Skip to content

Releases: adam-hanna/vibe-code

Release list

v1.0.1

Choose a tag to compare

@adam-hanna adam-hanna released this 17 Aug 19:47

v1.0.1

First npm release, published as @adam-hanna/vibe-code.

npm install -g @adam-hanna/vibe-code

The scope is not a rename: npm refuses the unscoped vibe-code as too similar
to the unrelated vibecode, and scoped names are exempt from that check. The
repo, the tags, the binary and the command are all still vibe-code / vibe.

Also in this release, the entry point moved from bin/vibe.ts to src/main.ts,
so the bin path is now dist/src/main.js and the bin/ directory is gone. It
held one TypeScript file while the executable npm actually invoked lived in
the ignored dist/bin/, which read as build output that had escaped .gitignore.

Requires the claude and codex CLIs to be installed and logged in - this
tool drives both and installs neither. Run vibe doctor first; it checks for
them and verifies each agent can run the tools it needs.

v1.0

Choose a tag to compare

@adam-hanna adam-hanna released this 17 Aug 19:28

v1.0

Automates the plan / critique / implement / review loop between Claude Code
and Codex, and stops only when it genuinely needs a human.

What is in this release:

  • Plan mode to bypass-permissions inside one Claude session, so the
    implementer already holds the plan and the investigation that produced it.
  • Codex critiques the plan and reviews the diff over one persistent thread,
    which is what lets the oscillation guard tell a repeated objection from a
    new one.
  • Claude's questions are schema-constrained data, not prose. Technical ones go
    to Codex; product ones escalate to a human with NEEDS-INPUT.md.
  • Explicit context compaction, since /compact does not work headless -
    overlapped with Codex turns so it costs no extra wall-clock.
  • A verification gate that runs the project's own test command three times,
    because a report of success is not success and one green run cannot rule out
    a race.
  • Severity model with P0 as the level no tolerance can swallow, and carried
    P1s that get fixed and recorded rather than dropped.
  • Brakes: round caps, oscillation fingerprinting, late-round convergence
    trend, a plan-phase share of the budget, and ceilings on both cost
    (Claude-side) and tokens (both agents).
  • Resumable runs that survive a terminal status without repeating work already
    paid for.

Known limitation: Codex reports no cost in any output mode and no account API
supplies one, so budget.maxCostUsd bounds the Claude side only. budget.maxTokens
is the ceiling that covers both.