Loading
Loading
Release · August 21, 2026 · All news
The end-of-task approval card is where a long audit run used to spill every command, route, and gated dependency onto the screen. This release compacts it into targeted proof lines, so the one card you read to approve stays readable.
Already installed? Ask your agent to “update Cognition” in plain English. Hosted behavior updates automatically with the same key — no new key needed.
01
The approval checkpoint is the one card you actually read at the end of a task, so it stops being a transcript. Summary is now capped to three high-level bullets, and long endpoint inventories are trimmed out of the summary text instead of being pasted in whole.
The card answers what happened, not what ran. A run that touched forty routes no longer prints forty routes at the top — the summary says what changed, and the proof lives in its own line below.
02
A task that ran tests, an eval harness, and endpoint smokes used to dump each command and its raw output onto the card. Those multiple proof receipts now collapse into a single Verified line, so the evidence reads as a conclusion, not a paste of your terminal.
Value claims render once, without a repeated Value label. Proven value still traces to an impact receipt, skill outcome, session receipt, or eval — the compaction changes how proof is shown, never whether it exists.
03
Safety-gated omissions now collapse into one Blocked safely line, so a blocked provider, write, or model path does not sprawl across the card. Skipped verification renders as an explicit Not checked, kept separate from memory that was deliberately left unused.
The three end-states stay legible at a glance: Verified is proof, Blocked safely is a gate doing its job, Not checked is an honest gap. None of them hide inside the others.
04
A long compliance-style audit is exactly where a card turns into noise. Coverage now locks that shape: no raw route dump, no sensitive agreement or key details rendered into the card, and no work silently sliding into an approval queue you never opened.
This is enforced by regression coverage, not a style guideline — the compacted card is the contract. The 1.6.77 polish stays active underneath: sections stay separated, residual notes stay secondary, and every skill or workflow draft still explains its source, why it arose, how it works, the pattern it saw, and its reuse context before you approve it.
The hosted behavior is live now. Ask your agent to update Cognition in plain English, then get back to work — the next checkpoint card will already be compact.
