v1.4.1

TL;DR

• Calculate current Claude and Codex API-equivalent costs accurately with cache-write TTLs, GPT-5.6 long-context tiers, fork-aware usage, and source-dated fallback pricing.
• Recover historical usage from real local transcripts even when JSONL records exceed 1 MiB, then recalculate eligible sessions with the latest bundled rates.
• Read spending and activity faster through a structured terminal dashboard, semantic color controls, complete command help, and explicit cached-token breakdowns.
• Use Git over SSH more consistently through policy-controlled session agents
  with cleaner remote-aware identity selection while preserving the existing
  private-key boundary.

Added

• Global color control: human-readable commands use the shared terminal
  palette by default and accept agentjail --no-color <command> as a
  process-wide color override while preserving Unicode structure.
• Terminal cost dashboard: agentjail cost renders project and model
  spending, impact bars, session totals, cache efficiency, and detailed token
  categories while retaining JSON output for automation.
• Policy-controlled Git SSH: the standard policy can bootstrap a
  session-only native OpenSSH agent through the same agent-neutral launch path
  for Claude Code, Codex, and Cursor.

Changed

• Colorized activity reports: stats, cost, and sessions list now use
  the shared semantic palette for totals, outcomes, spend, metadata, and impact
  bars; each also accepts a command-local --no-color flag.
• Cleaner SSH identity selection: session bootstrap follows Git push remote
  precedence and effective OpenSSH IdentityFile ordering, selects one matching
  identity by default, and asks only when multiple matches remain.
• Complete command help: legacy subcommands now expose their local flags in
  --help alongside inherited global options.

Fixed

• Current-model cost estimates: price GPT-5.6 Sol/Terra and Claude Opus
  4.8/5 from officially verified rates, and include cache pricing for Claude
  Opus 4.6 instead of reporting active sessions as $0.00.
• Real transcript ingestion: recover after oversized Claude and Codex JSONL
  records, retain valid later records, and surface bounded warnings instead of
  dropping the rest of a session.
• Cached and supplemental usage pricing: separate uncached input,
  cache-read, cache-write, and output charges; preserve Claude cache-write TTLs;
  apply GPT-5.6 long-context rates per request; and exclude copied Codex fork
  history from the child session.
• Pricing boundaries: attribute model changes to the request that produced
  each usage delta, keep synthetic zero-usage records out of human dashboards,
  and warn when incomplete history prevents a safe estimate.

Security

• Validated SSH delegation: accept only a clean, owned, live Unix agent
  socket, strip ambient delegation variables, and audit the capability without
  recording socket paths or key material.
• Existing private-key isolation preserved: SSH-agent delegation enables
  signing without granting the sandbox access to private-key files. This
  release improves selection and bootstrap consistency without changing that
  existing boundary; the strict policy still disables automatic Git SSH.