sarman/tftsr-devops_investigation
1
1
Fork 0
Code Issues 4 Pull Requests 1 Actions 34 Packages Projects 3 Releases 6 Wiki Activity

ci(pr-review): iterative batch review for large diffs
Some checks failed
Test / frontend-tests (pull_request) Successful in 1m41s
Details
Test / frontend-typecheck (pull_request) Successful in 1m58s
Details
PR Review Automation / review (pull_request) Successful in 10m22s
Details
Test / rust-fmt-check (pull_request) Failing after 12m51s
Details
Test / rust-clippy (pull_request) Successful in 14m40s
Details
Test / rust-tests (pull_request) Successful in 16m29s
Details

Browse Source 
Replaces the single monolithic LLM call (which timed out on PRs >~150KB)
with a batching strategy:

- Changed files are grouped into batches capped at 2500 lines each,
  always respecting file boundaries (no file is split across batches)
- Each batch is sent as an independent LLM call (max-time 600s)
- Verdicts are aggregated: harshest verdict across all batches wins
- Single-batch PRs produce identical output to the old workflow
- Multi-batch PRs get a combined review with per-batch sections and
  an Overall Verdict line at the end
- Partial failures are tolerated: if some batches succeed, the review
  is posted with a note about any failed batches
- Cleanup removes /tmp/pr_batch_*.txt alongside existing temp files
This commit is contained in:
Shaun Arman 2026-06-20 19:54:00 -05:00
parent 9b94e99e4d
commit c8399bcdeb
1 changed files with 232 additions and 152 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

384
.gitea/workflows/pr-review.yml
View File

@ -31,7 +31,7 @@ jobs:
git fetch --depth=1 origin ${{ github.head_ref }} git fetch --depth=1 origin ${{ github.head_ref }}
git checkout FETCH_HEAD git checkout FETCH_HEAD
- name: Build review context - name: Build review batches
id: context id: context
shell: bash shell: bash
run: | run: |
@ -49,13 +49,10 @@ jobs:
if [ "$FILE_COUNT" -eq 0 ]; then if [ "$FILE_COUNT" -eq 0 ]; then
echo "No reviewable files changed." echo "No reviewable files changed."
echo "diff_size=0" >> $GITHUB_OUTPUT echo "diff_size=0" >> $GITHUB_OUTPUT
echo "batch_count=0" >> $GITHUB_OUTPUT
exit 0 exit 0
fi fi
# Build context: full file content for each changed file.
# Files <= 500 lines: include complete content.
# Files > 500 lines: include the per-file diff with generous context (±50 lines).
#
# Secret scrubbing: match actual credential VALUES only — known API key formats, # Secret scrubbing: match actual credential VALUES only — known API key formats,
# or keyword="long_quoted_literal" (25+ chars). Never scrub on keyword alone, # or keyword="long_quoted_literal" (25+ chars). Never scrub on keyword alone,
# which would silently delete function signatures, variable declarations, and tests. # which would silently delete function signatures, variable declarations, and tests.
@ -63,35 +60,53 @@ jobs:
# Only strip lines that are ENTIRELY a long base64 blob (e.g. PEM cert bodies) # Only strip lines that are ENTIRELY a long base64 blob (e.g. PEM cert bodies)
B64_PATTERN='^[[:space:]]*[A-Za-z0-9+/]{60,}={0,2}[[:space:]]*$' B64_PATTERN='^[[:space:]]*[A-Za-z0-9+/]{60,}={0,2}[[:space:]]*$'
> /tmp/pr_context.txt # Split changed files into batches capped at MAX_BATCH_LINES each.
