Platform Support¶

Lore supports three coding agent platforms. All share the same knowledge base — skills, agents, docs, and work tracking persist across platforms.

Feature Matrix¶

Feature	Claude Code	Cursor	OpenCode
Session banner	Yes	Yes	Yes
Per-prompt reinforcement	Yes (conversation history)	No (compensated by MCP tools)	Yes (system prompt, zero accumulation)
MCP tools	Yes (`lore_search`, `lore_read`, `lore_health` via `.mcp.json`)	Yes (`lore_check_in`, `lore_context`, `lore_write_guard`)	No
MEMORY.md guard	Yes	Yes	Yes
Knowledge capture reminders	Yes	Yes	Yes
Bash escalation tracking	Yes	Yes	Yes
Context path guide	Yes	No	Yes
Compaction resilience	SessionStart re-fires	Flag + re-orientation on next cmd/MCP	Compacting event re-injects
Hook event logging	Yes	Yes	Yes
Instructions	`CLAUDE.md`	`.cursor/rules/lore-*.mdc`	`.lore/instructions.md`

Hook Directories¶

.lore/hooks/        → Claude Code (subprocess per event)
.cursor/hooks/      → Cursor (subprocess per event)
.cursor/mcp/        → Cursor MCP server (long-lived process)
.opencode/plugins/  → OpenCode (long-lived ESM modules)
.lore/lib/          → Shared logic (all platforms)

See Hook Architecture for module layout, lifecycle diagrams, and the full event reference.

Claude Code¶

Config: .claude/settings.json. Seven hooks cover all lifecycle events: session banner, per-prompt reminder, context path guide, memory guard, convention guard, framework guard, and knowledge tracker. SessionStart re-fires after compaction so the full banner is always present.

Foundry / Bedrock / Vertex Deployments¶

When running Claude Code through a cloud provider deployment (Microsoft Foundry, AWS Bedrock, Google Vertex), worker tier model selection requires additional env vars in ~/.claude/settings.json. Claude Code resolves the short aliases (haiku, sonnet, opus) through these env vars at runtime:

{
  "env": {
    "CLAUDE_CODE_USE_FOUNDRY": "1",
    "ANTHROPIC_FOUNDRY_RESOURCE": "your-resource-name",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "your-haiku-deployment-name",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "your-sonnet-deployment-name",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "your-opus-deployment-name"
  }
}

All three ANTHROPIC_DEFAULT_*_MODEL vars must be set. Without them, Claude Code uses its own hardcoded model IDs which may not match your deployment names, causing workers to either fall back to the orchestrator's model or return a 404 from the deployment API.

See Configuration: subagentDefaults for how to set matching deployment names in .lore/config.json.

Cursor¶

Config: .cursor/hooks.json + .cursor/mcp.json. Six hooks plus an MCP server (lore_check_in, lore_context, lore_write_guard) that compensates for Cursor's missing per-prompt hook. Hooks that Cursor silences (afterFileEdit, postToolUseFailure, preCompact) write state to disk; beforeShellExecution and the MCP server read it back. Gaps: No per-prompt hook, no context path guide.

OpenCode¶

Config: opencode.json + .opencode/plugins/. Six long-lived ESM plugins with system prompt injection via chat.system.transform. The first LLM call injects the full banner (~14K chars); subsequent calls are silent (no injection). After context compaction, the full banner is restored on the next call.

Setup¶

All platforms activate automatically after npx create-lore.

Platform	What loads
Claude Code	`CLAUDE.md` + `.claude/settings.json`
Cursor	`.cursor/rules/lore-*.mdc` + `.cursor/hooks.json` + `.cursor/mcp.json`
OpenCode	`opencode.json` + `.opencode/plugins/` + `.opencode/commands/`

CLAUDE.md is generated from .lore/instructions.md. Cursor .mdc rules are generated from multiple sources (instructions, agent-rules, conventions, agent-registry). Run bash .lore/scripts/sync-platform-skills.sh after editing any of those source files to keep platform copies in sync.

Sync Boundaries¶

/lore-update and sync-framework.sh overwrite framework infrastructure (.lore/hooks/, .lore/lib/, .lore/scripts/, .opencode/, .cursor/hooks/) and lore-* prefixed skills/agents. Operator content is never touched.

Category	Synced	Operator-owned
`lore-*` skills/agents	Yes	—
Operator skills/agents	—	Yes
`.lore/hooks/`, `.lore/lib/`, `.lore/scripts/`	Yes	—
`docs/context/conventions/system/`	Yes	—
`docs/knowledge/runbooks/system/`	Yes	—
`docs/` (except `system/` subdirs), `mkdocs.yml`, `.lore/config.json`	—	Yes

.lore/config.json controls subagent model defaults — see Configuration: subagentDefaults.

See Cross-Repo Workflow for how /lore-link generates and gitignores config files in work repos.