v0.16.6

Highlights

Local knowledge base:

• The vector-model download is recoverable now. A single transient failure used to pin the store on "embedding engine not initialized" for the rest of the session, and an interrupted download left an archive that made every later attempt read a corrupt file — clearing the cache removes both the model directory and the archive, and a failed init can be retried.
• Platform support is checked before any download: builds whose native tokenizer binding is not published for the platform get their own explanation instead of a generic "check your network" hint, and failures are classified (platform / archive / network / other) so the message matches the cause. Keyword search stays available in every case.

Status indicators:

• The footer status block reports what the session is doing: "compacting" while a compaction runs, "awaiting approval" / "awaiting answer" while an interactive panel is blocked on you (clock stopped, because nothing is progressing), and a short "retrying n/m" during step backoff. A prompt the engine refuses no longer leaves the block latched on "waiting" with the input silently queued.
• Rows inside the activity block share one prefix language: a reasoning row spins while it is live and takes the check mark when it ends, running rows keep spinning when the block is expanded instead of freezing, and a background-task row follows the block frame and stops when its task reports a terminal state.
• A plan is a message of its own: the block that led to it is sealed the moment the plan appears, so the work that follows starts a fresh block below the plan instead of appending rows above it, and every revision seals again. An approval outcome stays filed under the call it belongs to even when that block has already been sealed.

Version: all workspace packages bumped to 0.16.6.

Validation

• Full repository suite: 558 files, 7,109 tests passed, no failures; typecheck clean.
• Every change in this release passed an adversarial review pass, and the fixes are locked by mutation checks: reverting the fix turns the matching test red.
• The status and activity rendering was accepted against real rendered output at 60, 80 and 100 columns, and the plan/interaction shapes are covered by replay parity tests.