@jabworks/condux 0.1.1 → 0.2.0

package/README.md

@@ -1,13 +1,15 @@
 # @jabworks/condux
 
 OpenCode plugin for the [condux](https://github.com/jabworks/agentic-toolkit)
-workflow toolkit. Complements the condux skills (installed separately, see
-below) with the pieces OpenCode cannot pick up from a skill tree:
+workflow toolkit. It provides two pieces that OpenCode cannot pick up from a
+skill tree alone:
 
 - **Agents** — injects the four condux specialist agents (`coder`, `explorer`,
-  `planner`, `researcher`) as subagents via the `config` hook. If your
-  `opencode.json` already defines an agent with the same name, yours wins —
-  injection is skip-if-present.
+  `planner`, `researcher`) as subagents via the `config` hook, carrying the tool
+  restrictions their canonical definitions declare: `explorer` and `researcher`
+  are read-only (no `edit`, `write`, or `bash`) and `planner` cannot run shell
+  commands. If your `opencode.json` already defines an agent with the same name,
+  yours wins — injection is skip-if-present.
 - **Plan-review listener (opt-in)** — with `CONDUX_PLAN_REVIEW=1`, when the
   primary `plan` agent finishes a turn, spawns the plan-review annotate server
   from your installed condux skills. Best-effort: OpenCode does not await event
@@ -16,6 +18,21 @@ below) with the pieces OpenCode cannot pick up from a skill tree:
 
 ## Install
 
+### 1. Install skills
+
+Install the OpenCode-facing skill variants (trigger conditions folded into each
+description):
+
+```sh
+npx skills add https://github.com/jabworks/agentic-toolkit/tree/main/dist/opencode/skills -a opencode
+```
+
+This copies skills into `~/.config/opencode/skills/`.
+
+### 2. Add the plugin
+
+Add `@jabworks/condux` to your `opencode.json`:
+
 ```jsonc
 // opencode.json
 {
@@ -23,18 +40,20 @@ below) with the pieces OpenCode cannot pick up from a skill tree:
 }
 ```
 
-Install the condux skills alongside it (OpenCode-facing variants, with trigger
-conditions folded into each description):
-
-```sh
-npx skills add https://github.com/jabworks/agentic-toolkit/tree/main/dist/opencode/skills -a opencode
-```
+Restart OpenCode after adding the plugin.
 
 ## Notes
 
 - Injected agents carry no model pin — they inherit your session default. To
   restore the toolkit's tiering (e.g. a cheap model for `explorer`), override in
   `opencode.json`: `"agent": { "explorer": { "model": "anthropic/<model-id>" } }`.
+- Restrictions are expressed as OpenCode `permission` denials, not the deprecated
+  `tools` map — `tools` is folded into permissions while the config file is
+  parsed, which is over before a plugin's `config` hook runs, so a `tools` map
+  injected there has no effect. Only the `bash` and `edit` gates cross over
+  (`edit` covers edit/write/patch). Read-side tools keep OpenCode's defaults,
+  because the Claude allowlists omit `Grep`/`Glob` while the prompts still expect
+  to search — denying those would break the agents rather than constrain them.
 - The plan-review listener needs the `plan-review` skill installed in a
   discoverable skill tree (`.opencode/skills/`, `.agents/skills/`,
   `.claude/skills/`, or their global equivalents) — it runs

package/agents/explorer.md

@@ -1,6 +1,7 @@
 ---
 description: "Use this agent when you need to understand an unfamiliar codebase, trace types or symbols, identify module boundaries, discover conventions, map call graphs, or find where specific patterns are implemented — before planning or implementing any changes. Fire non-blocking when possible — delegate and continue while it traverses."
 mode: subagent
+permission: {"bash":"deny","edit":"deny"}
 ---
 
 You are a read-only codebase navigator and static-analysis specialist. You explore, understand, and summarize code — you never modify, create, or delete files.