# File boundaries are always respected — a file is never split across batches.
MAX_BATCH_LINES=2500
BATCH=1
BATCH_LINES=0
TOTAL_LINES=0
BATCH_FILE="/tmp/pr_batch_001.txt"
> "$BATCH_FILE"
while IFS= read -r file; do while IFS= read -r file; do
[ -f "$file" ] || continue [ -f "$file" ] || continue
lines=$(wc -l < "$file" | tr -d ' ') file_lines=$(wc -l < "$file" | tr -d ' ')
printf '\n════════ FILE: %s (%s lines) ════════\n' "$file" "$lines" >> /tmp/pr_context.txt
if [ "$lines" -le 500 ]; then # Build context for this individual file into a temp file
# Full file — model sees the complete implementation {
grep -v -E "$SECRET_PATTERN" "$file" \ printf '\n════════ FILE: %s (%s lines) ════════\n' "$file" "$file_lines"
| grep -v -E "$B64_PATTERN" \ if [ "$file_lines" -le 500 ]; then
>> /tmp/pr_context.txt || true grep -v -E "$SECRET_PATTERN" "$file" | grep -v -E "$B64_PATTERN" || true
else else
# Large file — emit annotated diff hunks (±50 lines of context each) printf '[File too large for full view (%s lines) — showing changed sections only]\n' "$file_lines"
printf '[File too large for full view (%s lines) — showing changed sections only]\n' "$lines" >> /tmp/pr_context.txt git diff -U50 origin/${{ github.base_ref }}..HEAD -- "$file" \
git diff -U50 origin/${{ github.base_ref }}..HEAD -- "$file" \ | grep -v -E "$SECRET_PATTERN" \
| grep -v -E "$SECRET_PATTERN" \ | grep -v -E "$B64_PATTERN" \
| grep -v -E "$B64_PATTERN" \ || true
>> /tmp/pr_context.txt || true fi
} > /tmp/fc_tmp.txt
FC_LINES=$(wc -l < /tmp/fc_tmp.txt | tr -d ' ')
# Start a new batch if this file would overflow the current one (and batch is not empty)
if [ "$BATCH_LINES" -gt 0 ] && [ $((BATCH_LINES + FC_LINES)) -gt $MAX_BATCH_LINES ]; then
BATCH=$((BATCH + 1))
BATCH_LINES=0
BATCH_FILE="/tmp/pr_batch_$(printf '%03d' $BATCH).txt"
> "$BATCH_FILE"
fi fi
cat /tmp/fc_tmp.txt >> "$BATCH_FILE"
BATCH_LINES=$((BATCH_LINES + FC_LINES))
TOTAL_LINES=$((TOTAL_LINES + FC_LINES))
done < /tmp/pr_files.txt done < /tmp/pr_files.txt
TOTAL=$(wc -l < /tmp/pr_context.txt | tr -d ' ') rm -f /tmp/fc_tmp.txt
echo "diff_size=${TOTAL}" >> $GITHUB_OUTPUT
# Cap at 6000 lines so we stay within the model's context window echo "batch_count=${BATCH}" >> $GITHUB_OUTPUT
if [ "$TOTAL" -gt 6000 ]; then echo "diff_size=${TOTAL_LINES}" >> $GITHUB_OUTPUT
head -n 6000 /tmp/pr_context.txt > /tmp/pr_context_capped.txt echo "Built ${BATCH} batch(es) from ${FILE_COUNT} files (${TOTAL_LINES} total lines)"
mv /tmp/pr_context_capped.txt /tmp/pr_context.txt
echo "[CONTEXT TRUNCATED at 6000 lines — ${TOTAL} total]" >> /tmp/pr_context.txt
fi
- name: Build codebase index - name: Build codebase index
id: index id: index
@ -175,7 +190,7 @@ jobs:
echo "comment_lines=${LINES}" >> $GITHUB_OUTPUT echo "comment_lines=${LINES}" >> $GITHUB_OUTPUT
echo "Fetched PR history: ${LINES} lines" echo "Fetched PR history: ${LINES} lines"
- name: Analyze with LLM - name: Analyze iteratively
id: analyze id: analyze
if: steps.context.outputs.diff_size != '0' if: steps.context.outputs.diff_size != '0'
shell: bash shell: bash
@ -185,137 +200,197 @@ jobs:
PR_TITLE: ${{ github.event.pull_request.title }} PR_TITLE: ${{ github.event.pull_request.title }}
