Adapt evidence-gated completion primitives for long-running agents #672

New issue

Open

opened 2026-07-24 00:10:17 +00:00 by coilyco-ops · 2 comments

coilyco-ops commented

2026-07-24 00:10:17 +00:00

Member

workflow: pull-request-and-merge

Source: https://github.com/anthropics/cwc-long-running-agents
License: Apache-2.0

Goal

AOS adapts the portable completion primitives from Anthropic's example harness for Claude, Codex, and other supported agents. AOS keeps the cross-harness contract and does not copy Claude-specific hooks blindly.

Acceptance

Every acceptance criterion begins unproven and can pass only after the agent inspects criterion-matched evidence.
The workflow supports an independent fresh-context evaluator that cannot edit the work it grades.
The agent records bounded checkpoints and a durable handoff for work that crosses sessions or context compaction.
Runtime-specific adapters use native goal, hook, or subagent surfaces where available. The shared skill does not pretend those surfaces are identical.
Operator stop and steering controls remain explicit where a runtime can enforce them.
The implementation maps this pattern against existing engineer, QA, completion-audit, and sealed-role doctrine before adding new eager instructions.
Each adapted artifact carries this source URL and the Apache-2.0 license in its own metadata.
AOS adds no centralized provenance registry or automatic upstream sync.

The source explicitly presents these files as ingredients rather than a turnkey harness, so this issue preserves that boundary.

workflow: pull-request-and-merge Source: https://github.com/anthropics/cwc-long-running-agents License: Apache-2.0 ## Goal AOS adapts the portable completion primitives from Anthropic's example harness for Claude, Codex, and other supported agents. AOS keeps the cross-harness contract and does not copy Claude-specific hooks blindly. ## Acceptance * Every acceptance criterion begins unproven and can pass only after the agent inspects criterion-matched evidence. * The workflow supports an independent fresh-context evaluator that cannot edit the work it grades. * The agent records bounded checkpoints and a durable handoff for work that crosses sessions or context compaction. * Runtime-specific adapters use native goal, hook, or subagent surfaces where available. The shared skill does not pretend those surfaces are identical. * Operator stop and steering controls remain explicit where a runtime can enforce them. * The implementation maps this pattern against existing engineer, QA, completion-audit, and sealed-role doctrine before adding new eager instructions. * Each adapted artifact carries this source URL and the Apache-2.0 license in its own metadata. * AOS adds no centralized provenance registry or automatic upstream sync. The source explicitly presents these files as ingredients rather than a turnkey harness, so this issue preserves that boundary.