row, so a backup that copies only usage.db loses nothing.
- */
+/** WAL + a checkpoint: the write path avoids an fsync per commit, and usage.db alone holds every row. */
 function pragmas(db) {
     run(db, 'PRAGMA journal_mode=WAL;' +
         'PRAGMA synchronous=NORMAL;' +
@@ -81,7 +73,7 @@ function storedParserVersion(db) {
         return null;
     }
 }
-// A store written by a newer tersio: this code has no forward migration, so a re-parse would swap rotated history for whatever transcripts are left.
+// A store written by a newer tersio: no forward migration, so a re-parse would swap rotated history for what is left.
 function isDowngrade(stored) {
     if (stored === null)
         return false;
@@ -109,6 +101,16 @@ function newestOpencodeMs(db) {
         return null;
     }
 }
+/** Whether the OpenCode tools backfill ran; unfilled rows stay empty, so this widens the scan once. */
+function opencodeToolsBackfilled(db) {
+    try {
+        const rows = query(db, `SELECT v FROM meta WHERE k='opencode_tools';`);
+        return rows.length > 0 && rows[0][0] === OPENCODE_TOOLS_VERSION;
+    }
+    catch {
+        return false;
+    }
+}
 // Same pre-migration state as the newest backup: another copy churns the 3-slot rotation for nothing.
 function newestBackupMatches(stored) {
     if (stored === null)
@@ -124,7 +126,7 @@ function newestBackupMatches(stored) {
         return false;
     }
 }
-// v9 shape (path-keyed rows) migrates in place with zero re-parse, so history survives even when transcripts are gone.
+// v9 shape (path-keyed rows) migrates in place with zero re-parse, so history survives even without transcripts.
 function canMigrateInPlace(db) {
     try {
         const names = query(db, `PRAGMA table_info(messages);`).map((r) => r[1]);
@@ -141,7 +143,7 @@ function migrateInPlace(db, preCount) {
             `CREATE TABLE files_new (id INTEGER PRIMARY KEY AUTOINCREMENT, path TEXT UNIQUE NOT NULL, mtime REAL NOT NULL, size INTEGER NOT NULL);` +
             `INSERT INTO files_new (path, mtime, size) SELECT path, mtime, size FROM files;` +
             `CREATE TABLE messages_new (file_id INTEGER NOT NULL REFERENCES files_new(id), t REAL, model TEXT NOT NULL,` +
-            ` i INTEGER NOT NULL, o INTEGER NOT NULL, d REAL, cr INTEGER NOT NULL, cw INTEGER NOT NULL,` +
+            ` i INTEGER NOT NULL, o INTEGER NOT NULL, d REAL, tf REAL, cr INTEGER NOT NULL, cw INTEGER NOT NULL,` +
             ` usd REAL, st TEXT NOT NULL, code REAL, note TEXT, tools TEXT NOT NULL DEFAULT '[]', h TEXT, a TEXT, id TEXT);` +
             `INSERT INTO messages_new (file_id, ${cols}) SELECT (SELECT id FROM files_new WHERE files_new.path = messages.file), ${cols} FROM messages;` +
             `DROP TABLE messages;DROP TABLE files;` +
@@ -160,9 +162,7 @@ function migrateInPlace(db, preCount) {
         return false;
     }
 }
-// v12 shape gains the agent kind (`main`, `subagent`) with a bare ALTER for the two names the
-// store already holds. No rows move and no re-parse is forced, so the existing rows read
-// `NULL` for it: a report then shows them under the main agent only, and new rows carry it.
+// v12 adds the agent kind with a bare ALTER; existing rows read NULL and report under the main agent.
 function canAddAgent(db) {
     try {
         const names = query(db, `PRAGMA table_info(messages);`).map((r) => r[1]);
@@ -172,8 +172,6 @@ function canAddAgent(db) {
         return false;
     }
 }
-// v12 shape gains the agent kind (`main`, `subagent`) with a bare ALTER: no rows move, so no
-// re-parse is forced. Existing rows read NULL, which a report folds under the main agent.
 function addAgentColumn(db) {
     try {
         run(db, `ALTER TABLE messages ADD COLUMN a TEXT;`);
@@ -183,6 +181,44 @@ function addAgentColumn(db) {
