Skip to content

Loki Mode v7.36.0

Choose a tag to compare

@github-actions github-actions released this 13 Jun 23:41
Immutable release. Only release title and notes can be modified.

Added

  • loki review --ultra (issue #168): an explicit, opt-in, on-demand cloud
    multi-agent code review that wraps the upstream claude ultrareview
    subcommand. Paid cloud operation, billed by Anthropic, separate from local
    model spend. Strictly opt-in: the disclosure prints every time, an interactive
    TTY prompts with default NO, and a non-interactive shell without --yes (or
    LOKI_ULTRAREVIEW=1) refuses with exit 2 and makes zero cloud calls. Findings
    are advisory and do not block the completion gate. Capability-gated: if the
    installed claude lacks ultrareview it prints an honest upgrade note and
    exits. No dollar amount is shown (no price API exists); the disclosure states
    the cost class only.
  • loki plan now states the trust moat (issue #166): a short, honest block
    ("Why this estimate is trustworthy") naming verified completion (the
    completion council blocks "done" without a real diff and green tests) and cost
    honesty (the quoted model is the model the run dispatches and the dashboard
    reports). Mirrored as a structured moat field in loki plan --json.
  • --allowedTools positive allowlist for reviewer/adversarial/council subcalls
    (issue #167), opt-in via LOKI_REVIEW_ALLOWLIST=1, default OFF and
    byte-identical default argv on both routes. Complements the v7.33.0
    --disallowedTools denylist with a least-privilege read/inspect allowlist
    (Read, Grep, Glob, read-only git and shell). Deny-precedence verified against
    the live CLI and the Claude Code permissions docs, so allowlist and denylist
    ship together. Gated on CLI support.

Fixed

  • Non-git codebase signature hardening (issue #171): trees over the content
    budget (or file-count cap LOKI_PRD_SIG_CONTENT_MAXFILES, default 20000) now
    use a sampled head+tail content hash (files-sampled:) instead of the old
    content-blind files-shallow: listing, so most same-size edits on large
    non-git trees are detected and a stale generated PRD is no longer silently
    reused. Batched (xargs -0 -n 64) to avoid per-file fork cost. A stored
    pre-upgrade files-shallow: signature with a matching listing + count still
    decides reuse on the first post-upgrade run. Git projects unaffected. Residual
    honest gap (a same-size edit confined to the middle of a file larger than 8KB)
    documented in the signature docblock.
  • Nightly parity-drift workflow (issue #173) now normalizes environment-only
    lines (disk space GB, the runtime-route block, the LOKI_LEGACY_BASH warning,
    dashboard pid, summary counts) before diffing the bash and Bun doctor output,
    matching the normalization already in scripts/local-ci.sh. This stops
    env-only false positives while still surfacing real route-logic divergence.