This is currentness metadata for adapter generation. It is not runtime proof on the user's local machine.

• Persistent guidance: AGENTS.md / AGENTS.override.md, hierarchical root -> CWD, closest guidance wins.
• Repo Skills: .agents/skills/<name>/SKILL.md; progressive disclosure via metadata then body/resources on use.
• Skill distribution: plugins; agents/openai.yaml may declare invocation policy and MCP dependencies.
• MCP: external tools/context.
• Subagents: project/user TOML agent definitions; explicit delegation, own context; thread/depth controls.
• Rule: keep AGENTS small; move procedures/reference into Skills/docs.

• CLAUDE.md: persistent every-session context.
• Skills: reusable on-demand knowledge/workflows; Agent Skills-compatible with Claude-specific fields such as disable-model-invocation, context: fork, agent/background/hooks.
• Subagents: isolated contexts with tool/permission/MCP/hook/skill controls.
• Hooks: lifecycle-triggered deterministic automation/enforcement; stronger than prompt-only guidance for invariants.
• MCP: external service/tool connections.
• Plugins: packaging layer for skills/hooks/subagents/MCP.

• GEMINI.md: hierarchical project context.
• Agent Skills: on-demand workflows.
• Hooks: lifecycle interception.
• Policy engine: TOML rules with allow/deny/ask_user, tool/MCP/subagent matching and priority.
• IMPORTANT 2026-09-07: workspace policy tier documented as currently disabled/non-functional; use User/Admin policies until current docs change.
• Subagents: local/remote, tool/MCP isolation and policy targeting.
• MCP + A2A supported surfaces.

• Instructions: user ~/.copilot/copilot-instructions.md, .github/copilot-instructions.md, modular *.instructions.md, AGENTS.md, CLAUDE.md, GEMINI.md.
• Multiple applicable instruction files are combined; docs do not define a general precedence across all of them. Avoid conflicts; never invent precedence.
• Skills: SKILL.md, auto or slash-invoked, resources/scripts; allowed-tools can preapprove tools.
• Custom agents: Markdown profiles at user/repo/org/enterprise scopes.
• Hooks: deterministic lifecycle logic.
• MCP: GitHub MCP preconfigured plus additional servers.
• Plugins: package agents/skills/hooks/MCP.

• Rules: .cursor/rules/*.mdc; plain .md in that directory is ignored by the rule system.
• AGENTS.md: plain Markdown alternative; nested directory support with specific guidance winning.
• Skills: .agents/skills/ and .cursor/skills/, plus compatibility reads for Claude/Codex skill locations; SKILL.md + paths/disable-model-invocation metadata.
• Subagents: .cursor/agents/ with compatibility for .claude/agents/ and .codex/agents/; Cursor scope wins naming conflicts.
• Plugins can bundle rules/skills/subagents/commands/MCP/hooks.
• MCP supported.

• Current official Devin docs describe Cascade Skills with progressive disclosure and .windsurf/skills/, plus .agents/skills/ compatibility.
• Skills != Rules != Workflows:
  • Skill: multi-step procedure + resources; auto or @mention.
  • Rule: behavioral constraint; activation modes.
  • Workflow: manual /slash-command prompt template.
• MCP supported; current branding/product lineage must be checked before emitting final Windsurf/Devin adapter names.

• Standalone agent management app, Projects/worktrees, synchronous/asynchronous agents.
• Tools: system commands, file IO, web, Chrome, Skills, MCP, subagents.
• Rules/Workflows, scoped settings/permissions and JSON Hooks are distinct surfaces.
• Artifacts are structured reviewable deliverables (plans, code diffs, diagrams, images, browser recordings).
• MCP configs vary across 2.0/IDE/CLI/SDK; adapter must target the actual product surface, not a generic name.

Documentation currentness != binary installed != authenticated != authorized != quota available != verified for task.

Firecrawl current session: surface connected but search returned HTTP 402 => QUOTA_BLOCKED, not unavailable.