[FEATURE] Add a 'propagate-the-fix' (downstream consistency propagation) pattern to the Workflow tool's quality-patterns library

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

The Workflow tool's built-in description documents a small library of named "quality patterns" for multi-agent orchestration — currently adversarial verify, perspective-diverse verify, judge panel, loop-until-dry, multi-modal sweep, and completeness critic. This library is the go-to reference an orchestrating session consults when composing a workflow's review/verification stage, so a common failure mode that isn't named there tends to go un-mitigated.

One such uncovered failure mode is what I'll call downstream consistency propagation: a review/consistency-check agent correctly identifies a root-cause inconsistency in an artifact, but then only reports the first place it noticed it — it never traces that same fact through every other place the artifact restates or assumes it. The result is a review that returns a confident "well-constructed, one item flagged for a human" verdict while several purely-mechanical, downstream consequences of the very inconsistency it already found remain un-caught.

Concretely (details anonymized): I ran a workflow with several drafting agents and one review agent explicitly instructed to check cross-file consistency and fix trivial issues directly. The review agent correctly discovered a real root-cause fact — a capability was structurally unavailable in some of the configurations being documented — and correctly flagged that one file still hedged this as "unconfirmed" while a sibling file had already resolved it confidently. Its overall verdict called the document set "internally consistent" and "ready to run."

On a subsequent manual pass I found four additional, purely mechanical inconsistencies in the same file that were direct downstream consequences of the exact fact the review agent had already surfaced: other sections still described the now-unavailable capability as present and testable, still budgeted resources for exercising it, and still referenced its (nonexistent) UI in step-by-step instructions. All four were trivially fixable once the root-cause fact was known — the review agent already knew it. They just required grepping the same file for every other reference to that fact, which the review pass never did even though its own prompt asked it to "read the whole document." The stated confidence ("internally consistent," "ready") therefore wasn't fully earned; a human had to redo trace-through work the review agent was already positioned to do.

None of the six existing patterns cover this: adversarial/perspective-diverse verify test whether a found finding is true; loop-until-dry and multi-modal sweep broaden discovery of new findings; the completeness critic asks "what modality/source did we skip?". None instructs an agent that, once it finds a root-cause fact that invalidates part of an artifact, it must reconcile every restatement of that fact within the same artifact before declaring "ready."

Proposed Solution

Add a seventh named pattern to the Workflow tool's quality-patterns library, e.g.:

Propagate-the-fix (downstream consistency propagation): when a review/consistency agent finds a root-cause fact that invalidates part of an artifact, its prompt should instruct it to search the SAME artifact(s) for every other place that fact is assumed or restated — not just the first place it was noticed — and reconcile all of them before returning a "ready" verdict. Reporting only the first instance leaves mechanically-fixable downstream consequences un-caught.

This mirrors the "simple counters miss the tail" spirit already baked into loop-until-dry, but applied to consistency reconciliation within an artifact rather than to discovery breadth. Being a named entry in the same library is what matters: orchestrators compose review stages by reaching for these named patterns, so an unnamed failure mode is a silently-unmitigated one.

Alternative Solutions

• Documenting the pattern only in a project-local skill/doc. This doesn't scale — the failure mode is generic to any review stage a session composes, and the Workflow tool's own library is the canonical, always-present reference. A per-project note is invisible to sessions that don't happen to load it.
• Relying on the orchestrator to hand-write the instruction each time. That's exactly what the pattern library exists to prevent: it exists so common review disciplines don't have to be re-derived per workflow.

Priority

Medium - Would be very helpful

Feature Category

API and model interactions

Use Case Example

1. A session orchestrates a workflow whose final stage is a single review agent checking a multi-section document (or set of related documents) for internal consistency, with authority to fix trivial issues directly.
2. The review agent finds a root-cause fact that makes part of the document invalid (e.g. "feature X is unavailable in configuration Y").
3. With this pattern named in the library, the orchestrator's review-agent prompt (built by reaching for "propagate-the-fix") instructs it to grep the same document for every other mention of feature X and reconcile each one before returning.
4. The review returns with all downstream restatements already reconciled, instead of a "ready" verdict that a human must still re-audit — saving a full redundant trace-through pass.

Additional Context

• Current library observed in Claude Code v2.1.217 (the six named patterns above).
• The requested pattern is deliberately scoped to reconciliation within an already-found finding, distinct from the existing discovery-breadth (loop-until-dry, multi-modal sweep) and finding-validation (adversarial/perspective-diverse verify) patterns.