     }
     return true;
 }
+// v13 shape gains the first-token latency with a bare ALTER, so no row moves and no wipe runs.
+function canAddTtft(db) {
+    try {
+        const names = query(db, `PRAGMA table_info(messages);`).map((r) => r[1]);
+        return names.includes('file_id') && !names.includes('tf');
+    }
+    catch {
+        return false;
+    }
+}
+// The column is additive, and a rotated transcript cannot be re-parsed, so the ALTER precedes the backfill.
+function addTtftColumn(db) {
+    try {
+        run(db, `ALTER TABLE messages ADD COLUMN tf REAL;`);
+    }
+    catch {
+        return false;
+    }
+    return !canAddTtft(db);
+}
+/** First-token-latency marker; unchanged transcripts are never revisited, so old stores stay null. */
+const TTFT_VERSION = '1';
+function ttftBackfilled(db) {
+    try {
+        const rows = query(db, `SELECT v FROM meta WHERE k='ttft_backfill';`);
+        return rows.length > 0 && rows[0][0] === TTFT_VERSION;
+    }
+    catch {
+        return false;
+    }
+}
+/** Clears ledger mtimes so every transcript re-reads; synthetic `db#row-id` OpenCode keys stay untouched. */
+function markTranscriptsDirty(db) {
+    try {
+        run(db, `UPDATE files SET mtime = 0 WHERE path NOT LIKE '%#%';`);
+    }
+    catch { /* a store with no files table yet, or a read-only file: nothing to backfill */ }
+}
 // v10 shape gains the message id with a bare ALTER: no rows move, so no re-parse and no guard trip.
 function canAlterInPlace(db) {
     try {
@@ -233,10 +269,14 @@ function ensureSchema(db, forceWipe) {
             run(db, `INSERT OR REPLACE INTO meta (k,v) VALUES ('parser_version','${PARSER_VERSION}');`);
             return;
         }
-        // The agent column is additive on the current shape: adding it keeps every row, and the
-        // unchanged transcripts that already hold their rows simply read NULL until a re-parse.
+        // The agent column is additive: every row survives and unchanged transcripts read NULL until a re-parse.
         if (stored !== null && canAddAgent(db) && addAgentColumn(db)) {
-            run(db, `INSERT OR REPLACE INTO meta (k,v) VALUES ('parser_version','${PARSER_VERSION}');`);
+            run(db, `INSERT OR REPLACE INTO meta (k, v) VALUES ('parser_version','${PARSER_VERSION}');`);
+            return;
+        }
+        // The ttft column is additive too; its values come from the backfill in `syncUsageDb`.
+        if (stored !== null && canAddTtft(db) && addTtftColumn(db)) {
+            run(db, `INSERT OR REPLACE INTO meta (k, v) VALUES ('parser_version','${PARSER_VERSION}');`);
             return;
         }
         run(db, `DROP TABLE IF EXISTS messages;DROP TABLE IF EXISTS files;` + SCHEMA_SQL +
@@ -347,22 +387,22 @@ function readFiles(db) {
     catch { /* fresh db */ }
     return known;
 }
-// OpenCode rows carry no status, code, note, tool list, or subagent marker: only tokens and the
-// msg id. They record as the main agent, so a store never holds a NULL kind.
-function opencodeStoredRow(oc, host, id) {
+// OpenCode rows carry only tokens, msg id, and tools; they record as the main agent, so no NULL kind.
+function opencodeStoredRow(oc, host, id, tools) {
     return {
         t: oc.ms ?? null,
         model: canonicalModelId(oc.model),
         i: intOf(oc.usage.input),
         o: intOf(oc.usage.output),
         d: durOf(oc.durMs),
+        tf: durOf(oc.ttftMs),
         cr: intOf(oc.usage.cacheRead),
         cw: intOf(oc.usage.cacheWrite),
         usd: costOf(oc.usage),
         st: 'completed',
         code: undefined,
         note: undefined,
-        tools: [],
+        tools,
         host,
         agent: 'main',
         id,
@@ -373,13 +413,14 @@ function parseFile(text, host, file, agent) {
