@AGENTS.md
- Treat this file as Claude-specific startup context only. Cross-agent doctrine remains in
AGENTS.md,.agents/skills/, and.llm/harness/. - For harnessed NetScript work, Claude coordinates. WSL Codex implements slice work when a
mobile-visible implementation agent is required. Evaluation follows
.llm/harness/workflow/lane-policy.md: native opposite-family sessions are the local default; OpenRouter is explicit third-opinion/native-limit escalation, and OpenHands is for explicitly cloud-driven work. - Before invoking a repo skill by name, check whether it exists in
.claude/skills/. If it does not, read the matching.agents/skills/<name>/SKILL.mddirectly. - Use
.llm/tools/agentic/claude/validate-claude-surface.tswhen Claude configuration, skills, hooks, or agent orchestration docs change. - Keep
.claude/skills/generated from.agents/skills/; do not hand-edit mirrored files.
- Select model/provider/effort from
.llm/harness/workflow/lane-policy.mdand pass the resulting identity explicitly to the launcher. Do not maintain a second routing table here.
- Treat Claude dynamic workflows / Ultracode as an expensive supervisor accelerator for hard planning, research synthesis, and orchestration design.
- Do not use Claude workflows as the default implementation lane for NetScript harness slices.
- Prefer WSL Codex subagents for implementation so the work remains mobile-visible, daemon-attached, and token-efficient.
- When a workflow is justified, cap its role to producing a compact plan, agent briefs, or evaluator prompts that are then handed to OpenHands or WSL Codex.
- Documentation authoring (Markdown/Lume content, per-package README + reference prose,
internal-doc consolidation) MAY use a Claude dynamic workflow as the implementation lane.
Rationale: doc quality is language-dominated and the canonical documentation route is optimized
for that work; authoring touches no
packages//plugins/source code — so the supervisor-does-not-write-framework-code boundary is not crossed. - Conditions on this exception:
- Authoring agents run under the harness SKILL (
netscript-harness+ the relevant domain skills:jsr-audit,netscript-doctrine,deno-freshas applicable) so output respects doctrine, the publish surface, and gates. - Route model/provider/effort per slice from
.llm/harness/workflow/lane-policy.md. - Validation stays in a separate opposite-family session with a per-package/per-domain verdict — the Claude workflow is the generator only; it does not self-certify.
- Any change to framework source (e.g. the
@netscript/fresh-ui*Namespacetype exports) remains a WSL Codex daemon-attached slice, never the Claude workflow.
- Authoring agents run under the harness SKILL (
- This exception is scoped to documentation. Framework/plugin implementation slices stay on WSL Codex per the rules above.