[Phase 3] Wire the DOWNSTREAM_IMPACT block into the triage + deep prompts and surface it in the verdict

[github-actions]

Story

As a PR reviewer,
I want the assembled DOWNSTREAM_IMPACT context fed to the cascade and reflected in the posted review,
so that I see which consumer repos a reusable-workflow change affects, right in the review, before it reaches ring-0.

Acceptance Criteria

1. scripts/review-one-pr.sh inlines the DOWNSTREAM_IMPACT block into TRIAGE_PROMPT_FILE (alongside the existing ADVISORY_BOT_FEEDBACK inlining) and exports DOWNSTREAM_IMPACT_FILE for the deep tier, when the feature flag is on.
2. prompts/triage.md and prompts/deep-review.md document the new input and instruct the model to treat downstream impact as a signal to weigh (annotate impacted consumers; escalate only per existing risk criteria — downstream impact alone is informational, not an auto-escalation trigger).
3. The posted review body surfaces an impacted-consumers note when impact is non-empty (e.g. "This change to <reusable> is consumed by N downstream repos: ..."), via the deep/synthesize prompt body field.
4. When the block is (none), no downstream note appears in the verdict and triage behavior is unchanged.
5. A bats/integration-style test asserts the prompt file contains the block when impact is present and omits the note when (none); shellcheck passes.

Tasks / Subtasks

• In scripts/review-one-pr.sh, call the Story 3 assembler before building TRIAGE_PROMPT_FILE, inline the block into the triage prompt, and export DOWNSTREAM_IMPACT_FILE for tiers 2/3 — all gated on the Story 5 feature flag. (AC: test issue from agent #1, Optimize review: small-PR and incremental fast paths #4)
• Update prompts/triage.md and prompts/deep-review.md: add the new input under their inputs sections and add guidance that downstream impact is informational (annotate, do not auto-escalate). (AC: Go-live improvements for PR review agent #2)
• Update the deep/synthesize prompt so the verdict body includes the impacted-consumers note when present. (AC: Add @claude delegation, auto-merge, and rebase handling #3)
• Add a test asserting prompt-assembly behavior for impact-present vs (none); register any new bats file in lint.yml. (AC: feat: add Copilot engine support via REVIEW_ENGINE toggle #5)

Dev Notes

• Triage has NO tools — every field must be inlined into TRIAGE_PROMPT_FILE (see the { cat prompts/triage.md; printf ... } > TRIAGE_PROMPT_FILE block at scripts/review-one-pr.sh:539). Add the DOWNSTREAM_IMPACT section there, next to ADVISORY_BOT_FEEDBACK.
• The deep tier reads enrichment from env-passed file paths (it lists e.g. $ADVISORY_BOT_FEEDBACK_FILE in prompts/deep-review.md Inputs and reads it in step 2). Add $DOWNSTREAM_IMPACT_FILE the same way.
• Do NOT make downstream impact an auto-escalation trigger — the triage decision criteria in prompts/triage.md are explicit and risk-based; downstream impact is an annotation/signal the model weighs, consistent with how ADVISORY_BOT_FEEDBACK is weighed. (Keeps AC free of a contested escalation-policy decision.)
• The verdict body is authored by the model in the body field consumed by scripts/post-pr-review.sh — surface the note via prompt instruction, not by post-processing the body in shell.

Project Structure Notes

Touches scripts/review-one-pr.sh and the prompts/ cascade files. Behavior remains gated by the Story 5 flag, so merging this story alone changes nothing until the flag is enabled.

References

• scripts/review-one-pr.sh#TRIAGE_PROMPT_FILE
• prompts/triage.md#decision-criteria
• prompts/deep-review.md#inputs
• scripts/post-pr-review.sh#body

Likely target surface

• scripts/review-one-pr.sh
• prompts/triage.md
• prompts/deep-review.md
• prompts/synthesize.md

Story prepared by the BMAD Scrum Master (Bob) for epic #748. Status: ready-for-dev.

[don-petry]

Planning note — implementation ordering (dependency on #753).

This story gates the downstream-impact wiring on the DOWNSTREAM_IMPACT_ENABLED feature flag, but that flag is introduced by #753 (Phase 3 — default-off flag + regression guard), which is sequenced after this story.

To avoid a half-wired intermediate state, implement #753 before #752 (or land the flag's definition as the first task of this story). Effectively: #752 should be treated as blocked_by #753 for the flag, even though the rest of the chain is #749 → #750 → #751 → wiring.

No scope change — just the build order. Flagged during a pre-activation DAG review.

---

Generated by Claude Code

[don-petry]

Dev-Lead Implementation Plan

Issue: #752 — [Phase 3] Wire the DOWNSTREAM_IMPACT block into the triage + deep prompts and surface it in the verdict

Scope

Behind the default-off DOWNSTREAM_IMPACT_ENABLED flag (Story 5), review-one-pr.sh assembles the Story 2/3 downstream-impact block, inlines it into the triage prompt (triage has no tools), exports DOWNSTREAM_IMPACT_FILE for the deep/audit tiers, and the prompts surface impacted consumers as an informational signal in the verdict body — without making downstream impact an auto-escalation trigger.

