@agfpd/iapeer 0.2.56 → 0.2.58

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@agfpd/iapeer",
-  "version": "0.2.56",
+  "version": "0.2.58",
   "description": "Foundation core for the iapeer multi-agent ecosystem: identity, registry, storage, codec.",
   "type": "module",
   "bin": {
@@ -33,6 +33,7 @@
   },
   "dependencies": {
     "@modelcontextprotocol/sdk": "1.29.0",
+    "@xterm/headless": "6.0.0",
     "proper-lockfile": "^4.1.2",
     "transliteration": "^2.6.1"
   },

package/src/cli/index.ts

@@ -936,6 +936,7 @@ const USAGE = `usage: iapeer <verb> [args]
   self-done                                                      (agent self-call, ephemeral) silent finish: arm own quiet-reap, wake no one
   native-memory <off|on> (--peer <p> | --all)                    gate/restore runtimes' native memory (canonized lever; контракт «Слот памяти»)
   trust-hooks <hooks.json> [--check]                             pre-seed codex hooks trust for a file-form hooks.json (no modal); --check = drift report
+  shadow [--once | --minutes <m>] [--interval <ms>]              READ-ONLY tmux→pty fidelity observer (migration burn-in): compares pty-model verdicts vs live tmux on the warm fleet → ~/.iapeer/logs/iapeer/shadow-fidelity.json. Never delivers/mutates.
 `
 
 export async function runCli(argv: string[], env: NodeJS.ProcessEnv = process.env): Promise<number> {
@@ -1750,6 +1751,27 @@ export async function runCli(argv: string[], env: NodeJS.ProcessEnv = process.en
         appendLifecycleEvent(attachCfg.eventLogDir, { ev: 'attach-end', identity: r.identity, rc: a.status ?? -1 }, { env })
         return a.status ?? 0
       }
+      case 'shadow': {
+        // Read-only fidelity OBSERVER for the tmux→pty migration burn-in (track «b»). Compares
+        // the pty-model verdicts (occupancy/ready-gate/liveness) vs live tmux verdicts on the
+        // warm fleet, logs to <eventLogDir>/shadow-fidelity.{jsonl,json}. STRICTLY read-only —
+        // nothing in delivery/lifecycle imports the shadow module; the dynamic import keeps
+        // @xterm OUT of the hot path (the daemon never loads it). --once | --minutes <m> |
+        // --interval <ms>; stop a long run with `touch <eventLogDir>/shadow-fidelity.STOP`.
+        const cfg = loadLifecycleConfig(env)
+        const { runShadowFidelity } = await import('../shadow/index.ts')
+        await runShadowFidelity({
+          logDir: cfg.logDir,
+          eventLogDir: cfg.eventLogDir,
+          sockDir: cfg.sockDir,
+          intervalMs: flags.interval ? Number(flags.interval) : 5000,
+          maxMinutes: flags.minutes ? Number(flags.minutes) : 0,
+          once: flags.once === true,
+          log: errOut,
+        })
+        out(`shadow-fidelity → ${cfg.eventLogDir}/shadow-fidelity.json\n`)
+        return 0
+      }
       default: {
         // `iapeer <runtime>` (launch) — folder-launch the cwd's peer, ALWAYS fresh.
         if (verb && isRuntime(verb)) {

package/src/lifecycle/index.ts

@@ -1001,6 +1001,22 @@ export async function wakeOrSpawn(args: WakeArgs, deps: WakeDeps = {}): Promise<
     const firstMessage = composeFirstMessage(cwd, args.task, fresh)
     const result = await launch(spec, adapter, firstMessage, launchCfg)
     if (result.status === 'FAILED') {
+      // enroll-on-FAILED (0.2.57): launch can return FAILED on a ready-gate false-negative
+      // (model-did-not-process-task) while the tmux session + its process actually SURVIVE —
+      // observed live (ttlverify, 13.06). Since 0.2.55 removed the self-TTL, such a remnant
+      // would be an UNBOUNDED orphan: no .session → invisible to superviseTick's idle-reap,
+      // and no wall-clock backstop. If the session is in fact ALIVE, enroll it so the
+      // activity-aware idle-reap stays the UNIVERSAL bound (a genuinely wedged session is
+      // idle-reaped after idleSecs; a false-negative-but-working one keeps working and is
+      // reaped only when truly idle, and its next message reaches it via the has-session fast
+      // path above). We do NOT flip to READY — the wake DID fail its ready-gate, the caller
+      // must still hear FAILED; we only make the live remnant supervised. Best-effort: a write
+      // hiccup must not mask the original FAILED reason.
+      if (sessionAlive(sock, identity)) {
+        try {
+          writeSessionState(cfg, { identity, runtime, personality: args.personality, cwd, wokeAt: Date.now() })
+        } catch { /* keep the original FAILED reason — enroll is best-effort */ }
+      }
       return { status: 'FAILED', woke: false, runtime, reason: result.reason }
     }
     // The session is up and the message delivered. Recording supervise-state must not

