gh-issues

name	gh-issues
description	Fetch GitHub issues, spawn sub-agents to implement fixes and open PRs, then monitor and address PR review comments. Usage: /gh-issues [owner/repo] [--label bug] [--limit 5] [--milestone v1.0] [--assignee @me] [--fork user/repo] [--watch] [--interval 5] [--reviews-only] [--cron] [--dry-run] [--model glm-5] [--notify-channel -1002381931352]
user-invocable	true
metadata	{"understudy":{"requires":{"bins":"[Truncated]"},"primaryEnv":"GH_TOKEN"}}

gh-issues — Auto-fix GitHub Issues with Parallel Sub-agents

You are an orchestrator. Follow these 6 phases exactly. Do not skip phases.

IMPORTANT — No gh CLI dependency. This skill uses curl + the GitHub REST API exclusively. When configured, Understudy should inject GH_TOKEN into the skill environment. Pass it as a Bearer token in all API calls:

curl -s -H "Authorization: Bearer $GH_TOKEN" -H "Accept: application/vnd.github+json" ...

---

Phase 1 — Parse Arguments

Parse the arguments string provided after /gh-issues.

Positional:

• owner/repo — optional. This is the source repo to fetch issues from. If omitted, detect from the current git remote: git remote get-url origin Extract owner/repo from the URL (handles both HTTPS and SSH formats).
  • HTTPS: https://github.com/owner/repo.git → owner/repo
  • SSH: [e-mail]:owner/repo.git → owner/repo If not in a git repo or no remote found, stop with an error asking the user to specify owner/repo.

Flags (all optional):

Flag	Default	Description
--label	(none)	Filter by label (e.g. bug, enhancement)
--limit	10	Max issues to fetch per poll
--milestone	(none)	Filter by milestone title
--assignee	(none)	Filter by assignee (@me for self)
--state	open	Issue state: open, closed, all
--fork	(none)	Your fork (user/repo) to push branches and open PRs from. Issues are fetched from the source repo; code is pushed to the fork; PRs are opened from the fork to the source repo.
--watch	false	Keep polling for new issues and PR reviews after each batch
--interval	5	Minutes between polls (only with --watch)
--dry-run	false	Fetch and display only — no sub-agents
--yes	false	Skip confirmation and auto-process all filtered issues
--reviews-only	false	Skip issue processing (Phases 2-5). Only run Phase 6 — check open PRs for review comments and address them.
--cron	false	Cron-safe mode: fetch issues and spawn sub-agents, exit without waiting for results.
--model	(none)	Model to use for sub-agents (e.g. glm-5, zai/glm-5). If not specified, uses the agent's default model.
--notify-channel	(none)	Telegram channel ID to send final PR summary to (e.g. -1002381931352). Only the final result with PR links is sent, not status updates.

Store parsed values for use in subsequent phases.

Derived values:

• SOURCE_REPO = the positional owner/repo (where issues live)
• PUSH_REPO = --fork value if provided, otherwise same as SOURCE_REPO
• FORK_MODE = true if --fork was provided, false otherwise

If --reviews-only is set: Skip directly to Phase 6. Run token resolution (from Phase 2) first, then jump to Phase 6.

If --cron is set:

• Force --yes (skip confirmation)
• If --reviews-only is also set, run token resolution then jump to Phase 6 (cron review mode)
• Otherwise, proceed normally through Phases 2-5 with cron-mode behavior active

---

Phase 2 — Fetch Issues

Token Resolution: First, ensure GH_TOKEN is available. Check environment:

echo $GH_TOKEN

If empty, stop and tell the user to provide a token. Understudy should expose GH_TOKEN directly or via the skill configuration.

Export as GH_TOKEN for subsequent commands:

export GH_TOKEN="<token>"

Build and run a curl request to the GitHub Issues API via exec:

curl -s -H "Authorization: Bearer $GH_TOKEN" -H "Accept: application/vnd.github+json" \
  "https://api.github.com/repos/{SOURCE_REPO}/issues?per_page={limit}&state={state}&{query_params}"

Where {query_params} is built from:

• labels={label} if --label was provided
• milestone={milestone} if --milestone was provided (note: API expects milestone number, so if user provides a title, first resolve it via GET /repos/{SOURCE_REPO}/milestones and match by title)
• assignee={assignee} if --assignee was provided (if @me, first resolve your username via GET /user)

IMPORTANT: The GitHub Issues API also returns pull requests. Filter them out — exclude any item where pull_request key exists in the response object.

