v0.10.1: the dashboard inside the package

pip install --upgrade agentmetry==0.10.1

The dashboard is inside the package. On 0.10.0 a pip install had none:
agentmetry dashboard signed the browser in and showed a JSON 404, and
agentmetry stats and agentmetry detections reported an empty trail because
they never sent the API token. Both are fixed, and CI now installs a built
wheel and checks the dashboard loads, so the gap cannot come back unnoticed.
Upgrade with pip install --upgrade agentmetry and restart the recorder.
Also new: agentmetry cloud cursor-cache, which reads Cursor cloud-agent tool
calls from the desktop app's local cache, and a fix for cloud claude, which
crashed before its first request. The ruleset fingerprint is unchanged.

Added

• agentmetry cloud cursor-cache: Cursor cloud-agent tool calls, read from
  the Cursor desktop app's own cache. A cloud agent runs on Cursor's VM,
  where no hook on this machine fires. When the desktop app streams one, it
  caches the conversation in its local state database, including every tool
  call: terminal commands, file reads and edits, searches, web fetches and MCP
  calls, with paths under the VM's /workspace. The command reads that
  database read-only, maps each finished call to the name the local Cursor
  hooks record (shell.run, cursor.Read, cursor.Write,
  cursor.MCP:<tool>), and runs it through the hook client's own mapper, so
  arguments are hashed and traits and ATT&CK labels computed in-process, as
  for a local event. Results and edit bodies are never forwarded. Events record
  as cursor_cloud under cursor-cloud-<agent id>, the correlation id the
  Cloud Bridge API path uses. A second run imports nothing twice, a call whose
  post failed is retried, and --dry-run counts without posting. No vendor
  key, no repository access and no request to Cursor. Run against a real
  database on 2026-10-21: 3 cloud-agent conversations, 1,251 finished tool
  calls, read in about six seconds from a 7.1 GB file. What it cannot see:
  agents the desktop app never streamed; it records after the fact and blocks
  nothing; the cache format is undocumented, so unrecognised rows are counted
  and skipped; and the cache is a file the user can edit, so it is a claim to
  check against Cursor's own run stream, not proof.

Fixed

• The package had no dashboard. The recorder looked for the dashboard's
  static export only at apps/dashboard/out in a git checkout, and the wheel
  shipped just the Python package. So on a pip install, agentmetry dashboard signed the browser in and landed it on {"detail":"Not Found"}.
  It worked on the maintainer's machine only because that recorder runs from a
  checkout. The release build now bundles the export into the package as
  agentmetry/dashboard_static, the recorder serves it when no checkout build
  exists, and with neither it serves a page saying so instead of a JSON 404.
  CI builds a wheel with the export on every change, installs it in a clean
  environment, serves it from a directory with no checkout above it, and fails
  if / is not the dashboard; the release build checks the same, and that the
  wheel carries index.html. Found by installing 0.10.0 from PyPI on a clean
  Python 3.11 environment and walking the README as a new user would.
• agentmetry stats and agentmetry detections printed nothing over a full
  trail. Neither sent the per-install token that every route but health has
  required since 0.9.3. The recorder answered 401, and the CLI read the error
  body as if it were the answer: stats printed "Events: 0" and detections
  printed "No detections" for a session holding a critical detection. Both now
  send the token, honour AGENTMETRY_URL as disposition and the hooks do,
  and report a refusal as an error with what to do about it. Tests pin the
  token and the refusal.
• doctor told a pip install to run scripts it does not have. For an agent
  installed without its hook it said to run scripts/install_<agent>_hooks.ps1,
  which exist only in a checkout. It now says agentmetry hooks install, and
  names the Antigravity script only when Antigravity is the agent missing,
  since that is the one hooks install does not write.
• One scan finding explained itself with its own code. An MCP server
  launched with -y printed "auto_confirm" as its explanation. It now says the
  package manager installs and runs the package without asking first.
• agentmetry cloud claude crashed before its first request. The command
  read args.remote to choose between local and remote sessions, and the
  parser never defined --remote, so every run with
  ANTHROPIC_COMPLIANCE_KEY set raised AttributeError. The flag now exists:
  local sessions by default, --remote for the API's remote ones. The tests
  covered the mappers and never ran the command; two new tests drive it end to
  end with the vendor API and the ingest stubbed, and both fail on the old
  parser.
• cloud claude does not see Claude Code cloud sessions, and the 0.9.7 entry
  implied it did. Anthropic's Compliance API covers Claude Code in the CLI and
  desktop app, and Cowork, and its announcement of 2026-08-11 says the beta
  "doesn't include Claude Code on the web". So cloud claude records sessions
  that ran on a developer's machine (which the hooks already see live) and
  Cowork, not agents running on Anthropic's infrastructure. The 0.9.7 entry is
  corrected in place.
• The 0.9.7 entry for Cloud Bridge claimed Cursor's live tool-call stream.
  map_cursor_stream_tool_call shipped, but no code fetches the stream, so
  agentmetry cloud cursor records agent lifecycle and runs (status, duration,
  pushed branches, PR) and no tool calls. The 0.9.7 entry is corrected in
  place. Neither path has yet been run against a live vendor key.

The detection ruleset is unchanged (fingerprint 15846a0915769d4a), as are the hash-chain record format and the event schema version (1.2.0).