v0.7.1 -- Context diet

Agent Stubs

12 agents extracted to lean frontmatter stubs -- just tool permissions, hooks, and a one-line body. Full methodology moved into skill directories as sibling files (methodology/gsp-{agent}.md), loaded on demand at spawn time. Session-start context dropped from ~1,536 lines to ~140 lines. Three agents removed entirely: gsp-ascii-artist, gsp-brand-syncer, gsp-campaign-director (plus gsp-scoper).

exec_context Refactor

Six skills were loading reference files into <execution_context> that the orchestrator never read -- they existed purely to be copy-pasted into agent prompts. Moved to explicit Read steps before agent spawn. ~1,579 lines removed from orchestrator context windows. Affected: brand-strategy (604 lines), brand-guidelines (278), brand-identity (239), brand-research (231), accessibility-audit (159), project-critique (68).

Benchmark System

New benchmark.sh captures JSON snapshots of token budget state -- per-skill scores, pipeline path totals, rate limit risk, test results. Release baselines freeze at each version; compare diffs against the baseline; trend shows trajectory across all releases. Shared scoring extracted to lib-scoring.sh, eliminating duplication between token-budget.sh and the benchmark. 8 historical snapshots retroactively captured from v0.5.0 through v0.7.0.

Token Budget Impact

| Metric | v0.6.3 | v0.7.1 | Delta | |--------|--------|--------|-------| | Total weight | 21,228 | 15,634 | -26% | | Red-zone skills | 7 | 4 | -3 | | Full E2E pipeline | 10,888 | 9,793 | -10% | | Brand diamond | 4,619 | 3,251 | -30% |

Model Selection

Skills no longer declare model: or effort: in frontmatter. Users control which model runs each phase. Pipeline skills include a hint in their description (e.g., "benefits from capable models") as passive guidance. The installer strips these fields for non-Claude runtimes.