omk-adaptorch-wpl
Status: Stable (v0.90.9). Wired into open-multi-agent-kit as a workspace dependency.
This package implements the AdaptOrch-native Work Packet Loop (WPL) design, an original
execution loop whose only work-producing action is AdaptOrch's adaptorch_run tool. See
AdaptOrch for the backend product this integrates with; AdaptOrch is
still under active development and this integration currently targets its free Start tier.
Design documents
The full design (state machine, verification layer, integration mapping, and the adversarial
review that shaped the safety gates below) lives outside this package at
.omk/runs/adaptorch-native-loop-algorithm-20260701/:
final-part1-core-algorithm.md — Work Packet state machine, termination conditions, cancellation policy
final-part2-verification-layer.md — the Outcome Adjudicator (5-state verification)
final-part3-integration.md — the projection/bridge between the two
Modules
src/types.ts — Work Packet, Dispatch Record, state enums (as string literal unions), on-disk schema types
src/state-machine.ts — the packet lifecycle state machine and transition guards
src/adaptorch-client.ts — thin typed wrapper around AdaptOrch's 10 MCP tools
src/adjudicator.ts + src/adjudicator-registry.ts — the Outcome Adjudicator and its per-kind registry
src/loop.ts — the integration layer wiring the state machine, AdaptOrch client, and adjudicator together
src/b2c-mapper.ts — B2C (Bridge-to-Code) mapping for patch-apply safety
src/b2c-verdict.ts — Verdict card schema for correctness wall decisions
src/deep-wall.ts — Deep verification wall for multi-runner evidence gating
src/receipt-signature.ts — Cryptographic receipt signing for verification evidence
Integration
The coding-agent consumes this package via:
packages/coding-agent/src/core/adaptorch-bridge.ts — advisory-only bridge (default-off,
circuit-breaker protected, TTL-cached). Provides getFreshHint (sync, cache-read-only) and
requestRefresh (fire-and-forget) for the reasoning router's turn-start path.
packages/coding-agent/docs/adaptorch-preview.md — read-first planning pipeline spec.
Safety notes (do not remove without updating the design docs)
ESCALATED packets never resume automatically; only an explicit external unblock signal moves them on.
- The first Dispatch Record of any payload that differs from the last human-approved baseline
(an
augmented_payload was applied, or a REROUTE changed topology) requires human approval
(AWAITING_APPROVAL), unless the loop is explicitly launched in pre-approved-batch mode.
- Loop-level budgets (
max_dispatch_attempts, max_loop_duration, dispatch-call budget) are
immutable for a loop instance's lifetime; raising them requires a new instance under the same review.
Preview spec: ../coding-agent/docs/adaptorch-preview.md