spexcode 0.4.1 → 0.4.3

package/spec-cli/src/pty-helper.mjs

@@ -0,0 +1,86 @@
+import * as pty from 'node-pty'
+import { execFileSync } from 'node:child_process'
+
+const [id, colsArg, rowsArg] = process.argv.slice(2)
+const cols = Number(colsArg)
+const rows = Number(rowsArg)
+const socket = process.env.SPEXCODE_TMUX || 'spexcode'
+
+if (!id || !(cols > 0 && rows > 0)) {
+  process.stderr.write('ERROR invalid helper arguments\n')
+  process.exit(2)
+}
+
+// Establish the outer-terminal contract before this process owns a PTY. tmux must preserve hyperlinks and
+// wrap client updates in DEC 2026 for xterm 6; doing the setup before forkpty preserves helper isolation.
+try {
+  // The dashboard renders the pane, not tmux's client chrome. Keeping this as a tmux session option makes
+  // the requested PTY grid the pane grid too; filtering a coloured final row in xterm would corrupt real
+  // pane content. SpexCode owns these sessions, so a later foreground attach sees the same status-free pane.
+  execFileSync('tmux', ['-L', socket, 'set-option', '-t', id, 'status', 'off'])
+  // Control follows the person actually driving: `latest` sizes the window to the most recently
+  // active client, so a small screen takes the grid the moment its user interacts and sees the WHOLE
+  // pane, while an idle larger viewer letterboxes. `largest` optimized the big display's real estate
+  // but left a small concurrent viewer a cropped corner it could never take back.
+  execFileSync('tmux', ['-L', socket, 'set-window-option', '-t', id, 'window-size', 'latest'])
+  execFileSync('tmux', ['-L', socket, 'set-option', '-g', 'mouse', 'on'])
+  execFileSync('tmux', ['-L', socket, 'set-option', '-g', 'history-limit', '50000'])
+  // Wheel routing stays tmux-default: copy-mode history for a plain pane, pass-through to a
+  // mouse-owning TUI. Claude's TUI virtual-scrolls its own transcript on those reports (it lives on
+  // the alternate screen and repaints in place, so tmux history cannot substitute — an in-place
+  // renderer never scrolls lines off the top). Restore the default bindings idempotently in case an
+  // earlier release rebound them server-wide; the unbind may already be clean, so it fails quietly.
+  execFileSync('tmux', ['-L', socket, 'bind-key', '-n', 'WheelUpPane', 'if', '-F', '-t=', '#{||:#{pane_in_mode},#{mouse_any_flag}}', 'send-keys -M', 'copy-mode -et='])
+  try { execFileSync('tmux', ['-L', socket, 'unbind-key', '-n', 'WheelDownPane']) } catch { /* already unbound */ }
+  const features = execFileSync('tmux', ['-L', socket, 'show-options', '-gsv', 'terminal-features'], { encoding: 'utf8' })
+  const xtermFeatures = new Set(features.split('\n').filter((line) => line.startsWith('xterm*:')).flatMap((line) => line.split(':').slice(1)))
+  const missing = ['sync', 'hyperlinks'].filter((feature) => !xtermFeatures.has(feature))
+  if (missing.length) {
+    execFileSync('tmux', ['-L', socket, 'set-option', '-as', 'terminal-features', `,xterm*:${missing.join(':')}`])
+  }
+} catch { /* attach below fails loudly if the tmux server/session is unavailable */ }
+
+const terminal = pty.spawn('tmux', ['-u', '-L', socket, 'attach-session', '-t', id], {
+  name: 'xterm-256color',
+  cols,
+  rows,
+  env: { ...process.env, LANG: process.env.LANG || 'en_US.UTF-8' },
+})
+
+terminal.onData((data) => process.stdout.write(Buffer.from(data, 'utf8')))
+terminal.onExit(({ exitCode }) => process.exit(exitCode === 0 ? 0 : 1))
+process.stderr.write(`READY ${terminal.pid}\n`)
+
+let input = ''
+process.stdin.setEncoding('utf8')
+process.stdin.on('data', (chunk) => {
+  input += chunk
+  let newline
+  while ((newline = input.indexOf('\n')) >= 0) {
+    const line = input.slice(0, newline)
+    input = input.slice(newline + 1)
+    if (!line) continue
+    try {
+      const message = JSON.parse(line)
+      if (message?.t === 'resize' && message.cols > 0 && message.rows > 0) {
+        const nextCols = Math.floor(message.cols)
+        const nextRows = Math.floor(message.rows)
+        terminal.resize(nextCols, nextRows)
+        process.stderr.write(`RESIZED ${nextCols} ${nextRows}\n`)
+      } else if (message?.t === 'input' && typeof message.data === 'string' && Buffer.byteLength(message.data, 'utf8') <= 64 * 1024) {
+        terminal.write(message.data)
+      }
+    } catch { /* malformed controls cannot become terminal input */ }
+  }
+})
+
+let closing = false
+function close() {
+  if (closing) return
+  closing = true
+  try { terminal.kill() } catch { /* already gone */ }
+  setTimeout(() => process.exit(0), 100).unref()
+}
+process.stdin.on('end', close)
+process.on('SIGTERM', close)
+process.on('SIGINT', close)

