Blog
Repo readiness, execution governance, and AI agent notes.
Browse product notes, engineering notes, field notes, and release essays about contract-first onboarding, CI alignment, and agent-safe repository operations.
AI Agent Contracts Cannot Define Repository Acceptance
An agent contract can govern how an AI changes code. It cannot by itself define the executable setup, verification, and evidence a repository needs for a change to count as accepted.
Testing the Proof System: Negative Controls and Dependency Evidence in Ota
How Ota separates dependency reachability, transaction-bound seam exercise, and validated negative controls without turning any red exit into causal proof.
Ota 1.6.24: Replayable Proof for Real Repository Execution
How Ota uses archived receipts, scoped baselines, input identity, and witnessed observations to distinguish a replay from a fresh derivation.
Why ok: true Is Not Enough for AI Agent Execution
How agents and CI should read Ota execution status, proof verdicts, scope, and not-proved boundaries without turning a narrow green into broad authorization.
Pressure-testing Ota on Open WebUI: proof cleanup ownership, bootstrap truth, and native vs Compose runtime boundaries
How Open WebUI forced Ota to tighten runtime-proof cleanup, keep CI bootstrap truth contract-owned, and distinguish host-owned native service teardown from Compose-owned runtime teardown.
Turning Ota Contracts into Merge-Gate Truth
How Ota turns contract-owned verification lanes into stable merge-check identities, CI drift findings, and an opt-in GitHub pull-request gate.
Ota v1.6.24 Now Available: Runtime Proof, Replay Trust, and Typed Execution
Ota v1.6.24 strengthens runtime proof with transaction-bound seam evidence and negative controls, adds replay-grade input provenance, and expands typed image, hydration, runtime, and artifact ownership.
Pressure-testing Ota on Bedrock: query identity as replay evidence
How Bedrock helped Ota separate declared replay inputs from witnessed SQL behavior, so a deterministic NL-to-SQL stability gate can name what it replayed without treating historical model output as a current execution input.