switchroom 0.11.1 → 0.12.1

package/telegram-plugin/tests/html-balanced.ts

@@ -1,63 +0,0 @@
-/**
- * Tiny HTML tag-balance validator for the Telegram allowlist.
- *
- * Replaces fast-check (not a current dep, see #662 P1) — paired with
- * vitest `it.each` it gives us property-style coverage across many
- * randomised renderer inputs without a new dependency.
- *
- * Telegram's HTML parser only accepts a small allowlist of tags
- * (https://core.telegram.org/bots/api#html-style). Anything outside
- * the allowlist is treated as raw text — we don't try to validate it.
- *
- * Self-closing tags inside Telegram's allowlist: only `<br/>` (and even
- * that is normalised to `\n`). We treat any `<x/>` as self-closing for
- * robustness, but emit nothing into the stack.
- */
-
-const ALLOWED = new Set([
-  'b', 'strong',
-  'i', 'em',
-  'u', 'ins',
-  's', 'strike', 'del',
-  'a',
-  'code',
-  'pre',
-  'tg-spoiler',
-  'span',
-  'tg-emoji',
-  'blockquote',
-  'br',
-])
-
-export interface BalanceResult {
-  balanced: boolean
-  openTags: string[]
-  extraCloses: string[]
-}
-
-export function isBalancedHtml(html: string): BalanceResult {
-  const stack: string[] = []
-  const extraCloses: string[] = []
-  // Match tags but skip HTML entities (`&lt;` etc) — those are NOT tags.
-  const re = /<\/?\s*([A-Za-z][A-Za-z0-9-]*)([^>]*)>/g
-  let m: RegExpExecArray | null
-  while ((m = re.exec(html)) !== null) {
-    const raw = m[0]
-    const name = m[1].toLowerCase()
-    if (!ALLOWED.has(name)) continue
-    const isClose = raw.startsWith('</')
-    const selfClose = raw.endsWith('/>') || name === 'br'
-    if (selfClose && !isClose) continue
-    if (isClose) {
-      const top = stack[stack.length - 1]
-      if (top === name) {
-        stack.pop()
-      } else {
-        extraCloses.push(name)
-      }
-    } else {
-      stack.push(name)
-    }
-  }
-  return { balanced: stack.length === 0 && extraCloses.length === 0, openTags: stack, extraCloses }
-}

package/telegram-plugin/tests/snapshot-serializer.ts

@@ -1,79 +0,0 @@
-/**
- * Normalizing serializer for HTML / progress-card snapshot tests.
- *
- * Problem: the raw progress-card render contains
- *   - elapsed-time strings like "12s elapsed" / "1m34s"
- *   - ISO timestamps like "2026-04-18T12:34:56.789Z"
- *   - message-ids, toolUseIds, runtime-generated UUIDs
- *
- * Any of these flip on every test run, so a naive snapshot rots the
- * moment it lands. This serializer canonicalizes those fields to stable
- * tokens (`<ELAPSED>`, `<TIME>`, `<ID>`, `<TOOL_USE_ID>`) so the
- * snapshot diff reflects *semantic* changes only.
- *
- * Use by passing `snapshotSerializer` to vitest's `expect.addSnapshotSerializer`
- * in a test's `beforeAll`, or globally via `vitest.config.ts`
- * (`test.snapshotSerializers`).
- */
-
-const PATTERNS: Array<{ re: RegExp; replacement: string }> = [
-  // ISO8601 timestamps
-  { re: /\b\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}(?:\.\d{3})?Z\b/g, replacement: '<TIME>' },
-  // Unix seconds (10 digits) — only within JSON-like contexts, prefixed by `date":` or `"date":`
-  { re: /"date"\s*:\s*\d{10}\b/g, replacement: '"date":<UNIX_SEC>' },
-  // Elapsed-time strings rendered by progress-card (hhh? mm ss).
-  // Match "NNs elapsed", "NNmNNs", "NNhNNm".
-  { re: /\b\d+m\d+s\b/g, replacement: '<ELAPSED>' },
-  { re: /\b\d+h\d+m\b/g, replacement: '<ELAPSED>' },
-  { re: /\b\d+s\s+elapsed\b/g, replacement: '<ELAPSED> elapsed' },
-  // Claude Code tool_use ids: `toolu_01ABC...` (base62, 20+ chars).
-  { re: /\btoolu_[A-Za-z0-9_]{10,}\b/g, replacement: '<TOOL_USE_ID>' },
-  // grammy/telegram file_ids: long alphanumerics starting with AgACAgI / BQACAgI etc.
-  { re: /\b(?:AgAC|BQAC|CQAC|DQAC)[A-Za-z0-9_-]{10,}\b/g, replacement: '<FILE_ID>' },
-  // UUIDs
-  { re: /\b[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}\b/g, replacement: '<UUID>' },
-]
-
-/**
- * Apply all normalizing replacements. Safe to call on any string —
- * non-matching content passes through unchanged.
- */
-export function normalizeSnapshot(input: string): string {
-  let out = input
-  for (const { re, replacement } of PATTERNS) {
-    out = out.replace(re, replacement)
-  }
-  return out
-}
-
-/**
- * Vitest snapshot serializer compatible with `expect.addSnapshotSerializer`.
- *
- * Only applies to string values that clearly look like HTML / Telegram
- * rendered content. Everything else passes through without interference
- * so we don't accidentally mangle unrelated assertions.
- */
-export const snapshotSerializer = {
-  test(val: unknown): boolean {
-    if (typeof val !== 'string') return false
-    // Cheap pre-filter: only normalize when at least one pattern would hit.
-    for (const { re } of PATTERNS) {
-      re.lastIndex = 0
-      if (re.test(val)) {
-        re.lastIndex = 0
-        return true
-      }
-      re.lastIndex = 0
-    }
-    return false
-  },
-  serialize(val: unknown): string {
-    return JSON.stringify(normalizeSnapshot(val as string))
-  },
-}
-
-/** Convenience: normalize + stringify with stable indent. Good for
- *  inline-snapshot-style assertions without invoking the serializer. */
-export function stableHtmlSnapshot(html: string): string {
-  return normalizeSnapshot(html)
-}