package/spec-cli/src/reaper.ts

@@ -1,11 +1,12 @@
-// @@@ socket-level connection reaper ([[spec-cli]]) - the ONE mechanism that actually reaps abandoned
-// sockets, because Node's own http.Server timeouts do not. MEASURED (eval server-reaps-abandoned-connections,
-// minimal http.createServer on Node 20/22/24): `headersTimeout` and `requestTimeout` DO NOT reap an
-// incomplete request — a slow-loris (TCP connect + partial headers, never completed) survives indefinitely
-// past the `connectionsCheckingInterval` sweep; only `keepAliveTimeout` (the idle-between-requests case) ever
-// fires. So the abandoned-connection pileup protection those options claim (the 135-conn starvation that
-// wedged the public port and triggered the mass-restore cascade) was NOT delivered. This helper is the real
-// mechanism: an explicit per-socket deadline at the server boundary, independent of the platform sweep.
+// @@@ socket-level connection reaper ([[spec-cli]]) - the ONE mechanism that reaps abandoned sockets (the
+// 135-conn starvation that wedged the public port and triggered the mass-restore cascade), and the SINGLE
+// OWNER of the deadlines it enforces. Node's own `headersTimeout`/`keepAliveTimeout` cover the same phases
+// (pre-request, idle-between-requests) and so are a second mechanism racing this one: MEASURED (eval
+// server-reaps-abandoned-connections, issue #65) a `headersTimeout: 20000` set beside the reaper won the
+// race on every reap at default config and silently capped SPEXCODE_REAP_HEADER_MS above 20s — the knob
+// went dead while the close still looked timely (Node's 408, not the reaper's destroy). So install()
+// DISABLES those overlapping Node timeouts on the server it guards; this helper is an explicit per-socket
+// deadline at the server boundary, with no platform machinery shadowing it.
 //
 // It keys on "no request has completed yet / idle between requests" — never on response DURATION — so a
 // long-lived ESTABLISHED stream (the /api/graph/stream SSE, a terminal WebSocket upgrade) is exempt for as
@@ -61,6 +62,13 @@ export function installConnectionReaper(server: HttpServer, opts: ReaperOptions
   const headerMs = resolveMs(opts.headerMs, process.env.SPEXCODE_REAP_HEADER_MS, 30000)
   const idleMs = resolveMs(opts.idleMs, process.env.SPEXCODE_REAP_IDLE_MS, 15000)
 
+  // claim single ownership of the phases the reaper covers (see header comment): Node's overlapping
+  // timeouts would otherwise race these deadlines and silently cap the env knobs. `requestTimeout` is
+  // deliberately LEFT at Node's default (~5 min): it bounds the in-flight request-body phase the reaper
+  // exempts (a silently-abandoned mid-body upload has no other reaper), and 5 min shadows no sane knob.
+  server.headersTimeout = 0
+  server.keepAliveTimeout = 0
+
   // per-socket tracking, on the SAME socket object 'request'/'upgrade' will report (see header comment).
   const track = (socket: Socket) => {
     const state: SocketState = {

