talont-org/open-integration-pr

Open Integration PR

Read pr_order_file JSON produced by analyze-prs, generate a rich PR description (per-PR details, complexity tags, merge outcomes, optional audit verdict), generate 1–3 arch-lens diagrams, carry forward Closes #N/Fixes #N references from original PR bodies so linked issues auto-close on merge, create the integration PR via gh pr create --body-file, close each collapsed PR with a comment referencing the new PR, and output pr_url=<url>.

Arguments

/autoskillit:open-integration-pr {batch_branch} {base_branch} {pr_order_file} [audit_verdict] [conflict_report_paths] [domain_partitions_path]

• batch_branch — integration branch name (e.g. pr-batch/pr-merge-20250228-143052)
• base_branch — PR target branch (e.g. main)
• pr_order_file — absolute path to JSON produced by analyze-prs
• audit_verdict (optional) — GO, NO GO, or empty string when audit was skipped
• conflict_report_paths (optional) — comma-separated list of absolute paths to conflict resolution report files, produced by resolve-merge-conflicts. When provided and non-empty, embed a "Conflict Resolution Decisions" section in the PR body.
• domain_partitions_path (optional) — absolute path to a JSON file containing pre-computed domain partitions (produced by compute_domain_partitions run_python step). When provided and present, read from file instead of computing via python3.

When to Use

• Called by merge-prs after all PRs have been merged into the integration branch
• Invoked via run_skill as the final step before CI watch

Critical Constraints

NEVER:

• Create files outside {{AUTOSKILLIT_TEMP}}/open-integration-pr/ (except the temp body file for gh pr create --body-file)
• Modify any source code
• Fail the pipeline if gh is unavailable or not authenticated — output pr_url= (empty) and exit successfully
• Close original PRs before the integration PR is successfully created
• Run subagents in the background (run_in_background: true is prohibited)

ALWAYS:

• Check gh auth status before attempting any GitHub operations
• Output pr_url=<url> on the last output line (empty string when GitHub unavailable)
• Carry forward ALL Closes #N and Fixes #N lines found in original PR bodies
• Use gh pr create --body-file (never inline body via --body)

Workflow

Step 1: Parse Arguments

Parse four positional args: batch_branch, base_branch, pr_order_file, audit_verdict (last one may be absent or empty string). Parse the optional fifth positional argument conflict_report_paths (may be absent or empty string). Split on , to get a list of paths; filter out any empty strings. Store as conflict_report_path_list. Parse the optional named argument domain_partitions_path (may be absent or empty string); store as domain_partitions_path.

Step 2: Read pr_order_file

Read the JSON file. Extract: prs array (each: number, title, branch, complexity, additions, deletions, overlap_with_pr_numbers, files_changed). Also read base_branch from JSON as confirmation. Store the PR list as pr_list.

Step 3: Fetch Closes/Fixes References from Original PR Bodies

Fetch all PR bodies in a single GraphQL alias query (1 API call instead of N sequential REST calls). Chunk into batches of 20 to stay within GraphQL complexity limits.

# Build GraphQL alias query for all PRs
PR_NUMS=($(echo "$pr_list" | jq -r '.[].number'))
BODIES_JSON="{}"
BATCH_SIZE=20
for batch_start in $(seq 0 $BATCH_SIZE $((${#PR_NUMS[@]} - 1))); do
    BATCH_NUMS=("${PR_NUMS[@]:$batch_start:$BATCH_SIZE}")
    QUERY="query { "
    for i in $(seq 0 $((${#BATCH_NUMS[@]} - 1))); do
        NUM="${BATCH_NUMS[$i]}"
        QUERY="${QUERY} pr${i}: repository(owner: \"${OWNER}\", name: \"${REPO}\") {
            pullRequest(number: ${NUM}) { number body }
        }"
    done
    QUERY="${QUERY} }"
    BATCH=$(gh api graphql -f query="${QUERY}" 2>/dev/null || echo "{}")
    BODIES_JSON=$(echo "${BODIES_JSON} ${BATCH}" | jq -s 'add // {}')
done

Extract every line matching (Closes|Fixes|Resolves)\s+#\d+ (case-insensitive) from each PR body in BODIES_JSON. Deduplicate across all PRs. Store as closing_refs (list of strings like Closes #42). Skip gracefully if gh is unavailable — closing_refs remains empty.