     // Guarded: a JSONL transcript must fall through, not abort the sync.
     let oc = null;
     try {
-        oc = classifyOpencodeMessage(JSON.parse(text));
+        const message = JSON.parse(text);
+        oc = classifyOpencodeMessage(message);
+        if (oc) {
+            rows.push(opencodeStoredRow(oc, host, file ? path.basename(file).replace(/\.[^.]+$/, '') : undefined, opencodeToolNames(message)));
+            return rows;
+        }
     }
     catch { /* JSONL */ }
-    if (oc) {
-        rows.push(opencodeStoredRow(oc, host, file ? path.basename(file).replace(/\.[^.]+$/, '') : undefined));
-        return rows;
-    }
     let codexProvider = null;
     let codexModel = null;
     for (const line of text.split('\n')) {
@@ -411,6 +452,7 @@ function parseFile(text, host, file, agent) {
                     i: intOf(usage.input),
                     o: intOf(usage.output),
                     d: undefined,
+                    tf: undefined,
                     cr: intOf(usage.cacheRead),
                     cw: intOf(usage.cacheWrite),
                     usd: costOf(usage),
@@ -436,6 +478,7 @@ function parseFile(text, host, file, agent) {
                 i: intOf(usage.input),
                 o: intOf(usage.output),
                 d: dur,
+                tf: durOf(parsed.ttftMs),
                 cr: intOf(usage.cacheRead),
                 cw: intOf(usage.cacheWrite),
                 usd: costOf(usage),
@@ -453,12 +496,12 @@ function parseFile(text, host, file, agent) {
     return rows;
 }
 function insertSql(file, r) {
-    return `INSERT INTO messages (file_id, t, model, i, o, d, cr, cw, usd, st, code, note, tools, h, a, id) VALUES (` +
+    return `INSERT INTO messages (file_id, t, model, i, o, d, tf, cr, cw, usd, st, code, note, tools, h, a, id) VALUES (` +
         `(SELECT id FROM files WHERE path=${esc(file)}),${r.t === null ? 'NULL' : String(r.t)},${esc(r.model)},${r.i},${r.o},` +
-        `${nullNum(r.d)},${r.cr},${r.cw},${nullNum(r.usd)},${esc(r.st)},` +
+        `${nullNum(r.d)},${nullNum(r.tf)},${r.cr},${r.cw},${nullNum(r.usd)},${esc(r.st)},` +
         `${nullNum(r.code)},${nullStr(r.note)},${esc(JSON.stringify(r.tools))},${nullStr(r.host)},${nullStr(r.agent ?? undefined)},${nullStr(r.id)});`;
 }
-// Unchanged transcripts are skipped via mtime+size; rows for deleted ones are kept so rotation never erases history. False when sqlite3 is unavailable.
+// Unchanged transcripts are skipped via mtime+size; rows for deleted ones are kept so rotation never erases history.
 export function syncUsageDb() {
     let db;
     guardReport = null;
@@ -477,6 +520,9 @@ export function syncUsageDb() {
         if (migrated && !forced)
             preWipe = countMessages(db);
         ensureSchema(db, forced);
+        // One rewrite so the additive column's NULL rows gain latency; runs while the marker is absent.
+        if (!downgrade && !ttftBackfilled(db))
+            markTranscriptsDirty(db);
         maybeScheduledBackup(db);
     }
     catch {
@@ -501,8 +547,10 @@ export function syncUsageDb() {
     const ocDbPath = opencodeDbPath();
     if (process.env.TERSIO_SESSIONS_DIR === undefined || process.env.TERSIO_OPENCODE_DB !== undefined) {
         const newest = newestOpencodeMs(db);
+        // A store predating the tool names re-reads the whole table once; the marker closes the window.
+        const since = newest === null || !opencodeToolsBackfilled(db) ? 0 : newest - OPENCODE_DB_OVERLAP_MS;
         if (newest !== null || fs.existsSync(ocDbPath)) {
-            ocDbRows.push(...readOpencodeDbRows(ocDbPath, newest === null ? 0 : newest - OPENCODE_DB_OVERLAP_MS));
+            ocDbRows.push(...readOpencodeDbRows(ocDbPath, since));
         }
     }
     const known = readFiles(db);
@@ -522,6 +570,7 @@ export function syncUsageDb() {
