v0.4.0

Fixed

• Output written to a pipe could be cut short. The CLI exited before an
  asynchronous pipe drained, so agent-tree --json | jq or a large
  --sessions --json catalog could receive only the first 64 KiB, or
  another multiple of it (65536 of 378774 bytes in one export measured on
  macOS; Linux showed the same for larger exports). The CLI now waits for
  stdout and stderr to flush. A stdout reader that stops early (| head)
  ends it with status 0 instead of an EPIPE error, and a closed stderr only
  drops diagnostics instead of ending the export.

Added

• Search across sessions. --search <text> and the MCP tool
  agent_tree_search find the session and the numbered step where a text
  appears in prompts, replies and tool inputs (tool results with
  --include-tool-output), newest session first, with redacted snippets.
  Every project and both agents by default; --cwd, --source,
  --since <days> and --limit narrow it, and --json prints the results.
  A raw-byte prefilter keeps a scan of every local session fast (19.9 GiB in
  about 15 s on one machine). Matching runs on redacted text, so a search
  can neither find nor confirm a secret. MCP replies are capped near 20 KB.
  The MCP server now has seven tools.
• Token usage per step. Every mindmap node carries usage whenever the
  session logged it: calls, prompt tokens with cache reads and writes,
  output and reasoning tokens, the largest context, the model's window for
  Codex, compactions, and subagent usage reported beside the main numbers.
  --usage adds these columns to the text tree, and agent_tree_list
  accepts usage: true. Claude Code writes one response as several records
  with the same usage, so each response counts once. Codex usage comes from
  token_usage_record where present, otherwise from token_count. Claude
  Code subagent transcripts count toward the step that started them.
• Open a session from any step. --open <step>, with --mode,
  --agent claude|codex and --open-dir, starts the agent with the
  continue or fork prompt in the step's directory, without a shell, and
  exits with the agent's status. For a --file session it first shows the
  directory and the instruction and waits for Enter; --yes skips that. It
  refuses without a terminal, inside an agent session (CLAUDECODE=1) and
  for prompts over 100,000 bytes, and exits 127 when the agent is not on
  PATH. No MCP tool starts an agent; agent_tree_snapshot and the
  interactive --snapshot suggest the --open command instead.
• agent-tree --help ends with the project's GitHub link.
• After the first successful resume from the interactive tree, the CLI
  prints a short note asking for a GitHub star, once per machine. It goes to
  stderr only and stays silent for piped output, CI, Claude Code, --json
  and --dump-json; AGENT_TREE_NO_STAR_HINT=1 turns it off. A marker under
  ~/.cache/agent-tree/ records that it was shown. The MCP server never
  prints it.

Known limits

• Search reads main transcripts only: not Claude Code subagent transcripts
  or Codex sessions archived under ~/.codex/session_archives/. Case
  folding covers ASCII letters only. Search skips its own agent-tree --search commands and agent_tree_search calls, but other text that
  quotes the query still matches.
• Claude Code logs no context window, so Claude rows show context size
  without a limit. Codex subagent sessions are not counted toward a step.
• --open runs on macOS and Linux, and its prompt is a command-line
  argument that other local users can see in ps.

Changed

• --cwd naming a missing directory exits 2, like other usage errors,
  instead of 1.
• The Claude Code reader no longer warns on stderr about a missing
  permission-mode line; current Claude Code omits it in many valid sessions
  (122 of the newest 300 on one machine). --verbose still shows it.

Docs

• The README covers search, token usage and --open, ticks the 0.4 roadmap
  items and adds the next ones; every list of MCP tools now has seven.
• The README is reorganized around the supported agents, features, a
  first-class "For AI agents" section (identity, isolated install and
  self-test, MCP wiring for any client, tools, recipes and rules) and a new
  roadmap: search across sessions, token usage per step, and opening a
  session from any step. The npm page now shows 0.2.1 and 0.3.0 as published
  instead of "publication pending".
• RELEASING.md records the npm web 2FA path for scripted publishes, that a
  prebuilt-tarball publish skips prepublishOnly, and that a new tarball can
  return 404 for minutes after npm view lists the version.
• RELEASING.md notes that an npm login lasts two hours, so the login and
  the publish approvals belong in one sitting, and that web login works
  without a TTY. CONTRIBUTING.md warns that building from a worktree whose
  node_modules is a symlink embeds the other checkout's paths in dist/.