Step 4: Get Changed Files

git diff --name-only {base_branch}..{batch_branch}
git diff --diff-filter=A --name-only {base_branch}..{batch_branch}
git diff --diff-filter=M --name-only {base_branch}..{batch_branch}

Store as changed_files, new_files, modified_files.

Step 4b: Load Conflict Resolution Reports

• If conflict_report_path_list is empty: skip — set conflict_resolution_table = "".
• For each path in conflict_report_path_list:
  • Read the file.
  • Extract the ## Per-File Resolution Decisions table (all lines from the | File | header through the last table row).
• Concatenate all extracted tables (one per report, separated by a blank line if multiple).
• Store as conflict_resolution_table.

This step is skipped gracefully if any path is missing — log a warning and exclude that file.

Step 4c: Partition Files by Domain

Read pre-computed domain partitions from disk when available:

DOMAIN_PARTITIONS="{}"
if [ -n "${domain_partitions_path:-}" ] && [ -f "$domain_partitions_path" ]; then
    DOMAIN_PARTITIONS="$(cat "$domain_partitions_path")"
fi

Store the parsed dict as domain_partitions (parse DOMAIN_PARTITIONS as JSON). Skip entirely and set domain_partitions = {} if changed_files is empty or domain_partitions_path is absent.

Step 4d: Fetch Domain Diffs (parallel)

For each domain name D in domain_partitions where domain_partitions[D] is non-empty, run the following in parallel (issue all Bash calls in a single message):

git diff {base_branch}..{batch_branch} -- {space-separated list of files in domain D}

Store results as domain_diffs: dict[str, str] mapping domain name → diff text. If a domain's diff text exceeds 12 000 characters, truncate to the first 12 000 characters and append \n... [truncated — diff exceeds 12 000 chars]. Domains with empty diffs are removed from domain_diffs.

Step 4e: Identify PRs per Domain

For each domain D in domain_partitions, find every PR in pr_list whose files_changed list intersects with domain_partitions[D].

domain_pr_numbers = {
    domain: [
        pr["number"]
        for pr in pr_list
        if set(pr.get("files_changed", [])) & set(domain_partitions[domain])
    ]
    for domain in domain_partitions
}

Store as domain_pr_numbers: dict[str, list[int]].

Step 4f: Fetch Domain Commits (parallel)

For each domain D in domain_diffs (domains that actually have diff content), run in parallel (all Bash calls in a single message):

git log {base_branch}..{batch_branch} --oneline -- {space-separated files in domain D}

Store as domain_commits: dict[str, list[str]] (each entry is a list of "sha message" strings). Empty results are stored as empty lists.

Step 4g: Run Parallel Domain Analysis Subagents

For each domain D in domain_diffs, spawn a Task subagent (model: sonnet) in a single parallel message. Issue all Task calls in a single message to maximize parallelism. Skip domains with no diff content (not in domain_diffs) — do not spawn subagents for them.

Each subagent receives a prompt with:

• The domain name
• The list of files changed in the domain
• The diff content (already truncated at 12 000 chars)
• The PR numbers and titles for PRs touching this domain (look up titles from pr_list)
• The commit one-liners for the domain

The subagent is instructed to return ONLY a JSON object with this exact shape:

{
  "domain": "Server/MCP Tools",
  "summary": "3-4 sentence description of what changed and why it matters",
  "key_changes": ["concise description of change 1", "concise description of change 2"],
  "pr_numbers": [42, 47],
  "commit_count": 5
}

Parse each subagent's output as JSON. If parsing fails for a domain, log a warning and omit that domain from domain_summaries. Store the collected results as domain_summaries: list[dict].

The 7 canonical domain names are: Server/MCP Tools, Pipeline/Execution, Recipe/Validation, CLI/Workspace, Skills, Tests, Core/Config/Infra.

Step 5: Select Arch-Lens Lenses

Spawn a subagent (Task tool, model: sonnet) with changed_files and the same lens menu as open-pr. Instruct it to return 1–3 lens names using the same development lens guard.