If in watch mode: Also filter out any issue numbers already in the PROCESSED_ISSUES set from previous batches.

Error handling:

• If curl returns an HTTP 401 or 403 → stop and tell the user:

  "GitHub authentication failed. Please check your GH_TOKEN configuration in Understudy and retry."

• If the response is an empty array (after filtering) → report "No issues found matching filters" and stop (or loop back if in watch mode).
• If curl fails or returns any other error → report the error verbatim and stop.

Parse the JSON response. For each issue, extract: number, title, body, labels (array of label names), assignees, html_url.

---

Phase 3 — Present & Confirm

Display a markdown table of fetched issues:

#	Title	Labels
42	Fix null pointer in parser	bug, critical
37	Add retry logic for API calls	enhancement

If FORK_MODE is active, also display:

"Fork mode: branches will be pushed to {PUSH_REPO}, PRs will target {SOURCE_REPO}"

If --dry-run is active:

• Display the table and stop. Do not proceed to Phase 4.

If --yes is active:

• Display the table for visibility
• Auto-process ALL listed issues without asking for confirmation
• Proceed directly to Phase 4

Otherwise: Ask the user to confirm which issues to process:

• "all" — process every listed issue
• Comma-separated numbers (e.g. 42, 37) — process only those
• "cancel" — abort entirely

Wait for user response before proceeding.

Watch mode note: On the first poll, always confirm with the user (unless --yes is set). On subsequent polls, auto-process all new issues without re-confirming (the user already opted in). Still display the table so they can see what's being processed.

---

Phase 4 — Pre-flight Checks

Run these checks sequentially via exec:

1. Dirty working tree check:

   git status --porcelain
   

   If output is non-empty, warn the user:

   "Working tree has uncommitted changes. Sub-agents will create branches from HEAD — uncommitted changes will NOT be included. Continue?" Wait for confirmation. If declined, stop.

2. Record base branch:

   git rev-parse --abbrev-ref HEAD
   

   Store as BASE_BRANCH.

3. Verify remote access: If FORK_MODE:

   • Verify the fork remote exists. Check if a git remote named fork exists:

     git remote get-url fork
     

     If it doesn't exist, add it:

     git remote add fork https://x-access-token:$[e-mail]/{PUSH_REPO}.git
     

   • Also verify origin (the source repo) is reachable:

     git ls-remote --exit-code origin HEAD
     

   If not FORK_MODE:

   git ls-remote --exit-code origin HEAD
   

   If this fails, stop with: "Cannot reach remote origin. Check your network and git config."

4. Verify GH_TOKEN validity:

   curl -s -o /dev/null -w "%{http_code}" -H "Authorization: Bearer $GH_TOKEN" https://api.github.com/user
   

   If HTTP status is not 200, stop with:

   "GitHub authentication failed. Please check your GH_TOKEN configuration in Understudy and retry."

5. Check for existing PRs: For each confirmed issue number N, run:

   curl -s -H "Authorization: Bearer $GH_TOKEN" -H "Accept: application/vnd.github+json" \
     "https://api.github.com/repos/{SOURCE_REPO}/pulls?head={PUSH_REPO_OWNER}:fix/issue-{N}&state=open&per_page=1"
   

   (Where PUSH_REPO_OWNER is the owner portion of PUSH_REPO) If the response array is non-empty, remove that issue from the processing list and report:

   "Skipping #{N} — PR already exists: {html_url}"

   If all issues are skipped, report and stop (or loop back if in watch mode).

6. Check for in-progress branches (no PR yet = sub-agent still working): For each remaining issue number N (not already skipped by the PR check above), check if a fix/issue-{N} branch exists on the push repo (which may be a fork, not origin):

---

Phase 5 — Spawn Sub-agents (Parallel)

Cron mode (--cron is active):

• Sequential cursor tracking: Use a cursor file to track which issue to process next:

  UNDERSTUDY_STATE_DIR="${UNDERSTUDY_HOME:-$HOME/.understudy}/gh-issues"
  CURSOR_FILE="$UNDERSTUDY_STATE_DIR/cursor-{SOURCE_REPO_SLUG}.json"
  # SOURCE_REPO_SLUG = owner-repo with slashes replaced by hyphens (e.g., nicepkg-understudy)
  

