open-codeasier v0.3.0

[0.3.0] - 2026-09-06

Added

• Add /cross-review-audit and the cross_review_audit tool to inspect a
  parent session's owned runs through the local run-store and SDK sessions,
  reporting protocol checks and role behavior without reading OpenCode
  internal storage.
• Accept an optional gitcodeCli absolute path in cross-review config and
  persist it from guided setup after npx open-codeasier detect-gitcode
  locates gitcode-cli (PATH, then ~/miniconda3 / ~/anaconda3 including env
  bin/Scripts; --version output must contain gitcode, so a Graphviz
  gc is ignored). GitCode targets (gitcode.com URLs or remotes) fetch
  PR/Issue details with that CLI in the parent session and always pass
  context; GitHub fetching still uses gh.

Changed

• Load project .opencode/cross-review.json exclusively when present, without
  reading or merging ~/.config/opencode/cross-review.json. Global fallback
  applies only when the project file is absent. A present-but-empty project
  config, including the model-free {} created by init, is no longer
  backfilled from global; tell callers to configure reviewers in the existing
  project file or pass --review-models.

Fixed

• Embed non-PR shared context up to the 1,000,000-character schema cap
  instead of silently clipping reviewer briefs (and judge briefs for
  parent-provided context) at 100,000 characters, set warning when
  gatherer output or an unenforced parent context still exceeds the
  embed limit, and tighten that budget from the tightest known
  reviewer/judge limit.context (4 ASCII chars/token, 1 non-ASCII
  char/token, 25% reserve) so a typical payload cannot blow a smaller
  known window. Models without limit.context still use the 1M schema
  cap.
• Treat never-dispatched cross-review roles as insufficient-evidence in
  /cross-review-audit, keep per-run protocol timelines attributed by
  runID, and let a later silent-model-replace fail win over an earlier
  truncated session.
• Fail /cross-review-audit silent-model-replace on a visible prompt-model
  mismatch even when that session is truncated, and treat cancelled/failed
  roles that never landed a prompt as insufficient-evidence.
• Continue gitcode-cli discovery when a PATH entry is unreadable or not a
  directory, and when listing conda env directories fails with EPERM, so
  detect-gitcode no longer crashes on those filesystem errors.
• Treat host type-default cross-review overrides (reviewModels: [] and 0
  for agents, maxConcurrency, and reviewerTimeoutMs) as omitted so they
  fall back to the loaded config instead of failing validation before it is
  applied. Truly illegal overrides still fail fast. Empty or zero values that
  still reach validation mention omitting the argument to use config. Blank
  judgeModel still means parent-session judging; empty focus is unchanged.
• Harden the asynchronous cross-review parent-session protocol. Finalize
  rejects while gathering, reviewers, the explicit judge, or a pending
  timeout decision are still active, so it cannot be used as a poll;
  readyToFinalize is true when reviewers are terminal, the explicit
  judge is terminal, or the run is already in a terminal phase. A
  timeoutAction on the poll that first observes an already-overdue
  session is applied instead of being ignored. Empty or
  whitespace-only context is omitted,
  including PR snapshot notes.md. A stray timeoutAction on an
  ordinary poll is ignored with a warning, including abort after a
  preserved session has already finished.
• Reject a non-PR cross_review_start or legacy cross_review when
  parent-session judging omits context, instead of dispatching
  reviewers with a warning. Read-only reviewer, gatherer, and judge
  briefs now forbid .git/** reads, guessed historical or host-absolute
  paths, whole-tree glob/re-read detours, and webfetch retries after
  403/404/429.