fix: make harness PR reviewer evaluate against the PR base branch

[notgitika]

Problem

The AI PR reviewer (agentcore-devx-automation "AgentCore Harness Review") reviews PRs against the default branch (main) rather than the PR's actual base branch. Its local context clones (/opt/workspace/agentcore-cli) sit on main, so PRs targeting refactor are analyzed against the main source layout.

This produces confidently-wrong findings: e.g. on #2105 it reported a "split-brain read/write" regression citing src/cli/commands/deploy/actions.ts, src/lib/schemas/io/path-resolver.ts, and src/cli/operations/init/files.ts — none of which exist on refactor (which uses src/handlers/ and src/core/).

Fix

The actual checkout logic lives in the reusable workflow in aws/agentcore-devx-devtools (not editable here), so this fixes it at the lever we control — the prompts referenced by pr-automation.yml:

• review.md: instruct the reviewer to determine the PR's base.ref and git fetch/git checkout it in the local clone before reading files for context, and to not raise findings premised on a file/path being absent without verifying against that base branch.
• system.md: document that agentcore-cli has a long-lived refactor branch that diverges from main, and that PRs must be analyzed against their own base branch.

Notes

• pull_request_target reads the workflow + prompts from the base branch, so this must land on refactor to fix refactor-targeted PRs. A parallel change on main keeps parity for main-targeted PRs.
• A more robust fix (checking out the base branch in the runner) belongs in aws/agentcore-devx-devtools; this prompt-level mitigation is the in-repo stopgap.

[agentcore-devx-automation]

Claude Security Review: no high-confidence findings. (run)

[agentcore-devx-automation]

AgentCore Harness Review

Verdict: Changes requested

Nice, well-scoped fix — the base-branch-vs-main mismatch is a real problem that will bite refactor-targeted PRs, and the layout facts in both prompts (src/handlers/+src/core/ on refactor vs. src/cli/+src/lib/ on main) match what's actually on those branches. One concrete concern before merging:

The example curl uses $CLONE_TOKEN, which isn't present at container runtime

In .github/harness/prompts/review.md (the new block, roughly lines 11–30) the recipe is:

curl -sH "Authorization: Bearer $CLONE_TOKEN" https://api.github.com/repos/aws/agentcore-cli/pulls/<number> | jq -r .base.ref

But per .github/harness/Dockerfile and .github/harness/README.md, CLONE_TOKEN is a build-time ARG used only to configure git config --global url."https://${CLONE_TOKEN}@github.com/".insteadOf .... The README explicitly states: "The token is never sent to the Harness runtime or persisted in this image." So at review time $CLONE_TOKEN will be empty, and curl -H "Authorization: Bearer " against api.github.com returns HTTP 401 (verified). jq -r .base.ref then emits null, and the follow-up git checkout null will fail — dropping the agent back to main, i.e. the exact failure mode this PR is trying to prevent.

Options:

1. Drop the auth header entirely. aws/agentcore-cli is public, so an unauthenticated call to /repos/aws/agentcore-cli/pulls/<n> succeeds and returns .base.ref. Simplest fix and matches how the review agent already fetches PR metadata today.
2. Rely on git ls-remote / a plain git fetch --all plus resolving the base ref from the diff/PR context the harness already passes in, rather than a fresh REST call.
3. If you actually want an authenticated call, arrange for the reusable workflow (aws/agentcore-devx-devtools/.github/workflows/reusable-pr-ai-review.yml) to inject a token env var into the runtime container, and reference that variable here — and update .github/harness/README.md to match.

Everything else (the system.md blurb, the "don't raise findings whose premise is a missing file until you've confirmed the base branch" guardrail, the file paths modified) looks good.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 97.42%. Comparing base (1d96d1f) to head (cf80f92).

Additional details and impacted files

@@            Coverage Diff            @@
##           refactor    #2106   +/-   ##
=========================================
  Coverage     97.42%   97.42%           
=========================================
  Files           429      429           
  Lines         26286    26286           
=========================================
  Hits          25609    25609           
  Misses          677      677           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[agentcore-devx-automation]

Claude Security Review: no high-confidence findings. (run)