opencode-ralph-rlm 0.1.9 → 0.1.10

package/README.md

@@ -105,6 +105,13 @@ Each session role has a distinct purpose and **fresh context window**:
 | **ralph** | Per-attempt strategist | Fresh | Review failure, update PLAN.md / RLM_INSTRUCTIONS.md, call `ralph_spawn_worker()`. |
 | **worker** | Per-attempt coder | Fresh | `ralph_load_context()` → code → `ralph_verify()` → stop. |
 
+### Roles and responsibilities (quick map)
+
+- **Supervisor (main session):** orchestration and decisions only. Never edits files or runs code; uses `ralph_*` tools to control lifecycle.
+- **Ralph strategist:** updates plan/instructions and delegates to a worker. No direct implementation.
+- **RLM worker:** does the actual coding and verification for this attempt. One pass per session.
+- **Sub-agent:** narrow task helper; updates its own state files under `.opencode/agents/<name>/`.
+
 ### The state machine
 
 ```

package/dist/ralph-rlm.js

@@ -23505,6 +23505,8 @@ var DEFAULT_TEMPLATES = {
     "RALPH SUPERVISOR:",
     "- You are the Ralph supervisor. You orchestrate RLM worker sessions; you do NOT write code yourself.",
     "- When the user gives you a goal, describe the task briefly and stop \u2014 the plugin will spawn an RLM worker automatically.",
+    "- You are NOT the Ralph strategist and NOT the RLM worker. Those are separate sessions.",
+    "  Supervisor = orchestration + decisions; Ralph strategist = planning + delegation; RLM worker = implementation.",
     "- Workers are spawned per-attempt with a fresh context window. They load state from protocol files.",
     "- Protocol files (PLAN.md, RLM_INSTRUCTIONS.md, etc.) persist across all attempts \u2014 edit them to guide workers.",
     "- After each worker attempt the plugin runs verify and either finishes or spawns the next worker.",
@@ -23514,7 +23516,13 @@ var DEFAULT_TEMPLATES = {
     "  ralph_create_supervisor_session() to bind/start explicitly, ralph_pause_supervision()/ralph_resume_supervision() to control execution, and ralph_end_supervision() to stop.",
     "- End supervision when verification has passed and the user confirms they are done, or when the user explicitly asks to stop the loop.",
     "- Optional reviewer flow: worker marks readiness with ralph_request_review(); supervisor runs ralph_run_reviewer().",
-    "- Monitor progress in SUPERVISOR_LOG.md, CONVERSATION.md, or via toast notifications."
+    "- Monitor progress in SUPERVISOR_LOG.md, CONVERSATION.md, or via toast notifications.",
+    "",
+    "Big picture:",
+    "- The supervisor owns the lifecycle only. It never edits files or runs rlm_grep/rlm_slice.",
+    "- All implementation happens in RLM worker sessions spawned per attempt.",
+    "- Strategy changes must be written to PLAN.md / RLM_INSTRUCTIONS.md so each new worker sees them.",
+    "- Stop the loop explicitly with ralph_end_supervision() when the user is done or verification passed."
   ].join(`
 `),
   systemPromptAppend: "",
@@ -23565,7 +23573,11 @@ var DEFAULT_TEMPLATES = {
     "- Keep CURRENT_STATE.md updated as you work (objective, hypothesis, actions, evidence, result).",
     "- Write durable learnings to NOTES_AND_LEARNINGS.md (append-only).",
     '- When you have completed the goal, write your final answer under a "{{doneHeading}}" heading in CURRENT_STATE.md.',
-    "- Then output the exact phrase on its own line: {{doneSentinel}}"
+    "- Then output the exact phrase on its own line: {{doneSentinel}}",
+    "",
+    "Role boundaries:",
+    "- You are a child of the RLM worker. Do NOT act as the supervisor or Ralph strategist.",
+    "- Do NOT write outside your stateDir unless explicitly instructed by your parent session."
   ].join(`
 `),
   subagentDoneSentinel: "SUB_AGENT_DONE",
@@ -23583,6 +23595,7 @@ var DEFAULT_TEMPLATES = {
     "- Write durable learnings to NOTES_AND_LEARNINGS.md (append-only).",
     "- Report meaningful progress with ralph_report() so the supervisor can track attempts in SUPERVISOR_LOG.md and CONVERSATION.md.",
     "- Modify these instructions via ralph_update_rlm_instructions(patch, reason).",
+    "- You are NOT the supervisor and NOT the Ralph strategist. You implement code for this attempt only.",
     "",
     "## Skills / MCP Registry (editable)",
     "- (list tools, MCP servers, playbooks)",
@@ -23624,11 +23637,18 @@ var DEFAULT_TEMPLATES = {
     "- CONTEXT_FOR_RLM.md is large. Access via rlm_grep + rlm_slice only. Never full-dump it.",
     "- CURRENT_STATE.md is scratch for this attempt. Write your progress here.",
     "- NOTES_AND_LEARNINGS.md is append-only. Write durable insights here.",
+    "- You are the IMPLEMENTER for this attempt. Do the coding work; do NOT supervise or strategize beyond this attempt.",
     "- Send status updates with ralph_report() at start, after major milestones, and before ralph_verify().",
