BrainPalace 26.6.31

[26.6.31] - 2026-06-09

Added

• Pre-index embedding-token estimate: brainpalace index <path> --estimate
  prints an approximate embedding-token count (files, code/doc split, tokenizer)
  without indexing, using the exact same file-selection rules as a real index
  (.gitignore, default excludes, nested-project exclusion, include/exclude
  patterns, file types). New server endpoint POST /index/estimate
  (IndexingService.estimate_tokens, tiktoken for OpenAI / chars-4 heuristic
  otherwise, chunk-overlap inflation). brainpalace init asks (opt-in, default
  no, interactive only) whether to show the estimate before the first index;
  after it, you can proceed, toggle code/docs scope and re-estimate, or skip
  indexing. New brainpalace init --include-code/--no-code (default on) threads
  the code/docs choice through both the estimate and the first index, so the
  estimate matches what is actually indexed (was previously hardcoded to include
  code).
• Dashboard instance detail header now carries inline Start / Stop / Restart
  (+ Open) controls for the selected instance, so bouncing the instance you're
  viewing no longer requires detouring through Server → Instances.
• Per-job chunk deltas: indexing jobs now record chunks_added / chunks_removed
  (server JobSummary + JobDetailResponse). The Jobs table gains +Chunks /
  −Chunks columns and a computed Duration column (right of Started); the
  job-detail drawer shows "Chunks added" / "Chunks removed" plus a distinct
  "Index total".

Changed

• Jobs table merges the old separate Status and Progress columns into one:
  the status badge always, with an inline progress bar shown only while a job is
  active (the standalone progress column was dead weight for finished jobs).

Fixed

• Job "Files" count no longer always reads 100 / 100. JobProgress now stores
  the pipeline's phase-weighted percent in a dedicated percent field, decoupled
  from files_processed / files_total, which now carry real document counts.
• Job detail no longer shows the same number for "Chunks" and "Chunks created":
  the per-job insert delta (chunks_added) is computed from the store count
  before/after plus eviction, distinct from the index-wide total_chunks.
• Dashboard "update available" banner no longer goes stale after an upgrade.
  update_check now caches only the PyPI latest (6h TTL) and reads the
  installed version live on every poll, so the banner self-clears the moment
  an in-place upgrade makes installed ≥ latest — even if the dashboard process
  was never restarted (previously the whole payload was cached, so a stale
  "26.6.x available" notice could linger for up to the TTL).

Changed

• brainpalace update now prints a pre-flight notice listing the running
  servers / control-plane dashboard before the upgrade runs (they keep
  serving old code until restarted). The restart itself still happens after the
  upgrade, on confirmation — stopping instances first would needlessly kill the
  dashboard mid-upgrade.

---