Create branch-scored delegated design improvement loop

[iainherd]

Goal

Create an automated/semiautomated delegated design-improvement loop for Lir Capital: identify design problems/opportunities, formulate solution candidates, implement candidates on isolated branches, score them with agent-profile reviews and browser/design evidence, deploy the best accepted branch to pre-prod, and repeat.

This is the operating model for moving from the current solid pre-prod baseline toward award-winning, high-end brand/design-agency quality.

Context

The current delegated issue workflow is working: bounded implementation subagents, assigned agent:* review profiles, parent/orchestrator verification, draft PRs, CI, and merge. The next level should turn this into a repeatable design-lab loop inspired by branch-based exploration/evaluation workflows: multiple candidate branches can compete, the strongest candidate becomes the new baseline/pre-prod, and weak ideas are discarded without contaminating main.

Related issues:

• Run agency-grade design critique loop #14 creates the current design critique evidence gate.
• Browser QA, accessibility, and performance closeout #16 performs browser QA/accessibility/performance closeout.
• Enable automatic pre-prod deploy after main merges #55 must enable automatic deployment of accepted main to lircap.pendragon.bot so the winning branch is visible without manual sync.

Primary implementation profile

agent:orchestrator

Required contributor/reviewer profiles

• agent:design-director — visual opportunity identification, agency-grade judgement, anti-template review.
• agent:brand-strategist — strategic fit, proof model, brand distinctiveness.
• agent:frontend-implementer — candidate implementation branches.
• agent:browser-qa-reviewer — screenshot/accessibility/responsive evidence.
• agent:code-security-reviewer — safe runtime/code review.
• agent:devops-deployer — integration with Enable automatic pre-prod deploy after main merges #55 auto-deploy and preview deployment policy.

Proposed loop

1. Baseline capture

   • Capture screenshots and metrics for current main / pre-prod.
   • Store baseline evidence under docs/design/loops/<run-id>/baseline/.

2. Opportunity identification

   • Dispatch design-director and brand-strategist scouts with only screenshots, issue context, design-language spec, critique rubric, and browser QA evidence.
   • Each scout returns ranked problems/opportunities with a measurable hypothesis.
   • Orchestrator selects 1–3 opportunities, not a broad redesign.

3. Candidate branch generation

   • Create one isolated branch/worktree per candidate, e.g. design-loop/<run-id>-a, design-loop/<run-id>-b.
   • Delegate each candidate to a bounded frontend implementer with strict constraints: no stock/AI imagery, no compliance-risky copy, no trackers, no secrets, no broad rewrites.

4. Scoring pass

   • Build each candidate and capture identical screenshots/metrics.
   • Score each candidate using a durable rubric with weighted dimensions:
     • strategic fit
     • crossing/ownable idea
     • composition
     • typographic tension
     • originality
     • restraint
     • institutional credibility
     • human warmth
     • imagery/placeholder discipline
     • accessibility
     • performance
     • compliance/security risk
     • AI-template smell
   • Include browser evidence: console, axe/equivalent, keyboard/focus, reduced motion, contrast, performance budget.

5. Selection gate

   • Only candidates with no blocking accessibility/security/compliance failures can win.
   • If no candidate clearly improves the baseline, close the run with “no promotion”.
   • If one wins, open a draft PR from the winning branch only; keep losing branches documented but unmerged.

6. Validation and promotion

   • Run assigned profile reviews on the winning branch.
   • Parent/orchestrator verifies tests/build/browser/deploy smoke and CI.
   • Merge winning PR to main.
   • Enable automatic pre-prod deploy after main merges #55 auto-deploys main to lircap.pendragon.bot.
   • Capture post-deploy screenshots and record the new baseline score.

7. Repeat

   • New loop starts from the deployed baseline, not from stale local branches.

Acceptance criteria

• A documented design-loop operating model exists in the repo, including roles, branch naming, prompts, scoring rubric, evidence locations, stop conditions, and promotion rules.
• The loop explicitly protects the parent/orchestrator context and iteration budget: parent prompts, verifies, selects, and merges; implementation/reviews happen in delegated profiles/subagents.
• A reusable scorecard template exists for candidate comparison.
• A reusable GitHub issue/PR comment template exists for loop runs.
• The model defines how Enable automatic pre-prod deploy after main merges #55 auto-deploy makes the winning main branch the pre-prod baseline.
• The model defines hard blockers: accessibility violations, secret leakage, external tracking, compliance-risky copy, stock/AI imagery, placeholder-to-production leakage, performance budget failure.
• The model allows “no promotion” when candidates fail to beat baseline.
• A first pilot loop issue is either created or documented as the next action after Enable automatic pre-prod deploy after main merges #55/Browser QA, accessibility, and performance closeout #16.

Non-goals

• Do not implement a fully autonomous production deploy/cutover.
• Do not bypass human/orchestrator promotion decisions.
• Do not let multiple agents write to the same branch simultaneously.
• Do not pass the entire artefact pack to every candidate agent.

Verification

• Parent verifies the docs/templates exist and reference the correct issue/profile workflow.
• Assigned profile reviews confirm the loop is safe, actionable, and aligned with the high-end agency standard.

[iainherd]

Adding score-trend adaptation to the loop design: promotion should be based on deltas over baseline, not just absolute score. The scorecard should include a material-improvement threshold, regression guards by dimension, explicit no-promotion outcomes, plateau detection after repeated low-delta runs, and a mode switch from broad exploration to surgical optimisation or new-input requests (e.g. photography, brand assets, client feedback) when scores stop improving. This prevents agent churn once the site approaches a local optimum.

[iainherd]

#56 is merged and verified.

PR: #59
Merge commit: 776afd25b8af5042060f25ef281908ceef056c83

Post-merge GitHub Actions:

• CI on main: PASS — https://github.com/lircap/lircap-com/actions/runs/28322637336
• Deploy pre-prod triggered by workflow_run: PASS — https://github.com/lircap/lircap-com/actions/runs/28322647121

External parent smoke after workflow completion:

• Unauthenticated /: 401 with WWW-Authenticate: Basic
• Authenticated /: home marker present
• Authenticated /team/: team marker present
• Authenticated /contact/: contact marker present
• Authenticated /robots.txt: sitemap reference present
• Authenticated /sitemap.xml: team canonical route present
• Active release symlink: /opt/lircap-preprod/releases/20260628T124500Z-28322647121-1-776afd25b8af

Review closeout:

• agent:code-security/process-safety: PASS
• agent:devops/browser-QA: PASS
• agent:design-director/brand-strategist: REQUIRED CHANGES → focused REQUIRED CHANGES → PASS

Parent verification before merge:

• npm run typecheck: PASS
• npm run lint: PASS
• npm test -- --run: PASS, 9 files / 43 tests
• npm run build: PASS
• python3 scripts/check-placeholders.py: PASS
• npm run perf:budget: PASS, total=48045B js=0B css=14991B
• git diff --check: PASS