Add ralph-github: plan execution with GitHub PR pipeline#1
Open
DanielGGordon wants to merge 12 commits intomasterfrom
Open
Add ralph-github: plan execution with GitHub PR pipeline#1DanielGGordon wants to merge 12 commits intomasterfrom
DanielGGordon wants to merge 12 commits intomasterfrom
Conversation
New skill that extends Ralph with full GitHub integration: - Per-task branches (each task branches off previous) - Codex review after each task (falls back to Claude Opus 4.6) - Automatic PR creation per task (triggers bugbot) - Bugbot polling + examination (cursor[bot] comments) - Auto-fix of review/bugbot findings - PR merge + rebase pipeline between tasks - Final PR handling (waits for bugbot on last task) Runs as standalone bash script: bash ~/dotfiles/claude/skills/ralph-github/ralph-github.sh [plan.md] Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
1. Git push stdout corrupts PR number extraction — redirect to stderr 2. Sed non-greedy .+? doesn't work in POSIX ERE — use two-step sed 3. log_phase in run_review captured into review output — redirect to stderr 4. set -e kills script on skip/stop countdown — use || rc=$? pattern Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
grep -oP uses Perl regex which is unavailable on macOS BSD grep, causing pr_num to always be empty and disabling the entire pipeline.
1. Remove 2>&1 from gh pr create to prevent stderr contaminating pr_url (bug 39c4a506): gh writes progress messages to stderr which polluted PREV_PR_URL and degraded downstream prompt quality. 2. Call handle_previous_pr on exit-code failure path (bug b9c7e8c1): Previously, non-zero exit from run_claude skipped handle_previous_pr entirely, leaving the previous PR open and unmerged. 3. Clean up dangling task branch on exit-code failure (bug ffef6275): Checkout base branch and delete the failed task branch, consistent with the no-changes cleanup path.
When a task fails (exit code path), handle_previous_pr was called with PREV_BRANCH as the return branch. Since merge_pr uses --delete-branch, PREV_BRANCH gets deleted, causing rebase_on_main to crash when it tries to checkout the deleted branch. Additionally, PREV_PR/PREV_PR_URL/PREV_BRANCH were never cleared after merge, so the next iteration would try to branch off a deleted ref. Fix: pass MAIN_BRANCH as the return branch (since there's nothing to rebase after a failed task) and clear pipeline state after handling.
- Add --review flag with codex/claude reviewer (--reviewer auto|codex|claude) - Add --model presets (opus-high, sonnet, haiku, etc.) with effort levels - Fix codex review: drop prompt arg incompatible with --base flag - Fix has_review_issues: grep full output instead of just first 3 lines - Consolidate ralph-github.sh to thin wrapper calling ralph.sh --review - Update README, SKILL.md, statusline, .gitignore Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…matching, captured progress messages - Add --no-review case to argument parser so the documented flag works (previously fell through to catch-all, breaking ralph-github.sh workflow) - Narrow has_review_issues to check only last 5 lines of output and remove overly broad 'clean' pattern to prevent false negatives - Redirect log_phase calls in run_review to stderr so progress messages display to user instead of being captured into review_out variable
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 1 potential issue.
Autofix Details
Bugbot Autofix prepared a fix for the issue found in the latest run.
- ✅ Fixed:
--no-reviewcannot override wrapper's appended--reviewflag- Moved
--reviewbefore"$@"so user-supplied flags like--no-revieware processed last and can override the default.
