Actualis 0.2.0

Actualis now reads Copilot CLI. Post your card.

Added

• GitHub Copilot CLI sessions are read, from
  $COPILOT_HOME/session-state/*/events.jsonl (default ~/.copilot): shell
  commands into the same audit, cost per model per session, supervision per
  command, refusals, subagents and premium requests. --agent copilot reads it
  alone. Copilot reports inputTokens including cache reads and writes for
  every provider, and Claude model ids with dots; both are handled, and
  --explain copilot says how. A person declining a prompt is recorded as
  denied-interactively-by-user, captured from a real denial, and is reported
  as a human refusal.
• --card writes a 1200×630 SVG and PNG to post: supervision, cost or
  volume, in a hero or terminal style. Standard library only. The SVG and
  PNG are drawn from one list of pixel runs, so they cannot disagree. Nothing
  identifying can reach either, and the leak test now checks both, plus the
  printed caption.
• --card refuses to combine with another mode -- --json, --fail-on,
  --diff, --why, --share, --watch, --mcp, --replay, --self-check,
  --explain, --agents, --suppress, --suppressions, --completions or
  --service -- so a card run can never silently skip a CI gate or quietly
  write nothing.
• The MCP server (--mcp) now sees Copilot sessions too.
• Copilot sessions with no shutdown record are reported, not dropped. The
  report says N Copilot sessions unpriced (no shutdown record) under the cost
  line, and --json carries copilot_unpriced_sessions. No cost is estimated
  for them.

Changed

• The card counts supervision per shell command, using the permission mode
  in force when each command ran, rather than per turn. --share still counts
  per turn, but --share and the card now share the one definition of
  "unsupervised" the AISVS mapping already used, so codex:never counts there
  too.
• MCP fleet_summary uses the same single "unsupervised" definition, so
  codex:never counts as unsupervised there as well.
• The vendor matrix (--explain vendors, --json) has a copilot column.

Fixed

• --share printed raw model ids. A fine-tune or private deployment id
  could reach a summary documented as containing nothing identifying. Model
  names now appear only when they are exact entries in the public price table;
  anything else is custom. The leak test seeds a private model id.
• A GitHub expression inside the Action's run: block killed the audit
  step. GitHub expands ${{ ... }} before bash sees the script, so a help
  message quoting one left a dangling quote and a syntax error on the runner.