v0.143.0

Fixed

• Switching between /concept iteration tabs could destroy every comment on the page. Four independent single lines caused it, and two of them fired precisely when things were already going wrong. restorePanelToReady() ended in localStorage.removeItem(STORAGE_KEY) — deleting every comment, rating and selection — and it runs both when the bridge answers 507 (nothing was persisted anywhere, so the local copy was the last one) and after the five-minute safety timeout, i.e. exactly when Claude has stopped responding because of a usage limit. The submit handler emptied the feedback dock, although the submitted round stays live until Claude appends the next section: a detour into an older tab came home to a blank dock on a round that had not even been answered yet. saveState() persisted frozen rounds and the frozen-painted dock, so browsing a past tab — which is what the tabs are for, and which ends every showScreen() in a saveState() — wrote the old round's submitted answers over the live round's unsent ones. And restoreState() deleted the whole blob on TTL expiry and on a page-version change.

  Typed keys are now namespaced per round (text:i3:d1-s1) — screen ids, {decisionId}-note keys and annotation ids all repeat across rounds, and the dock is one shared overlay outside section[data-iteration] — which is what makes keeping the text both safe and unambiguous. Nothing deletes the state blob any more: TTL and version changes prune it key by key, keep every text: entry, archive the original, and say so on the page. The dock is marked read-only on submit instead of emptied, and re-armed by unmarkDockSubmitted() on every path that hands control back.

Added

• A concept page's unsent comments are now mirrored to disk, fsynced, on every autosave. journal.jsonl only ever saw a payload the user pressed submit on; everything typed before that lived exclusively in localStorage, which does not survive a wiped profile, a private window, a quota error or a power cut — and which nothing else had a copy of. The bridge gains a draft store (POST /draft, GET /draft): an append-only revision log plus an atomically replaced snapshot per page, written with the same never-ack-what-is-not-on-disk rule as a submission. Because the log is append-only, GET /draft can answer recovered — the per-key union of the last non-empty value ever posted, minus the keys the page reported as deliberately cleared — so even a client that posts a blank blob, the shape every past data-loss bug took, cannot destroy anything. The page flushes with sendBeacon on pagehide and merges recovered back on load into any key it is missing or holds empty. The bridge is a plain Python process with no dependency on Claude, so it keeps accepting drafts long after a usage limit has silenced the session. GET /recovery lists the drafts, so a resumed session can tell "the user typed nothing" from "the user typed for an hour". Verified end to end in a browser: the bridge was SIGKILLed, localStorage wiped completely, the bridge restarted — every comment came back from disk. Gate entries 49–53 pin all of it as engine entries, so an existing page picks them up on its next iteration append.

• Two suites that execute rather than read: comment-durability.test.js runs the shipped engine text in jsdom and reproduces all four losses (19 of its 20 tests fail against the previous code), and draft-durability.test.js SIGKILLs a real bridge, corrupts its log and restarts it. A torn final journal line — the signature of a power cut mid-append — also used to swallow the next record; the store now terminates it once per process, and _durable_write fsyncs the directory after the atomic replace.

• The ship gate stopped failing on contention. These suites test hooks, the bridge server and git-sync by spawning them, so one vitest worker is never one process; the default (~cores-1) put 30+ runnable processes on a 12-core machine and the one that lost the scheduler was vitest's own main process, whose workers' onTaskUpdate RPC then timed out and exited the run 1 beside 2466 passing tests. Worker count is capped at a third of the cores.

---

Install: /plugin marketplace add Jerry0022/dotclaude