- Moved
Four performance improvements to the main task loop: 1. parse_stream: use bash pattern matching to skip frequent streaming events (text deltas, message lifecycle) without forking jq. Only fork jq for infrequent events (tool use, result). Reduces process spawns from hundreds to ~10-30 per task. 2. count_tasks: replace bash while-read loop with two grep -c calls. 3. find_next_task: replace bash while-read loop with grep -n | head -1. 4. trim_plan_for_task: new function that sends only the plan preamble plus the current phase section in the prompt, cutting completed phases. Reduces input tokens and time-to-first-token for large plans. Also fixes stale plan context bug — PLAN_CONTENT was only refreshed every 3 tasks despite the plan changing every task. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Change '|| return' to '|| return 0' in find_next_task so grep's exit code 1 (no matches) doesn't propagate to callers under set -e. Plain assignments like 'var=$(find_next_task)' would crash the script instead of returning an empty string. - Restore final_text fallback in parse_stream by accumulating content_block_delta lines to a temp file (no per-event jq fork). When the result event's .result field is empty, extract text from the accumulated deltas. This ensures RESULT_TMPFILE is populated for follow-up detection even when .result is absent.
Use jq -j (join) instead of jq -r when reconstructing text from text_delta fragments. jq -r appends a newline after each output value, causing deltas like ['Hello ', 'world'] to produce 'Hello \nworld' instead of the correct 'Hello world'. jq -j suppresses these inter-value newlines, matching the behavior of the old direct string concatenation approach.
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 2 potential issues.
Autofix Details
Bugbot Autofix prepared fixes for both issues found in the latest run.
- ✅ Fixed: SIGPIPE causes
find_next_taskto silently return empty- Replaced
|| return 0with|| truefollowed by a separate[[ -z "$match" ]] && return 0check, so SIGPIPE from the grep|head pipeline no longer causes the function to discard a valid match.
- Replaced
- ✅ Fixed: Inconsistent JSON matching may miss result events
- Added alternate patterns with a space after the colon (
"type": "assistant"and"type": "result") to the substring checks so both compact and pretty-printed JSON formats are matched.
- Added alternate patterns with a space after the colon (
…_stream - find_next_task: Use '|| true' instead of '|| return 0' to prevent SIGPIPE (exit 141) from pipefail causing the function to return empty when grep|head-1 finds a match. Check for empty match separately. - parse_stream: Add '"type": "assistant"' and '"type": "result"' (with space after colon) as alternate patterns so JSON with whitespace after colons is still matched correctly.
| ${review_output} | ||
|
|
||
| ## Working Directory | ||
| $(pwd)" 2>/dev/null || true |
There was a problem hiding this comment.
Fix agent ignores user-specified model flags
Low Severity
The fix_review_issues function calls claude -p without passing CLAUDE_MODEL_FLAGS, while the main task execution in run_claude does pass them. When a user specifies --model opus-max, task execution uses Opus 4.6 with max effort, but the fix agent silently falls back to whatever the user's default claude model is — potentially a much less capable model for a step that requires understanding and correctly fixing code issues identified by the reviewer.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
/ralph-githubskill andralph-github.shbash script that extends Ralph with full GitHub integrationcursor[bot]) automaticallyTest plan
bash ~/dotfiles/claude/skills/ralph-github/ralph-github.sh --helpto verify argument parsing--dry-runon a sample plan to verify task detection and flow🤖 Generated with Claude Code
Note
Medium Risk
Medium risk because it changes the core
ralph.shexecution loop to run git auto-commits and automated review/fix steps, which can affect developer repos and workflow. Also adds model-selection flags and prompt trimming that could alter task behavior and context.Overview
Adds a new
skills/ralph-githubskill +ralph-github.shwrapper that runsralph.sh --review, and updates the README to document the new review-enabled workflow.Enhances
skills/ralph/ralph.shwith--review/--no-review, reviewer selection (codexvs Claude fallback), auto-commit-before-review, and an automatic “fix findings” pass; also adds--modelpresets/effort flags, trims plan context to the current##phase to reduce prompt size, and reduces per-task overhead by optimizing stream-json parsing.Updates
statusline-command.shto show model effort level, render a true-color gradient context bar, add a remote-control indicator, and simplify the directory segment; adds.gitignorefor the clonedskills/excalidraw-diagram/directory.Written by Cursor Bugbot for commit ebcdc1f. This will update automatically on new commits. Configure here.