  Read the cursor file (create if missing):

  if [ ! -f "$CURSOR_FILE" ]; then
    echo '{"last_processed": null, "in_progress": null}' > "$CURSOR_FILE"
  fi
  

  • last_processed: issue number of the last completed issue (or null if none)
  • in_progress: issue number currently being processed (or null)

• Select next issue: Filter the fetched issues list to find the first issue where:

  • Issue number > last_processed (if last_processed is set)
  • AND issue is not in the claims file (not already in progress)
  • AND no PR exists for the issue (checked in Phase 4 step 5)
  • AND no branch exists on the push repo (checked in Phase 4 step 6)

• If no eligible issue is found after the last_processed cursor, wrap around to the beginning (start from the oldest eligible issue).

• If an eligible issue is found:

  1. Mark it as in_progress in the cursor file
  2. Spawn a single sub-agent for that one issue with cleanup: "keep" and runTimeoutSeconds: 3600
  3. If --model was provided, include model: "{MODEL}" in the spawn config
  4. If --notify-channel was provided, include the channel in the task so the sub-agent can notify
  5. Do NOT await the sub-agent result — fire and forget
  6. Write claim: After spawning, read the claims file, add {SOURCE_REPO}#{N} with the current ISO timestamp, and write it back
  7. Immediately report: "Spawned fix agent for #{N} — will create PR when complete"
  8. Exit the skill. Do not proceed to Results Collection or Phase 6.

• If no eligible issue is found (all issues either have PRs, have branches, or are in progress), report "No eligible issues to process — all issues have PRs/branches or are in progress" and exit.

Normal mode (--cron is NOT active): For each confirmed issue, spawn a sub-agent using sessions_spawn. Launch up to 8 concurrently (matching subagents.maxConcurrent: 8). If more than 8 issues, batch them — launch the next agent as each completes.

Write claims: After spawning each sub-agent, read the claims file, add {SOURCE_REPO}#{N} with the current ISO timestamp, and write it back (same procedure as cron mode above). This covers interactive usage where watch mode might overlap with cron runs.

Sub-agent Task Prompt

For each issue, construct the following prompt and pass it to sessions_spawn. Variables to inject into the template:

• {SOURCE_REPO} — source repo where the issue lives
• {PUSH_REPO} — repo to push branches to (same as SOURCE_REPO unless fork mode)
• {FORK_MODE} — true/false
• {PUSH_REMOTE} — fork if FORK_MODE, otherwise origin
• {number}, {title}, {url}, {labels}, {body} — from the issue
• {BASE_BRANCH} — from Phase 4
• {notify_channel} — Telegram channel ID for notifications (empty if not set). Replace {notify_channel} in the template below with the value of --notify-channel flag (or leave as empty string if not provided).

When constructing the task, replace all template variables including {notify_channel} with actual values.

You are a focused code-fix agent. Your task is to fix a single GitHub issue and open a PR.

IMPORTANT: Do NOT use the gh CLI — it is not installed. Use curl with the GitHub REST API for all GitHub operations.

First, ensure GH_TOKEN is set. Check: `echo $GH_TOKEN`. If empty, stop and report that the token is missing; do not attempt interactive auth inside the sub-agent.

Use the token in all GitHub API calls:
curl -s -H "Authorization: Bearer $GH_TOKEN" -H "Accept: application/vnd.github+json" ...

<config>
Source repo (issues): {SOURCE_REPO}
Push repo (branches + PRs): {PUSH_REPO}
Fork mode: {FORK_MODE}
Push remote name: {PUSH_REMOTE}
Base branch: {BASE_BRANCH}
Notify channel: {notify_channel}
</config>

<issue>
Repository: {SOURCE_REPO}
Issue: #{number}
Title: {title}
URL: {url}
Labels: {labels}
Body: {body}
</issue>

<instructions>
Follow these steps in order. If any step fails, report the failure and stop.

0. SETUP — Ensure GH_TOKEN is available:

if [ -z "$GH_TOKEN" ]; then echo "GH_TOKEN is missing" exit 1 fi

Verify: echo "Token: ${GH_TOKEN:0:10}..."

1. CONFIDENCE CHECK — Before implementing, assess whether this issue is actionable:
- Read the issue body carefully. Is the problem clearly described?
- Search the codebase (grep/find) for the relevant code. Can you locate it?
- Is the scope reasonable? (single file/function = good, whole subsystem = bad)
- Is a specific fix suggested or is it a vague complaint?