PR_NUMBER: ${{ github.event.pull_request.number }} PR_NUMBER: ${{ github.event.pull_request.number }}
PR_BODY: ${{ github.event.pull_request.body }} PR_BODY: ${{ github.event.pull_request.body }}
BATCH_COUNT: ${{ steps.context.outputs.batch_count }}
run: | run: |
set -euo pipefail set -euo pipefail
CHANGED_FILES=$(tr '\n' ' ' < /tmp/pr_files.txt) CHANGED_FILES=$(tr '\n' ' ' < /tmp/pr_files.txt)
> /tmp/pr_all_findings.txt
OVERALL_VERDICT="APPROVE"
BATCH_SUCCESS=0
BATCH_FAILED=0
# Build prompt file following anthropics/claude-code code-review pattern: for i in $(seq 1 "$BATCH_COUNT"); do
# - Multi-agent review (parallel analysis) BATCH_FILE="/tmp/pr_batch_$(printf '%03d' $i).txt"
# - High-signal issues only (no nitpicks, style, or speculative concerns) [ -f "$BATCH_FILE" ] || continue
# - Validate findings against codebase
# - Consider PR title/description for author intent # Build the prompt for this batch
# - Check for pre-existing issues {
{ printf '%s\n\n' 'You are a senior engineer performing a code review following the anthropics/claude-code code-review pattern.'
printf '%s\n\n' 'You are a senior engineer performing a code review following the anthropics/claude-code code-review pattern.' printf 'PR Title: %s\n' "$PR_TITLE"
printf 'PR Title: %s\n' "$PR_TITLE" printf 'PR Body: %s\n\n' "${PR_BODY:-No description provided}"
printf 'PR Body: %s\n\n' "${PR_BODY:-No description provided}" printf 'Files changed: %s\n\n' "$CHANGED_FILES"
printf 'Files changed: %s\n\n' "$CHANGED_FILES" if [ "$BATCH_COUNT" -gt 1 ]; then
printf '%s\n' '---' printf 'NOTE: This is a large PR split into %s review batches. You are reviewing BATCH %s of %s.\n' "$BATCH_COUNT" "$i" "$BATCH_COUNT"
printf '%s\n\n' '## CODEBASE INDEX' printf 'Focus ONLY on the files shown in this batch. Do not speculate about files not included here.\n\n'
printf '%s\n' 'These are the ONLY Tauri commands, TypeScript exports, Rust public functions,' fi
printf '%s\n' 'and database tables that exist in this project. Before raising any finding,'
printf '%s\n' 'confirm that every symbol you cite appears in this list or in the file'
printf '%s\n' 'contents below. If it does not appear in either, your finding is fabricated.'
printf '%s\n' '---'
cat /tmp/codebase_index.txt
printf '%s\n\n' '---'
printf '%s\n\n' '## CHANGED FILE CONTENTS'
printf '%s\n' 'Each section is the COMPLETE, FINAL file after PR changes (not a diff).'
printf '%s\n\n' 'Files over 500 lines show only changed sections with surrounding context.'
printf '%s\n' '---'
cat /tmp/pr_context.txt
printf '%s\n\n' '---'
if [ -s /tmp/pr_comments.txt ]; then
cat /tmp/pr_comments.txt
printf '%s\n\n' '---'
printf '%s\n' '## CRITICAL: PRIOR REVIEW CONTEXT ABOVE'
printf '%s\n' 'Before raising ANY finding, check the review history above.'
printf '%s\n' 'SILENTLY DISCARD any finding that has already been:'
printf '%s\n' ' - Marked as invalid or incorrect by a reviewer'
printf '%s\n' ' - Acknowledged as an intentional design decision or known limitation'
printf '%s\n' ' - Confirmed fixed in a prior commit'
printf '%s\n\n' 'Raising a previously-refuted finding is a critical error.'
