Repository navigation
Open
Labels
enhancementFeature requestsFeature requests
Description
Story
As a release manager,
I want the next shadow lane dispatched alongside stable on eligible PRs and the shadow_dual_run signal emitted from the real dual run,
so that a candidate release is validated against live production behaviour before any consumer sees it (initiative #495, SC5).
Split 2 of 2 from #1269. Blocked by the suppression story — dispatching a shadow lane before suppression is proven would post candidate output to real PRs.
Scope
With shadow_mode and total output suppression already landed and tested, turn the second lane on and wire the signal.
Acceptance Criteria
- On an eligible PR, the
nextlane is dispatched in parallel with thestablelane, withshadow_mode: true. Onlystableposts — guaranteed by the suppression from split 1, which this story must not weaken or bypass. - Lane concurrency is respected: the shadow lane must not evict or be evicted by the stable lane. The reusable's per-PR / per-SHA lanes exist precisely so unrelated work is not cancelled — give the shadow lane its own key rather than sharing stable's.
scripts/shadow-run.shis invoked with both lanes' results (STABLE_CONCLUSION,SHADOW_CONCLUSION, output files, run ids) and emits theshadow_dual_runJSON signal. Reuse it as-is — it is already onmainfrom PR feat: implement issue #605 — Shadow-mode dual-run for safe agent canary validation #1252.- Eligibility is explicit and opt-in: shadow dual-run happens only where deliberately enabled. With it off, behaviour is byte-identical to today.
- A shadow-lane failure is never allowed to fail the PR's checks or block the stable lane.
REGRESSIONis a promotion signal, not a PR gate — perdocs/release/shadow-mode.md, only the promotion gate consumes it. - Cost is bounded and stated: dual-run doubles agent spend on eligible PRs. The PR description names the eligibility scope and the expected multiplier.
- Tests: both lanes dispatched; only stable posts; the signal is emitted with the right status for at least
MATCH,REGRESSIONandNO_SHADOW; a shadow failure leaves PR checks green.
Tasks / Subtasks
- Dispatch the
nextlane withshadow_mode: truealongsidestable(AC: test issue from agent #1, Optimize review: small-PR and incremental fast paths #4) - Give the shadow lane its own concurrency key (AC: Go-live improvements for PR review agent #2)
- Collect both lanes' results and invoke
shadow-run.sh; emit the signal (AC: Add @claude delegation, auto-merge, and rebase handling #3) - Ensure a shadow failure cannot affect PR checks (AC: feat: add Copilot engine support via REVIEW_ENGINE toggle #5)
- State eligibility scope + cost multiplier in the PR (AC: feat: feature ideation framework — adversarial proposer/challenger/synthesizer #6)
- Tests (AC: Workflow should skip no-op PRs in MAX_PRS count #7)
Dev Notes
- Do not modify
shadow-run.shorshadow-compare.sh. They are the tested comparison + signal half (PR feat: implement issue #605 — Shadow-mode dual-run for safe agent canary validation #1252). This story is dispatch and wiring only. - Out of scope: wiring the signal into the promotion gate's consumption path. That gate lives in
petry-projects/.github/scripts/canary-rollout.sh— a different repo — and today readsstandards/canary-rings.json, with no consumer forgate.signals. Emitting the signal is this story; consuming it is tracked separately. State this exclusion in the PR rather than silently omitting it. - Design reference:
docs/release/shadow-mode.md— in particular that onlyREGRESSIONblocks promotion andDIVERGEDis advisory. - Budget: keep dispatch and signal wiring together (they are one behaviour) but nothing more. If it runs long, land dispatch + suppression verification and split the signal emission.
References
.github/workflows/dev-lead-reusable.yml;docs/release/shadow-mode.mdscripts/shadow-run.sh,scripts/lib/shadow-compare.sh(PR feat: implement issue #605 — Shadow-mode dual-run for safe agent canary validation #1252)- Parent Shadow-mode dual-run — dispatch half: run the shadow lane + suppress its PR output (follow-up to #605) #1269; Shadow-mode dual-run for safe agent canary validation #605 → [Phase 3] Automate the scheduled proposer behind the safe-rollout gates #587 → epic Initiative: Eval-gated, human-reviewed self-improving skills (SkillOpt-style) #581; initiative Initiative: Safe Release Strategy for Agentic Workflows (versioning · rings · canary) #495 SC5
Metadata
Metadata
Assignees
Labels
enhancementFeature requestsFeature requests
Activity
github-actions commentedon Sep 8, 2026
Thanks for the detailed write-up and for splitting this out cleanly from #1269! Since this is dispatch and wiring work building on suppression already landed, the main thing that would help reviewers most: which workflow file(s) and job steps are the intended insertion points for the parallel
next-lane dispatch, and is there a preferred mechanism (e.g., a newuses:step, a matrix, or explicitworkflow_dispatchcall) given the concurrency-key requirement in AC #2?