linux-riscv-contribute

Orchestrate an OpenClaw multi-agent pipeline to close Linux RISC-V gaps versus ARM/x86 (Linux tree + KVM lore), create and manage GitHub issues, generate des...

Overview

Use this skill to run a repeatable discover -> issue -> plan -> implement -> patch pipeline with OpenClaw as orchestrator and ACP agents (claude-code, codex) as workers.

Keep humans at exactly three gates:

1. Confirm gap triage and priorities.
2. Approve implementation plan.
3. Approve final patch email before sending.

Workflow

Step 0: Bootstrap workspace

Run scripts/bootstrap_openclaw_workflow.sh <docs_repo_root> <linux_repo_path> to create/update:

• kernel/openclaw/config/workflow.yaml
• kernel/openclaw/state/{gap_registry.yaml,issue_map.yaml,run_history/}
• kernel/openclaw/{plans,patches,logs}

If files already exist, do not overwrite without explicit user approval.

Step 1: Discover RISC-V gaps

Collect evidence from:

• Linux source tree (arch/riscv, arch/arm64, arch/x86, virt/kvm)
• KVM lore (https://yhbt.net/lore/kvm/)

Write structured entries to state/gap_registry.yaml with:

• gap_id, type (feature|performance|maintainability), summary
• evidence (paths, commits, lore URLs)
• severity (P0|P1|P2), confidence (high|medium|low)
• acceptance_hint

Pause for Gate-1 human triage before creating issues.

Step 2: Sync GitHub issues

For each approved gap:

• Create/update issue in configured repo.
• Add labels from severity/type.
• Save gap_id -> issue_number mapping to state/issue_map.yaml.

Use one issue per gap; avoid duplicate issues by matching gap_id.

Step 3: Plan with Claude Code (ACP)

Spawn ACP session explicitly:

• runtime: "acp"
• agentId: "claude-code"

Ask for:

• file-level design
• test matrix (kselftest, kvm-unit-tests, perf)
• rollback/risk notes
• upstreaming strategy

Save outputs under kernel/openclaw/plans/issue-<id>-plan.md. Pause for Gate-2 human plan approval.

Step 4: Implement and verify with Codex (ACP)

Spawn ACP session explicitly:

• runtime: "acp"
• agentId: "codex"

Run iterative loop until pass or policy limit:

1. Implement approved plan.
2. Build and run configured tests.
3. Parse failures and patch.

Record each iteration in state/run_history/*.json. If max iterations reached, return to Step 3 with failure summary.

Step 5: Generate patch and email package

Produce:

• git format-patch series
• checkpatch result
• suggested To/Cc (get_maintainer.pl, lore context)
• cover letter draft

Save artifacts in kernel/openclaw/patches/. Pause for Gate-3 human send approval.

Only send to mailing lists after explicit approval.

OpenClaw execution rules

• Prefer ACP sessions_spawn for agent work; set agentId explicitly.
• Limit parallel issues to 2-3 unless user changes policy.
• Never auto-send external email without user confirmation.
• Preserve auditability: every stage must have file artifacts.

Quick command prompts for operator

Use these ready prompts in OpenClaw chat:

1. 按 workflow.yaml 执行 Step-1，更新 gap_registry.yaml，并生成 Gate-1 审核表。
2. 基于已批准 gap 执行 Step-2，同步 issue 并输出映射表。
3. 对 issue #<n> 用 claude-code 执行 Step-3，生成详细方案和测试矩阵。
4. 对 issue #<n> 用 codex 执行 Step-4，直到验证通过或达到迭代上限。
5. 对 issue #<n> 执行 Step-5，先 dry-run 生成 patch 和发信草案，等待我确认。

References

• Workflow template: references/workflow-template.yaml
• Issue template: references/issue-template.md
• Human gate checklist: references/gate-checklist.md