printf '%s\n' '---' printf '%s\n' '---'
fi cat /tmp/codebase_index.txt
printf '%s\n\n' '## CODE REVIEW INSTRUCTIONS' printf '%s\n\n' '---'
printf '%s\n\n' 'You MUST follow this workflow precisely:' printf '%s\n\n' '## CHANGED FILE CONTENTS (THIS BATCH)'
printf '%s\n\n' '1. LAUNCH 4 PARALLEL ANALYSIS AGENTS to independently review the changes:' printf '%s\n' 'Each section is the COMPLETE, FINAL file after PR changes (not a diff).'
printf '%s\n\n' ' AGENT 1 (CLAUDE.MD COMPLIANCE): Audit changes for CLAUDE.md compliance' printf '%s\n\n' 'Files over 500 lines show only changed sections with surrounding context.'
printf '%s\n' ' - Only consider CLAUDE.md files that share a file path with the file or parents' printf '%s\n' '---'
printf '%s\n' ' - Quote exact rules being violated' cat "$BATCH_FILE"
printf '%s\n\n' ' AGENT 2 (CLAUDE.MD COMPLIANCE): Audit changes for CLAUDE.md compliance' printf '%s\n\n' '---'
printf '%s\n' ' - Same scope as Agent 1, parallel analysis' if [ -s /tmp/pr_comments.txt ]; then
printf '%s\n\n' ' AGENT 3 (BUG DETECTOR): Scan for obvious bugs in the diff itself' cat /tmp/pr_comments.txt
printf '%s\n' ' - Focus ONLY on the diff, no extra context' printf '%s\n\n' '---'
printf '%s\n' ' - Flag ONLY significant bugs, ignore nitpicks and likely false positives' printf '%s\n' '## CRITICAL: PRIOR REVIEW CONTEXT ABOVE'
printf '%s\n' ' - Do not flag issues that require context outside the git diff' printf '%s\n' 'Before raising ANY finding, check the review history above.'
printf '%s\n\n' ' AGENT 4 (BUG DETECTOR): Look for problems in introduced code' printf '%s\n' 'SILENTLY DISCARD any finding that has already been:'
printf '%s\n' ' - Security issues, incorrect logic, data loss' printf '%s\n' ' - Marked as invalid or incorrect by a reviewer'
printf '%s\n' ' - Only problems that fall within the changed code' printf '%s\n' ' - Acknowledged as an intentional design decision or known limitation'
printf '%s\n\n' '2. CRITICAL: Only flag HIGH SIGNAL issues where:' printf '%s\n' ' - Confirmed fixed in a prior commit'
printf '%s\n' ' - Code will fail to compile or parse (syntax errors, type errors)' printf '%s\n\n' 'Raising a previously-refuted finding is a critical error.'
printf '%s\n' ' - Code will definitely produce wrong results (clear logic errors)' printf '%s\n' '---'
printf '%s\n' ' - Clear, unambiguous violations with exact rule quoted' fi
printf '%s\n\n' ' DO NOT flag:' printf '%s\n\n' '## CODE REVIEW INSTRUCTIONS'
printf '%s\n' ' - Code style or quality concerns' printf '%s\n\n' 'You MUST follow this workflow precisely:'
printf '%s\n' ' - Potential issues that depend on specific inputs or state' printf '%s\n\n' '1. LAUNCH 4 PARALLEL ANALYSIS AGENTS to independently review the changes:'
printf '%s\n' ' - Subjective suggestions or improvements' printf '%s\n\n' ' AGENT 1 (CLAUDE.MD COMPLIANCE): Audit changes for CLAUDE.md compliance'
printf '%s\n' ' - Pre-existing issues' printf '%s\n' ' - Only consider CLAUDE.md files that share a file path with the file or parents'
printf '%s\n' ' - Issues that linters will catch' printf '%s\n' ' - Quote exact rules being violated'
printf '%s\n' ' - General security issues unless explicitly required in CLAUDE.md' printf '%s\n\n' ' AGENT 2 (CLAUDE.MD COMPLIANCE): Audit changes for CLAUDE.md compliance'
printf '%s\n\n' '3. FOR EACH ISSUE FOUND BY AGENTS 3 & 4:' printf '%s\n' ' - Same scope as Agent 1, parallel analysis'
