muse-tracker 0.0.12

• Stop recurring GitHub permission prompts. Hook polls, dry runs and
  export wakes use the installed skill without checking GitHub. Updates now
  happen only on request with self-update; version is offline unless
  called with --check-updates. Missing setup receipts and old wake payloads
  must not start background update attempts.
• An abandoned export now honors the retry timer after a recovery wake,
  preventing another worker from starting on every five-minute poll while
  the first one has not made progress.
• Release fetches use stable URLs with cache-revalidation headers instead of
  random query strings. Archive and file hash verification is preserved.
• Setup version 7 refreshes the hook prompt to prohibit background release
  checks and preserves a paused hook's disabled state. Setup explains the
  difference between Muse's site approval and a one-task approval, and stops
  when release access is denied or expires instead of creating more cards.

muse-tracker 0.0.11

• An agent that is still active keeps its root. The tasks query's end now
  stops at the window's end (created_at < :until_iso). Before, an agent active
  while the query ran (such as the conversation running the export) had its
  last activity past the window, dropped out of every window, and its root had
  no input until it went quiet.
• The CLI rejects a limited query whose rows are not in its paging column's
  order (QUERY_MISMATCH). Such a query lets LIMIT keep the wrong rows, and
  the page cut then drops the rest without a trace; a tasks query still ordered
  on start after setup version 5 is the case this catches.
• Setup version 6: bound end in queries.tasks, and guard each query's
  own-run filter against NULLs (with NOT IN, one NULL empties the result). Setup version 5 now queues
  its root re-send before exporting, and retries it in the same turn after a
  BUSY.

muse-tracker 0.0.10

• Task roots follow the agent's last activity. The tasks query now pages
  on end (tasksTimeColumn: "end"), so an agent's row comes back whenever
  it does something new. Roots of agents that started before the first export
  window now get their input, kind, session and status, and a root's status and
  final answer stay current instead of freezing at the agent's start.
• Langfuse: a child span waits until its task's row has been seen, so every span
  carries the task's session and tags (childrenHeld in export and status).
• Setup version 5: rewrite queries.tasks to page on end; Braintrust installs
  then re-send task roots once to repair those already sent.

muse-tracker 0.0.9

• Langfuse auth works with a standard Muse card. Verified against a real
  Langfuse v4.67.0: a bearer secret key has full project access (a bearer
  public key reaches only scores). Setup now requests the same api_key /
  bearer_header card as Braintrust, holding the secret key, with a
  Basic-header card as the fallback.
• Langfuse plan limits. Existence checks are one query per page window
  (by the agent's tag) instead of one per trace; reads are spaced to stay under
  Hobby's 30 a minute; batches are sized on the OTLP body (≤ 4 MB); a backfill
  older than the plan's read window (langfuseReadWindowDays, 30) needs
  --beyond-read-window.
• A send that failed without an answer is checked only once it is
  langfuseVisibilitySeconds old: real Langfuse makes spans readable a few
  seconds after accepting them.
• tests/e2e.py --langfuse-live URL runs the Langfuse phases against a real
  server through a stand-in for Muse's Secure Vault.

muse-tracker 0.0.8

• Langfuse destination ("destination": "langfuse"). Spans go to
  Langfuse's OTLP endpoint as agent, generation and tool observations
  with input, output, model, token usage, error level and trace-level session,
  tags and metadata. Langfuse stores every write, so each span is sent once:
  children on first sight, a task's root once the task is idle
  (rootSettleMinutes), and runs that may repeat earlier sends check Langfuse
  first. A send that fails without an answer is checked before its retry.
• More of muse.db in every trace, for both destinations: optional
  session_id, parent_task_id, status, kind, model, depth and
  output_text on tasks, output_text on model calls, server on tool calls.
• Setup version 4: saved queries add those columns.

muse-tracker 0.0.7

• Releases are GitHub releases of the public ericxtang/muse-tracker repo;
  the Railway server is gone. Installs read
  releases/latest/download/manifest.json and follow GitHub's redirects, but
  only over HTTPS to github.com and GitHub's download hosts; the archive's
  SHA-256 is still checked against the manifest.
• build.py --publish creates the release for VERSION and refuses one that
  already exists.