Saltar al contenido principal

Rapid OS v3.1.0 Release Notes

Release Version: 3.1.0
Artifact Schema Version: 1 (BehavioralRuleset versions 1 and 2; 2 is the default)
Python Requirement: Python >=3.10 (3.10, 3.11, 3.12, and 3.13 verified in CI)

1. Overview​

Rapid OS v3.1.0 adds the intent layer ("porcelain") on top of the v3 governance engine ("plumbing"). The plumbing (rapid spec, rapid run, rapid evidence, rapid eval…) is unchanged; the porcelain verbs (rapid start, rapid next, rapid check, rapid task, rapid attach, rapid finish, rapid approve, rapid status, rapid abort) drive a governed run on top of it, and /rapid-<mode> shortcuts teach each coding harness to follow that loop.

What does not change: Rapid OS does not invoke agents, does not use an LLM as a judge, and never approves a human gate. The plumbing runs nothing; the porcelain runs only the test_command of .rapid-os/config.json or the argv given to rapid check, creates a worktree only with rapid start --worktree, and uses read-only git commands to record the run's changes.

2. What's Included in v3.1.0​

Porcelain layer​

  • rapid start <mode>, rapid next, rapid check, rapid task, rapid attach, rapid finish, rapid approve, rapid status and rapid abort.
  • A pure Conductor (next_action) that asks the BehavioralEvaluator what is missing and returns a NextAction; .rapid-os/session.json remembers the active run.
  • Evidence producers rapid:check, rapid:start, rapid:finish and rapid:attach. See orchestration architecture.

Five engineering modes​

  • Playbooks in templates/modes/ for feature, bugfix, refactor, hardening and research.
  • bugfix: baseline before the red test and rapid check --expect fail on T001. refactor: blocked when passing tests drop below the baseline. hardening: isolated git worktree with rapid start --worktree. research: rapid attach, no mandatory tests.

Behavioral ruleset v2​

  • expectation: "fail" evidence produces a reproduction.<task> assertion and does not count as a failing task or as gate evidence (RAPID1320).
  • Review evidence produced by a harness:* producer cannot close a human gate.

Shortcuts​

  • rapid shortcuts install|list|remove and rapid init --shortcuts generate eight shortcuts (rapid-feature, rapid-bugfix, rapid-refactor, rapid-hardening, rapid-research, rapid-next, rapid-status, rapid-approve) for Claude Code, Codex, Antigravity, Cursor and VS Code from templates/shortcuts/. Idempotent, with .bak backups (RAPID1311–RAPID1318).
  • A RAPID1321 warning flags harness selections that leave the same shortcut in two folders one tool reads. In --json, warnings is a list of {code, message, ...} objects. See the shortcuts guide.

Claude Code hooks​

  • rapid hook stop|pre-tool, installed with rapid shortcuts install --hooks or rapid init --hooks: keep the agent working until a person is needed and deny reserved commands (rapid approve, rapid abort, rapid run gate, rapid evidence add) and edits inside .rapid-os/ (RAPID1337–RAPID1339).

Hardened human approval​

  • rapid approve requires a TTY, refuses CLAUDE_CODE_CHILD_SESSION=1 and refuses when any ancestor process is Claude Code (RAPID1310; RAPID1319 in rapid doctor when the ancestry cannot be read). Without arguments it lists pending human gates; --waive (waivable gates only) and --reject take a mandatory --reason (RAPID1330–RAPID1336).

Fixes​

  • rapid finish no longer fails in text mode, no longer attributes pre-existing untracked files to the run, and exits by verdict on a run that already ended.
  • Runs without a baseline gate and isolated runs can leave prepared; expected failures at rapid start and rapid approve are reported with a RAPID code instead of a traceback.

Documentation and portal​

  • Shortcuts are the main flow; the v3 commands are the advanced mode. New pages: shortcuts guide and orchestration architecture. Product boundaries state exactly what is executed. The unused default Docusaurus template site/ was removed.

3. New CLI Surface​

rapid start      (<mode> [--spec-file <path>|-] [--spec <id>] [--harness <id>] [--resume] [--worktree] [--json])
rapid next       ([--json])
rapid check      ([--task <id>] [--gate <id>] [--expect fail] [--json] -- <command>)
rapid task       (start | done | skip)
rapid attach     (<path> --label <text> [--task <id>])
rapid finish     ([--json])
rapid approve    ([<gate>] [--reviewer <name>] [--waive | --reject] [--reason <text>])
rapid status     ([--json])
rapid abort      ([--reason <text>])
rapid shortcuts  (install | list | remove) [--harness <id>] [--dry-run] [--hooks] [--json]
rapid hook       (stop | pre-tool)
rapid init       (... [--shortcuts] [--hooks])

4. Compatibility & Upgrade Notes​

  • v3 commands and flags are unchanged. Existing scripts that call rapid spec, rapid run, rapid evidence, rapid eval, rapid validate and the legacy v2 commands keep working without modification.
  • JSON output only gains optional fields. New fields (workspace_path, start_untracked, action.spec_template, warnings in rapid shortcuts --json) are omitted when unused.
  • Ruleset v2 is the default for rapid eval run; --ruleset 1 remains available.
  • v1 reports stay valid. A persisted EvaluationReport is replayed with the ruleset it was written with.
  • No schema or digest changes. Specs, contracts, RunState, RunEvidence and EvaluationReport keep schema_version = 1 and their digests; there is nothing to migrate in specs, runs, evidence or evals.
  • Package version is 3.1.0; all v3 JSON schemas remain at schema_version = 1. The governance core still uses only the Python standard library; optional integrations such as remote skills use Node.js/npx and need network access.

5. Known Limits​

  • Human approval is best-effort, not cryptographic. The ancestry check does not see a process that detaches from its parent (double fork, setsid, nohup … &), a renamed claude.exe or a reused pid; a TTY can be emulated; and the reviewer identity is the one declared (--reviewer or git config user.name). The rule is unchanged: a person approves, in their own terminal.
  • Ancestry detection covers Claude Code only. Codex and Cursor are not detected because their process names are not confirmed.
  • Hooks exist only for Claude Code, and the command analysis they perform is best-effort: it cannot stop a command built at runtime.
  • Duplicate shortcuts across tools. Cursor and VS Code also read .claude/skills and .agents/skills; their documentation does not say how same-named skills in different folders are shown, hence the RAPID1321 warning.

6. How to Upgrade​

See De 3.0 a 3.1. In short:

rapid --version                      # Rapid OS 3.1.0
rapid shortcuts install              # regenerate the shortcuts in your project (add --hooks for Claude Code)

Add .rapid-os/session.json and .rapid-os/scratch/ to .gitignore, and set test_command in .rapid-os/config.json.