Refactor changeset workflow to scheduled batch processing with cache-based deduplication

[Copilot]

Changeset generation previously required manual PR labeling and processed one PR at a time. Refactored to run every 2 hours, batch-process merged PRs, and track processed PRs via cache memory.

Changes

Trigger & Scheduling

• Changed from pull_request label-based trigger to schedule: cron: "0 */2 * * *"
• Retained workflow_dispatch for manual execution

PR Query & Filtering

• Added step to query merged PRs from last 4 hours using gh search prs
• Filters to 2-hour window for primary processing (4-hour window catches stragglers)
• Saves PR data to /tmp/gh-aw/pr-data/recent-merged-prs.json

Cache-Based Deduplication

• Added cache-memory with key changeset-processed-prs-${{ github.workflow }}
• Agent maintains /tmp/gh-aw/cache-memory/processed-prs.json as JSON array of processed PR numbers
• Skips already-processed PRs to prevent duplicate changesets

Batch Processing

• Agent processes all unprocessed PRs in single run
• Generates one changeset file per PR: <type>-pr-<number>-<slug>.md
• Implements early exit if no PRs need processing

Safe Output

• Changed from push-to-pull-request-branch to create-pull-request
• Creates PR with title prefix [changeset] and labels [changeset, automation]

Example

Workflow now processes multiple PRs automatically:

# Query merged PRs
gh search prs --repo ${{ github.repository }} \
  --merged --merged ">=${FOUR_HOURS_AGO}" \
  --json number,title,mergedAt,body,labels,url,author

# Filter to 2-hour window
jq '[.[] | select(.mergedAt >= $two_hours_ago)]' all-merged-prs.json

# Check cache and skip processed
jq '[.[] | select([.number] | inside($processed) | not)]' recent-prs.json

Agent creates changesets like patch-pr-456-update-docs.md and updates cache to prevent reprocessing.

Original prompt

Refactor the changeset agentic workflow to run every 2 hours and look at the latest merge pull requests and generate changeset for each of them.

• add a custom job that queries pull requests from the last 2 hours and creates an output whether there is work to do. The agent task "if" must check that outpjt
• only process at most the last 4h
• use cache memory to determine if PR was processed
• 1 changeset file per pull request
• create pull requests with files if any

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.

[Copilot]

Pull Request Overview

This PR refactors the changeset workflow from a manual, label-triggered single-PR processor to an automated batch processor that runs every 2 hours to generate changesets for recently merged PRs.

Key Changes:

• Trigger mechanism: Changed from pull_request with label filtering to scheduled cron execution (0 */2 * * *)
• Processing model: Shifted from single PR context to batch processing of all PRs merged in the last 2 hours
• Deduplication: Introduced cache-based tracking using cache-memory to prevent duplicate changeset generation
• Output mechanism: Replaced push-to-pull-request-branch with create-pull-request for submitting changeset files

Reviewed Changes

Copilot reviewed 1 out of 2 changed files in this pull request and generated 12 comments.

File	Description
.github/workflows/changeset.md	Removed PR trigger and label conditions; added schedule trigger; introduced cache-memory for tracking processed PRs; added pre-step to fetch merged PR data; updated agent instructions for batch processing; changed safe-output from push-to-pull-request-branch to create-pull-request; duplicated workflow content exists (lines 74-202 and 203-298)
.github/workflows/changeset.lock.yml	Compiled workflow reflecting all changes from the markdown source; removed pre_activation job and associated permission checks; removed push_to_pull_request_branch job; added create_pull_request job; added cache-memory setup with restore keys; simplified concurrency group; duplicated prompt content in lines 1288-1415 and 1417-1512

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

[nitpick] The restore-keys for the cache include keys with incomplete prefixes (changeset-processed-, changeset-) that don't match the actual cache key pattern changeset-processed-prs-${{ github.workflow }}.

The restore keys should follow a hierarchical fallback pattern that matches the actual key structure:

restore-keys: |
  changeset-processed-prs-${{ github.workflow }}-

The current restore-keys would only be useful if there were other cache entries with those prefixes, but based on the configuration, only keys starting with changeset-processed-prs- exist.

Suggested change

	changeset-processed-prs-
	changeset-processed-
	changeset-

[Copilot]

[nitpick] The instruction states "Format as a clear table or list" for the PR description but doesn't provide a specific format example. Given the structured nature of the data (PR numbers, titles, changeset files), providing a template would help ensure consistency.

Recommendation: Add an example PR description format:

## Changesets Generated

| PR | Title | Changeset File |
|----|-------|----------------|
| #123 | Fix rendering bug | `patch-pr-123-fix-rendering-bug.md` |
| #456 | Update documentation | `patch-pr-456-update-docs.md` |

Generated changesets for 2 merged PRs from the last 2 hours.

Suggested change

	- Format as a clear table or list
	- Format as a clear table or list
	- Example PR description format:
	```markdown
	## Changesets Generated
	| PR | Title | Changeset File |
	|----|-------|----------------|
	| #123 | Fix rendering bug | `patch-pr-123-fix-rendering-bug.md` |
	| #456 | Update documentation | `patch-pr-456-update-docs.md` |
	Generated changesets for 2 merged PRs from the last 2 hours.
	```

[Copilot]

The workflow markdown content contains a complete duplication of the agent instructions. Lines 74-202 are nearly identical to lines 203-298, with only minor variations in wording. This duplication significantly increases the token count sent to the AI agent and could cause confusion.

Recommendation: Remove the duplicate section (either lines 74-202 or 203-298) and keep only one consolidated version of the instructions.

