Wave 3 plugin build report#

Outcome#

The two Copilot plugins under copilot/ were revised in place so they are no longer shallow prompt scaffolds:

• miadi-adversarial-review-kit now behaves like a dissenting review surface with explicit outcome-drift and promotion-block logic.
• miadi-promotion-context-kit now behaves like a promotion-routing surface that separates provenance, spec, wiki, and context-layer concerns instead of flattening them.

Weaknesses found in the prior scaffold#

1. The adversarial kit named the right evidence layers, but it did not tell the agent how to issue a promotion veto, how to separate evidence from inference, or how to audit STCKin/context-layer outcomes.
2. The promotion kit described provenance/spec/wiki at a high level, but it did not encode a real decision matrix or a durable boundary between retrieval/context-layer material and wiki prose.
3. Both kits were too generic for replay: they lacked explicit output contracts, stronger included-surface descriptions, and artefact-friendly record instructions.
4. Manifest wording was serviceable but underspecified; it did not make the two plugin roles sharply different.

Changed files#

• copilot/miadi-adversarial-review-kit/README.md
• copilot/miadi-adversarial-review-kit/.github/plugin/plugin.json
• copilot/miadi-adversarial-review-kit/agents/miadi-adversarial-reviewer.md
• copilot/miadi-adversarial-review-kit/agents/rise-revision-critic.md
• copilot/miadi-adversarial-review-kit/agents/context-outcome-auditor.md
• copilot/miadi-adversarial-review-kit/skills/adversarial-rise-review/SKILL.md
• copilot/miadi-adversarial-review-kit/skills/artefacted-revision-report/SKILL.md
• copilot/miadi-adversarial-review-kit/skills/outcome-drift-audit/SKILL.md
• copilot/miadi-promotion-context-kit/README.md
• copilot/miadi-promotion-context-kit/.github/plugin/plugin.json
• copilot/miadi-promotion-context-kit/agents/miadi-promotion-architect.md
• copilot/miadi-promotion-context-kit/agents/wiki-promotion-curator.md
• copilot/miadi-promotion-context-kit/agents/context-layer-boundary-keeper.md
• copilot/miadi-promotion-context-kit/skills/context-layer-promotion/SKILL.md
• copilot/miadi-promotion-context-kit/skills/wiki-promotion-draft/SKILL.md
• copilot/miadi-promotion-context-kit/skills/promotion-decision-record/SKILL.md

Strongest improvements made#

1. Clear role split. The adversarial kit now blocks, audits, and revises; the promotion kit now routes, curates, and records promotion decisions.
2. Corpus alignment. The new wording explicitly preserves key corpus constraints: STCKin stays in the STC family, the context layer stays a downstream AvaLangStack consumption surface, and promotion still follows provenance -> spec -> wiki.
3. Replayability. Both kits now require concrete output structures such as evidence ledgers, promotion matrices, severity-grouped revisions, and artefact-friendly markdown records.
4. Boundary protection. The promotion kit now has a dedicated context-layer-boundary-keeper agent and explicit rules against flattening retrieval, lineage, ontology, and decision-memory concerns into generic wiki prose.
5. Outcome skepticism. The adversarial kit now has a dedicated context-outcome-auditor agent and outcome-drift-audit skill so plausible scaffolds do not pass as delivered outcomes.

Remaining weaknesses#

1. These plugins are still markdown orchestration surfaces, not executable enforcement layers; they depend on the quality of the active session and available evidence.
2. There is still no repo-level index or launcher update advertising these two kits alongside stckin-orchestration-kit.
3. The kits now define artefact-friendly outputs, but they still do not ship shared templates or automation that guarantees those reports are written the same way every time.