v0.2.1

v0.2.1

The honesty-and-hardening release for the second-brain layer, closing out the
2026-09-04 independent review. No new provider; the layer's claims now match
its evidence, the wind-down decision is data instead of protocol prose, and a
learning record finally has a writer.

Changes

• The wind-down decision becomes a command. studyloop brain wind-down --json [--connector NAME] answers the one question the end-of-session
  protocol needs — which second-brain offer to make, if any — as
  {channel, offer, sentence, reason}. Two rules, not one conjunction: the
  publish offer needs a configured, publish-capable provider; the xTiles offer
  needs provider: xtiles plus an xtiles connector the session can actually
  see. The offer sentences are pinned byte-identical across the CLI, the
  protocol, the skill and the guide.
• Learning records have a writer. studyloop plan record <plan-id> --title … [--body …] (and the record_plan_learning MCP tool) appends a learning
  record through the plan renderer — idempotent, numbered max+1, refusing
  bodies that would corrupt the document. The xTiles wind-down records into
  the plan first, so a second-brain page is a projection of a record the plan
  already has, never the only copy. ADR-0010 clause 1 amended to the rule the
  code obeys.
• xTiles guidance says what was actually proven (first hand-run of the
  prompts, Kiro CLI 2.21.0): the planner prompt creates a tile and returns a
  URL; the project prompt stops promising board views or collection-page
  refreshes; the wind-down prompt skips the Review task when nothing is due;
  data sent is your whole study state, not one plan's; cleanup boundaries are
  stated. studyloop install agents no longer double-links the skill for
  OpenCode.
• Obsidian writer hardening: no directory can be created outside the
  vault; --dry-run previews the "would replace your edits" warning; a note
  whose permissions cannot be read is refused rather than rewritten with
  defaults.
• Fixed: get_study_history (MCP) and studyloop plan evaluate no longer
  fail on the session-history search (ambiguous column + multiple FTS MATCH
  constraints); multi-word topics now match as phrases.
• Test honesty: the three tests that were red on main are root-caused
  and fixed; six wait_for_function async predicates that never awaited are
  replaced with a real polling helper; an acceptance suite (truth table, CLI
  contract, sentence pins, stub xTiles MCP server, live transcript gate
  checks) covers what previously needed a human; the release gate now fails on
  shipped-but-unarchived OpenSpec changes and stale ADR statuses, with an
  early-warning hook for three harnesses.

Verification

• just preflight — 4790 passed, 4 skipped
• just e2e — 503 passed, 20 skipped
• just gate-checks — all three wind-down gate checks pass on captured
  transcripts (offer-once/decline/silence at 3/3, zero flakes)
• GitHub CI green on the merged PR (#16), all 16 checks