You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The review cycle dispatches up to 4 parallel reviewers (core + 3 aspect) regardless of mode. In vibe mode this is disproportionately expensive — the core reviewer finds most unique issues, and aspect reviewers largely duplicate its findings. Token analysis from run 20260305-184232Z (#181) showed 4-way parallel reviewer fan-out consumed ~326K tokens across 2 rounds.
The review-cycle module already handles variable fan-out correctly. The gap: mode presets in orchestrate-dev can't feed into aspect reviewer activation — activation is resolved from preferences.yaml only, not from the mode cascade.
Scope
Two changes in this ticket:
Thread aspect_reviewers through the mode cascade so modes control which aspect reviewers run.
Remove --lite mode entirely. Replace orchestration mode with effort system #199's auto-sizing within accountability levels makes lite redundant — a small task at default accountability gets lite-like settings automatically.
Design decisions
Vibe mode: explicitly deactivates all aspect reviewers (false). Vibe intentionally trades thoroughness for speed.
Default/strict: no override (absent). The orchestrator retains discretion via the review-cycle module's file-pattern activation logic, which skips aspect reviewers when the changed files don't warrant them.
preferences.yaml override removed: The orchestration.aspect_reviewers boolean override mechanism in review-cycle.md is removed. It was never used in practice and would complicate Replace orchestration mode with effort system #199's cascade. Can be re-added when a concrete use case emerges.
review-cycle becomes a pure function of its inputs: it no longer reads preferences directly — the engine threads {aspect_reviewers} as a context variable.
Changes
File
Change
orchestrate-dev/SKILL.md
Remove lite (argument, preset column, pipeline spec). Add aspect_reviewers row to mode presets.
orchestrate/SKILL.md
Remove lite from wrapper reference. Add {aspect_reviewers} to review-cycle context preparation.
review-cycle.md
Add {aspect_reviewers} input. Replace preferences.yaml lookup with input-driven activation.
Acceptance criteria
--mode=vibe deactivates all three aspect reviewers (only core reviewer runs)
Default mode (no --mode) leaves aspect reviewer activation to the file-pattern logic in review-cycle
--mode=strict leaves aspect reviewer activation to the file-pattern logic in review-cycle
--lite is no longer accepted as a mode value
review-cycle.md no longer references preferences.yaml or orchestration.aspect_reviewers
{aspect_reviewers} appears in the review-cycle inputs table and the orchestrate engine's context preparation section
No other files reference lite in the orchestration context
Expected impact
For vibe mode (1 round): ~120K tokens (1 reviewer × 1 round) vs ~163K (4 reviewers × 1 round). ~63% savings on the review phase.
Problem
The review cycle dispatches up to 4 parallel reviewers (core + 3 aspect) regardless of mode. In vibe mode this is disproportionately expensive — the core reviewer finds most unique issues, and aspect reviewers largely duplicate its findings. Token analysis from run
20260305-184232Z(#181) showed 4-way parallel reviewer fan-out consumed ~326K tokens across 2 rounds.The review-cycle module already handles variable fan-out correctly. The gap: mode presets in
orchestrate-devcan't feed into aspect reviewer activation — activation is resolved frompreferences.yamlonly, not from the mode cascade.Scope
Two changes in this ticket:
aspect_reviewersthrough the mode cascade so modes control which aspect reviewers run.--litemode entirely. Replace orchestration mode with effort system #199's auto-sizing within accountability levels makes lite redundant — a small task at default accountability gets lite-like settings automatically.Design decisions
false). Vibe intentionally trades thoroughness for speed.absent). The orchestrator retains discretion via the review-cycle module's file-pattern activation logic, which skips aspect reviewers when the changed files don't warrant them.orchestration.aspect_reviewersboolean override mechanism inreview-cycle.mdis removed. It was never used in practice and would complicate Replace orchestration mode with effort system #199's cascade. Can be re-added when a concrete use case emerges.{aspect_reviewers}as a context variable.Changes
orchestrate-dev/SKILL.mdlite(argument, preset column, pipeline spec). Addaspect_reviewersrow to mode presets.orchestrate/SKILL.mdlitefrom wrapper reference. Add{aspect_reviewers}to review-cycle context preparation.review-cycle.md{aspect_reviewers}input. Replace preferences.yaml lookup with input-driven activation.Acceptance criteria
--mode=vibedeactivates all three aspect reviewers (only core reviewer runs)--mode) leaves aspect reviewer activation to the file-pattern logic in review-cycle--mode=strictleaves aspect reviewer activation to the file-pattern logic in review-cycle--liteis no longer accepted as a mode valuereview-cycle.mdno longer referencespreferences.yamlororchestration.aspect_reviewers{aspect_reviewers}appears in the review-cycle inputs table and the orchestrate engine's context preparation sectionlitein the orchestration contextExpected impact
For vibe mode (1 round): ~120K tokens (1 reviewer × 1 round) vs ~163K (4 reviewers × 1 round). ~63% savings on the review phase.
Related