borg 0.7.2

Corrections and answers reach Borg from every harness's own transcript, so a harness update can no
longer silence them unnoticed.

Fixed

• Rejected tool calls are captured again on Claude Code 2.1.312 and later, which stopped telling the
  hook about them. Borg now reads them from the transcript at each turn end.
  The harness's own transcript
• A session left open no longer keeps its corrections waiting: borg learn picks them up once it has
  been idle for a day.

Added

• Your answers to the agent's own questions (Claude Code, omp, opencode, Codex) become episodes, and
  lessons where they state a rule. Sources of lessons
• borg doctor warns, per harness version, when a harness's transcripts name more rejections or
  answers than Borg holds, and says when a harness stopped telling the hook. borg status and borg
  web show which harness's transcript a correction came from.
• borg assimilate reads Codex sessions and answers, and gains --session and --codex.

Upgrading

• Run borg doctor: it names the day capture broke and the borg assimilate --since <day> that
  recovers it. The dry run states the cost (one model call per episode); add --apply to learn it. A
  longer --since also recovers older answers, which no earlier release captured.
  BORG_TRANSCRIPT_CAPTURE=off turns transcript reading off.

borg 0.7.1

Fixes the borg skill for harnesses that read its header strictly.

Fixed

• The borg skill's description is valid YAML now. Its unquoted colon made strict parsers, such as
  Codex's and opencode's, reject the skill. Run borg update, or borg install <harness> --apply, to
  refresh the installed copies.

borg 0.7.0

Hand long-term knowledge over to another store, and Borg now tells you when it or the mesh moves on.

Added

• Memory hand-over.

  • borg memory export lists each lesson with its evidence.
  • borg memory absorb <id> marks a lesson as held elsewhere, such as a skills repo or a handbook,
    so Borg stops delivering it. If the mistake comes back, Borg counts a recurrence.
  • borg memory release undoes an absorb.

• borg update installs the latest release the way Borg was installed: cargo, or the release
  tarball with its sha256 checked. It then refreshes the harness hooks and runs borg doctor.
  borg update --check only checks.

• Update notices. Borg checks for a new release once a day in the background. Your agent tells
  you once per session and offers to run borg update. BORG_UPDATE_CHECK=off turns the check off.

• Mesh announcements. Nodes exchange their Borg versions. When a node starts using a newer mesh
  protocol, it tells the team once:

  • who runs what;
  • what no longer syncs;
  • how to update.

  BORG_ANNOUNCE=off turns this off. borg status, borg doctor, borg top and borg web show
  each member's version.

Changed

• The mesh protocol is borg/mesh/15. Older members keep syncing what they synced before; only
  absorbs need 0.7.

Fixed

• --help on any command prints its usage instead of running it. Before, borg learn --help ran a
  learn pass.

Upgrading

• From 0.6, run cargo install borg-ai borg-mesh borg-web --locked --force or
  install.sh. From now on,
  borg update does it.

borg 0.6.0

Messages pace themselves and keep their full text, and lesson selection is measured and learned.

Added

• Learned lesson selection is the default. It learns from outcomes, departs from the fixed rule
  only when the evidence supports it, and falls back for good if it sees harm. BORG_SELECT=fixed
  keeps the fixed rule.
  Evidence

• Selection measurements in borg status, borg top and borg web. Each check records:

  • what was delivered;
  • how it was judged;
  • its cost and latency.

  borg debug select --replay replays the logged checks offline.

• Message acknowledgements. A message counts as seen once it is shown. borg inbox --show <id>
  prints a long message whole.

• Optional per-collective risk and budget settings in collectives.json.

• Constants: every numeric
  constant, with its kind and reason.

Changed

• Messages are paced by acknowledgements instead of one per 10 minutes.
• Messages keep up to 16,384 characters (was 1,000).
• The message log is bounded by size instead of 500 lines.
• The mesh protocol is borg/mesh/14. Members on 0.5 still sync.
• The judge no longer sees how a lesson was delivered, so its verdicts compare fairly.
• borg listen learns each session's idle times and retries a refused wake with back-off.

Fixed

• A critical lesson that had already been handed over was dropped. It is now asked.
• Lessons cut by the early-delivery size limit disappeared without a record.
• A queued lesson could be delivered twice.
• omp and opencode could wake a session for messages it had already read.
• borg reach kept calling a model that was failing.
• The committed web page contained the build machine's home path.

Removed

• The fixed message rate and the 500-line message cap.
• The expert drone's $0.50 spend limit. Its time limit stays.

borg 0.5.0

The first release on crates.io. Sessions now talk to each other on any harness and across machines.
The hive and playbooks are gone. 0.4.0 was never published; its changes are part of this release.

Added

• Messages between sessions: borg tell, borg ask and borg inbox. A message can go to:

  • a drone;
  • a session on another machine;
  • everyone;
  • the engineer.

  Messages sync with the team.

• borg listen wakes an idle session when a message arrives.

• --version on borg, borg-mesh and borg-web.

Changed

• Steering is done with messages: borg pause, borg resume, and borg no across machines.
  borg steer is removed.
• The mesh protocol is borg/mesh/13. 0.3 members sync everything except messages.
• Releases are made with scripts/release.sh, as described in
  CONTRIBUTING.md.

Removed

• The hive (borg hive, goals, boards). Messages between sessions replace it.
• Playbooks, which were never promoted.
• borg doctor names any leftover hive or playbook data and the command that removes it.