Agent platform compatibility¶
The harness uses shared repository rules and a platform-neutral workflow while respecting each runtime's native discovery paths.
| Runtime | Durable project instructions | Repo skill entry | Notes |
|---|---|---|---|
| Claude Code | CLAUDE.md plus AGENTS.md authority |
.claude/skills/academic-writing-harness/ |
Existing phase skills remain under .claude/skills/ |
| Codex | AGENTS.md |
.agents/skills/academic-writing-harness/ |
.agents/skills is the current repo-scoped Codex root |
| OpenClaw | Workspace instructions plus AGENTS.md |
.agents/skills/academic-writing-harness/ |
OpenClaw scans project-agent skills from the same root |
.codex/skills and .cline/skills may still exist as integration-generated or
legacy compatibility assets, but they are not the authority for the shared
academic workflow. New cross-agent capabilities should be added to
.agents/skills and adapted into .claude/skills only when Claude-specific
discovery is required.
Capability mapping¶
The workflow depends on capabilities, not a provider-specific tool spelling:
flowchart TD
W[Academic writing skill] --> S[Workspace state]
W --> P[Project and output profile]
W --> L[Literature and full text]
W --> D[Draft and section briefs]
W --> Q[Hooks, gates, and review]
W --> E[Export and delivery]
L --> R[Verified reference store]
D --> A[Audit artifacts]
Q --> A
E --> A
An adapter may expose one facade or several fine-grained tools. It is compliant only if state recovery, evidence provenance, hard gates, and export validation remain observable.