Repository navigation
Description
Goal
Restructure the cascading PR-review pipeline so the deterministic PR context (metadata + diff) is gathered ONCE in a shell preflight step and passed to the agentic tiers as pre-populated context, instead of each tier re-fetching it via gh pr view / gh pr diff Bash tool calls inside its own reasoning loop. This eliminates redundant tool round-trips and the reasoning tokens spent deciding to issue them, reducing Effective Tokens (ET) on the highest-volume agent workflow. Motivated by GitHub's May-2026 result (62% ET reduction on comparable agentic CI workflows) and by the Token Cost Observatory (#464).
Grounding correction vs the idea
The triage tier already prefetches metadata + diff at the shell level and runs with NO tools (review-one-pr.sh tier 1, --disallowed-tools). The real remaining opportunity is the four agentic tiers — deep review, rubber-duck, security-audit, single-review — which the tier prompts explicitly instruct to run gh pr view --json <20+ fields> and gh pr diff in-loop (prompts/deep-review.md steps 3-4 and peers). The proven in-repo template is that same triage prefetch (not dev-lead-preflight.sh, which is only a secrets check). Note also: pre-feeding does NOT let us drop the Bash/Read/Grep/Glob tools — those tiers still need them for dynamic cross-file exploration and gh api standards lookups; the saving is the eliminated redundant fetch, not tool-schema removal (MCP is already opt-in / default-off).
Success metric (initiative-level)
Across a fixed A/B sample of real reviews, the deep + security-audit tiers show a measurable, sustained ET reduction vs a frozen pre-change baseline (directional target >=30% on those tiers' input+ET, reported as an exact delta — not a hard gate), with zero regression on the held-out deep-review eval suite (evals/deep-review/holdout/). ET is read from the existing per-call JSONL (scripts/lib/token-metrics.sh, ET = m*(1.0I + 0.1C + 4.0*O)) and aggregated by tier via the Token Cost Observatory (scripts/token_report.sh, #464).
Cost cap
The measurement + validation work is explicitly bounded: the A/B run covers a fixed sample of at most ~30 PRs per arm (on/off), the review-quality check reuses the EXISTING held-out deep-review eval suite (bounded case count, no new eval generation), and no story introduces an unbounded or self-tuning loop. No production default is flipped until the bounded A/B + eval gate pass and a human records a go/no-go.
Shape
Phase 1 audits + freezes a baseline. Phase 2 builds the shared prefetch plumbing (flag-gated, byte-identical when off) and rewires the tier prompts to consume it. Phase 3 adds a HEAD-SHA freshness safeguard against force-push staleness. Phase 4 runs the bounded A/B + eval regression gate and gates the default-on flip on a human decision.
Activation is a human decision — this epic is created inert.
Planned from idea discussion #563 by the BMAD Scrum Master initiative-planner. Inert until a maintainer adds initiative:auto.
Activity
don-petry commentedon Aug 20, 2026
Initiative driver — stale in-flight slot(s) detected
The following open sub-issue(s) still carry the
dev-leadlabel but havehad no dev-lead activity for ≥ 72h, so they are occupying an in-flight
slot without progressing: #1199, #1200.
No labels were changed. A human should decide whether to split-and-close, retry,
or park each one so the slot is freed for the rest of the epic.