[HugoXK]

Recompute plugin tool masks when a turn starts
Fixes #594
A Main Bot that onboarding creates while dsh web runs had no create_team / create_child_bot: its Agent resolved before @flybot/tree adopted it, so hidePluginTools computed its deny mask with the Bot outside the tree, and a direct user turn resolves no Agent, so nothing recomputed it (only agent/created and Bot-to-Bot delivery did). The mask is now also recomputed when a turn starts in the Session (session/event turn/start, before the model sees the message), the same seam the connector mask uses; an unchanged mask is left alone, so a masked tool still never flips (#276).
core/test/flybot.test.js resolves the Agent before adoption, adopts the Bot, starts a turn and asserts the mask lifts — it fails without the change. core/test/install.test.js's fixed listener list gains the new on session/event.
Ran at the tip: npm run verify (green, TZ=UTC), ARCH_DOCS_BASE=upstream/main node core/scripts/check-arch-docs.mjs, node .github/scripts/changelog.mjs check. Not run: smoke (this macOS box has no setsid, so dev/bench.sh cannot start dsh web), Desktop, real-model steward.

[Birfy]

PASS at b700a74c5ac9f4d84385f09049e1ed9ec81ec2d4 (maintenance self-review, Birfy; self-PASS counts under the 2026-10-21 owner decision).

Tier T2 (core/src/host/hooks.js; the tests, the patch files and the changelog fragment are lower). The branch was first updated with main 167ce83a (it still sat on dad73d63); the review and every run below are on the merged tip b700a74c, not on the old base.

Reviewed the diff at the tip: a session/event turn/start now recomputes the plugin-tool mask for the Session's Agents, through the same hidePluginTools the connector mask already uses on turn start (rt.connectorTurnChanged) — a Main Bot the tree adopts while the server runs gets create_team / create_child_bot on its first turn, with no restart. hidePluginTools keeps its "unchanged mask is left alone" guard (the key compare), so #276 still holds: a masked tool never flips and no Tool added / Tool removed row appears; its test still passes. Agents are tracked with WeakRefs and dead ones are dropped on the next visit; a turn that changes nothing makes no restrict call. Nothing here widens who may see a tool — the same visibleTo predicate decides, it is only recomputed at one more seam.

Ran at this content (TZ=UTC; proven to be this tree: all 871 blobs of b700a74c were recomputed locally byte-for-byte, modes included, and match):

• node --import ./core/test/harness-resolve.mjs --test core/test/flybot.test.js — 39/39, including Polish: Tool added/removed lines every turn, Chief self-intro wording differs, friend's Bot offered as a Main Bot, gear hover inconsistent #276 and the new A Main Bot created after the server starts has no create_team / create_child_bot until dsh web restarts #594 test.
• Red without the source change: with core/src/host/hooks.js reverted to main and the test kept, 38 ok + the new A Main Bot created after the server starts has no create_team / create_child_bot until dsh web restarts #594 test failing (the mask never lifts).
• Full core suite, one file at a time (125 files; the container has a 2 GiB memory cap and a 256 pid cap) — green. 29 files need esbuild/jsdom from a dev install; they pass once those are present.
• node scripts/ds-bot-patches.mjs check — OK (pin v0.1.5, 10 patches, core/ == ds-bot/ + patches/ds-bot/).
• ARCH_DOCS_BASE=HEAD~1 node core/scripts/check-arch-docs.mjs — OK (8 changed files).
• core lint (oxlint) — 0 errors, 42 pre-existing warnings; node scripts/lint-plugins.mjs — 0 errors.
• node core/scripts/build-client.mjs --check — OK; node .github/scripts/changelog.mjs check — 15 fragments; node --test ".github/scripts/*.test.mjs" — 93/93.

CI at the tip: tests-ok green (test, plugins tree/social/timeline, package), mergeable: clean.

Not run: one-shot npm run verify (2 GiB OOM here), smoke (this box has no setsid), Desktop, real-model steward.