[Phase 2] Rewire the deep-review + rubber-duck tier prompts to consume pre-fed context #1104

Story

As a pipeline maintainer,
I want update the deep-review and rubber-duck prompts to read the pre-fed metadata/diff files when present instead of running gh pr view / gh pr diff, while keeping dynamic exploration tools available,
so that the two tier-2 reviewers stop spending tool round-trips and reasoning tokens on deterministic re-fetching, capturing most of the ET saving.

Acceptance Criteria

1. When the prefetch flag is on and PR_CONTEXT_* files are present, prompts/deep-review.md and prompts/rubber-duck.md instruct the model to read the pre-fed metadata + diff files instead of running gh pr view / gh pr diff for the primary PR context.
2. When the flag is off (or the files are absent), both prompts fall back to the existing in-loop gh pr view / gh pr diff steps unchanged — behavior is byte-identical to today.
3. Both prompts explicitly retain permission to use Bash/Read/Grep/Glob and gh api for DYNAMIC needs (cross-file reads, org-standard lookups, secret-scan MCP) — only the deterministic metadata+diff fetch is replaced.
4. The pre-fed diff the deep tier consumes is the FULL diff (per Story 2), so the tier's reviewable surface is not reduced vs today.
5. The deep-review held-out eval is run via scripts/evals/run-eval.sh deep-review (5 cases, llm-judge scorer, pass_threshold 0.7) with the prompt change applied and the flag on, and its aggregate score is NOT lower than the flag-off baseline (no regression) — it is not required to improve.

Tasks / Subtasks

• Add a conditional 'pre-fed context' inputs section to deep-review.md and rubber-duck.md (AC: test issue from agent #1, Go-live improvements for PR review agent #2)
  • Reference PR_CONTEXT_METADATA_FILE / PR_CONTEXT_DIFF_FILE the same way $SAFETY_CHECKS_FILE is referenced
  • Keep the gh pr view/diff steps as the documented fallback when the files are absent
• Preserve dynamic-tool guidance (AC: Add @claude delegation, auto-merge, and rebase handling #3)
  • Keep the FORBIDDEN-enumeration guardrail and the gh api standards path intact
• Confirm full-diff consumption (AC: Optimize review: small-PR and incremental fast paths #4)
• Run the deep-review held-out eval and confirm no regression (AC: feat: add Copilot engine support via REVIEW_ENGINE toggle #5)
  • Use scripts/evals/run-eval.sh deep-review against evals/deep-review/holdout
  • Do NOT modify held-out cases (holdout-guard.sh enforces immutability)

Dev Notes

• The deep tier reads $SAFETY_CHECKS_FILE, $ADVISORY_BOT_FEEDBACK_FILE, $DOWNSTREAM_IMPACT_FILE today (prompts/deep-review.md Inputs). Add $PR_CONTEXT_METADATA_FILE / $PR_CONTEXT_DIFF_FILE the same way, with an explicit 'if set, read these INSTEAD of gh pr view/diff; if unset, run the gh commands as before' instruction — this keeps the flag-off path byte-identical (required for rollback + holdout-eval stability, per the SAFETY_CHECKS precedent).
• Do NOT weaken the tier's scope guardrails: keep the FORBIDDEN: gh search prs / gh pr list / enumeration block and the secret-scan MCP step. The model still needs Bash/Read/Grep/Glob (granted by engine.sh run_agentic/run_duck) for dynamic cross-file analysis — only the two deterministic fetches move to files.
• rubber-duck runs cross-engine (often a different model family, e.g. Gemini via run_duck); its prompt's gh pr view --json ...,repository,... field list differs slightly from deep's — ensure the pre-fed metadata superset (Story 2) covers repository too, or the duck prompt reads defensively.
• Regression gate: scripts/evals/run-eval.sh deep-review scores against evals/deep-review/holdout/cases.jsonl using the llm-judge scorer (pass_threshold in scorer.json). The held-out cases are immutable (scripts/lib/holdout-guard.sh, CODEOWNERS over evals/) — never edit them to make the number move.
• This story is independently reviewable: it touches only the two prompt files and their eval run; it depends on Story 2's plumbing but not on Story 4.
• Eval mechanics: scripts/evals/run-eval.sh executes the skill prompt once per held-out case via engine.sh run_triage with the case context INLINED (not fetched via tools). So the eval already exercises pre-fed-style context — a valid guard that the rewritten prompt still reasons correctly on inlined metadata/diff. It cannot catch a full-diff-vs-truncated mismatch; that is covered by Story 2's full-diff constraint and Story 6's A/B smoke.

Project Structure Notes

Touches prompts/deep-review.md and prompts/rubber-duck.md only. Eval invocation via existing scripts/evals/run-eval.sh; no eval-case edits (guarded).

References

• prompts/deep-review.md#inputs
• prompts/rubber-duck.md#steps
• scripts/evals/run-eval.sh
• evals/deep-review/holdout/cases.jsonl
• scripts/lib/holdout-guard.sh

Likely target surface

• prompts/deep-review.md
• prompts/rubber-duck.md

Story prepared by the BMAD Scrum Master (Bob) for epic #1101. Status: ready-for-dev.