Rate your confidence (1-10). If confidence < 7, STOP and report:
> "Skipping #{number}: Low confidence (score: N/10) — [reason: vague requirements | cannot locate code | scope too large | no clear fix suggested]"

Only proceed if confidence >= 7.

1. UNDERSTAND — Read the issue carefully. Identify what needs to change and where.

2. BRANCH — Create a feature branch from the base branch:
git checkout -b fix/issue-{number} {BASE_BRANCH}

3. ANALYZE — Search the codebase to find relevant files:
- Use grep/find via exec to locate code related to the issue
- Read the relevant files to understand the current behavior
- Identify the root cause

4. IMPLEMENT — Make the minimal, focused fix:
- Follow existing code style and conventions
- Change only what is necessary to fix the issue
- Do not add unrelated changes or new dependencies without justification

5. TEST — Discover and run the existing test suite if one exists:
- Look for package.json scripts, Makefile targets, pytest, cargo test, etc.
- Run the relevant tests
- If tests fail after your fix, attempt ONE retry with a corrected approach
- If tests still fail, report the failure

6. COMMIT — Stage and commit your changes:
git add {changed_files}
git commit -m "fix: {short_description}

Fixes {SOURCE_REPO}#{number}"

7. PUSH — Push the branch:
First, ensure the push remote uses token auth and disable credential helpers:
git config --global credential.helper ""
git remote set-url {PUSH_REMOTE} https://x-access-token:$[e-mail]/{PUSH_REPO}.git
Then push:
GIT_ASKPASS=true git push -u {PUSH_REMOTE} fix/issue-{number}

8. PR — Create a pull request using the GitHub API:

If FORK_MODE is true, the PR goes from your fork to the source repo:
- head = "{PUSH_REPO_OWNER}:fix/issue-{number}"
- base = "{BASE_BRANCH}"
- PR is created on {SOURCE_REPO}

If FORK_MODE is false:
- head = "fix/issue-{number}"
- base = "{BASE_BRANCH}"
- PR is created on {SOURCE_REPO}

curl -s -X POST \
  -H "Authorization: Bearer $GH_TOKEN" \
  -H "Accept: application/vnd.github+json" \
  https://api.github.com/repos/{SOURCE_REPO}/pulls \
  -d '{
    "title": "fix: {title}",
    "head": "{head_value}",
    "base": "{BASE_BRANCH}",
    "body": "## Summary\n\n{one_paragraph_description_of_fix}\n\n## Changes\n\n{bullet_list_of_changes}\n\n## Testing\n\n{what_was_tested_and_results}\n\nFixes {SOURCE_REPO}#{number}"
  }'

Extract the `html_url` from the response — this is the PR link.

9. REPORT — Send back a summary:
- PR URL (the html_url from step 8)
- Files changed (list)
- Fix summary (1-2 sentences)
- Any caveats or concerns

10. NOTIFY (if notify_channel is set) — If {notify_channel} is not empty, send a notification to the Telegram channel:

Use the message_send tool with:

• action: "send"
• channel: "telegram"
• recipient: "{notify_channel}"
• text: "✅ PR Created: {SOURCE_REPO}#{number}

{title}

{pr_url}

Files changed: {files_changed_list}"

</instructions>

<constraints>
- No force-push, no modifying the base branch
- No unrelated changes or gratuitous refactoring
- No new dependencies without strong justification
- If the issue is unclear or too complex to fix confidently, report your analysis instead of guessing
- Do NOT use the gh CLI — it is not available. Use curl + GitHub REST API for all GitHub operations.
- GH_TOKEN is already in the environment — do NOT prompt for auth
- Time limit: you have 60 minutes max. Be thorough — analyze properly, test your fix, don't rush.
</constraints>

Spawn configuration per sub-agent:

• runTimeoutSeconds: 3600 (60 minutes)
• cleanup: "keep" (preserve transcripts for review)
• If --model was provided, include model: "{MODEL}" in the spawn config

Timeout Handling

If a sub-agent exceeds 60 minutes, record it as:

"#{N} — Timed out (issue may be too complex for auto-fix)"

---

Results Collection

If --cron is active: Skip this section entirely — the orchestrator already exited after spawning in Phase 5.

After ALL sub-agents complete (or timeout), collect their results. Store the list of successfully opened PRs in OPEN_PRS (PR number, branch name, issue number, PR URL) for use in Phase 6.

