Executor plugins

Executors own the provider-specific LLM harness: tool loops, hook injection, session persistence, telemetry normalisation, and the model catalogue. The runner orchestrates which executor to call; it never imports Anthropic/OpenAI SDK code, never reads models.json itself, and never hardcodes model ids.

Scaffold a drop-in with coro plugin init my-llm --kind executor. That writes coro-plugin.json, a PhaseExecutorBase stub, and a starter models.json.

Model catalogue (models.json)

Each executor package ships a versioned models.json next to the package (published in "files"). Updating shipped models is a package bump, not a runner or dashboard change.

{
  "idPrefixes": ["claude-"],
  "idPatterns": ["^o\\d"],
  "extraAliases": {
    "planning": "tier:planning",
    "mini": "tier:coding"
  },
  "models": [
    {
      "id": "claude-opus-5",
      "displayName": "Claude Opus 5",
      "contextTokens": 1000000,
      "tier": "planning",
      "isDefault": true,
      "supportsThinking": true,
      "pricing": {
        "inputPerMTokens": 5,
        "outputPerMTokens": 25,
        "cacheReadPerMTokens": 0.5,
        "cacheCreationPerMTokens": 6.25
      }
    }
  ]
}

@coro-ai/plugin-sdk owns the schema and helpers:

• loadExecutorModelCatalogue(path) / parseExecutorModelCatalogue(json)
• supportsFromCatalogue — exact id, ${id}-… snapshots, idPrefixes, idPatterns
• defaultAliasesFromCatalogue — tier:* from isDefault plus extraAliases
• calculateCostFromCatalogue — per-million input / output / cache pricing

Pass the catalogue to PhaseExecutorBase so a drop-in author only implements init and executePhase. Shipped packages: packages/llm-anthropic/models.json and packages/llm-openai/models.json.

Core interface (PhaseExecutorRuntime)

Method	Notes
executePhase(req)	Returns an async iterable of PhaseExecutorEvent chunks (assistant deltas, tool proposals, usage metrics)
listModels()	Emits ExecutorModelDescriptor rows for dashboards + alias validation
supports(modelId)	Guards unknown models before network I/O

Optional:

• defaultAliases() — Seed settings.llm.aliases (tier:* plus provider-specific keys).
• calculateCost — Per-model USD from the catalogue (OpenAI); omit when the upstream reports total_cost_usd (Anthropic).
• classifyPhaseError(err) — Return 'stale-session' or 'recoverable-abort' so the runner can recover without importing an LLM package.
• mcpServer() — Attach executor-local MCP (uncommon).
• chat(req) — Plan-mode conversation (POST /intake/stream). Honor sessionState with the same strategy as executePhase: persist sessionId if supportsSessionResume, persist conversationHistory if supportsConversationReplay. messages is the textual fallback. Request also carries optional cwd (stable work root for Claude Code), systemPrompt, model, maxOutputTokens, signal, and read-only tools. Result returns output, usage, toolCalls, and the next sessionState.

PhaseExecutionRequest highlights

Carries HookPolicy (allowedTools, writeRoots, onPreToolUse), resume handles, combined MCP server list, workflow-scoped subagents, and working-directory metadata.

Capability bits (ExecutorCapabilities)

Flag	Runner effect
supportsNativeSubagents	Lets YAML subagents use the executor’s native Task tooling; otherwise Coro registers run_subagent.
supportsClaudeMdNativeWalkUp	When true, executor loads .claude/CLAUDE.md hierarchies itself.
supportsNativeFileTools	Suppresses fallback file_* + shell tools if the SDK already exposes them.
supportsSessionResume	Claude-style sessionId persistence.
supportsConversationReplay	Stateless models replay conversationHistory.
supportsThinking	Surfaces reasoningEffort hints from llm.aliases.
supportsImageInput	Enables multimodal blocks.
maxContextTokens	Hard ceiling enforced before calling upstream APIs.

Exactly one resume strategy should be true for a given plugin build.

Reference packages

Package	Role
@coro-ai/llm-anthropic	Wraps @anthropic-ai/claude-agent-sdk — native Tasks + CLAUDE.md walk-up. Catalogue in models.json.
@coro-ai/llm-openai	OpenAI Responses / chat-completions style loop with conversation replay. Catalogue in models.json.

Executors must remain workflow-agnostic: no references to campaign, lane, or business guardrails beyond what the runner already encodes in hooks.