package/spec-cli/src/reviewSnapshot.ts

@@ -0,0 +1,23 @@
+export type ReviewEvalNode = {
+  id: string
+  hue?: number
+  scenarios: any[]
+  evals: any[]
+  readings: any[]
+}
+
+export type ReviewSnapshot = {
+  issues: any[]
+  evalNodes: ReviewEvalNode[]
+}
+
+let current: ReviewSnapshot | null = null
+
+export function publishReviewSnapshot(snapshot: ReviewSnapshot): void {
+  current = snapshot
+}
+
+export function readReviewSnapshot(): ReviewSnapshot {
+  if (!current) throw new Error('review snapshot is unavailable before the first successful graph build')
+  return current
+}

package/spec-cli/src/reviews.ts

@@ -0,0 +1,347 @@
+import { createHash } from 'node:crypto'
+import { listSessions } from './sessions.js'
+import { getBoard } from './graphCache.js'
+import { buildSessionEvals, type SessionEvals } from '../../spec-eval/src/sessioneval.js'
+import { evalTimeline } from '../../spec-eval/src/evaltab.js'
+import { issuesEnabled as issuesEnabledForReview } from './localIssues.js'
+import { issueStores as issueStoresForReview } from './issues.js'
+import { readReviewSnapshot } from './reviewSnapshot.js'
+// @ts-expect-error The dashboard module is deliberately plain JS so the browser and server execute the
+// exact same tokenizer/matcher. It is shipped beside the built dashboard by the root package manifest.
+import { EVAL_FILTER_KIND, evalFilterModel, evalReviewState, issueFilterModel, tokenFilterState } from '../../spec-dashboard/src/reviewFilters.js'
+// @ts-expect-error See the shared-domain note above.
+import { EVAL_QUERY_DEFAULT, ISSUE_QUERY_DEFAULT, readToken } from '../../spec-dashboard/src/reviewQuery.js'
+
+export const REVIEW_PER_PAGE = 25
+
+type ReviewItem = Record<string, unknown>
+type ReviewOption = { value: string; label?: string; count?: number }
+type ReviewFacet = { key: string; label?: string; value: string; meaningful?: boolean; options: ReviewOption[] }
+type EvalNeighbor = { node: string; scenario: string; state: string }
+
+export type PagedReview<T extends ReviewItem = ReviewItem> = {
+  items: T[]
+  page: number
+  perPage: number
+  total: number
+  sourceTotal: number
+  pageCount: number
+  prev: number | null
+  next: number | null
+  revision: string
+  counts: Record<string, number>
+  facets: Record<string, ReviewFacet>
+  section: { key: string; value: string; options: ReviewOption[] } | null
+}
+
+export type EvalDetailReview = {
+  scope: string | null
+  selected: ReviewItem | null
+  history: ReviewItem[]
+  neighbors: {
+    prev: EvalNeighbor[]
+    next: EvalNeighbor[]
+    total: number
+    index: number | null
+    order: 'default'
+  }
+  revision: string
+  summary?: SessionEvals['summary']
+  evalRevision?: SessionEvals['evalRevision']
+}
+
+const revisionOf = (value: unknown): string => createHash('sha256').update(JSON.stringify(value)).digest('hex')
+
+export function reviewPageNumber(value: unknown): number {
+  const raw = typeof value === 'string' && /^\d+$/.test(value) ? Number(value) : Number.NaN
+  return Number.isSafeInteger(raw) && raw > 0 ? raw : 1
+}
+
+function responseModel(model: any): Pick<PagedReview, 'counts' | 'facets' | 'section'> {
+  const facets = Object.fromEntries(Object.entries(model.facets ?? {}).map(([key, raw]) => {
+    const facet = raw as any
+    return [key, {
+      key,
+      ...(facet.label != null ? { label: String(facet.label) } : {}),
+      value: String(facet.value ?? ''),
+      ...(facet.meaningful != null ? { meaningful: !!facet.meaningful } : {}),
+      options: (facet.options ?? []).map((option: any) => ({
+        value: String(option.value ?? ''),
+        ...(option.label != null ? { label: String(option.label) } : {}),
+        ...(Number.isFinite(option.count) ? { count: Number(option.count) } : {}),
+      })),
+    }]
+  })) as Record<string, ReviewFacet>
+  const section = model.section ? {
+    key: String(model.section.key),
+    ...(model.section.label != null ? { label: String(model.section.label) } : {}),
