capture: populate classification.yaml at brain root (untagged repos fail-safe to confidential, block claude.ai capture) #67

Open
opened 2026-06-23 08:39:47 +00:00 by mathias · 1 comment
Owner

Symptom

Sibling to #66, surfaced in the same 2026-06-23 capture dogfood. The classification mechanism shipped in #50/#56 (server reads an optional classification.yaml at brain root; defaultFor() returns internal for hyperguild/homelab, confidential for client-*, and confidential fail-safe for anything untagged). But no classification.yaml has ever been written, so only defaultFor() governs.

Consequence: every brain wing / gitea repo except hyperguild and homelab resolves to confidential, and since effective classification = strictest across all capture targets, a capture from claude.ai (us-nexus) that touches any other repo (brain, ai-sessions, infra, tapir, …) is refused by the I1 gate. This is technically correct (fail-safe), but it means capture is unusable from claude.ai for most real sessions until the homelab-internal repos are explicitly tagged.

What to do

Create classification.yaml at the brain repo root, explicitly tagging the homelab-internal repos and wings as internal so they stop fail-safing to confidential. Confirm the exact schema against the #56 loader (Load/classificationFor in ingestion/internal/classification/), but it is a wings: + repos: map of name → level.

Tag as internal (homelab-sovereign, non-client): the homelab repos — brain, ai-sessions, infra, hyperguild, homelab, tapir, agentsquad, jepa-fx-risk, swedsl, and the other non-client mathias/* repos. Leave any client-* work to default confidential. Treat genuinely public repos (if any) as public.

Important — this is a classification policy artifact, not a code change. It declares which repos are safe to handle through us-nexus harnesses (claude.ai). That is a sovereignty/trust-zone decision (I1) and should be reviewed deliberately, not auto-generated. Mathias should eyeball the final repo→level map before it merges. When in doubt, leave untagged (fail-safe confidential) rather than tag permissively.

Acceptance

  • classification.yaml at brain root, schema matching the #56 loader
  • Homelab-internal repos + wings tagged internal; client-* left to default confidential
  • Repo→level map reviewed by Mathias before merge (sovereignty decision, not a mechanical fill)
  • Brain pod picks up the file (confirm load path — does it need a reload/restart, or is it read per-request?)
  • Re-dogfood from claude.ai: a capture touching ai-sessions + hyperguild resolves internal and is NOT refused
  • Untagged/unknown repo still fail-safes to confidential (regression check — do not break the fail-safe)

Related

  • #66 (SummaryWriter wiring — sibling; both needed for capture to fully work from claude.ai)
  • #50/#56 (the classification mechanism this populates)
  • brain note wiki/hyperguild/failures/first-live-dogfood-of-braincapture-surfaced-two-real-operat.md (the finding)
## Symptom Sibling to #66, surfaced in the same 2026-06-23 capture dogfood. The classification **mechanism** shipped in #50/#56 (server reads an optional `classification.yaml` at brain root; `defaultFor()` returns `internal` for `hyperguild`/`homelab`, `confidential` for `client-*`, and **`confidential` fail-safe for anything untagged**). But **no `classification.yaml` has ever been written**, so only `defaultFor()` governs. Consequence: every brain wing / gitea repo except `hyperguild` and `homelab` resolves to **confidential**, and since effective classification = strictest across all capture targets, a capture from claude.ai (us-nexus) that touches *any* other repo (`brain`, `ai-sessions`, `infra`, `tapir`, …) is **refused by the I1 gate**. This is technically correct (fail-safe), but it means capture is unusable from claude.ai for most real sessions until the homelab-internal repos are explicitly tagged. ## What to do Create `classification.yaml` at the brain repo root, explicitly tagging the homelab-internal repos and wings as `internal` so they stop fail-safing to confidential. Confirm the exact schema against the #56 loader (`Load`/`classificationFor` in `ingestion/internal/classification/`), but it is a `wings:` + `repos:` map of name → level. Tag as **internal** (homelab-sovereign, non-client): the homelab repos — `brain`, `ai-sessions`, `infra`, `hyperguild`, `homelab`, `tapir`, `agentsquad`, `jepa-fx-risk`, `swedsl`, and the other non-client mathias/* repos. Leave any `client-*` work to default `confidential`. Treat genuinely public repos (if any) as `public`. **Important — this is a classification policy artifact, not a code change.** It declares which repos are safe to handle through us-nexus harnesses (claude.ai). That is a sovereignty/trust-zone decision (I1) and should be reviewed deliberately, not auto-generated. Mathias should eyeball the final repo→level map before it merges. When in doubt, leave untagged (fail-safe confidential) rather than tag permissively. ## Acceptance - [ ] `classification.yaml` at brain root, schema matching the #56 loader - [ ] Homelab-internal repos + wings tagged `internal`; `client-*` left to default confidential - [ ] Repo→level map reviewed by Mathias before merge (sovereignty decision, not a mechanical fill) - [ ] Brain pod picks up the file (confirm load path — does it need a reload/restart, or is it read per-request?) - [ ] Re-dogfood from claude.ai: a capture touching `ai-sessions` + `hyperguild` resolves `internal` and is NOT refused - [ ] Untagged/unknown repo still fail-safes to confidential (regression check — do not break the fail-safe) ## Related - #66 (SummaryWriter wiring — sibling; both needed for capture to fully work from claude.ai) - #50/#56 (the classification mechanism this populates) - brain note `wiki/hyperguild/failures/first-live-dogfood-of-braincapture-surfaced-two-real-operat.md` (the finding)
Author
Owner

Drafted — brain PR #12 (held for your sovereignty review)

classification.yaml at the mathias/brain root (where classification.Load reads it). Schema matches the #56 loader (wings: / repos: → level).

Tagged internal (from the dispatch's list, nothing beyond it):

  • wings: hyperguild, homelab
  • repos: brain, ai-sessions, infra, hyperguild, homelab, tapir, agentsquad, jepa-fx-risk, swedsl

client-* left absent → confidential by default; fail-safe preserved for anything untagged.

STOP point — this is yours to confirm: the repo→level map is a sovereignty/trust-zone decision (which repos are safe through us-nexus). I drafted; review every entry before merging #12. When you were unsure I left it out.

Load path (acceptance Q): read once at startup, not per-request → after merge + brain-sync to the koala hostPath, the ingestion pod must restart/roll to pick it up.

## Drafted — brain PR #12 (held for your sovereignty review) `classification.yaml` at the `mathias/brain` root (where `classification.Load` reads it). Schema matches the #56 loader (`wings:` / `repos:` → level). **Tagged `internal`** (from the dispatch's list, nothing beyond it): - wings: `hyperguild`, `homelab` - repos: `brain`, `ai-sessions`, `infra`, `hyperguild`, `homelab`, `tapir`, `agentsquad`, `jepa-fx-risk`, `swedsl` `client-*` left absent → confidential by default; fail-safe preserved for anything untagged. **STOP point — this is yours to confirm:** the repo→level map is a sovereignty/trust-zone decision (which repos are safe through us-nexus). I drafted; **review every entry before merging #12**. When you were unsure I left it out. **Load path (acceptance Q):** read **once at startup**, not per-request → after merge + `brain-sync` to the koala hostPath, the ingestion pod must **restart/roll** to pick it up.
Sign in to join this conversation.
No labels
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: mathias/hyperguild#67