Release 0.165.1: Codex-Exec Unifies the Automation Stack

Introduction

Release 0.165.1 is all about tightening the automation spine that keeps Screeps GPT operating around the clock. We replaced the aging bespoke copilot-exec implementation with a thin compatibility layer that now delegates directly to codex-exec, our wrapper around the official openai/codex-action@v1. The goal is simple: eliminate redundant maintenance while keeping every workflow, agent, and regression suite humming without extra configuration.

Key Features

• Unified every Copilot-powered workflow on top of the new codex-exec composite action.
• Preserved the legacy copilot-exec interface so existing prompts, tokens, and cache strategies continue to work.
• Hardened regression coverage to ensure delegation, caching, and token plumbing behave exactly like before (tests/regression/copilot-exec-force-response.test.ts).
• Synchronized documentation across the automation knowledge base (packages/docs/source/docs/automation/overview.md, .github/copilot-instructions.md, and AGENTS.md) so future agents understand the new stack.

Technical Details

The heart of the release lives in .github/actions/copilot-exec/action.yml. Instead of carrying a full implementation, the action now contains a single step named “Delegate to codex-exec” that forwards every input—prompt-path, caching flags, MCP configuration, and the recently-added force-response switch—into .github/actions/codex-exec/action.yml. We intentionally kept parameter names untouched and map copilot-token to codex-token inside the wrapper. That small detail lets fifty-plus workflows continue using secrets.GITHUB_TOKEN or secrets.OPENAI_API_KEY without mechanical edits while still unlocking Codex-specific features such as official caching semantics.

The new codex-exec action brings several quality-of-life improvements straight from the upstream OpenAI tooling. It lazily checks whether the repository has already been cloned before invoking actions/checkout@v4, so matrix workflows no longer pay an extra checkout cost when one job has already cloned the repo. Prompt rendering happens via envsubst, and we hash the rendered prompt to create deterministic cache keys stored under .codex-cache/output.txt. A Restore result cache step uses actions/cache@v4 and short-circuits the run when the exact prompt/model pair has already been answered, meaning routine automation (issue triage, changelog digestion, PTR monitoring) now hits cached responses far more often. When a cache miss occurs, the action shells out to openai/codex-action@v1 with the resolved working directory, keeping execution logs and artifacts consistent with upstream expectations.

From an architecture standpoint, this migration aligns with the zero-tolerance policy for obsolete code that we documented in AGENTS.md. Maintaining a custom CLI required shadowing upstream API changes and re-implementing token handling. Delegating to Codex collapses that surface area into a single wrapper while still letting us enforce guardrails like sandbox mode (allow-all-paths) and verbose telemetry. It also standardizes how we pass MCP configuration via the additional-mcp-config parameter, which is critical because workflows such as .github/workflows/copilot-review.yml and .github/workflows/copilot-issue-triage.yml rely on the same MCP servers (GitHub + Playwright). Instead of each workflow owning its own integration glue, they all inherit the logic from codex-exec through the compatibility wrapper.

Bug Fixes

This release did not ship discrete bug fixes. However, the automation overhaul effectively resolves the maintenance drag caused by divergent CLI implementations. By relying on the official Codex action, we no longer hit drift where local changes lag behind upstream API behavior, and the force-response parameter in tests/regression/copilot-exec-force-response.test.ts guarantees cache busting works when workflows demand fresh data.

Breaking Changes

None. Every workflow keeps calling copilot-exec under the same contract. The compatibility layer translates inputs behind the scenes, so there is no need to rename steps, update secrets, or tweak outputs. Specialized agents located under .github/actions/copilot-*-agent/ still expose their own ergonomics but ultimately call codex-exec under the hood, which the regression suite explicitly verifies.

Impact

Operationally, consolidating on codex-exec cuts the automation maintenance footprint while improving reliability:

• Consistent Behavior: Workflows such as copilot-changelog-to-blog.yml, copilot-email-triage.yml, and copilot-strategic-planner.yml now share the exact same execution engine, eliminating subtle drift that previously showed up when the custom CLI adopted a feature later than another workflow needed it.
• Faster Runs: Conditional checkout and prompt result caching shave tens of seconds off high-frequency workflows. Because the cache key is now the prompt SHA plus model, unrelated changes elsewhere in the repo no longer invalidate AI outputs.
• Simpler Secret Management: Documentation updates in packages/docs/source/docs/automation/overview.md and .github/copilot-instructions.md reiterate that workflows map copilot-token to codex-token, so rotating secrets or onboarding a new agent involves changing values in one place.
• Auditability: The regression suite ensures .github/workflows/* either call codex-exec directly or depend on a specialized agent that does. The tests also check that every migrated workflow reads codex-token: \\${{ secrets.OPENAI_API_KEY }}, giving us high confidence that future contributors won’t accidentally fall back to the deprecated path.

What’s Next

With the execution layer modernized, the next releases will focus on two fronts. First, we want to expose Codex-specific telemetry (latency, token counts, cache hits) directly inside the guard workflows so maintainers can spot drift before it shows up in PTR monitoring. Second, we plan to document an end-to-end recipe for writing new agents that plug into codex-exec without duplicating logic—think of it as a “starter kit” that lives beside packages/docs/source/docs/automation/overview.md. Expect additional instrumentation and authoring guidance as we continue refining the autonomous development pipeline.