+    value: String(model.section.value ?? ''),
+    ...(model.section.meaningful != null ? { meaningful: !!model.section.meaningful } : {}),
+    options: (model.section.options ?? []).map((option: any) => ({
+      value: String(option.value ?? ''),
+      ...(option.label != null ? { label: String(option.label) } : {}),
+      ...(Number.isFinite(option.count) ? { count: Number(option.count) } : {}),
+    })),
+  } : null
+  return { counts: { ...(model.sections ?? {}) }, facets, section }
+}
+
+export function paginateReview<T extends ReviewItem>(
+  source: T[],
+  shown: T[],
+  model: any,
+  requestedPage: unknown,
+  revisionInputs: unknown,
+): PagedReview<T> {
+  const page = reviewPageNumber(requestedPage)
+  const total = shown.length
+  const pageCount = Math.ceil(total / REVIEW_PER_PAGE)
+  const start = (page - 1) * REVIEW_PER_PAGE
+  const overflow = pageCount > 0 ? page > pageCount : page > 1
+  return {
+    items: shown.slice(start, start + REVIEW_PER_PAGE),
+    page,
+    perPage: REVIEW_PER_PAGE,
+    total,
+    sourceTotal: source.length,
+    pageCount,
+    prev: page > 1 ? page - 1 : null,
+    next: page < pageCount || overflow ? page + 1 : null,
+    revision: revisionOf(revisionInputs),
+    ...responseModel(model),
+  }
+}
+
+const issueOrder = (a: any, b: any): number => String(b.created ?? '').localeCompare(String(a.created ?? ''))
+  || String(a.id ?? '').localeCompare(String(b.id ?? ''))
+
+export async function issuesReview(query: string | undefined, requestedPage: unknown) {
+  const [, sessions] = await Promise.all([getBoard(), listSessions()])
+  const issues = readReviewSnapshot().issues.slice().sort(issueOrder)
+  const text = String(query ?? '').trim() || ISSUE_QUERY_DEFAULT
+  const model = issueFilterModel(issues, tokenFilterState(text, 'issue'), { sessions, defaultSection: '' })
+  return {
+    enabled: issuesEnabledForReview(),
+    stores: issueStoresForReview(),
+    ...paginateReview(issues, model.shown, model, requestedPage, {
+      domain: 'issues', issues, sessions: sessions.map((session) => session.id),
+    }),
+  }
+}
+
+const byNewest = (a: any, b: any): number => String(b.ts ?? '').localeCompare(String(a.ts ?? ''))
+  || String(a.node ?? '').localeCompare(String(b.node ?? ''))
+  || String(a.scenario ?? '').localeCompare(String(b.scenario ?? ''))
+
+export function trunkEvalReviewItems(nodes: any[]): ReviewItem[] {
+  const blind: any[] = []
+  const items: any[] = []
+  for (const node of nodes ?? []) {
+    const latest = new Map<string, any>()
+    for (const reading of node.evals ?? []) if (!latest.has(reading.scenario)) latest.set(reading.scenario, reading)
+    for (const scenario of node.scenarios ?? []) {
+      const reading = latest.get(scenario.name)
+      if (!reading) {
+        blind.push({
+          scenario: scenario.name,
+          expected: scenario.expected,
+          tags: scenario.tags,
+          node: node.id,
+          hue: node.hue,
+          filterKind: EVAL_FILTER_KIND.BLIND,
+        })
+        continue
+      }
+      items.push({
+        ...reading,
+        expected: scenario.expected ?? reading.expected,
+        tags: scenario.tags,
+        state: evalReviewState(reading),
+        node: node.id,
+        hue: node.hue,
+        filterKind: EVAL_FILTER_KIND.RESULT,
+      })
+    }
+  }
+  blind.sort((a, b) => String(a.node).localeCompare(String(b.node)) || String(a.scenario).localeCompare(String(b.scenario)))
+  return [...blind, ...items.sort(byNewest)]
+}
+
+export function scopedEvalReviewItems(model: SessionEvals): ReviewItem[] {
+  const blind: any[] = []
+  const own: any[] = []
+  const inherited: any[] = []
+  for (const node of model.nodes ?? []) {
+    const latest = new Map<string, any>()
+    for (const reading of node.evals ?? []) if (!latest.has(reading.scenario)) latest.set(reading.scenario, reading)
+    for (const scenario of node.scenarios ?? []) {
+      const reading = latest.get(scenario.name)
