Align eval-harness Python version with CI (3.14)

[itsmiso-ai]

What

Pinned Python version to 3.14 in eval-harness.yaml to match CI and resolve duplicate-key regression test failure.

Why

Version skew between local (3.9) and CI (3.x) caused the duplicate-key regression test to ship broken; pinning to 3.14 ensures consistent behavior.…

Fixes #496

Opened by foreman on review GO (workload wl-misospace-pr-reviewer-action-496).

[its-saffron]

AI Automated Review

Full PR review.

Analysis engine: dsv4f@https://litellm.jory.dev/v1 (openai) — primary route

Recommendation

Approve. This is a minimal, well-targeted fix that aligns the eval-harness workflow's Python interpreter with CI (3.14), directly satisfying the linked issue's primary acceptance criterion ("eval-harness.yaml pins the same Python version as ci.yaml ... matching literal"). The validate CI check passes on this commit, and repo history/grep confirm ci.yaml pins python-version: '3.14' as the reference. No blockers or material issues found.

Change-by-change findings

.github/workflows/eval-harness.yaml

• python-version changed from "3.12" to '3.14', byte-for-byte the same literal as ci.yaml:24 (python-version: '3.14'), so the weekly scheduled eval run and CI now exercise the identical interpreter on the same pr_reviewer code. This was the one gap Renovate couldn't close (the Renovate PRs PR 491–493 already converted the workflow's actions to SHA-pinned v7 forms, which I confirmed are present in the current file at checkout 3d3c42e5..., 5fda3b95..., and 043fb46d...).
• The new explanatory comment matches the spirit and content of the ci.yaml comment about the duplicate-key regression and the local 3.9 vs. CI 3.x skew; it's accurate and mirrors the accepted rationale from commit 112cd5a.
• The quoted scalar ('3.14') matches the exact style used in ci.yaml, which aids the stated goal of keeping the two workflows visually consistent.
• No security, secret-handling, or path concerns: this file only touches workflow interpreter pinning, uses the existing secrets.* references unchanged, and carries no risk flags.

Sources

• Full contents of .github/workflows/eval-harness.yaml (already SHA-pinned actions v7 + new 3.14 pin)
• git grep for 3.14/python-version: matches in both .github/workflows/ci.yaml:24 and .github/workflows/eval-harness.yaml:62, with identical co-located comments
• Repository history: a5f5814 (this PR's head), 84f0e6c (eval harness CI workflow PR 488), 112cd5a (ci: catch duplicate YAML keys + pin CI Python to 3.14 PR 256), Renovate PRs PR 491–493 (action SHAs)
• Issue PR 496 body (acceptance criteria and evidence)
• CI check result for a5f5814: validate success

Standards Compliance

AGENTS.md contains no workflow-interpreter or action-pinning convention that conflicts with this change. The change is consistent with the documented CI-pinning precedent (SHA-pinned v7 actions + 3.14 interpreter) and with the repository's overall stance that the review-quality gate should exercise the same interpreter as shipping CI. No prohibition on workflow comment style or quoted-scalar style exists in the standards file.

Linked Issue Fit

Issue PR 496 asks that the eval-harness workflow use the same pinned Python interpreter as ci.yaml and the same SHA-pinned action style, so the review-quality gate exercises the same interpreter as production CI. Verified against the issue's acceptance criteria:

1. "pins the same Python version as ci.yaml (single source for the version, e.g. a repo-level default or matching literal)" — Met. python-version: '3.14' is now an exact matching literal to ci.yaml:24. While a repo-level default would be the ideal single source, the issue explicitly names "matching literal" as an acceptable form.
2. "uses the same SHA-pinned action style" — Met. The current workflow file pins actions/checkout, actions/setup-python, and actions/upload-artifact to v7 SHAs (delivered by the in-flight Renovate PRs PR 491–493 that the issue anticipated); this PR completes the remaining interpreter gap.
3. "The weekly scheduled eval run and CI both exercise the identical interpreter for the same pr_reviewer code" — Met. Both workflows now run Python 3.14.

No acceptance criterion is missed or misinterpreted; the PR's scope is exactly the interpreter alignment the issue identified as remaining.