Ideas · Builder Notes

Kognita 0.5.0, and the review that kept going all day

Kognita alpha 0.5.0 shipped just after midnight on 19 October with its open items listed. The rest of the day went on closing what review found: approvals, client scope, taint laundering, replay, races and in-flight actions.

I released 0.5.0 at 12:30 am with the verified items ticked and the open items and known review findings in the changelog. Two fixes went in first: MCP calls naming a client outside the delegation were denied, and a caller's classification header became a floor. The classifier always runs, and the stricter label wins. Then I kept reviewing, and so did the reviewers. Over the next nineteen hours, forty fixes landed against 0.5. They fell into a handful of themes.

Approvals bound to one exact request

Approval continuations could be replayed, reused for a changed request, or released without the checks an immediate allow gets. An approval now binds to the envelope hash, plus the body hash at the gateway, and is spent once in a transaction. Signers must be named and distinct, a blank citation escalates, and a resumed release re-checks the credential, sandbox and screening. Agent edits to documents became public propose_document and apply_proposal_tool: the approval binds to the exact diff, and the apply fails if the target has changed.

One client scope for every check

A delegation for client A could authorise a tool call whose arguments named client B. Tools now declare argument paths per dimension, conflicting subjects deny, and the effective subject set is resolved once and read by authority, limits, anomaly detection and taint.

Taint that can't be washed off

Review found client-data taint could be dropped by starting a fresh run, calling without a run, or nesting content in tool arguments. Taint is now remembered per agent and delegation, runless delegated calls need a bound run, and nested arguments are classified recursively within bounds, failing closed past them.

The gateway, in both directions

Responses are now screened against the run's ceiling and field allow-list, and the gateway re-decides on the response label. Run interventions apply to model calls, token and cost caps are enforced once usage is known, and decoded JSON values are redacted. The gateway now holds the provider key itself, and an agent presenting its own key is denied and flagged.

Concurrency, replay and stopping mid-action

Quota, delegation budget and anomaly counters are admitted under a write lock, tested with threads. Export verification checks event count and head hash, and replay reconstructs child delegations and successful gateway decisions. The MCP transport really closes the socket on cancellation, an unconfirmed cancellation stays uncertain, and recovery waits for in-flight actions and stays retryable until every compensation succeeds.

Identity and adapters

Incomplete agent identity denies, and the model is pinned into the deployment hash, so a model swap needs re-approval. Daily transaction and product limits joined the blast radius. System triggers are admitted only when the client configuration lists them, and they skip only the credential check. Adapters carry the credential and delegation, keep tool declarations and bind callbacks per call.

Looking back

Most of these were laundering routes and races I could have hunted for in the same release. Listing the open items in the 0.5.0 changelog is what made the day manageable. Every fix had a line to close.

All ideasEmail Ahmed