Add .loop-build harness and IntelliJ project configuration files

.loop-build/prompts/executor.md

@@ -0,0 +1,34 @@
+# loop-build Executor Prompt
+
+## Inputs
+
+- `current_task`: contents of `.loop-build/state/current_task.json`
+- `step_id`: current step id
+- `context_snippets`: short file snippets around affected areas
+- `verify_summary`: latest verify summary if any
+
+## Task
+
+Produce a strictly scoped execution plan for exactly **one** step.
+
+Output must be either:
+1) a unified diff
+2) or a list of shell commands (`apply_patch`, `cat`, `cp`, etc.)
+
+No commands or patch may touch more than 2 files.
+
+If returning commands:
+- they are for the single step only
+- include only minimal verification command needed for this step
+- all non-read actions require explicit approval in harness policy
+
+Example accepted outputs:
+
+- `git diff` style block
+- command list with one command per line
+
+Disallowed:
+- Full file dumps
+- Multiple-step execution beyond the selected step
+- Extra unrelated refactors
+- Implicitly chaining future step actions

.loop-build/prompts/planner.md

@@ -0,0 +1,63 @@
+# loop-build Planner Prompt
+
+## Inputs
+
+- `goal`: user goal for the task
+- `constraints`: hard constraints and guardrails
+- `repo_context_hint`: optional repo context (size, language, architecture, risk areas)
+- `current_state_summary`: optional summary from previous run
+
+## Task
+
+Generate a strict JSON object for `TASKS` creation only (no implementation code).
+
+Output MUST be valid JSON with these top-level fields:
+
+- `task_id`
+- `goal`
+- `constraints`
+- `plan_steps`
+- `success_criteria`
+- `risk_notes`
+- `required_files`
+- `verify_targets`
+
+Each element in `plan_steps` must include:
+
+- `step_id`
+- `action`
+- `verification`
+- `rollback`
+- `risk_level`
+- `expected_output`
+
+Use only 3~7 steps.
+
+## Output format (example)
+
+```json
+{
+  "task_id": "task-2026-001",
+  "goal": "...",
+  "constraints": ["..."],
+  "plan_steps": [
+    {
+      "step_id": "1",
+      "action": "Read minimal context and add patch for file X",
+      "verification": "unit",
+      "rollback": "git checkout -- fileX || git checkout .",
+      "risk_level": "LOW",
+      "expected_output": "Target test starts passing or diff is constrained."
+    }
+  ],
+  "success_criteria": ["..."],
+  "risk_notes": ["..."],
+  "required_files": ["path/to/file1", "path/to/file2"],
+  "verify_targets": ["unit", "lint"]
+}
+```
+
+Rules:
+- Do not return implementation code or full diffs.
+- Keep each step actionable and minimal for one-step incremental execution.
+- One step must map to one incremental change and one verification check.

.loop-build/prompts/reviewer.md

@@ -0,0 +1,31 @@
+# loop-build Reviewer Prompt
+
+## Inputs
+
+- `diff_summary`: short unified diff summary
+- `verify_result`: result of the most recent `verify.sh` execution
+
+## Output JSON
+
+Return JSON with fields:
+
+- `risk_points`
+- `rollback_suggestions`
+- `approve_next` (true/false)
+- `next_step_recommendation`
+
+Optional context:
+- whether diff touched >2 files
+- whether verification passed
+- any open questions for the next step
+
+Example:
+
+```json
+{
+  "risk_points": ["..."],
+  "rollback_suggestions": ["git checkout -- path"],
+  "approve_next": true,
+  "next_step_recommendation": "continue"
+}
+```