Skip to main content

Agent Workflow Runtime Parity v0.1

Purpose

Enforce that agent reference workflows stay replay-deterministic across:

  • native selfhost CLI runtime (genesis)
  • WASI/wasm-host bridge runtime posture (genesis_wasi)

This is a strict parity gate, not a smoke test.

Runner

  • Read-only check: scripts/check_agent_workflow_runtime_parity.sh
  • Explicit producer: scripts/update_agent_workflow_runtime_parity_report.sh
  • Optional primary report: .genesis/perf/agent_workflow_runtime_parity_report.json
  • Optional history: .genesis/perf/agent_workflow_runtime_parity_history.jsonl
  • Optional mutation parity companion report: .genesis/perf/agent_generative_workloads_parity_report.json
  • Default minimum history floor for p95 enforcement: GENESIS_AGENT_PARITY_P95_MIN_SAMPLES=8

Inputs

For a fresh run, the parity renderer executes the reference-workflow renderer twice in sequence:

  1. native lane (runtime_profile = "native")
  2. wasi lane (runtime_profile = "wasi-wasm-host-bridge")

Each lane receives a distinct temporary root. Sequential execution is normative because this gate compares semantic runtime behavior and per-workflow timing against fixed baselines; simultaneous lanes would let cross-runtime CPU or I/O contention change the verdict. Concurrent-load behavior is qualified independently by the task-concurrency, host-bridge fault-injection, and AI stress gates.

Reference workflow set includes browser-runtime, XR runtime, and deployment lanes, so parity checks cover all platform runtime additions automatically via shared gauntlet inputs.

After native/wasi gauntlets complete, the parity renderer invokes the generative-workload renderer with both lane reports to verify replay-digest parity across deterministic mutation cases derived from the shared workflow pool.

Validated retained native/WASI reports may be reused when their kind, profile, runtime profile, status, and freshness satisfy the reuse contract. Checks never refresh those retained reports; only the explicit producer retains fresh lane, generative, aggregate, and history outputs.

Each lane emits a gauntlet report with per-workflow:

  • replay_hash (raw workflow replay payload hash)
  • replay_hash_normalized (parity signal hash when domain-level normalization is required)

Pass Criteria

  • Every workflow exists in both lanes.
  • Every workflow passes in both lanes.
  • Every workflow parity hash matches between lanes
    • parity hash = replay_hash_normalized when present, else replay_hash
  • Domain success counts/ok-status match between lanes.
  • Total parity lane elapsed time stays within budget.

Report Contract

  • kind = "genesis/agent-workflow-runtime-parity-v0.1"
  • ok boolean
  • gauntlet_profile
  • native_bin, wasi_bin
  • native_report, wasi_report
  • workflow_count, domain_count
  • elapsed_ms, budget_ms
  • history_samples, history_p95_ms, history_p95_enforced, history_p95_ok
  • p95_min_samples
  • fail_reasons (string list)
  • missing_native_workflows, missing_wasi_workflows
  • workflow_mismatches (includes raw replay hash, parity hash, and lane ok states)
  • domain_mismatches

The SLO schema contract is validated by scripts/check_slo_report_contracts.sh.