Present a summary table:

Issue	Status	PR	Notes
#42 Fix null pointer	PR opened	https://github.com/.../pull/99	3 files changed
#37 Add retry logic	Failed	--	Could not identify target code
#15 Update docs	Timed out	--	Too complex for auto-fix
#8 Fix race condition	Skipped	--	PR already exists

Status values:

• PR opened — success, link to PR
• Failed — sub-agent could not complete (include reason in Notes)
• Timed out — exceeded 60-minute limit
• Skipped — existing PR detected in pre-flight

End with a one-line summary:

"Processed {N} issues: {success} PRs opened, {failed} failed, {skipped} skipped."

Send notification to channel (if --notify-channel is set): If --notify-channel was provided, send the final summary to that Telegram channel using the message_send tool:

Use the `message_send` tool with:
- action: "send"
- channel: "telegram"
- recipient: "{notify_channel}"
- text: "✅ GitHub Issues Processed

Processed {N} issues: {success} PRs opened, {failed} failed, {skipped} skipped.

{PR_LIST}"

Where PR_LIST includes only successfully opened PRs in format:
• #{issue_number}: {PR_url} ({notes})

Then proceed to Phase 6.

---

Phase 6 — PR Review Handler

This phase monitors open PRs (created by this skill or pre-existing fix/issue-* PRs) for review comments and spawns sub-agents to address them.

When this phase runs:

• After Results Collection (Phases 2-5 completed) — checks PRs that were just opened
• When --reviews-only flag is set — skips Phases 2-5 entirely, runs only this phase
• In watch mode — runs every poll cycle after checking for new issues

Cron review mode (--cron --reviews-only): When both --cron and --reviews-only are set:

1. Run token resolution (Phase 2 token section)
2. Discover open fix/issue-* PRs (Step 6.1)
3. Fetch review comments (Step 6.2)
4. Analyze comment content for actionability (Step 6.3)
5. If actionable comments are found, spawn ONE review-fix sub-agent for the first PR with unaddressed comments — fire-and-forget (do NOT await result)
   • Use cleanup: "keep" and runTimeoutSeconds: 3600
   • If --model was provided, include model: "{MODEL}" in the spawn config
6. Report: "Spawned review handler for PR #{N} — will push fixes when complete"
7. Exit the skill immediately. Do not proceed to Step 6.5 (Review Results).

If no actionable comments found, report "No actionable review comments found" and exit.

Normal mode (non-cron) continues below:

Step 6.1 — Discover PRs to Monitor

Collect PRs to check for review comments:

If coming from Phase 5: Use the OPEN_PRS list from Results Collection.

If --reviews-only or subsequent watch cycle: Fetch all open PRs with fix/issue- branch pattern:

curl -s -H "Authorization: Bearer $GH_TOKEN" -H "Accept: application/vnd.github+json" \
  "https://api.github.com/repos/{SOURCE_REPO}/pulls?state=open&per_page=100"

Filter to only PRs where head.ref starts with fix/issue-.

For each PR, extract: number (PR number), head.ref (branch name), html_url, title, body.

If no PRs found, report "No open fix/ PRs to monitor" and stop (or loop back if in watch mode).

Step 6.2 — Fetch All Review Sources

For each PR, fetch reviews from multiple sources:

Fetch PR reviews:

curl -s -H "Authorization: Bearer $GH_TOKEN" -H "Accept: application/vnd.github+json" \
  "https://api.github.com/repos/{SOURCE_REPO}/pulls/{pr_number}/reviews"

Fetch PR review comments (inline/file-level):

curl -s -H "Authorization: Bearer $GH_TOKEN" -H "Accept: application/vnd.github+json" \
  "https://api.github.com/repos/{SOURCE_REPO}/pulls/{pr_number}/comments"

Fetch PR issue comments (general conversation):

curl -s -H "Authorization: Bearer $GH_TOKEN" -H "Accept: application/vnd.github+json" \
  "https://api.github.com/repos/{SOURCE_REPO}/issues/{pr_number}/comments"

Fetch PR body for embedded reviews: Some review tools (like Greptile) embed their feedback directly in the PR body. Check for:

• <!-- greptile_comment --> markers
• Other structured review sections in the PR body

curl -s -H "Authorization: Bearer $GH_TOKEN" -H "Accept: application/vnd.github+json" \
  "https://api.github.com/repos/{SOURCE_REPO}/pulls/{pr_number}"

Extract the body field and parse for embedded review content.

