fix: have headless Claude write review report via Write tool

claude -p only captures the final text response to stdout, not
intermediate tool output. 7/9 reviews produced summaries instead of
full reports. Fix: instruct Claude to write the report to a specific
file path via the Write tool, which reliably captures the full content.

- Add {{REVIEW_OUTPUT_PATH}} placeholder to prompt template
- Add --review-output arg to hydrate-review-prompt.py
- Redirect claude stdout to .review.out (debug), report comes from file

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: ZviBaratz

scripts/field-test-runner.sh

@@ -387,6 +387,7 @@ if [[ "$OPT_REVIEW" == true || "$OPT_REVIEW_CHANGED" == true ]]; then
             --lint-json "$local_lint"
             --plugin-dir "$PLUGIN_DIR"
             --template "$REVIEW_TEMPLATE"
+            --review-output "$review_file"
         )
         [[ -f "$local_diff" ]] && hydrate_args+=(--diff-json "$local_diff")
         [[ -f "$local_ann" ]] && hydrate_args+=(--annotations "$local_ann")
@@ -403,13 +404,15 @@ if [[ "$OPT_REVIEW" == true || "$OPT_REVIEW_CHANGED" == true ]]; then
         echo "  Launching review: $name"
 
         # Launch claude -p in background subshell
+        # Claude writes the report to $review_file via Write tool (instructed in prompt).
+        # Stdout goes to .review.out for debugging; stderr to .review.err.
         (
             timeout 600 claude -p \
                 --plugin-dir "$PLUGIN_DIR" \
                 --add-dir "$ext_path" \
                 --dangerously-skip-permissions \
                 --max-budget-usd 2.00 \
-                "$prompt" > "$review_file" 2>"$err_file"
+                "$prompt" > "$RESULTS_DIR/$name.review.out" 2>"$err_file"
         ) &
         REVIEW_PIDS+=($!)
         REVIEW_NAMES+=("$name")

scripts/hydrate-review-prompt.py

@@ -30,6 +30,7 @@ def main():
     parser.add_argument("--plugin-dir", required=True, help="Plugin directory path")
     parser.add_argument("--diff-json", default="", help="Path to diff results JSON")
     parser.add_argument("--annotations", default="", help="Path to annotations YAML")
+    parser.add_argument("--review-output", default="", help="Path for review output file")
     parser.add_argument("--template", required=True, help="Path to prompt template")
     args = parser.parse_args()
 
@@ -80,6 +81,7 @@ def main():
     output = output.replace("{{DIFF_JSON_SECTION}}", diff_section)
     output = output.replace("{{ANNOTATIONS_SECTION}}", ann_section)
     output = output.replace("{{PLUGIN_DIR}}", args.plugin_dir)
+    output = output.replace("{{REVIEW_OUTPUT_PATH}}", args.review_output)
 
     sys.stdout.write(output)
 

scripts/review-prompt.md

@@ -1,6 +1,10 @@
 Run /ego-review on the extension at {{EXT_PATH}}.
 
-Your ENTIRE text output will be saved as the review report file. Do NOT summarize at the end — your full output IS the deliverable. Write the complete report inline as you go, with all six sections in full detail. Every word you output goes into the report.
+After completing the review, you MUST write the full report to a file using the Write tool:
+
+Write the report to: {{REVIEW_OUTPUT_PATH}}
+
+The report MUST contain ALL six sections in full detail. Do NOT summarize. Do NOT abbreviate. If a section has no findings, write "None" under it — do not omit the section.
 
 ## Phase 0 Override
 
@@ -14,11 +18,10 @@ Skip Phase 0 (ego-lint). It has already been run. Use these pre-computed results
 
 {{ANNOTATIONS_SECTION}}
 
-## Required Output
+## Required Report Format
 
-Write the FULL ego-review report with ALL sections below. Do NOT abbreviate or skip any section. If a section has no findings, write "None" under it — do not omit it.
+The report file MUST start with:
 
-Start your output with:
 ```
 ## EGO Review Report — {{NAME}}
 ```