OpenAgent v1.1.0-beta.34

OpenAgent v1.1.0-beta.34: grounded memory authorization and durable recovery

Product beta 34 pairs with Core beta 19 to repair two failures reproduced from
Friday's real Telegram conversation about watches and Patek Philippe.

The visible Not authorized for memory.manage and Not authorized for memory.read messages were not generated by the authorization boundary: that
turn contained no tool invocation or tool-error receipt, while a subsequent
Vault call under the same Telegram identity succeeded. Core now keeps a bounded
same-turn tool receipt and permits an authorization claim only when it matches
an actual tool failure. An unsupported claim is replaced with a truthful
correction, including when the optional prose reply guard is disabled. The
mandatory identity prompt also requires the Agent to invoke the Vault for
memory operations instead of describing an attempt it did not make.

The same incident exposed an independent SQLite continuity defect. The durable
journal contained the conversation, but projection into the normalized session
tables could be deferred by cannot open savepoint - SQL statements in progress. Recovery projection now opens a detached connection and transaction
after the journal commit, so active statements on the request connection cannot
drop the conversational context presented to the next turn.

Regression coverage exercises grounded and ungrounded authorization reports,
streamed terminal correction, empty-turn receipt isolation, and detached
recovery projection under an active SQLite statement. Product beta 34 is paired
with Core beta 19 and Tools beta 4. The signed desktop App remains beta 15
because these repairs are server-side.

Production acceptance — 2026-10-10

Core commit acc3ec4dd9325d07e7e8ce0148bef9f5eaa8a5dd is published as
v1.1.0-beta.19; Product commit
b8a51befb483fa3eb881e62928480f7e9e46d524 is published as
v1.1.0-beta.34. Both tags and both origin/main branches resolve to those
exact commits. The public Core release has 23 wheels plus its clean manifest;
the Product release has 9 wheels plus its clean manifest. Downloaded asset
sizes and SHA-256 digests matched the local release artifacts, and the public
installer verified all 38 first-party wheels.

The release gates passed as follows:

• Core: 164 tests and 23 subtests passed.
• Product: 88 tests and 21 subtests passed from both source and the exact clean
  wheel set.
• CLI: 99 tests passed with one skip from the exact clean wheel set.
• Documentation: VitePress production build and the GitHub Pages workflow
  passed.
• Packaging: pip check, self-check, clean-manifest validation, source digest
  validation, and workspace ancestry validation passed.

Private online SQLite backups and configuration copies were verified before
cutover. Mixout, Friday, Performa, Ribel, Yoanna, and ICS then started on
Product beta 34 with Core, Gateway, SDK, and SQLite storage beta 19. Every
service reached active/running with zero automatic restarts and returned HTTP
200 from its loopback health endpoint. Every post-cutover database returned
quick_check=ok, remained in WAL mode with a bounded WAL, had no unclaimed task
or workflow request, and emitted fresh scheduler start/beat events.

The affected Friday Telegram conversation received a real end-to-end test in
the existing Patek session. The Agent wrote token OA-MEM-B19-1845 with
vault_write_note, read it back with vault_read_note, and displayed the token
in Telegram. A second message beginning with Continua caused another visible
Vault recall and returned only the same token. The durable journal contains
successful write and read receipts (tool_call_error=false, isError=false),
and both turns plus the final answer are complete in the normalized session
tables. Since the cutover there are no matching cannot open savepoint,
database is locked, or Not authorized for memory.* events.

The other five deployments received runtime, scheduler, health, and SQLite
acceptance rather than a channel-level Vault mutation. No desktop rebuild was
required: the signed desktop remains beta 15 because the corrected execution,
projection, and reply-grounding paths run on the Agent servers.