Implementation Checklist

• scripts/lib/downstream-impact.sh — add a sourced downstream_impact_triage_section helper that emits the triage inlining (nothing when the flag is off → byte-identical; the block when impact present; (none) otherwise).
• scripts/review-one-pr.sh — source the lib; when DOWNSTREAM_IMPACT_ENABLED=true, derive changed paths from PR_METADATA, call assemble_downstream_impact against scripts/lib/consumer-manifest.json (exports DOWNSTREAM_IMPACT_FILE), and inline the section into TRIAGE_PROMPT_FILE next to ADVISORY_BOT_FEEDBACK.
• prompts/triage.md — document the DOWNSTREAM_IMPACT input + guidance: annotate impacted consumers, informational only, not an auto-escalation trigger.
• prompts/deep-review.md — document $DOWNSTREAM_IMPACT_FILE, read it in steps, same informational guidance.
• prompts/cascade-action.md — when the file is non-empty and not (none), add an impacted-consumers note to the verdict body (the prompt that authors the posted body; synthesize.md is not in this cascade's live path).

Tests to Write

• tests/test_downstream_impact_triage_inline.bats — asserts the triage section contains the block when impact is present, shows (none)/omits the consumer list when (none), and emits nothing when the flag is off.
• Register the new bats file in .github/workflows/lint.yml.

Test Command

bash scripts/run-bats.sh tests/test_downstream_impact_triage_inline.bats (full suite via lint.yml), plus bash .dev-lead/scripts/dev-lead-lint.sh (shellcheck + frontmatter).

Note

The issue lists prompts/synthesize.md as a likely target, but that prompt is not in this cascade's live path — scripts/review-one-pr.sh authors the posted body via prompts/cascade-action.md. I will update cascade-action.md instead (the real synthesizer), per the Dev Note "the verdict body is authored by the model in the body field consumed by scripts/post-pr-review.sh."

[don-petry]

Dev-Lead: Implementation Complete

Plan Execution

• Added downstream_impact_triage_section helper to scripts/lib/downstream-impact.sh (inlines the block; no-op when the flag is off).
• Wired the flag-gated assembly + triage inlining into scripts/review-one-pr.sh (exports DOWNSTREAM_IMPACT_FILE for tiers 2/3).
• Documented the new input + informational/non-escalation guidance in prompts/triage.md and prompts/deep-review.md.
• Surfaced the impacted-consumers note in the verdict body via prompts/cascade-action.md.
• Added tests/test_downstream_impact_triage_inline.bats and registered it in .github/workflows/lint.yml.

Acceptance Criteria

1. AC#1 — review-one-pr.sh assembles the block, inlines it into TRIAGE_PROMPT_FILE next to ADVISORY_BOT_FEEDBACK, and exports DOWNSTREAM_IMPACT_FILE for the deep/audit tiers — all under DOWNSTREAM_IMPACT_ENABLED. ✓
2. AC#2 — triage.md + deep-review.md document the input and instruct the model to annotate impacted consumers, escalating only per existing risk criteria (downstream impact is informational, never an auto-escalation trigger). ✓
3. AC#3 — cascade-action.md (the prompt that authors the posted body) inserts a "Downstream impact / consumed by N downstream repos" note when the block is non-empty. ✓
4. AC#4 — (none) ⇒ no consumer list in triage and no body note; flag-off ⇒ byte-identical triage prompt and zero gh fetches. ✓
5. AC#5 — new bats asserts block-present vs (none) behavior; shellcheck passes. ✓

Test Results

shellcheck -S warning scripts/lib/downstream-impact.sh scripts/review-one-pr.sh  → PASS

bats tests/test_downstream_impact_triage_inline.bats
1..8
ok 1 emits nothing when the feature flag is off
ok 2 emits nothing when the feature flag is explicitly false
ok 3 inlines the block when impact is present and the flag is on
ok 4 the inlined section labels downstream impact as informational, not auto-escalation
ok 5 shows (none) and omits any consumer list when the block is (none)
ok 6 treats a missing impact file as (none) without failing
ok 7 reads DOWNSTREAM_IMPACT_FILE when no path argument is passed
ok 8 appended to a triage prompt file, the block is present for impact and absent for (none)

Full lint.yml bats suite (all registered files): 560/560 ok, 0 failures.
bash .dev-lead/scripts/dev-lead-lint.sh → all checks passed (67 shellcheck files, 4 agent profiles).

Files Changed

• scripts/lib/downstream-impact.sh: new downstream_impact_triage_section helper (flag-gated triage inlining).
• scripts/review-one-pr.sh: source the lib; flag-gated assembly from PR_METADATA changed files; inline the section into the triage prompt.
• prompts/triage.md: DOWNSTREAM_IMPACT input + informational/non-escalation guidance.
• prompts/deep-review.md: $DOWNSTREAM_IMPACT_FILE input + read step + guidance.
• prompts/cascade-action.md: surface the impacted-consumers note in the verdict body.
• tests/test_downstream_impact_triage_inline.bats: new test (registered in .github/workflows/lint.yml).

Notes

• synthesize.md vs cascade-action.md: the issue listed prompts/synthesize.md as a likely target, but that prompt is not in this cascade's live path — review-one-pr.sh authors the posted body via prompts/cascade-action.md (escalated) and prompts/single-review.md (triage-approved). Per the Dev Note ("the verdict body is authored by the model in the body field consumed by scripts/post-pr-review.sh"), I implemented the body note in cascade-action.md.
• Scope of the body note: per the issue's "deep/synthesize" framing, the note is surfaced on the deep→synthesize (escalated) path. The triage-approved quick-confirm path (single-review.md) was left untouched to avoid scope creep; surfacing the note there too is a small, reasonable follow-up if desired.
• Pre-existing failure: tests/test_batch_fallback.bats test "Gemini trust error (55) falls back to Copilot" fails on the base branch as well (unrelated to this change; that file is not in the lint.yml bats list).
• Behavior remains fully gated by DOWNSTREAM_IMPACT_ENABLED (Story 5, [Phase 3] Default-off feature flag + regression guard + operator docs for the downstream-impact pass #753), so merging this story alone changes nothing until an operator enables the flag.