Step 6.3 — Analyze Comments for Actionability

Determine the bot's own username for filtering:

curl -s -H "Authorization: Bearer $GH_TOKEN" https://api.github.com/user | jq -r '.login'

Store as BOT_USERNAME. Exclude any comment where user.login equals BOT_USERNAME.

For each comment/review, analyze the content to determine if it requires action:

NOT actionable (skip):

• Pure approvals or "LGTM" without suggestions
• Bot comments that are informational only (CI status, auto-generated summaries without specific requests)
• Comments already addressed (check if bot replied with "Addressed in commit...")
• Reviews with state APPROVED and no inline comments requesting changes

IS actionable (requires attention):

• Reviews with state CHANGES_REQUESTED
• Reviews with state COMMENTED that contain specific requests:
  • "this test needs to be updated"
  • "please fix", "change this", "update", "can you", "should be", "needs to"
  • "will fail", "will break", "causes an error"
  • Mentions of specific code issues (bugs, missing error handling, edge cases)
• Inline review comments pointing out issues in the code
• Embedded reviews in PR body that identify:
  • Critical issues or breaking changes
  • Test failures expected
  • Specific code that needs attention
  • Confidence scores with concerns

Parse embedded review content (e.g., Greptile): Look for sections marked with <!-- greptile_comment --> or similar. Extract:

• Summary text
• Any mentions of "Critical issue", "needs attention", "will fail", "test needs to be updated"
• Confidence scores below 4/5 (indicates concerns)

Build actionable_comments list with:

• Source (review, inline comment, PR body, etc.)
• Author
• Body text
• For inline: file path and line number
• Specific action items identified

If no actionable comments found across any PR, report "No actionable review comments found" and stop (or loop back if in watch mode).

Step 6.4 — Present Review Comments

Display a table of PRs with pending actionable comments:

| PR | Branch | Actionable Comments | Sources |
|----|--------|---------------------|---------|
| #99 | fix/issue-42 | 2 comments | @reviewer1, greptile |
| #101 | fix/issue-37 | 1 comment | @reviewer2 |

If --yes is NOT set and this is not a subsequent watch poll: ask the user to confirm which PRs to address ("all", comma-separated PR numbers, or "skip").

Step 6.5 — Spawn Review Fix Sub-agents (Parallel)

For each PR with actionable comments, spawn a sub-agent. Launch up to 8 concurrently.

Review fix sub-agent prompt:

You are a PR review handler agent. Your task is to address review comments on a pull request by making the requested changes, pushing updates, and replying to each comment.

IMPORTANT: Do NOT use the gh CLI — it is not installed. Use curl with the GitHub REST API for all GitHub operations.

First, ensure GH_TOKEN is set. Check: echo $GH_TOKEN. If empty, stop and report that the token is missing; do not attempt interactive auth inside the sub-agent.

<config>
Repository: {SOURCE_REPO}
Push repo: {PUSH_REPO}
Fork mode: {FORK_MODE}
Push remote: {PUSH_REMOTE}
PR number: {pr_number}
PR URL: {pr_url}
Branch: {branch_name}
</config>

<review_comments>
{json_array_of_actionable_comments}

Each comment has:
- id: comment ID (for replying)
- user: who left it
- body: the comment text
- path: file path (for inline comments)
- line: line number (for inline comments)
- diff_hunk: surrounding diff context (for inline comments)
- source: where the comment came from (review, inline, pr_body, greptile, etc.)
</review_comments>

<instructions>
Follow these steps in order:

0. SETUP — Ensure GH_TOKEN is available:

if [ -z "$GH_TOKEN" ]; then echo "GH_TOKEN is missing" exit 1 fi

Verify: echo "Token: ${GH_TOKEN:0:10}..."

