Make code-review a dedicated category served by the BC PR Review engine - #774

main tightened EvaluationContext.agent_name to AgentHarness and now reads .expected_metrics off it, so the engine runner's raw 'BC Review Engine' string was both a type error and a latent AttributeError at result-build time. Add a dedicated ENGINE harness (execution_time-only metrics, matching what run_engine_review reports) and use it in 'evaluate engine'.

Retire the bespoke /review prompt path that copilot/claude used: reject CODE_REVIEW in 'evaluate/run copilot|claude' (model comparison now flows through 'evaluate engine --model'), and remove the code-review-template from the shared prompt config.

De-pollute shared config: move the engine settings into an engine-owned agent/engine/config.yaml (read by _load_engine_settings), dropping the 'engine:' block from agent/shared/config.yaml.

Workflows: drop code-review from copilot/claude evaluation dispatch options and add engine-evaluation.yml as the sole code-review entry point, exposing engine-ref plus bcquality-ref/-repo overrides so BCQuality can be evaluated from a private branch.

…uched

Move the code-review engine knobs (path, agent_version, min_severity, bcquality) back into agent/shared/config.yaml under an 'engine:' block so they are easy to tweak alongside the other agent settings, and delete the separate engine-owned config.yaml. Restore claude-evaluation.yml to main: this PR no longer changes it (the 'claude' command still rejects code-review in code, so the category option stays as-is).

…null

CI supplies BC_REVIEW_ENGINE_ROOT (which takes precedence); a local run sets that env var or fills engine.path on a private branch.

agent_version and min_severity already have hard code defaults ('0.0.0' / 'Low') in run_engine_review, so listing them in config was redundant; drop them and keep only path and bcquality.{repo,ref,local_path}.

Drop the 'code-review always runs the engine' delegate comments (the branch condition + helper name already say it; model flow is documented in the engine agent docstring), the '2b integration' jargon in that docstring, and the redundant BC_REVIEW_ENGINE_ROOT inline note. Keep the comments that explain why (engine checkout, engine config knobs).

…ir, record engine revision

- _resolve_pwsh: drop the powershell (5.1) fallback so only PowerShell 7+ is used, matching the error message and the engine scripts' requirement.

- Prepare-BCQualityRoot.ps1: remove a reused fetch destination before init/fetch so a leftover root cannot carry stale (filtered-away) content or fail 'git remote add origin'.

- ExperimentConfiguration now records the resolved BC-ALAgents engine revision (with a dirty marker) as a bc-review-engine@<sha> plugin, so results from different engine commits are distinguishable.

…ng a clean review

_write_review_json now raises AgentError when agent-output.txt is empty or not a valid findings report (report is None), or when the report has no findings list (missing/non-list). Previously both silently produced an empty review.json, which the scorer credits as a correct 0-findings review (precision 1.0), hiding generation/dispatch failures. A legitimately empty review still carries a real findings: [] list and is preserved. Adds tests/test_engine_agent.py.

…_agent

The four run_* exports are distinct backends; run_engine_review spawns the PROD PowerShell orchestrator (Copilot lives inside the engine) and shares almost no parameters with run_copilot_agent. code-review is routed to the engine at the command layer, so it still lives under the copilot command for users. Adds notes at the __all__ export and the function docstring to pre-empt the recurring 'why not a copilot option?' question.

…5219)

The BC-Bench execution backend that runs the PROD BC-ALAgents review engine was named asymmetrically (run_engine_review) vs its three peers run_bcal_agent / run_claude_code / run_copilot_agent, which repeatedly drew reviewer questions about why it is a separate top-level backend.

Rename the backend org naming only: agent/engine -> agent/pr_review; run_engine_review -> run_pr_review_agent; _run_engine(_evaluation) helpers; CLI commands 'engine' -> 'pr-review'; AgentHarness.ENGINE member -> PR_REVIEW.

Deliberately keep the domain term 'engine' where it names the external PROD BC-ALAgents review engine: AgentHarness value stays 'BC Review Engine', BC_REVIEW_ENGINE_ROOT env, config engine.path, _resolve_engine_* helpers, engine_report_* functions, and copilot-evaluation.yml are untouched.

…ons block

When the dead code-review-template prompt was removed from config.yaml, the line immediately after it (the '# 1. whether to copy custom instructions' bullet of the instructions: control block) was deleted too, leaving an orphaned '# 2.' with no '# 1.'. Restore the bullet so the numbering matches the skills:/agents: blocks.

Align the remaining user-facing names with the pr_review backend:

- AgentHarness label value 'BC Review Engine' -> 'BC PR Review' (result agent_name + summarize dashboard); synced in copilot-evaluation.yml.

- Env var BC_REVIEW_ENGINE_ROOT -> BC_PR_REVIEW_ROOT (agent.py, workflow, config comment, docstrings).

- Config key engine.path -> pr_review.path (config block + accessors); helpers _load_engine_settings/_resolve_engine_root -> _load_pr_review_settings/_resolve_pr_review_root.

- Fixed stale 'bcbench evaluate engine' hint -> 'bcbench evaluate pr-review'.

Kept engine-provenance internals (bc-review-engine plugin, load_engine_report, product prose) as they name the external BC-ALAgents engine.

code-review is always served by the fixed BC-ALAgents review engine, so it no longer masquerades as a peer of the copilot/claude harnesses:

- Rename 'run/evaluate pr-review' -> 'run/evaluate code-review' (category-named, not harness-named).

- Remove the copilot code-review reroute (asking for copilot no longer silently runs the engine) and the claude cross-pointer error; both now reject code-review uniformly via a shared reject_code_review() helper that points to the dedicated command.

- CI: split into a code-review engine step and a Copilot step, so the workflow mirrors the command split instead of routing code-review through 'evaluate copilot --category'.

The PR-review engine backend is the Copilot-powered BC-ALAgents review engine, not a fourth AI harness. Move agent/pr_review -> agent/copilot/pr_review so the package tree reflects that, and drop run_pr_review_agent from the top-level bcbench.agent __all__ so it is no longer exported as a peer of run_copilot_agent/run_claude_code/run_bcal_agent. It is reached only via the dedicated code-review command, importing from bcbench.agent.copilot.pr_review.

…w entrypoints