carlok — zsh — 88×30

cat _posts/2026-10-15-roundstorm-0-1-1-claude-s-tools-become-an-allow-list-and-a-cancelled-turn-leaves-nothing-running.md

roundstorm 0.1.1: Claude's tools become an allow-list, and a cancelled turn leaves nothing running

roundstorm shipped 0.1.1, seven pull requests of hardening and fixes on the local-first deliberation app, and a .dmg for Apple Silicon with its SHA-256 in SHA256SUMS.txt — on top of the per-launch token and the cleared advisories described earlier in the week.

The security change is how the Claude adapter gates tools. It built a --disallowed-tools deny-list from a hardcoded snapshot, so a tool added by a later Claude Code release was available at every tier, reasoning included; it now passes --tools with the tier’s grant, and an explicit empty string for reasoning (#50). The grant was checked against the installed CLI rather than assumed — --tools "" leaves exactly StructuredOutput, and a real turn per tier used only what it was granted (reasoning: none, research: WebSearch, full: Bash + WebSearch) — while argv tests assert that no tier uses a deny-list and that an unknown tier gets nothing. Cancellation got the same attention: brains are spawned detached on Unix and terminate() now signals the whole process group, with a fallback to the direct signal when the child is not a group leader so it cannot take the daemon’s own group down (#51). That is verified by a grandchild test that fails rather than hangs once the group kill is removed, and by aborting a real turn mid-Bash: sleep 40 and finding no sleeper left behind. One behaviour change is worth knowing: because brains now sit in their own group, a terminal Ctrl-C no longer reaches them by itself, so the daemon’s shutdown and the CLI’s SIGINT handler both call terminateAll.

Two long-standing growth issues closed with it. Each turn now carries at most about 60,000 characters of transcript, configurable as ROUNDSTORM_HISTORY_CHARS: over budget, the oldest messages are dropped behind one visible marker while the newest round and the human’s messages are pinned, and the ledger is untouched (#52) — simulated at twelve rounds, total prompt volume falls about 18% and each turn stays bounded where it used to grow with the transcript. Semantic search became room-scopable, streaming and model-aware (#53): rows are streamed and reduced to the best limit as they arrive instead of materialising every vector, /api/search?roomId= scopes a query to a room, only rows stored under the current embedding model count as indexed or pending so a model change re-embeds, and a message with no text — the actual cause of a hanging pending count — no longer counts as awaiting embedding. The Claude adapter also moved onto the shared process pump, removing the duplicated one behind thirteen new adapter-translation tests (#49).

The desktop shell was modernised in the same batch: the Tauri family to 2.12 together, with cargo update moving about 150 lockfile entries (#54), and React 19, Vite 8 and @vitejs/plugin-react 6 with no source changes required (#55). The Markdown component was checked before the upgrade: under React 18 and React 19 it renders byte-identical markup, including an injected <script> that is escaped rather than emitted as a tag, and the built page carries no inline script, so the CSP is unaffected. Vite 8 needs Node 22.12, so the app’s floor is now 22.16.