Deep mode fires many create_asset calls at once. UpsertAssets used a
DEFERRED transaction (SELECT then INSERT) on a 25-connection pool, so
writers hit SQLITE_BUSY immediately and returned "database is locked"
in a few milliseconds instead of waiting.
Open SQLite with _txlock=immediate and a 30s busy timeout, shrink the
pool, and retry locked upserts. Add a 32-way concurrent upsert test.
A node claiming something ran must cite a tool_call_id the host issued in
the trajectory. A citation absent from the trajectory is invented, so the
node is dropped along with its edges; a missing citation counts as invented
only when every real ID survived budget fitting, so a trimmed payload cannot
silently empty a graph. Judging is skipped where the evidence cannot be
enumerated: the process-detail supplement dumps raw event JSON and the
messages-table fallback carries no call IDs. Verified citations land in node
metadata rather than tool_execution_id, which is foreign-keyed to
tool_executions.
Also close two smaller holes: fact text can no longer carry the host index
delimiters, which would close the block early and spill the remaining index
into the system prompt; and the console fact endpoints reject the reserved
note/context-* prefix, matching the MCP tool layer.
Attack chain: failure text goes to textContent, so opt out of i18next
interpolation escaping that rendered upstream errors as //";
give regenerate its own message key; keep the summary card expanded across
a 409 refresh; probe for a cached graph before parsing last_react_input so
polling no longer re-reads a multi-megabyte column; hide the card in the
node-details exclusive state and bound it on narrow screens; drop the dead
hardcoded Chinese title from the API payload.
Blackboard: stop mirroring note/context-* keys into the fact body links
section, which get_project_fact returns verbatim without scope filtering;
single-line category and confidence in index rows so a newline cannot inject
a fake row; make the path overview fail closed on an empty index set.
Provenance: reuse fitUserIntentLedgerText when sanitizing a ledger entry so
a single entry cannot exceed its cap by the three ellipsis runes.
The .cjs harness now escapes interpolated values like real i18next, so the
existing XSS test also pins the escapeValue opt-out.
Mark generated execution prompts, populate deterministic user sections from actual human entries, preserve canonical ledger on overflow/transient restarts, fit long human entries under their formatted budget, and keep unchanged historical keys editable. Add real reduction-to-snapshot Cookie recovery and production entry-point regression tests.
Carry allowlisted summary origin and host user ledger across the trace recovery boundary. Keep immutable snapshot links after alias failures, scope legacy padded keys, remove stale indexes under small budgets, reuse retrieval clues for labels, and add cross-platform memory CI.
Ed1s0nZ