[Copilot]

The workflow markdown content is duplicated in the compiled workflow. Lines 1288-1415 are nearly identical to lines 1417-1512. This duplication wastes token budget when the prompt is sent to the AI agent and could lead to conflicting instructions.

Recommendation: Remove the duplicate section in the source markdown file to prevent it from appearing twice in the compiled workflow.

See below for a potential fix:

# (Lines 1417-1474 removed; no replacement needed. The workflow continues with the next step.)

[Copilot]

[nitpick] The concurrency group has changed from gh-aw-${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }} to gh-aw-${{ github.workflow }}. This means only one instance of the entire workflow can run at a time, rather than allowing concurrent runs for different PRs.

While this may be intentional for a scheduled workflow that processes all PRs in batch, the cancel-in-progress: true setting was also removed. This could lead to multiple scheduled runs queuing up if a previous run takes longer than 2 hours.

Recommendation: Consider adding cancel-in-progress: true to prevent queue buildup, or document why queuing multiple runs is acceptable.

Suggested change

	group: "gh-aw-${{ github.workflow }}"
	group: "gh-aw-${{ github.workflow }}"
	cancel-in-progress: true

[Copilot]

[nitpick] The filename pattern instruction includes <short-slug> but then the example shows patch-pr-123-fix-bug.md where "fix-bug" is the slug. However, line 130 says "Use format from the changeset format reference above" but doesn't specify what the slug should be derived from.

Recommendation: Clarify that the slug should be derived from the PR title (as mentioned in line 175: "Use the PR title") and provide clearer guidance:

• Filename: <type>-pr-<pr-number>-<short-slug>.md
• Slug: Derived from PR title, lowercase, hyphens, 2-4 words max
• Example: PR Network Permissions Testing Report - MCP Fetch Tool Analysis #123 "Fix rendering bug" → patch-pr-123-fix-rendering-bug.md

Suggested change

	- Filename: `<type>-pr-<pr-number>-<short-slug>.md` (e.g., `patch-pr-123-fix-bug.md`)
	- Use lowercase and hyphens for the slug
	- Keep the slug short (2-4 words max)
	- Filename: `<type>-pr-<pr-number>-<short-slug>.md`
	- **Derive `<short-slug>` from the PR title**:
	- Take the PR title, convert to lowercase
	- Replace spaces and punctuation with hyphens
	- Use 2-4 words max for brevity
	- Example: PR #123 "Fix rendering bug" → `patch-pr-123-fix-rendering-bug.md`

[Copilot]

The "Important Notes" section (line 295) states "If there are no unprocessed PRs, simply do nothing (the job won't run due to conditional)" but there is no conditional check in the workflow that would prevent the job from running. The workflow relies on the agent to perform early exits, not on workflow-level conditionals.

This statement is misleading and inconsistent with lines 97-100 and 110-113 which correctly describe early exit behavior within the agent execution.

Recommendation: Update line 298 to match the accurate description: "If there are no unprocessed PRs, just exit without doing anything - this is normal and expected"

Suggested change

	- If there are no unprocessed PRs, simply do nothing (the job won't run due to conditional)
	- If there are no unprocessed PRs, just exit without doing anything - this is normal and expected

[Copilot]

The activation job no longer has an if condition to control when it runs. Previously it checked needs.pre_activation.outputs.activated == 'true' and various event conditions. Now the job always runs on schedule, which may be intentional but could lead to unnecessary executions.

Recommendation: Consider adding a condition like if: github.event_name == 'schedule' || github.event_name == 'workflow_dispatch' to explicitly control when the activation job runs, or document why the unconditional execution is desired.

Suggested change

	runs-on: ubuntu-slim
	runs-on: ubuntu-slim
	if: github.event_name == 'schedule' || github.event_name == 'workflow_dispatch'

[Copilot]

The instruction to add /tmp/gh-aw/cache-memory/processed-prs.json to the git commit (line 148) is incorrect. The cache-memory directory is outside the repository working directory and should not be committed to the PR. Only the changeset files in .changeset/ should be committed.

The cache is persisted via GitHub Actions cache mechanism, not via git commits.

Recommendation: Remove line 148 from the instructions. The correct git operations should be:

git add .changeset/*.md
git commit -m "Add changesets for merged PRs"

Suggested change

git add /tmp/gh-aw/cache-memory/processed-prs.json

[Copilot]

The jq filter for removing already-processed PRs uses the inside function incorrectly. The expression [.number] | inside($processed) checks if the array [.number] exists inside the processed array, but this is backwards.

Current (incorrect): [.number] | inside($processed) - checks if single-element array is inside processed
Should be: .number as $n | $processed | contains([$n]) | not - checks if processed contains the number

Alternatively, use: select(.number as $n | ($processed | index($n)) == null)

Example of the bug:

• If processed = [123, 456] and .number = 123
• [123] | inside([123, 456]) returns false (because the array [123] is not a subset)
• The correct check should return true (because 123 is in the processed list)

Suggested change

	- Use jq to filter: `jq --argjson processed "$(cat /tmp/gh-aw/cache-memory/processed-prs.json)" '[.[] | select([.number] | inside($processed) | not)]' /tmp/gh-aw/pr-data/recent-merged-prs.json`
	- Use jq to filter: `jq --argjson processed "$(cat /tmp/gh-aw/cache-memory/processed-prs.json)" '[.[] | select(.number as $n | ($processed | index($n)) == null)]' /tmp/gh-aw/pr-data/recent-merged-prs.json`