+      if (!reading) {
+        blind.push({
+          scenario: scenario.name,
+          expected: scenario.expected,
+          tags: scenario.tags,
+          impact: scenario.impact,
+          node: node.id,
+          hue: node.hue,
+          filterKind: EVAL_FILTER_KIND.BLIND,
+        })
+        continue
+      }
+      const item = {
+        ...reading,
+        expected: scenario.expected ?? reading.expected,
+        tags: scenario.tags,
+        state: evalReviewState(reading),
+        node: node.id,
+        hue: node.hue,
+        filterKind: EVAL_FILTER_KIND.RESULT,
+      }
+      ;(reading.inSession ? own : inherited).push(item)
+    }
+  }
+  return [...blind, ...own.sort(byNewest), ...inherited.sort(byNewest)]
+}
+
+const evalItemKey = (item: any): string => `${String(item?.node ?? '')}\0${String(item?.scenario ?? '')}`
+
+function evalNeighbor(item: any): EvalNeighbor {
+  return {
+    node: String(item.node),
+    scenario: String(item.scenario),
+    state: String(item.state ?? evalReviewState(item)),
+  }
+}
+
+export function boundedEvalNeighbors(items: ReviewItem[], node: string, scenario: string, want = 5) {
+  const key = `${node}\0${scenario}`
+  const index = items.findIndex((item) => evalItemKey(item) === key)
+  if (index < 0) return { prev: [], next: [], total: items.length, index: null, order: 'default' as const }
+  const before = index
+  const after = items.length - index - 1
+  const take = Math.min(want, before + after)
+  const nextN = Math.min(after, Math.max(Math.ceil(take / 2), take - before))
+  const prevN = Math.min(before, take - nextN)
+  return {
+    prev: items.slice(index - prevN, index).reverse().map(evalNeighbor),
+    next: items.slice(index + 1, index + 1 + nextN).map(evalNeighbor),
+    total: items.length,
+    index,
+    order: 'default' as const,
+  }
+}
+
+export function projectEvalDetail(
+  items: ReviewItem[],
+  historySource: ReviewItem[],
+  node: string,
+  scenario: string,
+  metadata: { scope?: string | null; summary?: SessionEvals['summary']; evalRevision?: SessionEvals['evalRevision'] } = {},
+): EvalDetailReview {
+  const results = items.filter((item: any) => item.filterKind === EVAL_FILTER_KIND.RESULT)
+  const selected = results.find((item) => evalItemKey(item) === `${node}\0${scenario}`) ?? null
+  const history = historySource.filter((reading: any) => String(reading.scenario) === scenario)
+  const neighbors = boundedEvalNeighbors(results, node, scenario)
+  const scope = metadata.scope ?? null
+  return {
+    scope,
+    selected,
+    history,
+    neighbors,
+    revision: revisionOf({ scope, selected, history, neighbors, summary: metadata.summary, evalRevision: metadata.evalRevision }),
+    ...(metadata.summary ? { summary: metadata.summary } : {}),
+    ...(metadata.evalRevision ? { evalRevision: metadata.evalRevision } : {}),
+  }
+}
+
+export async function evalDetailReview(node: string, scenario: string, scope?: string | null): Promise<EvalDetailReview | null> {
+  if (scope) {
+    const model = await buildSessionEvals(scope)
+    if (!model) return null
+    const sourceNode = model.nodes.find((candidate) => candidate.id === node)
+    return projectEvalDetail(scopedEvalReviewItems(model), sourceNode?.evals ?? [], node, scenario, {
+      scope,
+      summary: model.summary,
+      evalRevision: model.evalRevision,
+    })
+  }
+  await getBoard()
+  const snapshot = readReviewSnapshot()
+  const sourceNode = snapshot.evalNodes.find((candidate) => candidate.id === node)
+  return projectEvalDetail(trunkEvalReviewItems(snapshot.evalNodes), sourceNode?.readings ?? [], node, scenario)
+}
+
+async function timelineEvalReview(text: string, requestedPage: unknown) {
+  const node = readToken(text, 'node')
+  if (!node) return null
+  const [timeline, sessions] = await Promise.all([evalTimeline(node), listSessions()])
+  const measured = new Set(timeline.readings.map((reading) => reading.scenario))