package/src/shadow/index.ts

@@ -0,0 +1,170 @@
+// Read-only fidelity OBSERVER for the tmux→pty migration (prod burn-in track «b»).
+//
+// For each warm claude/codex peer it builds an @xterm/headless model from the daemon's
+// EXISTING pipe-pane log (`<identity>.log` under cfg.logDir) — the same raw child bytes
+// tmux's emulator gets — and compares the PTY-model verdicts (composer-occupancy / ready-gate
+// / liveness) to the live AUTHORITATIVE tmux verdicts, logging divergences + rolling fidelity
+// to `<eventLogDir>/shadow-fidelity.{jsonl,json}`.
+//
+// STRICTLY READ-ONLY: it tails log FILES and runs read-only tmux (`capture-pane`/`display`/
+// `has-session`). It NEVER delivers, sends keys, spawns, or touches delivery/lifecycle/
+// registry/launch. Nothing in the delivery/lifecycle hot paths imports this module — it runs
+// ONLY via the `iapeer shadow` CLI verb (dynamic-imported, so @xterm is never loaded in the
+// daemon). With it off, the daemon's behavior is byte-identical. This is the prod burn-in:
+// the supervisor's model proven alongside tmux on real traffic BEFORE any delivery cutover.
+import { Terminal } from '@xterm/headless'
+import { openSync, closeSync, fstatSync, readSync, existsSync, appendFileSync, writeFileSync, mkdirSync } from 'node:fs'
+import { join } from 'node:path'
+import { tmux, composerCaptureHasHumanInput, listOnlinePeers } from '../transport/index.ts'
+import { getAdapter } from '../launch/index.ts'
+import { buildSocketPath, buildProcessAddress } from '../core/socket.ts'
+import type { Runtime } from '../core/constants.ts'
+
+const SEED_BYTES = 4 * 1024 * 1024
+const promptGlyph = (rt: string): string => (rt === 'claude' ? '❯' : '›')
+const pidAlive = (pid: number): boolean => {
+  try { process.kill(pid, 0); return true } catch (e) { return (e as NodeJS.ErrnoException).code === 'EPERM' }
+}
+
+/** PTY composer-occupancy verdict: getCell on the model's viewport composer row (the row whose
+ *  first visible cell is the prompt glyph); human input = a non-ghost (non-dim, non-grey246)
+ *  visible cell after it. Mirrors composerCaptureHasHumanInput — validated 0-divergence vs it. */
+export function composerOccupancyFromModel(term: Terminal, cols: number, rows: number, runtime: string): boolean {
+  const b = term.buffer.active, top = b.baseY, bot = b.baseY + rows, g = promptGlyph(runtime)
+  let py = -1, px = -1
+  for (let y = top; y < bot; y++) {
+    const l = b.getLine(y); if (!l) continue
+    let fx = -1
+    for (let x = 0; x < cols; x++) { if ((l.getCell(x)?.getChars() || ' ').trim() !== '') { fx = x; break } }
+    if (fx < 0) continue
+    if (l.getCell(fx)!.getChars() === g) { py = y; px = fx }
+  }
+  if (py < 0) return false
+  const l = b.getLine(py)!
+  for (let x = px + 1; x < cols; x++) {
+    const c = l.getCell(x); if ((c?.getChars() || ' ').trim() === '') continue
+    if (!(c!.isDim() || (c!.isFgPalette() && c!.getFgColor() === 246))) return true
+  }
+  return false
+}
+
+/** The model's viewport as PLAIN text (mimics `capture-pane -p`, the input isInputReady takes). */
+export function modelToPlainText(term: Terminal, cols: number, rows: number): string {
+  const b = term.buffer.active, lines: string[] = []
+  for (let y = b.baseY; y < b.baseY + rows; y++) {
+    const l = b.getLine(y); if (!l) { lines.push(''); continue }
+    let s = ''
+    for (let x = 0; x < cols; x++) s += l.getCell(x)?.getChars() || ' '
+    lines.push(s.replace(/\s+$/, ''))
+  }
+  return lines.join('\n')
+}
+
+interface Obs { addr: string; sock: string; runtime: Runtime; cols: number; rows: number; term: Terminal; fd: number; offset: number }
+
+export interface ShadowOptions {
+  logDir: string        // daemon pane-logs live here (<identity>.log)
+  eventLogDir: string   // shadow-fidelity.{jsonl,json} written here
+  sockDir: string
+  intervalMs?: number
+  maxMinutes?: number   // 0/undefined + once=false → run until <eventLogDir>/shadow-fidelity.STOP
+  once?: boolean
+  log?: (s: string) => void
+}
+
+function feed(o: Obs): Promise<void> {
+  const sz = fstatSync(o.fd).size
+  if (sz <= o.offset) return Promise.resolve()
+  const buf = Buffer.alloc(sz - o.offset)
+  readSync(o.fd, buf, 0, buf.length, o.offset)
+  o.offset = sz
+  return new Promise(res => o.term.write(buf, () => res()))
+}
+
+function enumerate(models: Map<string, Obs>, opts: ShadowOptions): void {
+  let peers: { runtime: Runtime; personality: string }[] = []
+  try { peers = listOnlinePeers(opts.sockDir).filter(p => p.runtime === 'claude' || p.runtime === 'codex') } catch { return }
+  const live = new Set<string>()
+  for (const p of peers) {
+    const addr = buildProcessAddress(p.runtime, p.personality); live.add(addr)
