fix(ci): make /security-review fail loudly when the model never runs

[tejaskash]

Summary

• Fix the silent failure mode that produced a false "no high-confidence findings" comment on fix: resolve create harness Dockerfile paths from command cwd #1474.
• Root cause: the bundled /security-review skill's SessionStart hook runs git diff --name-only origin/HEAD... as its first command. actions/checkout@v6 with ref: <fork-head-sha> and fetch-depth: 0 fetches the PR head's history but does not always fetch origin/<base>, so the diff resolves to fatal: no merge base. The hook errors, the SDK exits cleanly with num_turns=0 and zero tokens, the inline-comment buffer is empty, and the workflow's summary step still posts "no findings."
• Verified in run 27159048605 — the SDK transcript shows <local-command-stderr>fatal: origin/HEAD...HEAD: no merge base</local-command-stderr> followed by subtype: success, num_turns: 0.

Changes

1. Make the merge base actually exist. Prepare base ref for /security-review skill now explicitly git fetches origin/<base> and runs git merge-base origin/<base> HEAD as a sanity check. Fails the step if the merge base can't be resolved, instead of letting the skill silently bail downstream.
2. Catch silent skill bailouts in general. New Verify model actually ran step reads the action's execution_file (a JSON array; final element is the SDK result envelope) and requires num_turns > 0, output_tokens > 0, and is_error == false. If the model never productively ran, the step fails and the summary comment posts "the review did not actually analyze this PR (model took 0 turns — the skill likely failed during setup)" instead of "no high-confidence findings."

Together: change 1 fixes the immediate bug; change 2 ensures any future cause of "SDK exits clean without the model doing real work" no longer reports a false-clean review.

Test plan

• Re-run this workflow on a fork PR (e.g. fix: resolve create harness Dockerfile paths from command cwd #1474) and confirm the Prepare base ref step logs the merge base + file count.
• Confirm Verify model actually ran reports a sensible num_turns / output_tokens on a successful run.
• Manually break it (delete the git fetch line locally) and confirm the workflow now fails with the explicit error and the summary comment posts the failure message rather than "no findings."

[agentcore-devx-automation]

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

[github-actions]

Package Tarball

aws-agentcore-0.18.0.tgz

How to install

gh release download pr-1482-tarball --repo aws/agentcore-cli --pattern "*.tgz" --dir /tmp/pr-tarball
npm install -g /tmp/pr-tarball/aws-agentcore-0.18.0.tgz

[agentcore-cli-automation]

LGTM — this is a clean fix to the silent-failure bug.

Quick summary of what I checked:

1. git fetch + merge-base sanity: confirmed the diagnosis. actions/checkout@v6 with an explicit ref: only fetches that ref's history; on fork PRs it doesn't necessarily land origin/<base> in the local clone, so git diff origin/HEAD... fails the SessionStart hook with "no merge base". The explicit refspec fetch + git merge-base precondition closes that gap and fails loudly when it can't resolve.
2. execution_file shape: verified against anthropics/claude-code-action@v1 (base-action/src/run-claude-sdk.ts). The action writes the full SDK message array via writeExecutionFile(messages) on both the happy path and the SDK-error catch block, and the result envelope is the last type: 'result' message in the stream. jq '.[-1].num_turns // 0' is the right shape, and usage.output_tokens matches SDKResultSuccess / SDKResultError in @anthropic-ai/claude-agent-sdk's type definitions.
3. Failure-mode flow: when model-ran exits 1, the implicit success() on Count buffered findings skips it, leaving FINDING_COUNT empty → defaults to 0 in the summary script. The if (modelRan === 'false') branch wins and posts the "did not actually analyze" message instead of "no findings". Remove claude-security-reviewing label runs under if: always(), so the label still gets cleaned up. ✓
4. ran=unknown path: only triggers if claude-execution-output.json is missing entirely, which the action only does on catastrophic failure before the SDK loop — at which point steps.review.conclusion is almost certainly 'failure', and the summary correctly falls through to the "review run failed" branch rather than lying about findings.

No blocking issues found.

[agentcore-devx-automation]

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