Skip to content

Loki Mode v7.29.0

Choose a tag to compare

@github-actions github-actions released this 10 Jun 12:46
Immutable release. Only release title and notes can be modified.

Added (the quickstart trio, from UX-QUICKSTART-DESIGN)

  • loki quickstart: a guided 4-step first build (setup check, one-line idea,
    template pick, plan review). Every step defaults so Enter-Enter-Enter-Enter
    builds the sample Todo app. Template matching is a deterministic offline
    keyword scorer over the bundled templates (top 3 shown, no LLM, no
    network). The plan step quotes the real estimator's figures before any
    spend; declining costs nothing. Lands the PRD at ./prd.md with no-clobber
    guards (existing files are never silently overwritten; the fallback walks
    numbered suffixes). Non-TTY/CI invocations exit 2 with an automation hint,
    never hang.
  • Inline provider install offer: when no AI provider CLI is found, doctor
    (and the pre-flight gate on start/demo/quick/quickstart) offers to install
    Claude Code. Strictly consent-gated: the only command ever executed is
    npm install -g @anthropic-ai/claude-code, printed before running, only on
    an interactive TTY; non-TTY/CI/--json paths never prompt. Auth handoff
    uses claude auth login, and "Provider ready" is claimed only after
    claude auth status confirms loggedIn=true; anything unconfirmed gets an
    honest check-it-yourself message. Opt out with LOKI_NO_INSTALL_OFFER=1.
    Note: --yes / LOKI_ASSUME_YES now also consents to this install on an
    interactive terminal (non-TTY and CI still never install).
  • loki demo cost confirm: the estimate (cost, time, iterations, labeled as
    an estimate) always prints before spending; interactive runs confirm
    [Y/n]; --yes skips the prompt but never the estimate; non-TTY/CI without
    --yes refuse with exit 2. Declining prints "Cancelled. Nothing was spent."
  • LOKI_COMPLEXITY is honored by loki plan (the same env var the runner
    honors), with an honest "forced via LOKI_COMPLEXITY" note in formatted
    output; invalid values are ignored. This makes the demo quote the figures
    for the tier the demo actually runs.

Fixed

  • loki stop --all exercised by the test suite no longer kills unrelated
    live loki runs on the same machine: the stop-scoping suite now scopes the
    machine-wide kill to its own uniquely marked test runners via a test-only
    pattern knob (user-facing --all semantics are unchanged), and local-ci
    permanently proves harmlessness by spawning a foreign-mimic sentinel
    before the stop suites and asserting it survives by PID.
  • Stale "claude login" instructions corrected to "claude auth login" across
    the wiki (the old command starts an interactive AI session on current
    Claude Code versions instead of authenticating).

Tests

  • New: tests/cli/test-quickstart.sh (11 cases incl. interview-abuse and
    no-clobber), tests/cli/test-provider-offer.sh (15 cases incl. argv proof
    that only the documented npm command runs). Extended:
    tests/test-plan-command.sh 16 to 25, tests/test-cli-commands.sh 23 to 25
    on both routes. All wired into scripts/local-ci.sh.