+    if (models.has(addr)) continue
+    try {
+      const sock = buildSocketPath(p.runtime, p.personality, opts.sockDir)
+      const log = join(opts.logDir, `${addr}.log`); if (!existsSync(log)) continue
+      const geo = tmux(sock, 'display', '-p', '-t', addr, '#{pane_width}x#{pane_height}'); if (!geo.ok) continue
+      const [cols, rows] = geo.out.trim().split('x').map(Number); if (!cols || !rows) continue
+      const term = new Terminal({ cols, rows, allowProposedApi: true, scrollback: 3000 })
+      const fd = openSync(log, 'r'); const offset = Math.max(0, fstatSync(fd).size - SEED_BYTES)
+      models.set(addr, { addr, sock, runtime: p.runtime, cols, rows, term, fd, offset })
+    } catch { /* a peer racing spawn/teardown — skip this round */ }
+  }
+  for (const [addr, o] of models) if (!live.has(addr)) { try { closeSync(o.fd) } catch { /* */ }; models.delete(addr) }
+}
+
+/** Run the read-only fidelity observer. Resolves when once/maxMinutes/STOP completes. */
+export async function runShadowFidelity(opts: ShadowOptions): Promise<void> {
+  mkdirSync(opts.eventLogDir, { recursive: true })
+  const jsonl = join(opts.eventLogDir, 'shadow-fidelity.jsonl')
+  const statsFile = join(opts.eventLogDir, 'shadow-fidelity.json')
+  const stopFile = join(opts.eventLogDir, 'shadow-fidelity.STOP')
+  const interval = opts.intervalMs ?? 5000
+  const log = opts.log ?? (() => { /* */ })
+  const models = new Map<string, Obs>()
+  const tot = { samples: 0, occDiv: 0, occNt: 0, rgDiv: 0, livDiv: 0 }
+  const startedISO = new Date().toISOString(), t0 = Date.now()
+  enumerate(models, opts)
+  for (const o of models.values()) await feed(o)  // seed each model
+  log(`shadow-fidelity: ${models.size} warm peers, interval ${interval}ms\n`)
+  let i = 0
+  for (;;) {
+    if (opts.maxMinutes && Date.now() - t0 >= opts.maxMinutes * 60000) break
+    if (existsSync(stopFile)) break
+    if (i > 0 && i % 15 === 0) enumerate(models, opts)
+    for (const o of models.values()) {
+      try {
+        await feed(o)
+        const adapter = getAdapter(o.runtime)
+        const capE = tmux(o.sock, 'capture-pane', '-e', '-p', '-t', o.addr); if (!capE.ok) continue
+        const occTmux = composerCaptureHasHumanInput(capE.out, adapter.deliveryMarkers)
+        const occPty = composerOccupancyFromModel(o.term, o.cols, o.rows, o.runtime)
+        const capP = tmux(o.sock, 'capture-pane', '-p', '-t', o.addr)
+        const rgTmux = capP.ok ? adapter.isInputReady(capP.out) : null
+        const rgPty = adapter.isInputReady(modelToPlainText(o.term, o.cols, o.rows))
+        const livTmux = tmux(o.sock, 'has-session', '-t', o.addr).ok
+        const ppid = tmux(o.sock, 'display', '-p', '-t', o.addr, '#{pane_pid}')
+        const pid = ppid.ok ? Number(ppid.out.trim()) : NaN
+        const livPty = Number.isFinite(pid) ? pidAlive(pid) : false
+
+        tot.samples++
+        const occDiv = occPty !== occTmux, rgDiv = rgTmux !== null && rgPty !== rgTmux, livDiv = livPty !== livTmux
+        const occNt = occPty || occTmux
+        if (occNt) tot.occNt++
+        if (occDiv) tot.occDiv++
+        if (rgDiv) tot.rgDiv++
+        if (livDiv) tot.livDiv++
+        if (occDiv || rgDiv || livDiv) {
+          appendFileSync(jsonl, JSON.stringify({
+            ts: new Date().toISOString(), peer: o.addr, geom: `${o.cols}x${o.rows}`,
+            occupancy: { pty: occPty, tmux: occTmux, diverged: occDiv, nontrivial: occNt },
+            readyGate: { pty: rgPty, tmux: rgTmux, diverged: rgDiv },
+            liveness: { pty: livPty, tmux: livTmux, diverged: livDiv },
+            ...(occDiv ? { capE: capE.out.slice(-2500), model: modelToPlainText(o.term, o.cols, o.rows).slice(-2500) } : {}),
+          }) + '\n')
+          log(`DIVERGE ${o.addr} occ=${occDiv} rg=${rgDiv} liv=${livDiv}${occNt && occDiv ? ' *** NON-TRIVIAL occ ***' : ''}\n`)
+        }
+      } catch { /* an observer error must NEVER escape — it is a passive sidecar */ }
+    }
+    i++
+    writeFileSync(statsFile, JSON.stringify({
+      startedISO, lastISO: new Date().toISOString(), elapsedMin: +((Date.now() - t0) / 60000).toFixed(1),
+      peers: models.size, samples: tot.samples,
+      occupancy: { divergences: tot.occDiv, nonTrivialSamples: tot.occNt },
+      readyGate: { divergences: tot.rgDiv },
+      liveness: { divergences: tot.livDiv },
+      fidelity: tot.samples ? +(1 - (tot.occDiv + tot.rgDiv + tot.livDiv) / (tot.samples * 3)).toFixed(6) : 1,
+    }, null, 2))
+    if (opts.once) break
+    await new Promise(r => setTimeout(r, interval))
+  }
+  for (const o of models.values()) { try { closeSync(o.fd) } catch { /* */ } }
+}