+  const items = [
+    ...timeline.scenarios.filter((scenario) => !measured.has(scenario.name)).map((scenario) => ({
+      ...scenario,
+      scenario: scenario.name,
+      node,
+      filterKind: EVAL_FILTER_KIND.UNMEASURED,
+    })),
+    ...timeline.readings.map((reading, index) => ({
+      ...reading,
+      state: evalReviewState(reading),
+      node,
+      filterKind: EVAL_FILTER_KIND.RESULT,
+      filterKey: `${EVAL_FILTER_KIND.RESULT}:${index}`,
+    })),
+    ...(timeline.dangling ?? []).map((track) => ({
+      ...track,
+      node,
+      filterKind: EVAL_FILTER_KIND.DANGLING,
+      filterKey: `${EVAL_FILTER_KIND.DANGLING}:${track.threadId}`,
+    })),
+  ]
+  const filtered = evalFilterModel(items, tokenFilterState(text, 'eval'), { sessions, defaultKind: 'all', defaultSection: '' })
+  return {
+    scope: null,
+    view: 'timeline',
+    node,
+    hasEvalFile: timeline.hasEvalFile,
+    gates: [],
+    unknown: 0,
+    ...paginateReview(items, filtered.shown, filtered, requestedPage, {
+      domain: 'evals', view: 'timeline', node, timeline, sessions: sessions.map((session) => session.id),
+    }),
+  }
+}
+
+export async function evalsReview(query: string | undefined, requestedPage: unknown, options: { view?: string } = {}) {
+  const text = String(query ?? '').trim() || EVAL_QUERY_DEFAULT
+  if (options.view === 'timeline') return timelineEvalReview(text, requestedPage)
+  const scope = readToken(text, 'scope') || null
+  if (scope) {
+    const model = await buildSessionEvals(scope)
+    if (!model) return null
+    const items = scopedEvalReviewItems(model)
+    const sessions = await listSessions()
+    const filtered = evalFilterModel(items, tokenFilterState(text, 'eval'), { sessions, defaultKind: 'all', defaultSection: '' })
+    return {
+      scope,
+      gates: model.gates,
+      unknown: model.nodes.reduce((count, node) => count + (node.unknownCoverage?.length ?? 0), 0),
+      summary: model.summary,
+      evalRevision: model.evalRevision,
+      ...paginateReview(items, filtered.shown, filtered, requestedPage, {
+        domain: 'evals', scope, items, gates: model.gates, summary: model.summary,
+        evalRevision: model.evalRevision, sessions: sessions.map((session) => session.id),
+      }),
+    }
+  }
+  const board = await getBoard()
+  const items = trunkEvalReviewItems(readReviewSnapshot().evalNodes)
+  const filtered = evalFilterModel(items, tokenFilterState(text, 'eval'), { sessions: board.sessions, defaultKind: 'all', defaultSection: '' })
+  return {
+    scope: null,
+    gates: [],
+    unknown: 0,
+    ...paginateReview(items, filtered.shown, filtered, requestedPage, {
+      domain: 'evals', items, sessions: board.sessions.map((session) => session.id),
+    }),
+  }
+}

package/spec-cli/src/session-timeline.ts

@@ -118,6 +118,20 @@ export function superviseTimeline(): void {
   scan()
 }
 
+// the channel of the LAST HUMAN send (from == null): 'note' when the note-reply hint rode along, else null.
+// This is what makes the reply-channel hints SYMMETRIC ([[sessions-core]] sendText): a human send with no
+// note flag arriving after a note-send is the "back at a terminal" transition, and the delivery gets the
+// counter-insert. Derived from the durable log — no new state, and it survives a server restart. Agent
+// senders (`from` set) say nothing about where the HUMAN is reading, so they neither set nor clear it.
+export function lastHumanSendVia(id: string): 'note' | null {
+  const evs = readEvents(id)
+  for (let i = evs.length - 1; i >= 0; i--) {
+    const e = evs[i]
+    if (e.kind === 'sent' && e.from == null) return e.replyVia === 'note' ? 'note' : null
+  }
+  return null
+}
+
 // record a CONFIRMED prompt delivery (called by sendText after the harness accepted it). `text` is the
 // caller's message BEFORE any mechanism insert (the note-reply hint is transport, not conversation);
 // `replyVia` marks that the hint rode along so a surface can badge it.