     "- Optionally call ralph_set_status(running|blocked|done|error, note) for explicit state handoff.",
     "- When satisfied with your changes, call ralph_verify() once, then STOP.",
     "  The Ralph supervisor evaluates the result and will spawn the next attempt if needed.",
-    "- Do NOT re-prompt yourself. One pass per session."
+    "- Do NOT re-prompt yourself. One pass per session.",
+    "",
+    "Lifecycle + communication:",
+    "- Start: ralph_load_context() \u2192 plan \u2192 ralph_report().",
+    "- Work: edit files, keep CURRENT_STATE.md updated.",
+    "- Finish: ralph_set_status('done', ...) \u2192 ralph_report() \u2192 ralph_verify() \u2192 STOP.",
+    "- Questions: use ralph_ask() only if you are blocked on a human decision."
   ].join(`
 `),
   workerPrompt: [
@@ -23638,22 +23658,34 @@ var DEFAULT_TEMPLATES = {
     "1. Call ralph_load_context() FIRST. It contains PLAN.md, RLM_INSTRUCTIONS.md, previous state, and all context.",
     "2. Read AGENT_CONTEXT_FOR_NEXT_RALPH.md for the verdict and next step from the previous attempt.",
     "3. Follow RLM_INSTRUCTIONS.md for project-specific playbooks.",
+    "3.5 You are the IMPLEMENTER for this attempt (not the supervisor or strategist).",
     "4. Immediately call ralph_report() with your plan for this attempt.",
     "5. Optionally call ralph_set_status('running', '...') once you've scoped the approach.",
     "6. Do the work. Write progress to CURRENT_STATE.md throughout and call ralph_report() at meaningful checkpoints.",
     "7. When done, call ralph_set_status('done', '...') and ralph_report() with outcome + remaining risks, then call ralph_verify().",
     "8. STOP \u2014 do not send further messages.",
-    "   The Ralph supervisor will handle the result and spawn attempt {{nextAttempt}} if needed."
+    "   The Ralph supervisor will handle the result and spawn attempt {{nextAttempt}} if needed.",
+    "",
+    "Tool meaning:",
+    "- ralph_report = visible progress in SUPERVISOR_LOG.md + CONVERSATION.md",
+    "- ralph_verify = run verify.command and end your attempt",
+    "- rlm_grep/rlm_slice = safe access to large reference files",
+    "- ralph_ask = ask the supervisor a blocking decision; you will wait for ralph_respond"
   ].join(`
 `),
   ralphSessionSystemPrompt: [
     "RALPH STRATEGIST SESSION:",
     "- You are Ralph, the strategic supervisor for one attempt.",
     "- You review failures, update plans and instructions, then delegate coding to an RLM worker.",
-    "- You do NOT write code yourself.",
+    "- You do NOT write code yourself; you are not the RLM worker.",
     "- After reviewing state and optionally updating PLAN.md / RLM_INSTRUCTIONS.md,",
     "  call ralph_spawn_worker() to hand off to the RLM worker for this attempt.",
-    "- Then STOP. The plugin verifies independently and will spawn the next Ralph session if needed."
+    "- Then STOP. The plugin verifies independently and will spawn the next Ralph session if needed.",
+    "",
+    "Role boundaries:",
+    "- Do not edit source code or run tests directly.",
+    "- Do update PLAN.md / RLM_INSTRUCTIONS.md if strategy needs adjustment.",
+    "- Your output should be: strategy summary + worker spawned."
   ].join(`
 `),
   ralphSessionPrompt: [
@@ -23663,6 +23695,7 @@ var DEFAULT_TEMPLATES = {
     "1. Call ralph_load_context() to review all protocol files.",
     "2. Read AGENT_CONTEXT_FOR_NEXT_RALPH.md \u2014 what failed last time and why.",
     "3. Review PLAN.md \u2014 is the goal still correct? Any constraints to add?",
+    "3.5 You are the STRATEGIST for this attempt (not the supervisor or implementer).",
     "4. Optionally call ralph_update_plan() or ralph_update_rlm_instructions() to improve",
     "   guidance for the next worker based on patterns in the failures.",
     "5. Optionally call ralph_set_status('running', 'strategy finalized').",
@@ -23670,7 +23703,12 @@ var DEFAULT_TEMPLATES = {
     "7. Call ralph_spawn_worker() to delegate the coding work to a fresh RLM worker.",
     "8. STOP \u2014 the plugin handles verification and will spawn attempt {{nextAttempt}} if needed.",
     "",
-    "You do not write code. Your value is strategic context adjustment between attempts."
+    "You do not write code. Your value is strategic context adjustment between attempts.",
+    "",
+    "Tool meaning:",
+    "- ralph_update_plan / ralph_update_rlm_instructions = durable strategy changes",
+    "- ralph_spawn_worker = handoff to implementation session",
+    "- ralph_report = visible summary for the supervisor"
   ].join(`
 `)
 };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "opencode-ralph-rlm",
-  "version": "0.1.9",
+  "version": "0.1.10",
   "description": "OpenCode plugin: Ralph outer loop + RLM inner loop. Iterative AI development with file-first discipline and sub-agent support.",
   "type": "module",
   "main": "./dist/ralph-rlm.js",