Eve library

V6-002 — In-run takeover co-authoring and mixed-lineage publication evidence

Evidence

Revision 1

On this page
  • Owner: Codex implementation agent
  • Accountable product: V6
  • Priority: P0 live operational seam
  • Linked task: V6-002, bind human authoring to in-run takeover and mixed-lineage publication
  • Source commit: 9ddf36de1b5a492ef9c3bf4a36b7476921361acc
  • Evidence captured: 2026-07-25

Implemented boundary#

The durable AgentRun lifecycle now exposes a read-only oshun.agentic-run.authoring-takeover-binding.v1 after validating all of the following:

  • the authenticated tenant can see the run;
  • the run and Yemaya override system are paused;
  • the requested override is active, targets human-takeover, and belongs to the authenticated operator;
  • the takeover owner matches that operator;
  • the checkpoint exists, is restorable, and passes checksum verification; and
  • the run payload names the exact autonomous production run being edited.

The binding contains the AgentRun and common-run identities, operator, tenant, takeover override, integrity-checked checkpoint, exact autonomous subject run, common-run digest, and evidence references.

Studio Authoring now requires that binding when the human draft opens and requires the same still-active binding again when a branch is merged and when the result is published. A takeover for a different production run cannot open the workspace. A missing or inactive override cannot merge the branch. Handback before publication prevents release.

The exact takeover binding is embedded in the human replacement output before the variant is hashed. The common transactional intervention therefore signs a mixed human/agent lineage over content that includes the takeover identity, rather than attaching an unauthenticated label afterward.

Publication boundary#

The new authoring publication route fails closed unless:

  • the authenticated operator still owns the original active takeover;
  • the authoring lifecycle is approved;
  • the review item has passed its V1 compellingness rerun;
  • dependency-bound review approval is current;
  • source evidence remains inside its declared half-life;
  • the committed review variant matches the merged version digest;
  • the corresponding common artifact has a valid signed mixed-lineage attestation; and
  • a production publisher is configured.

Only after those checks does the route call the configured publisher, transition the review item to published, and durably record the real publication reference, operator, rationale, signed lineage, variant digest, and takeover binding. The record survives autonomy-plane snapshot and restore.

Adversarial and restart coverage#

The owning tests prove:

  • a takeover binding is issued only for an active operator-owned takeover with a valid restorable checkpoint;
  • a different operator cannot bind the takeover;
  • checkpoint-backed handback invalidates the binding;
  • a takeover whose subject is a different autonomous production run cannot open an authoring workspace;
  • replacement without the bound active takeover is rejected;
  • the human output carries the exact run, override, and checkpoint binding;
  • mixed lineage is signed and verifies with the configured key;
  • gate reruns pass and the human replacement is incorporated into the review queue and preference learner;
  • current human approval is required before release;
  • the configured publisher returns the publication reference recorded in the authoring workspace;
  • publication and its rationale survive snapshot/restore; and
  • another publication attempt after handback is rejected.

Verification#

Command / gate Result
Focused autonomy route and AgentRun lifecycle Vitest PASS — 2 files, 21 tests
Broader autonomy bindings, persistence/privacy, full cycle, and AgentRun route Vitest PASS — 5 files, 63 tests
Common transactional-intervention contract Vitest PASS — 1 file, 13 tests
Direct autonomy TypeScript check PASS
Direct BFF production build PASS — 34.7 MB server bundle
Strict ESLint over all changed TypeScript PASS
Prettier and git diff --check PASS

Artifact hashes#

Artifact SHA-256
apps/oshun/bff/src/agentic/run-lifecycle-store.ts 704b11e12261a4c6d08dd8a4deb94f010234060542244a42a4d3036d501a3f77
apps/oshun/bff/src/agentic/run-lifecycle-store.test.ts 52864d0d390788587c955be0f6806a61ab8312a6be9c15d8dae4ee4c0731112c
apps/oshun/bff/src/agentic/autonomous-authoring-store.ts 9087012d5407b05dfea7523b61ada13a51fe60df5bfd84c654c75b615e34ac9b
apps/oshun/bff/src/agentic/autonomy-route.ts 7c5df743a4d816e58b261ce87ead1ab169e0b49bfd22d9356dc8ed596b001859
apps/oshun/bff/src/agentic/autonomy-route.test.ts a24524d15580f1ff1408638e61f6b8c8d178d7c7c182b2635832f710f4dd54ef

Honest boundary#

This evidence proves the authenticated, durable, fail-closed code path with controlled production-seam doubles. It does not claim a live external publisher transaction, deployed provider traffic, a human usability study, or the remaining V6-002 live-provider and continuity/consent/latency/failover/ safety/incident-response exercises. It also does not constitute the V6-003 flagship multi-session companion slice.