Fix: use chat.message hook for context injection

[niklas-r]

Problem

ELF's context injection (golden rules, heuristics, past learnings into the system prompt) is not working against the current OpenCode (v1.2.6) plugin API.
Two issues with the current chat.params usage:

1. Signature mismatch. OpenCode hooks use (input, output) — two separate arguments. The plugin uses async (params) and accesses params.input.message.text. Since params is the input object, params.input resolves to undefined, hitting the if (!userMessage) return bail-out on every call.
2. Wrong hook for content injection. chat.params only exposes { temperature, topP, options } in its output — there is no systemPrompt or message text to modify. Per the current @opencode-ai/plugin type definitions, this hook is for tweaking LLM parameters, not injecting content.
   Symptoms:

• elf metrics returns [] — no injection metrics recorded
• All golden rules have hitCount: 0
• No [ELF MEMORY] block visible in the agent's system prompt

Fix

Switch from chat.params to chat.message, which provides:

• output.message.system — the mutable system prompt (UserMessage.system)
• output.parts — the user's message as Part[] (filtered for TextPart)

"chat.message"?: (input: {
    sessionID: string;
    agent?: string;
    model?: { providerID: string; modelID: string };
    messageID?: string;
}, output: {
    message: UserMessage;  // has system?: string
    parts: Part[];         // TextPart has text: string
}) => Promise<void>;

Changes

• src/index.ts — Replace chat.params with chat.message, fix signature to (input, output), read user text from output.parts, inject into output.message.system
• package.json — Update declared hook from chat.params to chat.message
• README.md — Update architecture diagram to reflect chat.message

Verified

After this fix, the [ELF MEMORY] block is correctly injected into the system prompt on every message, containing golden rules and any matched heuristics:

[ELF MEMORY]
Golden Rules:
- Always validate user inputs before processing to prevent security vulnerabilities
- Use TypeScript strict mode for better type safety and fewer runtime errors
...
Applicable Heuristics:
- When creating e2e tests: use Playwright with the page object pattern... [project]