Skip to content

kyb v0.3.4

Latest

Choose a tag to compare

@github-actions github-actions released this 09 Nov 09:41
· 22 commits to main since this release
5f31630

kyb v0.3.4 for aarch64-apple-darwin (Apple Silicon), Docker (linux/amd64, linux/arm64) and crates.io. Intel macOS is not supported.

A patch release. It adds a decision journal for agent teams and a bounded brief before the work, plus a one-line installer.

Decision journal: kyb confident (#55)

When an agent asks its owner to decide something (a release, a deploy, a push, a deletion, a public post, a design choice), it records the question, the answer and later the outcome. Before the next question it asks what was decided before about the same thing.

  • kyb confident check "<one English line>" returns up to three earlier decisions from other sessions (who decided, the verdict, the owner's words, the outcome), hazards and lint on the draft question.
    • Hazards are the same kind of step on the same target that was corrected, rejected, reverted, dropped or never answered.
    • The lint flags several decisions bundled into one question, and facts left unchecked.
  • kyb confident ask opens a record when the agent asks. It runs the same check and returns an id.
  • kyb confident decided <id> records the answer: approve_plain, approve_with_additions, choose_option, defer, correction, reject, confused, counter_question or redirect. The owner may choose an option nobody listed. A retry with the same answer commits nothing.
  • kyb confident outcome <id> appends what came of the decision (kept, corrected, reverted, dropped). A ref to the evidence is required.
  • kyb confident abandon <id> marks a question nobody answered.
  • kyb confident open lists records still waiting for an answer; one open for more than 48 hours is marked stale.
  • Storage. Records are JSON files under confident/ in the same git repository, so every change is a version. Ordinary search, the brief and the export never rank them.
  • Matching.
    • It runs on the one-line gist with the embedding model the server already uses; each gist is embedded once, when it is written.
    • Without a model it falls back to word overlap.
    • The same kind of step on the same target counts as a precedent however it was worded.
    • Records from the caller's own session younger than an hour are left out: that is the same conversation asking again.
  • Refused workspaces. KYB_CONFIDENT_DENY lists workspace name fragments whose writes are refused.
  • API, MCP and the skill.
    • Routes /confident, /confident/check and /confident/{id}/decided|outcome|abandon; capability confident.
    • MCP tools kyb_confident_check and kyb_confident_open are read-only. kyb_confident_ask, kyb_confident_decided and kyb_confident_outcome write and are left out with ?readonly=1.
    • The skill and the MCP instructions tell an agent to check before every question to its owner. A precedent is evidence for the agent to judge, never permission.

Session brief: kyb brief (#53, #54)

kyb brief (GET /brief, MCP kyb_brief) shows what is live in the base, as pointers only. It lists:

  • deadlines;
  • open incidents by severity;
  • active work (in progress, blocked, urgent);
  • knowledge retracted or changed recently, with who wrote it.

Each section is capped, so the brief fits in about 40 lines. The skill and the MCP tool tell an agent to call it before infrastructure work, as context and never as a to-do list. Nothing calls it automatically, so sub-agents do not each load it into their context.

Install

  • One-line installer (#46, #49). https://kybmemory.com/install.sh picks the install path:
    • Apple Silicon with Homebrew: Homebrew;
    • a machine with usable Docker: the Docker image (rootless included);
    • macOS without Homebrew: the release archive under launchd.
      Running it again upgrades in place. KYB_MODE=client installs only the CLI and skill; KYB_DRY_RUN=1 prints the plan.
  • Public compose file. https://kybmemory.com/docker-compose.yml.
  • Skill installer (#50). On a fresh machine it creates the agent's instructions file instead of skipping it.

Fixed

  • The CLI refuses a key that is not a slug with exit 1 (rejected) rather than exit 3 (unreachable) (#41).
  • MCP refuses such a key instead of panicking the handler (#42).

Build and CI (#47, #48)

  • Builds and tests run through Mitts in CI and in the Docker image.
  • CI runs the installer for real on Linux amd64 and arm64 and on macOS, including an upgrade from the previous release.
  • CI installs Mitts from crates.io when the release download fails.