package/agents/planner.md

@@ -1,6 +1,7 @@
 ---
 description: "Use this agent when planning a feature, architecture decision, or file structure before any implementation begins. Invoke whenever a task requires breaking down work into steps, evaluating architectural trade-offs, proposing folder/file structures, or producing ADR-style decision records — especially before delegating to the coder agent."
 mode: subagent
+permission: {"bash":"deny"}
 ---
 
 You are a senior engineering planner — a principal-level architect. You produce task breakdowns, ADR-style decision records, and file/folder proposals. You never write code; every output feeds a downstream `coder` delegation and must be precise enough to execute without ambiguity.

package/agents/researcher.md

@@ -1,6 +1,7 @@
 ---
 description: "Use this agent when any external library or framework API needs to be looked up before implementation. Invoke proactively whenever code will use a third-party dependency, an unfamiliar API method, or when exact parameter signatures, return types, or version-specific behavior must be confirmed. Fire non-blocking — delegate and continue working while it runs."
 mode: subagent
+permission: {"bash":"deny","edit":"deny"}
 ---
 
 You are a library and framework research specialist. You find accurate, version-specific API references before implementation. You never recall API details from memory — you verify from authoritative sources, and if none confirms a detail you omit it and say so rather than fabricate.

package/index.js

@@ -3,7 +3,8 @@
 // Injects the four condux specialist agents (coder, explorer, planner,
 // researcher) into the loaded config via the `config` hook, reading their
 // definitions from the bundled agents/*.md (generated from the canonical
-// Claude-dialect sources by scripts/build-opencode.mjs in the toolkit repo).
+// Claude-dialect sources by scripts/build-opencode.mjs in the toolkit repo),
+// including the tool restrictions those sources declare.
 // User-defined agents with the same name always win — injection is skip-if-present.
 //
 // Optional plan-review listener (CONDUX_PLAN_REVIEW=1): when the primary
@@ -22,13 +23,15 @@ const PKG_DIR = path.dirname(fileURLToPath(import.meta.url));
 const AGENTS_DIR = path.join(PKG_DIR, 'agents');
 
 // Bundled agent files use exactly the frontmatter the toolkit generator emits:
-// a JSON-quoted `description`, a plain `mode`, body = system prompt.
+// a JSON-quoted `description`, a plain `mode`, an optional `permission` deny map
+// (JSON object), body = system prompt.
 function parseAgent(text, name) {
-  const match = text.match(/^---\ndescription: (".*")\nmode: (\S+)\n---\n/);
+  const match = text.match(/^---\ndescription: (".*")\nmode: (\S+)\n(?:permission: (\{.*\})\n)?---\n/);
   if (!match) throw new Error(`condux-opencode: malformed bundled agent ${name}`);
   return {
     description: JSON.parse(match[1]),
     mode: match[2],
+    permission: match[3] ? JSON.parse(match[3]) : undefined,
     prompt: text.slice(match[0].length).trim(),
   };
 }
@@ -74,6 +77,13 @@ export const ConduxPlugin = async ({ worktree }) => {
           description: def.description,
           mode: def.mode,
           prompt: def.prompt,
+          // Only present for agents whose canonical definition restricts them
+          // (explorer/researcher are read-only, planner cannot run shell) —
+          // without it OpenCode's defaults grant bash/edit/write to all four.
+          // Must be `permission`, not the deprecated `tools`: `tools` is folded
+          // into permissions while the config file is parsed, which has already
+          // happened by the time this hook runs.
+          ...(def.permission ? { permission: def.permission } : {}),
         };
       }
     },

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@jabworks/condux",
-  "version": "0.1.1",
+  "version": "0.2.0",
   "description": "OpenCode plugin for the condux workflow toolkit — injects the condux specialist agents (coder, explorer, planner, researcher) and an opt-in plan-review listener.",
   "type": "module",
   "publishConfig": {