`inputs.*` expressions in prompt body not resolved when called via `workflow_call`

[benvillalobos]

🤖 AI Assisted bug report

Summary

When a gh-aw compiled workflow is invoked as a reusable workflow (workflow_call) instead of workflow_dispatch, ${{ inputs.* }} expressions in the .md prompt body are not resolved. The agent sees the literal text ${{ inputs.errors }} instead of the actual value.

Reproduction

1. Create an agentic workflow (.md) with both workflow_call and workflow_dispatch triggers, and ${{ inputs.some_input }} in the prompt body
2. Compile with gh aw compile
3. Call it from another workflow using uses: + with: (reusable workflow pattern)
4. The agent receives the literal ${{ inputs.some_input }} text — the value is never substituted

Dispatching the same workflow via workflow_dispatch works correctly.

Root Cause

In actions/setup/js/runtime_import.cjs, the evaluateExpression() function resolves expressions differently depending on their prefix:

• needs.* and steps.* (line ~275): Converted to GH_AW_* env var names and looked up in process.env. This works because the compiler sets these env vars in the lock file's "Create prompt" step.
• inputs.*: Does NOT get the env var lookup. Falls through to the GitHub context path (line ~306): inputs: context.payload?.inputs || {}. For workflow_dispatch, context.payload.inputs is populated by GitHub Actions. For workflow_call, it is not — the inputs are delivered differently.

The compiler already generates the correct env vars (e.g., GH_AW_INPUTS_ERRORS: ${{ inputs.errors }}), and GitHub Actions resolves them in YAML for both trigger types. The values are available in process.env — the runtime just never reads them for inputs.* expressions.

Suggested Fix

Add inputs.* to the env var lookup condition at line ~275 of runtime_import.cjs:

// Before:
if (trimmed.startsWith("needs.") || trimmed.startsWith("steps.")) {

// After:
if (trimmed.startsWith("needs.") || trimmed.startsWith("steps.") || trimmed.startsWith("inputs.")) {

This makes inputs.errors resolve to process.env.GH_AW_INPUTS_ERRORS — the same pattern that already works for needs.* and steps.*.

Why It Was Not Caught

The existing workflow_call smoke test (smoke-workflow-call.md) declares no inputs and uses no ${{ inputs.* }} expressions in its prompt body, so this code path was never exercised.

Context

We discovered this while converting the a workflow from workflow_dispatch (fire-and-forget) to workflow_call (reusable workflow) so two workflows would appear as nodes in a single GitHub Actions run graph.

[benvillalobos]

Just wanted to add a note here. I think the actual issue is that my workflow did not pass any value for errors, which is probably why the template variable didn't get substituted. Which makes me wonder if there should be some kind of artifact left there explicitly stating that there was no value, maybe like an empty array or something?
Workflow run that was passed an empty errors value

Currently testing this by enforcing errors have a value before being passed, will report back.