Weekly Update – July 6, 2026

Jul 6, 2026

It was a productive week in github/gh-aw — with dozens of pull requests landing across the compiler, linters, JavaScript setup scripts, and documentation. Here’s a look at the highlights.

Notable Pull Requests

fix(compiler): auto-add `pre_activation` to `safe_outputs`/`conclusion` needs

A sneaky compiler bug was generating skillet.lock.yml files with broken actionlint expressions: safe_outputs and conclusion jobs referenced ${{ needs.pre_activation.outputs.skill_name }} without actually declaring pre_activation as a dependency. This fix auto-wires the dependency whenever a message template references pre_activation outputs — no more cryptic expression errors in generated lock files.

refactor(linters): consolidate AST/context helpers into `internal/astutil`

The linter suite had quietly grown several copies of the same helper functions — enclosingFuncType, context-type resolution, OS-call detection — scattered across individual analyzers. This PR gathers them all into a single pkg/linters/internal/astutil package and rewires the affected analyzers, eliminating drift risk and making future linter work easier to reason about.

ambient-context: reduce copilot-agent-analysis first-request size by ~28%

copilot-agent-analysis was the largest ambient-context payload at 27,299 characters — most of it content that’s rarely needed at runtime. By gating cold-start rebuild content behind an optional import, this PR trims the first-request size to 11,876 characters, cutting token costs on every agent activation that uses this analysis path.

Add shared prompt quality gate for plateaued agent-review workflows

Agent effectiveness scores had been stuck around 61–62 for several weeks — a signal that prompt design, not runtime bugs, was the limiting factor. This PR introduces a reusable quality rubric shared across analyzer and reviewer workflows, giving those workflows a concrete target for what “good” looks like and a path out of the plateau.

fix(setup/js): numeric coercion, setOutput stringification, and async entrypoint cleanup

A sweep across 23 files in actions/setup/js replaced global isNaN (which silently coerces inputs) with Number.isNaN, fixed core.setOutput value types, and cleaned up unhandled async rejections. Small correctness improvements that prevent subtle runtime surprises in CI steps.

Agent of the Week: Weekly Issue Summary

Your Monday morning data journalist — scans all issue activity from the past week and compiles trends, charts, and resolution statistics into a single digest comment.

weekly-issue-summary has been running quietly every Monday around 3 PM UTC, pulling 30 days of issue data, generating CSV trend files, and rendering two charts: one for issue open/close velocity and one for resolution time distributions. In its last three runs it made 13 GitHub API calls each time and burned through roughly 59 AI credits — efficient for a workflow that touches every open and closed issue in the repo. Two of the three runs succeeded without any write-side effects, posting the full digest to a tracking issue, while one run hit a timeout on the data preparation phase and bailed cleanly.

The June 15th failure is the fun part: the observability report flagged it with the note “this run consumed a heavy execution profile for its task shape” and gently suggested the team might want to swap in a smaller model. The workflow took the feedback in stride and came back the following Monday working perfectly.

Usage tip: Pair weekly-issue-summary with a label strategy — the chart breakdowns are most useful when issues are consistently labeled, since resolution-time distributions get interesting when you can split them by category.

→ View the workflow on GitHub

Try It Out

All of this week’s changes are already on main — pull the latest and run gh aw compile to pick up the compiler and linter improvements. Got feedback or spotted something worth fixing? Contributions are always welcome at github/gh-aw.