package/telegram-plugin/tool-error-filter.ts

@@ -1,89 +0,0 @@
-/**
- * tool-error-filter.ts — severity classifier for tool_result isError events.
- *
- * Acceptance item 5: benign tool failures (no-match, file-not-found,
- * recoverable Telegram errors) don't surface as raw debug in the progress
- * card checklist. Real failures (auth, crash, network) still escalate.
- *
- * Design: pure function, no side effects, no dependencies. The gateway
- * calls isBenignToolError() on every tool_result with isError=true; when
- * it returns true, the progress-card item is marked 'done' (✅) rather
- * than 'failed' (❌) and no operator notification is raised.
- *
- * Important: this classification applies to the DISPLAY only. The agent
- * session transcript is never mutated — the raw error text still reaches
- * the model so it can reason about what happened.
- */
-
-/**
- * Pattern groups for benign tool errors.
- *
- * These are errors that represent "no results" or "resource absent" — the
- * tool ran correctly but found nothing. Surfacing them as ❌ in the checklist
- * is noise: the user can't act on "grep found nothing" and the agent will
- * handle it in context.
- */
-
-// File-not-found patterns (common across Bash, Read, Edit tools)
-const FILE_NOT_FOUND_RE =
-  /no such file or directory|file not found|path does not exist|enoent/i
-
-// No-match patterns (grep, find, search tools)
-const NO_MATCH_RE =
-  /no match(es)? found|returned no results?|not found in|0 result/i
-
-// Recoverable Telegram API patterns (message deleted, not modified, etc.)
-const TELEGRAM_RECOVERABLE_RE =
-  /message (is not modified|to edit not found|can't be deleted|was deleted)|MESSAGE_ID_INVALID|message not found/i
-
-// "Not a git repository" — narrow tool-setup pattern. Earlier drafts also
-// matched `command not found` and `permission denied` but those were too
-// broad: a real EACCES on /etc/passwd, a real "kubectl not found" during
-// a deploy, are genuine failures the user must see. Kept tight to the one
-// truly-benign case (running git outside a repo).
-const TOOL_SETUP_RE =
-  /not a git repository/i
-
-// Timeout / cancellation that the agent will retry. The bare `aborted`
-// substring was previously included but matched DB transaction aborts,
-// git merge aborts, and policy-rejection messages — all real failures.
-// Dropped in favor of explicit timeout and operation-cancelled phrasing.
-const TIMEOUT_RE =
-  /timed? ?out|operation cancelled/i
-
-/**
- * Returns true when a tool error text matches a known benign pattern (no
- * results / resource absent / recoverable Telegram error / explicit timeout).
- * Returns false for empty input, unknown text, or any text that doesn't
- * match a pattern.
- *
- * The function is fail-closed: empty / undefined input → false. Callers
- * that want to suppress only on positive evidence should call this directly;
- * callers that need an extra short-circuit on missing input should guard at
- * the call site with `text && isBenignToolError(text)`.
- *
- * The text parameter is the raw tool result content; the first ~500 chars
- * are sufficient for pattern matching, and callers should truncate before
- * calling for performance / event-size reasons.
- */
-export function isBenignToolError(text: string): boolean {
-  if (!text) return false
-  return (
-    FILE_NOT_FOUND_RE.test(text) ||
-    NO_MATCH_RE.test(text) ||
-    TELEGRAM_RECOVERABLE_RE.test(text) ||
-    TOOL_SETUP_RE.test(text) ||
-    TIMEOUT_RE.test(text)
-  )
-}
-
-/**
- * Severity of a tool error for routing purposes.
- * - `benign`: no-match / not-found / recoverable — suppress from UI
- * - `real`: auth failure, crash, unexpected error — surface in UI
- */
-export type ToolErrorSeverity = 'benign' | 'real'
-
-export function classifyToolError(text: string): ToolErrorSeverity {
-  return isBenignToolError(text) ? 'benign' : 'real'
-}