bd3e3c0512dff847fcdaaec5d6c71c11cd84d090
note tag <coord> <cat> (in-place recategorize); close out the notes -> model CUBE integration + error-tracing from prior run data saved in CUBE
note tag
- New `cube note tag <coord> <cat>` subcommand: reclassify an existing note in place by
setting its doc_type to `note:<cat>` and put_record under the same coord. Used to move
mis-tagged notes into the right category (e.g. a finding accidentally logged as a debug
or action). Puts are WAL+delta+checkpoint durable. Note: `delete_raw` is NOT durable here
(the checkpoint/delta records Puts of surviving keys only — no tombstones — so a deleted
key resurrects from the base snapshot on reload), so `tag` is the reliable in-place
reclassify path. 5 notes tests pass.
Notes -> model CUBE integration (message-save-path)
- cubesys:🎶 WordFlags-tagged CZYX note log (doc_type=note:<category>); project threads
via linked_records; project resume (<cubecoords> project resume <name> <text>) and context
injection (<project context <name>>). Durable ConcurrentStore (WAL + delta + checkpoint);
short-lived processes must checkpoint() before exit.
- CLI: cube note add|list|search|show; cube project list|show|resume|context.
- cube-notes-mcp (stdio MCP server) exposes mcp__cubenotes__note_*/project_* tools to the
DeepSeek Harness; harness registers the mcp-client at the BUNDLE layer (inject: [tools])
so the plugin actually mounts (profile-layer rows can't inject the tools service — the
silent no-mount bug).
How we used it: trace errors from prior run data saved in CUBE
- The headless agent (cube-agent) treats CUBE as a searchable, replayable log: every action
is written as a note (category=action) so a run can be replayed/fine-tuned; each review run
reads ONLY uncovered findings (category=finding) and marks them covered with a checkpoint
note (category=checkpoint) — true incremental, non-destructive review.
- The deterministic, grounded report tool reads finding notes verbatim (no model
hallucination), assigns severity (HIGH/MEDIUM/LOW) + priority order, and writes a review
.txt. It surfaced 7 previously un-categorized logs (GIT-SHALLOW-CLONE, POSTFIX-TLS-CERTS,
SELINUX-RESTORECON, DSH-WEB-EADDRINUSE, CONCURRENT-STORE-CHECKPOINT, MCP-BUNDLE-LAYER,
CUBES-TWO-TREES) that were logged but never tagged finding.
- Coverage is precise: the report marks exactly the findings it reviewed, so the next run
reports only new ones. The notes store is also the harness GUI's CUBE surface
(mcp-cubenotes note_list), so everything the agent writes is visible and reviewable.
Description
No description provided
536 KiB
Languages
Rust
94.5%
Python
2.8%
Shell
2.7%