         if (!prev || prev.mtime !== entry.mtime || prev.size !== entry.size)
             changed.push(file);
     }
+    // `ocDbRows` is non-empty whenever the backfill window is open, so this never skips it.
     if (!changed.length && !ocDbRows.length)
         return true;
     const chunks = [];
@@ -571,7 +620,7 @@ export function syncUsageDb() {
             if (!push(`INSERT INTO files (path, mtime, size) VALUES (${esc(key)},${row.created},0)` +
                 ` ON CONFLICT(path) DO UPDATE SET mtime=excluded.mtime;` +
                 `DELETE FROM messages WHERE file_id=(SELECT id FROM files WHERE path=${esc(key)});` +
-                insertSql(key, opencodeStoredRow(parsed, 'opencode', row.id))))
+                insertSql(key, opencodeStoredRow(parsed, 'opencode', row.id, opencodeToolNames(row.row, row.tools)))))
                 return false;
         }
     }
@@ -580,6 +629,16 @@ export function syncUsageDb() {
     }
     if (!push(`INSERT OR REPLACE INTO meta (k, v) VALUES ('last_sync',${esc(String(Date.now()))});`))
         return false;
+    // Only once the rows are written, so a sync that fails midway widens the scan again.
+    if (ocDbRows.length > 0) {
+        if (!push(`INSERT OR REPLACE INTO meta (k, v) VALUES ('opencode_tools','${OPENCODE_TOOLS_VERSION}');`))
+            return false;
+    }
+    // Set only after the rewrite commits: a crash mid-backfill leaves mtimes cleared and the marker absent.
+    if (changed.length > 0) {
+        if (!push(`INSERT OR REPLACE INTO meta (k, v) VALUES ('ttft_backfill','${TTFT_VERSION}');`))
+            return false;
+    }
     if (!flush())
         return false;
     if (preWipe > 0) {
@@ -617,18 +676,18 @@ export function readUsageDb() {
     }
     let rows;
     try {
-        // The `tools` column is the widest by far and the common case is the literal `[]`, so it is
-        // read as one column and skipped in JS rather than parsed per row. A store written before
-        // the agent column does not hold it, so the row is read without that column.
+        // The wide `tools` column is read as one string and skipped in JS; a pre-agent store is read without it.
         const names = query(db, `PRAGMA table_info(messages);`).map((r) => r[1]);
         const agentCol = names.includes('a') ? 'a' : "''";
-        rows = query(db, `SELECT t, model, i, o, d, cr, cw, usd, st, code, note, tools, h, ${agentCol}, id FROM messages;`);
+        // A store predating first-token latency has no `tf` column; read NULL so the row still loads.
+        const tfCol = names.includes('tf') ? 'tf' : 'NULL';
+        rows = query(db, `SELECT t, model, i, o, d, cr, cw, usd, st, code, note, tools, h, ${agentCol}, id, ${tfCol} FROM messages;`);
     }
     catch {
         return null;
     }
     const accum = newSessionAccum();
-    for (const [t, model, i, o, d, cr, cw, usd, st, code, note, tools, h, a, id] of rows) {
+    for (const [t, model, i, o, d, cr, cw, usd, st, code, note, tools, h, a, id, tf] of rows) {
         const ts = t === '' ? undefined : Number(t);
         let toolNames = [];
         if (tools && tools !== '[]' && tools !== '') {
@@ -654,7 +713,7 @@ export function readUsageDb() {
             st: st === 'error' || st === 'aborted' ? st : 'completed',
             code: code === '' ? undefined : Number(code),
             note: note === '' ? undefined : note,
-        }, toolNames, h === '' ? undefined : h, id === '' || id === undefined ? undefined : id, a === '' ? undefined : a);
+        }, toolNames, h === '' ? undefined : h, id === '' || id === undefined ? undefined : id, a === '' ? undefined : a, tf === undefined || tf === '' ? undefined : Number(tf));
     }
     accum.recent.sort((a, b) => b.t - a.t);
     let syncedAt = 0;

