Eve library

V6-001 — Durable cognition and concurrent-worker evidence

Evidence

Revision 1

On this page
  • Owner: Codex implementation agent
  • Accountable product: V6
  • Priority: P0 product acceptance slice
  • Linked task: V6-001, replace in-process cognition state and preserve its execution and authority limits across restart and concurrent workers
  • Source commits:
    • durable kill/quarantine control plane: d33a98f2cf35549c1eedcfbdf38906ddaabdd7da
    • durable run, budget, failure, and child state: 98ebbf320a11dd40cee3fad479a62e5811dd2461
    • durable post-loss operator recovery: 50a3a76031f41f8485434694f7a887ae47254659
  • Evidence captured: 2026-07-24

Implemented boundary#

The canonical createDurableAgentRunRuntime composes AgentRunManager, DurableAgentRunLedger, DurableAgentControlPlane, and DurableKillSwitchRegistry over one RUN-002 DurableRunStore. Every cognition run is projected into a RUN-001 common aggregate and checkpointed before the first provider call, after every iteration and tool boundary, and at terminal success or failure. Provider failures persist a failed terminal state with a hashed error receipt rather than a raw provider body.

The common aggregate retains reconstructible bindings for agent family, provider, each tool grant, token/cost/call budgets, maximum iterations, per-call token ceiling, and anomaly thresholds. Fresh managers can therefore reserve and start a child from a durable parent that is absent from local memory. The ledger validates tenant ownership, refuses authority widening, requires bounded children under cost-capped parents, and uses optimistic RUN-002 transitions so concurrent workers cannot both oversubscribe the same remaining parent budget. Reservation and settlement are idempotent, and stale parent checkpoints merge authoritative remote child state instead of erasing it.

Kill switches remain scoped to agent, family, tenant, or provider. Kill activation/release and anomaly quarantine/release are durable control-thread interventions, so a newly constructed worker refuses the affected run before calling the provider. The run loop also polls the durable switch at each iteration boundary.

The production Ori Presence BFF no longer constructs an in-memory-only AgentRunManager. It mounts the canonical durable runtime with FileDurableRunStore and fails closed when OSHUN_V10_ORI_RUN_STORE_PATH is absent. The Hetzner deployment supplies a dedicated BFF run-state volume and includes it in the off-box backup inputs.

The runtime also composes DurableAgentRunOperator over that exact store. Its pause, claim, takeover, plan-step edit, approval, and resume commands commit the RUN-001 aggregate, optimistic run head, RUN-005 lifecycle, HITL-003 intervention journal, claim lease, approval invalidation, and receipt in the file store's single lock/fsync/rename transaction. A new process derives the same stable thread, reconstructs the current plan and lifecycle, and continues from the committed revisions instead of any local object.

The production BFF exposes the recovery authority at /v1/admin/rail/ori/cognition-runs/:runId and its pause, claim, takeover, edit-step, approve, and resume mutations. The BFF never accepts tenant or actor identity from the payload: tenant comes from the validated request context and actor comes from the authenticated principal. Only admin:*, admin:studio, or admin:workspace:model can enter the surface. Every mutation requires a caller-stable command ID, stable timestamp, and rationale; future timestamps are bounded, payloads are strict, and reads return control metadata rather than raw prompts or provider bodies. Missing durable authority, missing tenant context, scope failure, and tenant-header mismatch all fail closed.

Adversarial and restart coverage#

The owning tests prove:

  • an executing checkpoint exists before the first provider invocation;
  • terminal completion, immutable limits, authority, and usage can be resumed from a new store instance;
  • a fresh manager can trace and execute a child against a durable parent that it has never held in process;
  • concurrent workers permit only one of two oversubscribing child reservations;
  • a stale parent write cannot erase a remote child reservation;
  • widened child tools are rejected after restart;
  • repeated settlement is idempotent;
  • provider failure persists failed without retaining the raw error body;
  • tenant kill and quarantine survive worker reconstruction and prevent provider execution until durable release; and
  • missing production run-store configuration fails closed;
  • a fresh BFF and fresh file-store instance can inspect, pause, take over, version-edit, reapprove, resume, and inspect the same run after the first process closes;
  • editing creates a human-authored contiguous plan version, abandons the stale claim, and invalidates the dependency-bound approval;
  • resume before renewed approval is rejected, then succeeds after approval;
  • stable command retries return the original receipt without a second event;
  • concurrent recovery workers serialize competing claims so exactly one wins;
  • terminal durable runs cannot be bootstrapped back into executing state; and
  • operator scope, tenant mismatch, and absent production authority are rejected at the HTTP boundary.