printf '%s\n' ' - Launch a VALIDATION AGENT to verify the issue is real' printf '%s\n\n' ' AGENT 3 (BUG DETECTOR): Scan for obvious bugs in the diff itself'
printf '%s\n' ' - Validation agent checks: issue is truly an issue, not false positive' printf '%s\n' ' - Focus ONLY on the diff, no extra context'
printf '%s\n' ' - Use full codebase to validate (not just diff)' printf '%s\n' ' - Flag ONLY significant bugs, ignore nitpicks and likely false positives'
printf '%s\n' ' - If validation fails, discard the issue silently' printf '%s\n' ' - Do not flag issues that require context outside the git diff'
printf '%s\n\n' '4. OUTPUT FORMAT (strict):' printf '%s\n\n' ' AGENT 4 (BUG DETECTOR): Look for problems in introduced code'
printf '%s\n\n' ' **Summary** (2-3 sentences)' printf '%s\n' ' - Security issues, incorrect logic, data loss'
printf '%s\n' ' **Findings**' printf '%s\n' ' - Only problems that fall within the changed code'
printf '%s\n' ' - [SEVERITY] file:line - description' printf '%s\n\n' '2. CRITICAL: Only flag HIGH SIGNAL issues where:'
printf '%s\n' ' Evidence: quoted line' printf '%s\n' ' - Code will fail to compile or parse (syntax errors, type errors)'
printf '%s\n\n' ' Fix: concrete change' printf '%s\n' ' - Code will definitely produce wrong results (clear logic errors)'
printf '%s\n\n' ' (Write "No findings." if none.)' printf '%s\n' ' - Clear, unambiguous violations with exact rule quoted'
printf '%s\n' ' **Verdict**: APPROVE / APPROVE WITH COMMENTS / REQUEST CHANGES' printf '%s\n\n' ' DO NOT flag:'
printf '%s\n\n' '5. SEVERITY DEFINITIONS:' printf '%s\n' ' - Code style or quality concerns'
printf '%s\n' ' - BLOCKER: fails to compile, corrupts data, or security vulnerability' printf '%s\n' ' - Potential issues that depend on specific inputs or state'
printf '%s\n' ' - WARNING: real risk to address before merge' printf '%s\n' ' - Subjective suggestions or improvements'
printf '%s\n' ' - SUGGESTION: minor improvement, follow-up PR fine' printf '%s\n' ' - Pre-existing issues'
printf '%s\n\n' '6. FOCUS AREAS:' printf '%s\n' ' - Issues that linters will catch'
printf '%s\n' ' - Security bugs, logic errors, data loss, injection, unhandled errors' printf '%s\n' ' - General security issues unless explicitly required in CLAUDE.md'
printf '%s\n\n' '7. IGNORE:' printf '%s\n\n' '3. FOR EACH ISSUE FOUND BY AGENTS 3 & 4:'
printf '%s\n' ' - Style, missing comments, speculative future concerns' printf '%s\n' ' - Launch a VALIDATION AGENT to verify the issue is real'
printf '%s\n\n' '8. FALSE POSITIVES TO AVOID:' printf '%s\n' ' - Validation agent checks: issue is truly an issue, not false positive'
printf '%s\n' ' - Pre-existing issues' printf '%s\n' ' - Use full codebase to validate (not just diff)'
printf '%s\n' ' - Something that appears buggy but is actually correct' printf '%s\n' ' - If validation fails, discard the issue silently'
printf '%s\n' ' - Pedantic nitpicks that senior engineers would not flag' printf '%s\n\n' '4. OUTPUT FORMAT (strict):'
printf '%s\n' ' - Issues that linters will catch' printf '%s\n\n' ' **Summary** (2-3 sentences)'
printf '%s\n' ' - General code quality concerns unless explicitly required in CLAUDE.md' printf '%s\n' ' **Findings**'
printf '%s\n' ' - Issues mentioned in CLAUDE.md but explicitly silenced in code' printf '%s\n' ' - [SEVERITY] file:line - description'
} > /tmp/prompt.txt printf '%s\n' ' Evidence: quoted line'
printf '%s\n\n' ' Fix: concrete change'
printf '%s\n\n' ' (Write "No findings." if none.)'