package/dist/extensions/shared/utils.js

@@ -1,4 +1,4 @@
-// Shared helpers for the installer and the AI add-ons updater. Node built-ins only — this module must stay dependency-free.
+// Shared helpers for the installer and AI add-ons updater. Node built-ins only — must stay dependency-free.
 import { createHash } from 'node:crypto';
 import { accessSync, constants, createWriteStream, existsSync, mkdirSync, renameSync, statSync } from 'node:fs';
 import fs from 'node:fs/promises';
@@ -87,7 +87,7 @@ function redirectNext(res, url) {
     }
     return null;
 }
-// pi sets PI_CODING_AGENT; OMP does not. OMP does read PI_CODING_AGENT_DIR as its own relocation variable, so the override only counts under that marker.
+// pi sets PI_CODING_AGENT; OMP does not. It reads PI_CODING_AGENT_DIR to relocate, so the override only counts there.
 export function isPiProcess() {
     return process.env.PI_CODING_AGENT === 'true';
 }
@@ -95,7 +95,7 @@ export function piAgentDir() {
     const override = process.env.PI_CODING_AGENT_DIR;
     return isPiProcess() && override ? override : path.join(homeDir(), '.pi', 'agent');
 }
-// Ponytail is a Tersio dependency that npm hoists, so it can sit in Tersio's node_modules, a host's plugin dir, or beside it. Walking up covers all.
+// Ponytail is hoisted, so it sits in Tersio's node_modules, a host's plugin dir, or beside it; walking up covers all.
 export function findHoistedPackage(pkg, fromDir, ...relInside) {
     let dir = fromDir;
     for (;;) {
@@ -235,7 +235,7 @@ export async function readTextIfExists(p) {
         return null;
     }
 }
-// Chunked so a 200 MiB host binary is never held in memory. `tail` keeps the last needle.length-1 bytes of the previous chunk, because a match can begin near the end of one chunk and finish inside the next.
+// Chunked: a 200 MiB binary never fits memory. `tail` keeps needle.length-1 bytes in case a match spans chunks.
 export async function fileContains(filePath, needleText) {
     const needle = Buffer.from(needleText, 'utf8');
     if (needle.length === 0)
@@ -266,7 +266,7 @@ export async function fileContains(filePath, needleText) {
         await handle?.close().catch(() => { });
     }
 }
-// Tersio-owned data home: ~/.tersio. First use moves legacy ~/.omp/plugins/tersio-* files over so existing installs keep their history.
+// Tersio data home: ~/.tersio. First use moves legacy ~/.omp/plugins/tersio-* files over so installs keep history.
 const migratedTersioFiles = new Set();
 export function tersioHome() {
     // Matches the other TERSIO_* overrides so a test never reads the real home.

package/dist/extensions/timing/index.js

@@ -0,0 +1,31 @@
+export default function timingExtension(pi) {
+    // First token per request, keyed by the request's own start stamp, which survives every stream event.
+    const firstTokenAt = new Map();
+    pi.on('message_update', (event) => {
+        const message = event.message;
+        if (!message || message.role !== 'assistant')
+            return;
+        const anchor = typeof message.timestamp === 'number' ? message.timestamp : 0;
+        if (anchor <= 0 || firstTokenAt.has(anchor))
+            return;
+        // `start` opens the stream with no token yet; the first delta is the first token.
+        if (!String(event.assistantMessageEvent?.type ?? '').endsWith('_delta'))
+            return;
+        firstTokenAt.set(anchor, Date.now());
+    });
+    pi.on('message_end', (event) => {
+        const message = event.message;
+        if (!message || message.role !== 'assistant')
+            return;
+        const anchor = typeof message.timestamp === 'number' ? message.timestamp : 0;
+        const at = anchor > 0 ? firstTokenAt.get(anchor) : undefined;
+        if (anchor > 0)
+            firstTokenAt.delete(anchor);
+        // A host that recorded its own latency wins; this only fills the gap.
+        if (at === undefined || message.ttft !== undefined)
+            return;
+        const ttft = at - anchor;
+        if (ttft > 0)
+            message.ttft = ttft;
+    });
+}

