[bai-uipath]

Motivation

Picks up UiPath/coder_eval#225: the codex harness copies skills into .agents/skills/ instead of symlinking them. With symlinks, codex lists every skill as uipath:<skill>, models sometimes open the nonexistent .agents/skills/uipath/<skill>/SKILL.md, and the rg/find search they fall back on skips symlinks, so the agent can lose its skill.

Merge order

Step	What	Status
1	Merge UiPath/coder_eval#225	Done (1895a2c)
2	Dispatch the coder_eval release workflow with bump: patch: coder-eval==0.12.14 on PyPI, ghcr.io/uipath/coder-eval-agent:0.12.14	Done
3	Re-run this PR's checks	Done, all green
4	Merge this PR

• After merge, the ADO nightly (coderEvalVersion=pinned) and the GHA workflows that read tests/.coder-eval-version use 0.12.14 on their next run.

Changes

• Pin coder-eval 0.12.12 → 0.12.14. What that brings in for this repo:

version	PR	effect here
0.12.13	UiPath/coder_eval#220	Pricing rows for Haiku 5.5, Sonnet 5.5 and GPT-6.1 Sol. Runs on those models no longer book cost: None.
0.12.13	UiPath/coder_eval#221	claude-code multi-turn tasks book each resumed turn's own usage. Cost on multi-turn tasks drops to the real figure (it was double counted).
0.12.13	UiPath/coder_eval#219	Opt-in agent.context_window cap. Unset by default, no change here.
0.12.13	UiPath/coder_eval#222	Delegate: batched tool calls count once under max_turns. Fewer false MAX_TURNS_EXHAUSTED on delegate runs.
0.12.14	UiPath/coder_eval#225	Codex skills are real copies in .agents/skills/, present during each turn and removed before grading.

• No test, checker or workflow here reads .agents/skills or sets context_window, so nothing else needs to change.

Validation

Check	Setup	Result
Run skill smoke tests	codex, --tags smoke over every skill, coder-eval 0.12.14 from PyPI	Pass
ADO dry run, Linux slice	codex · gpt-5.6-luna, skillsRef = this PR's head, coderEvalVersion=pinned (resolved 0.12.14, image coder-eval-agent:0.12.14), suites skills,activation, docker driver	skills 26/28 (gate ≥90%), activation 26/29, 0 errors

Skill install in the Linux dry-run log:

Signal	Count
Copied 32 skill(s) (one per codex turn)	63
Linked skill / symlink lines	0
.agents/skills/uipath/ paths or uipath:<skill> names	0

• Both skills misses activated their skill (skill_triggered 1.00) and missed on the task's own checks.
• The dry run's overall gate reports studio-web: host-sdk MISSING in harness smoke. Dry runs on 0.12.12 report the same line, so it does not come from this bump.
• Before/after on a live codex task (tempdir and docker): fix(codex): copy skills into .agents/skills instead of symlinking coder_eval#225.

🤖 Generated with Claude Code

[github-actions]

Claude finished @bai-uipath's task in 49s —— View job

---

Review of PR #3982

• Gather context and read review criteria
• Read the full diff
• Read changed files in full
• Check for consistency with project conventions
• Post review feedback

---

Summary

Bumps the pinned coder-eval version from 0.12.12 to 0.12.14 in tests/.coder-eval-version. This picks up five upstream changes (pricing rows, cost-accounting fix, context-window cap opt-in, delegate turn-counting fix, and codex skill-copy fix).

Change-by-Change Review

1. tests/.coder-eval-version

Severity: OK

Single-line change 0.12.12 → 0.12.14. No stale references to 0.12.12 remain anywhere in the repo (confirmed via grep). The file is the single source of truth for the coder-eval pin — consumed by smoke-skills.yml, smoke-rpa-skills.yml, run-coder-eval.yml, and the ADO nightly. All workflows read this file at runtime, so no other files need updating.

What's Missing

Nothing identified.

Area Ratings

Area	Status	Notes
Frontmatter	N/A	No skill files changed
E2E Tests	N/A	No test tasks changed
Skill Body	N/A	No skill files changed
References & Assets	N/A	No references changed
Repo Hygiene	OK	No secrets, no stale references, scoped change

Issues for Manual Review

• As noted in the PR description, coder-eval==0.12.14 is not yet published to PyPI. CI checks will fail until the upstream release workflow completes. The merge-order table in the PR description covers this clearly.

Conclusion

Clean, minimal version-pin bump. No issues found — approve once CI passes after the upstream release.

---