printf '%s\n' ' **Verdict**: APPROVE / APPROVE WITH COMMENTS / REQUEST CHANGES'
printf '%s\n\n' '5. SEVERITY DEFINITIONS:'
printf '%s\n' ' - BLOCKER: fails to compile, corrupts data, or security vulnerability'
printf '%s\n' ' - WARNING: real risk to address before merge'
printf '%s\n' ' - SUGGESTION: minor improvement, follow-up PR fine'
printf '%s\n\n' '6. FOCUS AREAS:'
printf '%s\n' ' - Security bugs, logic errors, data loss, injection, unhandled errors'
printf '%s\n\n' '7. IGNORE:'
printf '%s\n' ' - Style, missing comments, speculative future concerns'
printf '%s\n\n' '8. FALSE POSITIVES TO AVOID:'
printf '%s\n' ' - Pre-existing issues'
printf '%s\n' ' - Something that appears buggy but is actually correct'
printf '%s\n' ' - Pedantic nitpicks that senior engineers would not flag'
printf '%s\n' ' - Issues that linters will catch'
printf '%s\n' ' - General code quality concerns unless explicitly required in CLAUDE.md'
printf '%s\n' ' - Issues mentioned in CLAUDE.md but explicitly silenced in code'
} > /tmp/prompt_batch.txt
# Write body to file — passing 100KB+ JSON as a shell arg hits ARG_MAX. jq -cn \
jq -cn \ --arg model "qwen3.5-122b-think" \
--arg model "qwen3.5-122b-think" \ --rawfile content /tmp/prompt_batch.txt \
--rawfile content /tmp/prompt.txt \ '{model: $model, messages: [{role: "user", content: $content}], stream: false}' \
'{model: $model, messages: [{role: "user", content: $content}], stream: false}' \ > /tmp/body.json
> /tmp/body.json
echo "[$(date -u +%Y-%m-%dT%H:%M:%SZ)] PR #${PR_NUMBER} - Calling liteLLM API ($(wc -c < /tmp/body.json) bytes)..." echo "[$(date -u +%Y-%m-%dT%H:%M:%SZ)] PR #${PR_NUMBER} - Batch ${i}/${BATCH_COUNT} ($(wc -c < /tmp/body.json) bytes)..."
HTTP_CODE=$(curl -s --max-time 600 --connect-timeout 30 \
--retry 2 --retry-delay 15 --retry-connrefused --retry-max-time 1200 \ HTTP_CODE=$(curl -s --max-time 600 --connect-timeout 30 \
-o /tmp/llm_response.json -w "%{http_code}" \ --retry 2 --retry-delay 15 --retry-connrefused --retry-max-time 1200 \
-X POST "$LITELLM_URL/chat/completions" \ -o /tmp/llm_response.json -w "%{http_code}" \
-H "Authorization: Bearer $LITELLM_API_KEY" \ -X POST "$LITELLM_URL/chat/completions" \
-H "Content-Type: application/json" \ -H "Authorization: Bearer $LITELLM_API_KEY" \
--data @/tmp/body.json) -H "Content-Type: application/json" \
echo "HTTP status: $HTTP_CODE" --data @/tmp/body.json)
echo "Response file size: $(wc -c < /tmp/llm_response.json) bytes"
if [ "$HTTP_CODE" != "200" ]; then echo "Batch ${i} HTTP status: $HTTP_CODE"
echo "ERROR: liteLLM returned HTTP $HTTP_CODE" echo "Batch ${i} response size: $(wc -c < /tmp/llm_response.json) bytes"
cat /tmp/llm_response.json
if [ "$HTTP_CODE" != "200" ]; then
echo "ERROR: Batch ${i} failed (HTTP $HTTP_CODE)"
cat /tmp/llm_response.json
{
echo "## Batch ${i} of ${BATCH_COUNT}"
echo ""
echo "Review unavailable — LLM returned HTTP ${HTTP_CODE}."