1. CHECKOUT — Switch to the PR branch:
git fetch {PUSH_REMOTE} {branch_name}
git checkout {branch_name}
git pull {PUSH_REMOTE} {branch_name}

2. UNDERSTAND — Read ALL review comments carefully. Group them by file. Understand what each reviewer is asking for.

3. IMPLEMENT — For each comment, make the requested change:
- Read the file and locate the relevant code
- Make the change the reviewer requested
- If the comment is vague or you disagree, still attempt a reasonable fix but note your concern
- If the comment asks for something impossible or contradictory, skip it and explain why in your reply

4. TEST — Run existing tests to make sure your changes don't break anything:
- If tests fail, fix the issue or revert the problematic change
- Note any test failures in your replies

5. COMMIT — Stage and commit all changes in a single commit:
git add {changed_files}
git commit -m "fix: address review comments on PR #{pr_number}

Addresses review feedback from {reviewer_names}"

6. PUSH — Push the updated branch:
git config --global credential.helper ""
git remote set-url {PUSH_REMOTE} https://x-access-token:$[e-mail]/{PUSH_REPO}.git
GIT_ASKPASS=true git push {PUSH_REMOTE} {branch_name}

7. REPLY — For each addressed comment, post a reply:

For inline review comments (have a path/line), reply to the comment thread:
curl -s -X POST \
  -H "Authorization: Bearer $GH_TOKEN" \
  -H "Accept: application/vnd.github+json" \
  https://api.github.com/repos/{SOURCE_REPO}/pulls/{pr_number}/comments/{comment_id}/replies \
  -d '{"body": "Addressed in commit {short_sha} — {brief_description_of_change}"}'

For general PR comments (issue comments), reply on the PR:
curl -s -X POST \
  -H "Authorization: Bearer $GH_TOKEN" \
  -H "Accept: application/vnd.github+json" \
  https://api.github.com/repos/{SOURCE_REPO}/issues/{pr_number}/comments \
  -d '{"body": "Addressed feedback from @{reviewer}:\n\n{summary_of_changes_made}\n\nUpdated in commit {short_sha}"}'

For comments you could NOT address, reply explaining why:
"Unable to address this comment: {reason}. This may need manual review."

8. REPORT — Send back a summary:
- PR URL
- Number of comments addressed vs skipped
- Commit SHA
- Files changed
- Any comments that need manual attention
</instructions>

<constraints>
- Only modify files relevant to the review comments
- Do not make unrelated changes
- Do not force-push — always regular push
- If a comment contradicts another comment, address the most recent one and flag the conflict
- Do NOT use the gh CLI — use curl + GitHub REST API
- GH_TOKEN is already in the environment — do not prompt for auth
- Time limit: 60 minutes max
</constraints>

Spawn configuration per sub-agent:

• runTimeoutSeconds: 3600 (60 minutes)
• cleanup: "keep" (preserve transcripts for review)
• If --model was provided, include model: "{MODEL}" in the spawn config

Step 6.6 — Review Results

After all review sub-agents complete, present a summary:

| PR | Comments Addressed | Comments Skipped | Commit | Status |
|----|-------------------|-----------------|--------|--------|
| #99 fix/issue-42 | 3 | 0 | abc123f | All addressed |
| #101 fix/issue-37 | 1 | 1 | def456a | 1 needs manual review |

Add comment IDs from this batch to ADDRESSED_COMMENTS set to prevent re-processing.

---

Watch Mode (if --watch is active)

After presenting results from the current batch:

1. Add all issue numbers from this batch to the running set PROCESSED_ISSUES.
2. Add all addressed comment IDs to ADDRESSED_COMMENTS.
3. Tell the user:

   "Next poll in {interval} minutes... (say 'stop' to end watch mode)"

4. Sleep for {interval} minutes.
5. Go back to Phase 2 — Fetch Issues. The fetch will automatically filter out:
   • Issues already in PROCESSED_ISSUES
   • Issues that have existing fix/issue-{N} PRs (caught in Phase 4 pre-flight)
6. After Phases 2-5 (or if no new issues), run Phase 6 to check for new review comments on ALL tracked PRs (both newly created and previously opened).
7. If no new issues AND no new actionable review comments → report "No new activity. Polling again in {interval} minutes..." and loop back to step 4.
8. The user can say "stop" at any time to exit watch mode. When stopping, present a final cumulative summary of ALL batches — issues processed AND review comments addressed.

Context hygiene between polls — IMPORTANT: Only retain between poll cycles:

• PROCESSED_ISSUES (set of issue numbers)
• ADDRESSED_COMMENTS (set of comment IDs)
• OPEN_PRS (list of tracked PRs: number, branch, URL)
• Cumulative results (one line per issue + one line per review batch)
• Parsed arguments from Phase 1
• BASE_BRANCH, SOURCE_REPO, PUSH_REPO, FORK_MODE, BOT_USERNAME Do NOT retain issue bodies, comment bodies, sub-agent transcripts, or codebase analysis between polls.