Feature: Add Declarative Forced Evaluation Hooks (Markdown/Bash-Only) for Reliable Skill Activation #38

Activity

1. malhashemi commented on Dec 10, 2025

   

   malhashemi

   on Dec 10, 2025

   Owner

   More actions

   Thanks for the detailed proposal and the research behind it! Let me explain why I think this falls outside the scope of opencode-skills, and offer some alternatives that might work better for you.

   1. Architectural Mismatch

   <available_skills> is per-agent, and OpenCode allows users to swap primary agents mid-session. A session-start injection would be stale the moment someone switches agents. This makes the proposed hook mechanism fundamentally incompatible with OpenCode's multi-agent architecture. The noReply api also send user messages only, I do not think there is an already made API for the plugins to manipulate system prompts, but would have to confirm.

   Additionally, OpenCode already injects available tools into each agent's system message automatically. The information is already there, the question is whether the agent uses it.

   2. This Is a Generic Prompt Injection Plugin

   What you're describing: reading markdown files and injecting them at session start, is general-purpose functionality that has nothing to do with skills specifically. OpenCode's plugin system already supports this pattern. Here's a minimal implementation:

   // .opencode/plugin/forced-eval.ts
   import type { Plugin } from "@opencode-ai/plugin"
   import { readFileSync, readdirSync } from "fs"
   import { join } from "path"
   
   export const ForcedEvalPlugin: Plugin = async ({ client }) => {
     return {
       event: async ({ event }) => {
         if (event.type === "session.created") {
           const hooksDir = ".opencode/prompt-hooks/"
           try {
             const files = readdirSync(hooksDir)
               .filter(f => f.endsWith(".md"))
               .sort()
             
             for (const file of files) {
               const content = readFileSync(join(hooksDir, file), "utf-8")
               await client.session.chat({
                 path: { id: event.properties.info.id },
                 body: { 
                   content,
                   noReply: true
                 }
               })
             }
           } catch (e) {
             // No hooks directory, skip silently
           }
         }
         
         // Re-inject after context compaction (optional)
         if (event.type === "session.compacted") {
           // Same logic here if you want persistence
         }
       }
     }
   }

   That's it. Put your forced-eval.md in .opencode/prompt-hooks/ and you're done. This keeps concerns separated: opencode-skills handles tool registration, your injection plugin handles prompting strategy.

   3. The Real Lever Is Skill Descriptions

   The blog's results are compelling, but I'd argue they're solving the wrong problem. Forced evaluation is a mechanical workaround for poorly described skills. The model doesn't activate skills because it doesn't understand when they're relevant.

   I'd challenge you to test this hypothesis. Create a skill with two different descriptions:

   Boring description:

   ---
   description: "Provides Svelte 5 runes patterns"
   ---

   Engaging description:

   ---
   description: "CRITICAL for any Svelte 5 work. Contains the reactive state patterns ($state, $derived, $effect) that replaced Svelte 4's let declarations. Using old patterns WILL break. Call this before writing ANY Svelte component."
   ---

   Run identical prompts against both. I'd bet the second version approaches that 80% activation rate without any injection hooks, because you're giving the model the context it needs to make good decisions, rather than forcing a mechanical checklist.

   4. Why I'm Resistant to Baking This In

   The philosophy of this plugin is feature parity with Claude Code's skills system, exposing skills as callable tools. Prompt injection strategies are orthogonal to that. They're also:

   • Highly personal: What works for one workflow may annoy another
   • Model-dependent: As models improve, aggressive prompting becomes unnecessary (and potentially counterproductive)
   • Already possible: Via the plugin pattern above

   I'd rather keep opencode-skills focused and let users compose their own prompting strategies.

   ---

   Happy to help you refine descriptions or debug your injection plugin if you go that route. Would be curious to see your results if you try the description experiment! I am also keeping this issue open for further discussions
2. taqtiqa-mark commented on Dec 10, 2025

   

   taqtiqa-mark

   on Dec 10, 2025

   Author

   More actions

   The philosophy of this plugin is feature parity with Claude Code's skills system

   Without this Claude Code skills feature, would the plugin have feature parity? IIUC the author does not extend CC or Skills to respect the hook. It is a native feature available by default.

   <available_skills> is per-agent, and OpenCode allows users to swap primary agents mid-session. A session-start injection would be stale the moment someone switches agents. This makes the proposed hook mechanism fundamentally incompatible with OpenCode's multi-agent architecture.

   I believe this can be addressed by choosing the OC event that gives you as-close-to-CC-parity as possible.

   The blog's results are compelling, but I'd argue they're solving the wrong problem.

   I think you have a point, but limited in the following way. Unfortunately the LLM is not autonomous. It is conditioned on the user input. Hence the efficacy of the skill description will always be conditional on the same user input. The forced eval-hook should rarely lower hit rates, but how much it improves will always be a conditional value.

   You may be right that the author has inflated the hit-rate improvement by using skill descriptions that are tangential (weakly correlated) to the user input (and what the LLM interpolates and extrapolates from that input).

   Why I'm Resistant to Baking This In

   No complaints there. I agree no default hook should be distributed until CC does that.

   That's it. Put your forced-eval.md in .opencode/prompt-hooks/ and you're done.

   Hmmm, I use linux. But I could start by bootstrapping the kernel, then iterating over: "putting x in y" until I have a functional system..... or I could install a distro and use my time on things like this feature request.

   Happy to help you refine descriptions or debug your injection plugin if you go that route.

   When i find the time, or have a urgent need i'll re-open.
3. taqtiqa-mark

   closed this as completedon Dec 10, 2025