echo ""
echo "---"
echo ""
} >> /tmp/pr_all_findings.txt
BATCH_FAILED=$((BATCH_FAILED + 1))
continue
fi
if ! jq empty /tmp/llm_response.json 2>/dev/null; then
echo "ERROR: Invalid JSON in batch ${i} response"
BATCH_FAILED=$((BATCH_FAILED + 1))
continue
fi
BATCH_REVIEW=$(jq -r '.choices[0].message.content // empty' /tmp/llm_response.json)
if [ -z "$BATCH_REVIEW" ]; then
echo "ERROR: Empty content in batch ${i} response"
BATCH_FAILED=$((BATCH_FAILED + 1))
continue
fi
echo "Batch ${i} review length: ${#BATCH_REVIEW} chars"
BATCH_SUCCESS=$((BATCH_SUCCESS + 1))
# Track harshest verdict across batches
if echo "$BATCH_REVIEW" | grep -q "REQUEST CHANGES"; then
OVERALL_VERDICT="REQUEST CHANGES"
elif echo "$BATCH_REVIEW" | grep -q "APPROVE WITH COMMENTS" && [ "$OVERALL_VERDICT" != "REQUEST CHANGES" ]; then
OVERALL_VERDICT="APPROVE WITH COMMENTS"
fi
if [ "$BATCH_COUNT" -eq 1 ]; then
echo "$BATCH_REVIEW" >> /tmp/pr_all_findings.txt
else
{
echo "## Batch ${i} of ${BATCH_COUNT}"
echo ""
echo "$BATCH_REVIEW"
echo ""
echo "---"
echo ""
} >> /tmp/pr_all_findings.txt
fi
done
if [ "$BATCH_SUCCESS" -eq 0 ]; then
echo "ERROR: All ${BATCH_COUNT} batches failed"
exit 1 exit 1
fi fi
if ! jq empty /tmp/llm_response.json 2>/dev/null; then
echo "ERROR: Invalid JSON response from liteLLM" # Assemble final review
cat /tmp/llm_response.json if [ "$BATCH_COUNT" -eq 1 ]; then
exit 1 cp /tmp/pr_all_findings.txt /tmp/pr_review.txt
else
{
echo "_This PR was reviewed in ${BATCH_COUNT} batches (${BATCH_FAILED} batch(es) failed)._"
echo ""
cat /tmp/pr_all_findings.txt
echo "---"
echo ""
echo "**Overall Verdict**: ${OVERALL_VERDICT}"
echo ""
echo "_Overall verdict reflects the most critical finding across all batches._"
} > /tmp/pr_review.txt
fi fi
REVIEW=$(jq -r '.choices[0].message.content // empty' /tmp/llm_response.json)
if [ -z "$REVIEW" ]; then
echo "ERROR: No content in liteLLM response"
exit 1
fi
echo "Review length: ${#REVIEW} chars"
echo "$REVIEW" > /tmp/pr_review.txt
- name: Verify findings against codebase - name: Verify findings against codebase
if: steps.analyze.outcome == 'success' if: steps.analyze.outcome == 'success'
@ -424,4 +499,9 @@ jobs:
- name: Cleanup - name: Cleanup
if: always() if: always()
shell: bash shell: bash
run: rm -f /tmp/pr_diff.txt /tmp/pr_context.txt /tmp/codebase_index.txt /tmp/pr_comments.txt /tmp/prompt.txt /tmp/body.json /tmp/llm_response.json /tmp/pr_review.txt /tmp/review_post_response.json /tmp/pr_files.txt run: |
rm -f /tmp/pr_files.txt /tmp/pr_context.txt /tmp/codebase_index.txt \
/tmp/pr_comments.txt /tmp/prompt_batch.txt /tmp/body.json \
/tmp/llm_response.json /tmp/pr_review.txt /tmp/pr_all_findings.txt \
/tmp/review_post_response.json /tmp/fc_tmp.txt
rm -f /tmp/pr_batch_*.txt