v3.15.8: Memory writes survive dreams; workers honour AQE_WORKERS_ENABLED

What's New

Your acknowledged memory writes are no longer lost when a dream is running,
and starting the MCP server no longer deletes old dream insights.
AQE_WORKERS_ENABLED=false now really turns the background workers off.
Learning pattern lookups and stats, test retries through symlinks and
per-file coverage from a subfolder also got fixes. Every change was
reviewed adversarially before merge.

Fixed

• Acknowledged memory writes are no longer lost during a dream
  (#1060, #1068): while a dream was running, a memory_store that
  returned persisted: true, or any other foreground write, could be
  silently rolled back if the dream was cancelled, failed or was still
  running when the MCP client closed. Dreams no longer hold a database
  savepoint, so those writes stay. Dream results are the same as before. A
  dream that is still running at shutdown is now recorded as interrupted
  instead of being left at running. Thanks to @pacphi for the report and
  the repro. This is a mitigation: the full rework stays tracked in
  #1060.
• Starting the MCP server no longer deletes old dream insights
  (#1061, #1067): every aqe-mcp start used to delete unapplied
  dream_insights older than 30 days within about 30 seconds, with no log
  line, even with AQE_WORKERS_ENABLED=false. Now:
  • AQE_WORKERS_ENABLED=false (also 0, off, no) starts neither the
    background workers nor the QualityDaemon.
  • The retention sweep never runs at startup. The first one runs after one
    full worker interval (30 minutes).
  • Every deletion is logged on stderr with the count and the age window.
  • Thanks to @pacphi for the report and the numbers from real stores.
• Learning pattern lookups and stats (#1055, #1072): looking up
  patterns by tag no longer throws a TypeError, and qe/learning/optimize
  pattern totals and averages count only active patterns. Provenance and
  archive records used to be counted (averages came out as null) and
  could push real patterns out of the search. Pattern export writes each
  pattern once, and pattern dates are real dates again. Thanks to @pacphi
  for the diagnosis.
• Test retries through a symlinked folder (#1047, #1053): a Vitest
  retry of a test file reached through a symlink now clears when the real
  file passes. Thanks to @rudycelekli.
• Per-file coverage from a subfolder (#1030, #1047, #1054):
  relative per-file coverage paths now resolve against the project root
  (AQE_PROJECT_ROOT, then the repository root), so quality assessment and
  defect prediction find the file's measurement when the MCP server starts
  in a subfolder. Thanks to @rudycelekli.

Added

• AQE_DISABLED_WORKERS: a comma-separated list of background worker
  ids to skip, for example learning-consolidation. Unknown ids are
  reported on stderr.
• AQE_DREAM_INSIGHTS_RETENTION_DAYS: how old an unapplied dream
  insight must be before the sweep deletes it. The default is 30. 0 or
  off turns the sweep off. Any other value that is not a whole number from
  1 to 36500 prints a warning and the default applies.

Both are documented in docs/v3/reference/configuration.md.

Changed

• Installs made with aqe init --no-database no longer start background
  workers. Those installs already write AQE_WORKERS_ENABLED=false, and
  the MCP server now honours it. Remove that line from your MCP config if
  you want the workers.
• Dream branching is off by default. Dreams no longer run inside a
  savepoint, so they no longer emit branch events or DREAM_MERGE /
  DREAM_DISCARD witness entries. Library users can still set
  DreamConfig.enableBranching: true, but that keeps the original risk: a
  discarded dream rolls back foreground writes made during it (#1060).
  If a dream fails part-way, what it already wrote (activations and
  insights for a cycle marked failed or interrupted) is kept.
• Relative coverage paths are project-root relative. For library users:
  a relative path passed to the per-file coverage readers now resolves
  against the project root instead of the current folder. Absolute paths
  and explicit roots behave as before.
• qe/learning/optimize totals change. patterns and dashboard now
  report active patterns only, so totals can drop and averages become real
  numbers. Readers receive Date objects for createdAt and lastUsedAt.
• For library users: the exported LearningConsolidationProtocol can
  now archive low-value patterns in the namespace it reads. Before, it never
  matched anything. Nothing in AQE runs it by default (#1071).

Known issues

• Dreams still need the full rework: compute first, then commit in one
  short transaction, and cancel on shutdown. A dream killed abruptly
  (SIGKILL, crash) can still be left at running (#1060).
• The retention sweep still hard-deletes rows instead of archiving them,
  AQE_DREAM_INSIGHTS_RETENTION_DAYS=false is treated as invalid (so the
  30-day default applies), worker ids in AQE_DISABLED_WORKERS are
  case-sensitive, and the explicit aqe daemon start command is not gated by
  AQE_WORKERS_ENABLED (#1066).
• quality_assess leaves its Queen task slot in use (#1070).
• Learning export, import and the consolidation protocol read a different
  namespace than the one patterns are learned in, and the pattern scan stops
  at about 5,000 patterns with only a log warning (#1071).
• Workspace member dependencies are not scanned from a monorepo root yet
  (#1057).
• MCP test_execute_parallel runs the test suite twice (#1059, fix in
  review in #1065).
• Windows paths are unit-tested only, not run on Windows (#971, #982).

Getting Started

npx agentic-qe init --auto

See CHANGELOG for full details.