package/dist/extensions/updater/index.js

@@ -1,4 +1,4 @@
-// /ai-addons manual updater. Node built-ins only. RTK ships SHA256 checksums with no signature, so checksum-only is the best available.
+// /ai-addons manual updater. Node built-ins only. RTK ships unsigned SHA256 checksums, so checksum-only it is.
 import { execFileSync } from 'node:child_process';
 import { createHash } from 'node:crypto';
 import fsSync from 'node:fs';
@@ -12,9 +12,7 @@ const IS_WINDOWS = process.platform === 'win32';
 const EXTENSION_DIR = path.dirname(fileURLToPath(import.meta.url));
 const PONYTAIL_REMOTE = 'https://raw.githubusercontent.com/DietrichGebert/ponytail/main/package.json';
 const RTK_BINARY = resolveRtkBinary();
-// Beside this file, not a host path: these are the rules /caveman reads. A retired
-// agent-tree copy has no `caveman` sibling, so the OMP plugin install is the
-// fallback: it is the one the plugin manifest loads from.
+// Beside this file: a retired agent-tree copy has no `caveman` sibling, so fall back to the OMP plugin.
 const CAVEMAN_SIBLING = (name) => path.join(EXTENSION_DIR, '..', 'caveman', name);
 function cavemanRule(name) {
     const beside = CAVEMAN_SIBLING(name);
@@ -30,7 +28,7 @@ const CAVEMAN_TRACKED = [
 ];
 // Host-neutral: this updater runs inside pi, OMP, and OpenCode alike.
 const RELOAD_MSG = 'Reminder: restart your agent (or reload extensions) for updates to take effect.';
-// Ponytail is a hoisted dependency, so the installed copy is found by walking up rather than by naming a host directory.
+// Ponytail is hoisted, so the installed copy is found by walking up, not by naming a host directory.
 function ponytailLocal() {
     return findHoistedPackage('@dietrichgebert/ponytail', EXTENSION_DIR, 'package.json');
 }
@@ -149,7 +147,6 @@ async function checkAddons(ctx) {
     }
     return lines.join('\n');
 }
-// Ponytail is a hoisted dependency; replace its package dir from the npm tarball.
 async function updatePonytail(pi, ctx, dryRun = false) {
     void pi;
     const localPkg = ponytailLocal();

package/dist/tersio.js

@@ -1,8 +1,8 @@
 #!/usr/bin/env node
 // Install Tersio add-ons on any OMP device; see help for commands.
-import { PACKAGE_BIN, PACKAGE_VERSION, applyUpdate, commandArg, dashboard, dashboardExport, dashboardOpen, dashboardPort, doctor, reset, settings, showHelp, showVersion, uninstall, unknownCommand, update, usage, currency, } from './cli/common.js';
+import { PACKAGE_BIN, PACKAGE_VERSION, applyUpdate, commandArg, dashboard, dashboardExport, dashboardOpen, dashboardPort, doctor, reinstall, reset, settings, showHelp, showVersion, uninstall, unknownCommand, update, usage, currency, } from './cli/common.js';
 import { closeRL } from './cli/interactive.js';
-import { runInstall } from './cli/install.js';
+import { runInstall, runReinstall } from './cli/install.js';
 import { runDoctor } from './cli/doctor.js';
 import { runLatestUpdate } from './cli/update.js';
 import { runUninstall } from './cli/uninstall.js';
@@ -15,6 +15,7 @@ function printHelp() {
 
 Commands:
   install      Install the add-ons for one agent (--host omp|pi|opencode, or ask at a terminal)
+  reinstall    Refresh one agent's install: pick it, reinstall, then verify (--host picks without asking)
   update       Refresh the CLI and add-ons (RTK binary, Caveman rule, Ponytail; --host limits to one agent)
   doctor       Check the current installation (--fix repairs, --dry-run previews)
   usage        Ledger-backed usage + savings report
@@ -26,7 +27,7 @@ Commands:
   help         Show this help
 