package/src/shadow/shadow.test.ts

@@ -0,0 +1,42 @@
+import { describe, expect, test } from 'bun:test'
+import { Terminal } from '@xterm/headless'
+import { composerOccupancyFromModel, modelToPlainText } from './index.ts'
+
+// The observer's pty-side verdicts are pure functions of the @xterm model — hermetic, no tmux.
+// These mirror the verdicts validated 0-divergence vs the prod tmux verdicts in the migration
+// fidelity shadows (occupancy: getCell dim/ghost; ready-gate: isInputReady on plain text).
+const build = (frame: string, cols = 80, rows = 24): Promise<Terminal> =>
+  new Promise(res => { const t = new Terminal({ cols, rows, allowProposedApi: true, scrollback: 200 }); t.write(frame, () => res(t)) })
+
+describe('composerOccupancyFromModel (pty occupancy verdict)', () => {
+  test('human non-dim text after the prompt glyph → TRUE', async () => {
+    const t = await build(`\x1b[24;1H❯ the quick brown fox`)
+    expect(composerOccupancyFromModel(t, 80, 24, 'claude')).toBe(true)
+  })
+  test('dim/ghost placeholder after the glyph → FALSE (the critical discriminator)', async () => {
+    const t = await build(`\x1b[24;1H❯ \x1b[2m\x1b[38;5;246mTry "fix typecheck errors"\x1b[0m`)
+    expect(composerOccupancyFromModel(t, 80, 24, 'claude')).toBe(false)
+  })
+  test('empty composer (glyph alone) → FALSE', async () => {
+    const t = await build(`\x1b[24;1H❯ `)
+    expect(composerOccupancyFromModel(t, 80, 24, 'claude')).toBe(false)
+  })
+  test('codex glyph › with human text → TRUE', async () => {
+    const t = await build(`\x1b[24;1H› reviewer note`)
+    expect(composerOccupancyFromModel(t, 80, 24, 'codex')).toBe(true)
+  })
+  test('no prompt glyph at all → FALSE', async () => {
+    const t = await build(`\x1b[24;1Hjust some output line`)
+    expect(composerOccupancyFromModel(t, 80, 24, 'claude')).toBe(false)
+  })
+})
+
+describe('modelToPlainText (feeds isInputReady)', () => {
+  test('renders the viewport as plain text (no SGR), trailing-trimmed', async () => {
+    const t = await build(`\x1b[1;1Hhello\x1b[2;1H\x1b[1mbold\x1b[0m world`)
+    const plain = modelToPlainText(t, 80, 24)
+    expect(plain.split('\n')[0]).toBe('hello')
+    expect(plain).toContain('bold world')
+    expect(plain).not.toContain('\x1b')
+  })
+})