v0.1.0 — initial public release

Install (Windows x86_64)

Download a2a-v0.1.0-windows-x86_64.exe from the assets below. Rename it to a2a.exe and either:

• Place it on a directory already on your PATH and run a2a to launch the welcome wizard, or
• Place it anywhere and double-click / run it once — the welcome wizard offers to add its directory to your user PATH (Windows: HKCU\Environment\Path; Unix: prints a manual recipe).

Then run a2a (welcome wizard) → cd <project> → a2a init → a2a auth add ... to register your first Cursor API key.

For full instructions see GUIDE.en_us.md / GUIDE.zh_cn.md.

What's in v0.1.0

CLI

• Subcommands: init, ask, auth, doctor, list, clean, status, models, reset.
• No-subcommand entry points:
  • a2a (typed in a terminal or double-clicked from Explorer) → welcome wizard. Detects whether the binary's directory is on user PATH, optionally appends it (HKCU\Environment\Path on Windows, manual recipe on Unix), and detects the cursor-agent CLI. Stale a2a directories on PATH (other locations whose a2a.exe differs from the running binary) are flagged and removable in one Y/n.
  • a2a --agent → structured health report ([health] block of key: value lines an AI agent can regex-parse) plus an [next-step for you, the agent] block worded as direct imperatives. Never pauses, never prompts for Y/n — safe to capture stdout in a single read from Cursor's terminal tool.
• Non-interactive credential entry via --from-stdin on a2a auth add / a2a auth update: the API key is read from stdin's first non-empty line, with UTF-8 BOM stripped (PowerShell pipe-safe). Tokens never appear in shell history.

Storage (single SQLite file, no on-disk config)

• ~/.a2a/credentials.db (Unix 0600 / Windows user-private ACL):
  • profiles table — plaintext API keys (per the trust-the-local-user threat model; no encryption / KDF / master password).
  • meta table — currently only default_profile.
  • model_aliases table — user-global model alias registry, shared across all a2a projects on this machine.
• No project-level .a2a/config.toml, no bundled default-config.toml, no --config / A2A_CONFIG override. Every value that pre-r20 lived in [defaults] is now a build-time constant in crate::defaults.

Consultation (ask)

• Parallel cursor-agent subprocesses (one per model alias), each launched with the resolved profile's API key via CURSOR_API_KEY.
• Workspace isolation via readonly_mirror only: a temp dir containing only the prompt frontmatter's context_files. There is no scratch mode and no --scratch flag.
• Per-call profile chain via --profiles a,b,c (comma-separated, ordered, deduped). When omitted, a single-element chain is used (resolved default profile). There is no --profile (singular) and no --no-fallback flag — pass --profiles <single> for the equivalent.
• Three-class error routing (hard-coded keyword match against cursor-agent stderr; no regex config file): KeyDead (401 / billing / quota / etc.) deletes the profile and advances the chain; ModelUnavailable / Unknown skips the alias entirely; Transient (rate-limit / TLS / DNS / timeout) retries the same profile up to 3× with 1 s / 3 s / 10 s back-off, using --resume <session_id> to keep the same Cursor backend chat.
• Cross-account --resume is never issued — Cursor backend silently drops history on cross-account resume but reports success, so the KeyDead profile-switch path always sends a fresh prompt with a continuation prefix.
• Hard 15-minute timeout per cursor-agent invocation (covers wedged HTTP/2 streams / TLS half-open). Timeout preserves any session_id and stderr captured so far via shared Arc<Mutex> slots.
• "Store drained" guard: when KeyDead deletes the last profile, every in-flight alias bails immediately and the orchestrator prints a recovery banner instead of letting peer aliases waste more cursor-agent calls.
• Raw-answer audit: every consultation persists each model's markdown answer plus a meta.toml (alias, cursor_model, mode, profile_used, fallback_chain, success, elapsed_ms, session_ids[], last_session_id, fallback_attempts[], optional [models.budget] from --log-budget).

Templates / Cursor IDE integration

• a2a init [--path <project>] [--force] writes five templates to two locations each:
  • <project>/.a2a/template/<stage_rel> — staged audit copy.
  • <project>/<dst_rel> — live copy under .cursor/skills/{a2a,a2a-operator,a2a-setup-guide}/SKILL.md, .cursor/rules/40-a2a-protocol.mdc, and .cursor/templates/a2a-prompt-template.md.
• All templates are baked into the binary via include_str! (no on-disk install dir, no tool_install_dir() walk). A {{A2A_VERSION}} placeholder is substituted at install time.
• The three skills together cover the agent flow:
  • a2a-multi-ai-consult — when to consult multiple models, how to synthesize, the AskQuestion-or-bust rule.
  • a2a-operator — natural-language → CLI translation for everyday operations, with confidentiality rules around tokens.
  • a2a-setup-guide — first-time-setup wizard triggered by the literal user message a2a_guide.
• a2a init --force is the single-shot way to refresh installed templates after upgrading the binary. There is no a2a sync subcommand.

Per-command stderr nudge

• When cursor-agent is missing from PATH, every subcommand except doctor / status / no-subcommand emits a one-line stderr warning pointing at https://cursor.com/cli. doctor and the welcome wizard print their own dedicated reachability block, so the warning is suppressed there.

Housekeeping

• a2a list / a2a clean [--older-than <duration>] [--yes] for consultation-history hygiene. Each ask also kicks off a detached, best-effort 7-day prune of stale consultation dirs.
• a2a reset models [--yes] wipes the SQLite model_aliases table; a2a reset credentials [--yes] deletes ~/.a2a/credentials.db entirely.

Known limitations

• No streaming output forwarded to the user — raw answer files are written after each cursor-agent subprocess exits. Per-alias progress lines (received first response (streaming...), still receiving... +N chars, still alive (...)) are printed during the run.
• Budget estimation is character-count-based (proxy for tokens); per-model token-aware estimation is not implemented.
• Tested primarily on Windows. macOS / Linux paths are written cross-platform but have had less wall-clock testing.

---

License: MIT OR Apache-2.0