 Options:
-  --fix (doctor: repairs all; --fix=<scope> repairs one of extensions, registrations, rtk, ponytail, cli)
+  --fix (doctor: repairs all; --fix=<scope> repairs one of extensions, registrations, rtk, ponytail, cli, service)
   --scope user (legacy; accepted and ignored — user scope is the only scope)
   --host omp|pi|opencode (install|uninstall|update: target agent; a terminal asks instead)
   --keep-ponytail (uninstall: keep the bundled Ponytail copy — removed by default)
@@ -97,6 +98,10 @@ async function main() {
         closeRL();
         return;
     }
+    if (reinstall) {
+        await runReinstall();
+        return;
+    }
     await runInstall();
 }
 main().catch((e) => { closeRL(); console.error(e); process.exitCode = 1; });

package/extensions/caveman/index.ts

@@ -3,13 +3,17 @@ import { fileURLToPath } from 'node:url';
 import { getSharedComboState, hasModeState, isOmpSubagentPrompt, lastCustomValue, normalizeInputCommand, normalizeMode, reconcileSharedComboEntries, sessionEntries, setSharedComboListener, setSharedComboMode, systemPromptIncludes } from '../shared/session-state.ts';
 import { dirname, join } from 'node:path';
 import { injectPromptText, onHostEvent, setExtensionLabel } from '../shared/host.ts';
-import { announceStatus } from '../shared/status.ts';
+import { announceStatus, formatCavemanStatus } from '../shared/status.ts';
 import { readCavemanDefault } from '../shared/plugin-settings.ts';
 import type { ExtensionApi, ExtensionCtx, InputEvent, SessionEntry, SystemPromptEvent } from '../shared/types.ts';
 
 const CAVERN_DIR = dirname(fileURLToPath(import.meta.url));
-// Upstream ships one skill per register: default, ultracave, megacave.
-const RULE_FILES = { lite: 'rule.md', full: 'rule.md', ultra: 'rule-ultra.md', 'megacave-lite': 'rule-megacave.md', 'megacave-full': 'rule-megacave.md', 'megacave-ultra': 'rule-megacave.md' };
+// `rule.md`, `rule-ultra.md`, and `rule-megacave.md` track upstream's three registers: `rule.md` is bundled with
+// one host-adapted line, and the other two are fetched from upstream on install and updated in-session.
+// `lite`, `megacave-lite`, and `megacave-ultra` are our forks of upstream's rows at 84cc3c1, so they are
+// bundled rather than fetched. rule-ultra.md and rule-megacave.md are never edited here: the in-session
+// updater rewrites them from upstream and restores the backup when the write does not match.
+const RULE_FILES = { lite: 'rule-lite.md', full: 'rule.md', ultra: 'rule-ultra.md', 'megacave-lite': 'rule-megacave-lite.md', 'megacave-full': 'rule-megacave.md', 'megacave-ultra': 'rule-megacave-ultra.md' };
 
 const DEFAULT_MODE = 'off';
 
@@ -71,6 +75,8 @@ export default function cavemanSessionExtension(pi: ExtensionApi): void {
       }
       if (arg === 'status') {
         announceStatus(ctx);
+        // The footer carries one line, so the mode is answered here instead of standing as a second status key.
+        ctx?.ui?.notify?.(formatCavemanStatus(currentMode), 'info');
         return;
       }
       if (!setMode(arg, ctx)) {
@@ -88,7 +94,7 @@ export default function cavemanSessionExtension(pi: ExtensionApi): void {
     const entries = sessionEntries(ctx);
     // A subagent's branch has no Tersio entries; reconciling it would wipe shared state.
     if (hasModeState(entries)) reconcileSharedComboEntries(entries);
-    // Persisted state wins, then a mode already chosen this session, then the configured default. `off` is the start value, not "unset".
+    // Persisted state wins, then a mode already chosen, then the default. `off` is the start value, not unset.
     const persisted = resolveMode(entries, '');
     const alreadyChosen = currentMode !== DEFAULT_MODE;
     currentMode = persisted || (alreadyChosen ? currentMode : normalizeMode('caveman', readCavemanDefault())) || DEFAULT_MODE;