Verification#

Command Result
pnpm exec vitest run --config libs/iris/agents/core/vite.config.ts PASS — 16 files, 223 tests
pnpm exec tsc -p libs/iris/agents/core/tsconfig.lib.json --pretty false --noEmit PASS; repeated after both rebases
pnpm exec vitest run --config libs/v6/moirai-kernel/vitest.config.ts PASS — 2 files, 15 tests
pnpm exec vitest run --config testing/prompt-injection/vitest.config.ts PASS — 5 files, 232 tests
Focused BFF operator route and Ori runtime tests PASS — 2 files, 7 tests; repeated after rebase
Focused transactional-intervention contract PASS — 1 file, 13 tests
pnpm exec vitest run --config libs/contracts/vitest.config.ts PASS — 258 files, 5,094 tests
Direct contracts, Iris core, and BFF TypeScript checks PASS
node scripts/build.mjs from apps/oshun/bff PASS — 34.6 MB production server bundle
docker compose -f infra/hetzner/docker-compose.yml config --no-interpolate --quiet PASS
Docs render, freshness, integrity, Python, and Playwright gates PASS — 3,138 files; 55 Python; 188 browser pass / 16 skip
Strict ESLint over all changed TypeScript files PASS — zero errors/warnings after ignored-file notices were suppressed
Prettier and git diff --check PASS

The first full contracts run exposed an unrelated repository-baseline omission: the alignment ratchet did not classify eight newly landed governance contracts. The ratchet now explicitly records that pending ownership debt, its focused 5-test suite passes, and the repeated full contracts run passes all 5,094 tests. No V6 implementation imports those governance contracts.

The BFF ratcheted typecheck completed but reported four unrelated app-own errors in admin-isis-cli-surface-route.test.ts, autonomous-content-metrics-store.ts, autonomy-route.test.ts, and tenant-content-graph-store.spec.ts; no changed file was reported. The Iris Nx build reached its unchanged tracing dependency and stopped on five implicit-any diagnostics in libs/shared/tracing/src/genai.ts; the owning Iris TypeScript check is green. Nx expanded the BFF build into a 332-project dependency matrix, so it was stopped under the repository's small-host safety rule and replaced with the BFF's direct production build script.

Artifact hashes#

Artifact SHA-256
libs/iris/agents/core/src/agentic/agent-run.ts 1c8d89645d9681a1eca5198b9702e6f55589a61df4332654e8e94cc0e7eae2ed
libs/iris/agents/core/src/agentic/common-run-adapter.ts 4529dac247a2f1a19ff05bfaadaac4e33eb4c3632715464510b70a80c97e1ff9
libs/iris/agents/core/src/agentic/durable-control-plane.ts 571441997aea1bfba693a890c41bee80e465fc43a10e11ea7968c04f0c5f474f
libs/iris/agents/core/src/agentic/durable-run-ledger.ts 642f0c84272bfe93e4855478b887b0eb52d83595b4e0d920dc03225d4b254648
libs/iris/agents/core/src/agentic/durable-runtime.ts 10980fab1c5fac1c9130a1142f1124edf3b3ef44466e599fda95380fa96b3ee6
libs/iris/agents/core/src/agentic/durable-operator.ts 94de5b28391d6fbe8a4b8ce5874d8c0a3c2ea6658821edeb801cb369c66ed98a
libs/iris/agents/core/src/agentic/durable-operator.spec.ts f250f8882b10c9dda392296e82d64be4b419c90d0849b86450998a3583c219f9
libs/contracts/src/run-protocol/transactional-intervention.ts b8c4cbef8c3b4ffaab0e76eea949272565c6152081518ae39d80a39776d166f5
libs/contracts/src/run-protocol/transactional-intervention.spec.ts bc6d9b28988474005eadde8a56b7b5dfc26093546d346f6f553c83569ff0de93
libs/iris/agents/core/src/agentic/durable-control-plane.spec.ts 3b6c3318a03051782ac28a9d707c6777671ad165caffca14849dfd626f3d85dd3
libs/iris/agents/core/src/agentic/durable-run-ledger.spec.ts e9dfeef7e71d9359f5a83968a7e68cd9f6a7102540728f3994bbd95bb58a5bd3
apps/oshun/bff/src/routes/v10-ori-cognition-operator.ts af0051ddc0c2f8666178d1740cb7098ed1c4f6bcfb2bc8e4affc64b77d46fe89
apps/oshun/bff/src/routes/v10-ori-cognition-operator.spec.ts 9a04eaf7d89d53810c7d6c392198781b0474fba279f8bdcb3f8eb103162fef8e
apps/oshun/bff/src/v10/ori-presence-runtime.ts e6f5eb5952a5fb8089701779bc2d7ea880c10a9a93a2e03cadd6697003c880e9
apps/oshun/bff/src/v10/ori-presence-runtime.test.ts 947154edaea5671cab78e6c23f7f0ece9ccc5dba0c03421d89ddc41e4bfb7b44
infra/hetzner/docker-compose.yml 1c1e204d939583e933c28ac0644313490c48ac025e9e4085cc463a358c378de4
V6/V6_ARCHITECTURE.md 780997a6d56b50b1d054642bc5e1e5006f71612240b89b3aad5eda09011ca290

Honest boundary#

This evidence closes durable cognition state, preservation of execution and authority limits, and the authenticated post-loss operator control plane across restart and concurrent workers. “Resume” here means an atomic transition back to execution-eligible lifecycle state after dependency-bound approval; it does not claim that this evidence ran a live provider dispatcher after the transition. It also does not claim multi-host load testing, a live provider incident, restore-from-off-box rehearsal, a browser consumer for the new recovery routes, V6-002 live in-run co-authoring/mixed-lineage publication, or the V6-003 flagship slice.