Repository navigation
Replies: 2 comments 1 reply
|
More implementation experience for this proposal: separating these fields has proved useful:
Two failure modes are worth explicit acceptance tests. First, reviewed groups and older bundles can diverge: retain current member clauses even when a combined question becomes stale, and count source observations separately from answer requests. Grouping should never imply closure. Second, a recently checked question can gain its answer in another page: bounded rotating retrieval can notice material new evidence without waiting out the cooldown or rechecking everything. Record a completed check only after the evidence was actually judged; rotate failed probes without treating them as an absence of evidence. The existing pinned-question evidence/watermark machinery looks useful to reuse here, while preserving its owner-private boundary and adding the separate team answer/review lifecycle. Partial answers must retain their exact residual, and a rejected proposal must restore its visibility. A review acceptance check should also compare actual saved dated provenance, not just successful write receipts or a lower list count. These would make useful acceptance cases for a native implementation. |
|
One thing worth pinning down in part 1: "page slug plus an anchor for the source line" drifts the same way model-written titles do, just more slowly. Insert a paragraph above the gaps section and every line anchor below it shifts; rename the heading and a heading anchor breaks. Pending answers and hand-overs keyed on that id then detach exactly like the reworded-title case in your lessons. What holds up better is a two-part identity plus an explicit rebind step:
On incremental rebuild of a changed page: The Optionally, writing the minted id back into the page as an HTML comment ( From keeping knowledge-entry ids stable across re-extraction in SwarmAI's DDD layer. Discussion: DDD Knowledge Governance |
Uh oh!
There was an error while loading. Please reload this page.
Moved from #7969, opened by @Walliiee on 2026-10-29. Feature requests now live in Discussions > Ideas.
What problem does this solve?
Pages in a team brain often list what they don't know: a "Missing information" section, a "Where the records disagree" section, "Open questions", or a decision with no owner. The page author has already written these down, and they are often the most useful unknowns in the brain to act on.
gbrain has no native way to:
Teams build it outside gbrain instead, with an LLM pass over those sections. That works for a while, then breaks in predictable ways (see the lessons below).
What does the solution look like?
This is a sketch, and the design is yours to make. It has five parts:
Gap extraction from designated sections
missing,disagreeorno_ownerdisagreegap, the records that disagreefind_expertscould feed this)Incremental rebuild
MCP and CLI surface, for example:
gaps_list: filter by person, area, kind or page, with paginggap_answer: a member writes one of three things into their own source for review: a proposed answer, a hand-over to another person, or "close as not needed"gap_status: where a gap is in its lifecycleLifecycle
A pending answer hides the gap from lists only until the review acts. A partly answered gap comes back.
Optional: gap bundles. These group gaps that one question or one document would answer together, so a person can close several in one conversation or by attaching one file.
Lessons from running this outside gbrain
disagreegaps are most useful with the conflicting records attached, so the person answering sees both sides without searching.Alternatives considered
Here is how this differs from what gbrain already has:
find_contradictionswanted_pages,find_orphansopen_loopsquestions)A gap is a known unknown that the page itself declares, with a loop for the team to answer it. It could reuse pieces of the features above: probe findings as evidence on a
disagreegap, andfind_expertsfor routing. None of them covers the lifecycle.The other alternative is the one we run today: an external LLM pass plus a side store keyed by title. All of the lessons above come from it.
I'm happy to share more detail on these failure modes, or to test a design against a working team brain.
All reactions