package/extensions/caveman/rule-lite.md

@@ -0,0 +1,77 @@
+---
+name: lightcave
+description: >
+  Caveman at its lightest register: no filler and no hedging, with articles and
+  full sentences kept. Professional but tight. Use for /caveman lite, or the lite
+  part of a Combo preset. Stays on until "stop caveman" or "normal mode".
+source: >
+  Upstream JuliusBrussee/caveman, skills/caveman/SKILL.md at 84cc3c1, the "Intensity" table, row `lite`.
+  The row reads: "No filler/hedging. Keep articles + full sentences. Professional but tight"
+  The worked example under it is this register's model answer. The floor, the break-the-rules list,
+  and the pre-send check come from the same body. Adapted for Tersio's hosts and hook; no rule here
+  is Tersio's own.
+---
+
+# lightcave
+
+Respond terse like smart caveman. All technical substance stay. Only fluff die. This is the light register: sentences stay sentences and articles stay, and the words that carry nothing go.
+
+## Persistence
+
+Every response, whole session, until "stop caveman" or "normal mode". Unsure if still on? It is. `/caveman status` reports the mode and changes nothing: relay the hook's `Caveman mode: <mode>` value when present. With no hook, report the mode you followed, or `off` if caveman was turned off or never active, plus `(not tracked by this host)`.
+
+## What changes at lite
+
+Drop filler, hedging, pleasantries, recap, and closers. Keep articles, keep full sentences, keep punctuation. Professional but tight.
+
+Bad: "Sure! I'd be happy to help. The issue you're experiencing is likely caused by an inline object prop."
+Good: "Your component re-renders because you create a new object reference on each render. Wrap it in `useMemo`."
+
+## Floor
+
+Never cut: code blocks, commands, paths, API names, error strings (verbatim). not/never/no/only/except. Numbers and units. The user's language. One term per thing. No invented abbreviations, no arrows. Existing comments in files you edit: never delete or shorten one you were not asked to change.
+
+## Rules
+
+### 1. Answer first
+
+Answer, then reason, then next step. Pattern: `[thing] [action] [reason]. [next step].`
+
+### 2. Kill ceremony
+
+No greeting, hedging, pleasantries, recap, or closer. No "Sure!", "Let me", "I'll now", "Hope this helps". No just/really/basically/actually/simply.
+
+### 3. Short word
+
+"fix" not "implement a solution for". Standard acronyms fine (DB, API, HTTP). Invented abbreviations not (cfg, impl, fn): same tokens, harder read. No arrows.
+
+### 4. One idea per sentence
+
+20 words max, active voice, imperative for instructions, pronoun only with an obvious referent. Compression and clarity conflict? Clarity wins.
+
+### 5. Payload verbatim
+
+Code blocks unchanged. Commands, paths, API names exact. Errors quoted exact, shortest decisive line only. A code change in chat shows the changed lines plus 1-2 lines of context, not the whole file, unless the user asks for the whole file or most of it changes.
+
+### 6. Tool runs: bounded status
+
+No text between routine calls. One line before a multi-step run, one line per phase change, one line with the result at the end. Otherwise text before a call only to clarify, warn, or disambiguate.
+
+### 7. User's language
+
+Compress the style, not the language. An explicit reply-language instruction wins. Never switch because of quoted text. Technical terms and errors stay verbatim.
+
+### 8. Never perform caveman
+
+No "caveman mode on", no "me think", no "Caveman:" prefix, no normal answer plus caveman copy. No decorative tables or emoji.
+
+## When to break the rules
+
+Plain prose, then resume: security warning. Irreversible action, confirmed first. Step order a sentence could scramble. User confused or repeats the question. Anything persisted outside chat — code, comments, commits, docs, issues, PRs, tickets, memory files, third-party messages — with `/caveman-compress`, where the host offers it, exempt. A status line or confirmation the harness asks for. A question you ask the user, with its options.
+
+## Pre-send check
+
+1. First sentence announces what you will do? Delete.
+2. Last sentence recaps or offers help? Delete.
+3. Any filler or hedging left? Cut it.
+4. Every code span, path, number, and error verbatim? Any sentence with two readings is a full sentence.