Step 6: Generate Arch-Lens Diagrams

For each selected lens, follow this exact sequence:

1. Write the PR context to a file using the Write tool:

• Path: {{AUTOSKILLIT_TEMP}}/open-integration-pr/pr_arch_lens_context_{YYYY-MM-DD_HHMMSS}.md
• Content: The following PR context block, with placeholders filled in:

# PR Context — Changed Files

This diagram is for a Pull Request. Focus the diagram on the areas of the codebase affected by these changes. Do not create a generic whole-project diagram.

## New files (use ★ prefix on these nodes):
{list of new_files from Step 4, or "None"}

## Modified files (use ● prefix on these nodes):
{list of modified_files from Step 4, or "None"}

## Instructions:
- Focus exploration and the diagram on the architectural areas these files belong to
- Use `★` prefix on nodes representing new files/components
- Use `●` prefix on nodes representing modified files/components
- Leave unchanged components unmarked (include them only if needed for context/connectivity)
- The diagram should help PR reviewers understand the architectural impact of these specific changes

2. Immediately call the Skill tool to load the arch-lens skill (e.g., /autoskillit:arch-lens-module-dependency). The loaded skill will read the PR context file written in step 1 above.

If the Skill tool returns an error containing "disable-model-invocation" or "cannot be used", do NOT write a diagram freehand. Discard this lens iteration silently. If ALL arch-lens invocations fail this way, set validated_diagrams = [] (the Architecture Impact section is omitted per Step 7 behavior).

3. Follow the loaded skill's instructions to explore the codebase and generate the diagram. Using ONLY classDef styles from the mermaid skill (no invented colors).

The arch-lens skills write their output to {{AUTOSKILLIT_TEMP}}/arch-lens-{lens-name}/ (relative to the current working directory). After each skill runs, read the generated markdown file and extract the mermaid code block(s).

After extracting the mermaid block, inspect its content for ★ or ● characters:

• If the block contains at least one ★ or ● → add it to validated_diagrams.
• If the block contains neither → discard this diagram; do not add it to the list.

Step 7: Compose PR Body

Write to {{AUTOSKILLIT_TEMP}}/open-integration-pr/pr_body_{timestamp}.md. (relative to the current working directory)

## Integration Summary

Collapsed {N} PRs into `{batch_branch}` targeting `{base_branch}`.

## Merged PRs

| # | Title | Complexity | Additions | Deletions | Overlaps |
|---|-------|-----------|-----------|-----------|---------|
| #{number} | {title} | {complexity} | +{additions} | -{deletions} | {overlap_with_pr_numbers or "—"} |
...

{If domain_summaries is non-empty:}
## Domain Analysis

{For each entry in domain_summaries (ordered by domain name):}
### {entry.domain}

{entry.summary}

**Key changes:**
{For each item in entry.key_changes:}
- {item}

**Contributing PRs:** {comma-separated #{N} for each N in entry.pr_numbers, or "—" if empty}
**Commits:** {entry.commit_count} commit(s)

{If audit_verdict is non-empty:}
## Audit

**Verdict:** {audit_verdict}

{If conflict_resolution_table is non-empty:}
## Conflict Resolution Decisions

The following files had merge conflicts that were automatically resolved during pipeline integration.

{conflict_resolution_table}

{If validated_diagrams non-empty:}
## Architecture Impact

{For each validated diagram:}
### {Lens Name} Diagram

```mermaid
{diagram content}

{For each item in closing_refs:} {Closes #N}

🤖 Generated with Claude Code via AutoSkillit

### Step 8: Check GitHub Availability

```bash
gh auth status 2>/dev/null

If exit code non-zero: output pr_url= and exit successfully.

Step 9: Create Integration PR

gh pr create \
  --base {base_branch} \
  --head {batch_branch} \
  --title "Integration: collapsed PRs #{numbers} into {base_branch}" \
  --body-file {{AUTOSKILLIT_TEMP}}/open-integration-pr/pr_body_{timestamp}.md

{numbers} = comma-separated PR numbers (e.g., #42, #47, #51). Capture the new PR URL as new_pr_url. Extract the PR number from the URL as new_pr_number.

Step 10: Close Original PRs

For each PR in pr_list:

gh pr close {number} --comment "Collapsed into integration PR #{new_pr_number} ({new_pr_url})"
sleep 1  # Rate-limit discipline: 1s between mutating calls

Continue even if individual close operations fail (log warning, do not exit).

Output

pr_url = {new_pr_url}