Decisions needed after the v3.52.0 mod hardening run

Why this issue

The overnight mod-system hardening run (release v3.52.0) finished. A few things need a person to decide. None block the release.

Decisions needed

PR DRAFT fix(mods): domain plugins screen memory_store only in their own namespace (narrows a guard, needs a decision) #3751 (guard overlap, draft). Stops six domain plugins from naming themselves as owner of another plugin's memory write. But it also means a setup with, say, only ruflo-music enabled would no longer refuse a secret written to a non-music namespace. Options: accept, or build the safer version where non-owners defer the message instead of dropping the guard.
T11: any program that can write a file can fake a "guarded" status. Relabel the status as unauthenticated, or accept the risk.
T15: which provenance values should recall accept?
Defaults: keep modelConfirm = auto and modTrust = observe, or tighten?
New file reads: compactCarry reads swarm and claims state; memory_import guard reads the file it is pointed at. Confirm both are acceptable.
T19: an http.fetch hook can see other plugins' requests. Recorded, no fix yet.
rvf_ingest is still unguarded.
ruvector 0.3.3 vs the pinned ^0.2.27: conditional go. Old intelligence.json needs ruvector hooks reembed. No pin was changed.
ADR-449 (guidance and learning loop on mods) stays Proposed until someone approves phase 1.

Not yet built

ADR-451 items 4, 8, 9, 10.

Details: ADR-450 (threat model) and ADR-451/452 in the docs folder.

Nightly triage — 2026-10-11

Status: needs human input
This tracking issue lists 9 decision items from the v3.52.0 mod-hardening run: (1) whether to accept PR #3751's guard-overlap narrowing or build a safer deferred-message variant, (2) relabeling the "guarded" status as unauthenticated (T11), (3) which provenance values recall should accept (T15), (4) whether to tighten the modelConfirm/modTrust defaults, (5) whether new file reads in compactCarry and the memory_import guard are acceptable, (6) the recorded-but-unfixed http.fetch cross-plugin visibility (T19), (7) rvf_ingest remaining unguarded, (8) whether to bump the pinned ruvector ^0.2.27 to 0.3.3, and (9) approving phase 1 of ADR-449 (guidance/learning loop on mods) — plus a "not yet built" note on ADR-451 items 4/8/9/10. Item 1 already appears resolved: the safer "defer to owner, don't narrow the guard" variant merged as #3767 on 2026-10-05, and the draft #3751 was closed as superseded on 2026-10-09 with that evidence cited directly in its own comments. Item 5's memory_import half also shipped (PR #3764, fail-open file screen, merged 2026-10-05). Items 2, 3, 4, 6, 7, 8 and 9 show no merged follow-up — #3726 (ruvector 0.3.3 validation) explicitly left the pin decision to the maintainer, and ADR-449 is still marked Proposed per the ADR index refresh in #3752 — so they remain open product/security/design calls only the repo owner can make.

Generated by Claude Code

Navigation Menu

Repository navigation

Decisions needed after the v3.52.0 mod hardening run #3785

Why this issue