examples: live ContextLevy demo (do not merge)

[nonlooped]

Summary

Permanent showcase PR for ContextLevy. Adds /examples/agent-debt-pr/ fixtures (lockfile, build output, agent config) so visitors can inspect a live PR comment, workflow check, and job summary.

Do not merge. Maintainers: rebase this branch onto \main\ when the demo needs refreshing.

Test plan

• ContextLevy workflow runs on this PR
• PR comment posted when thresholds exceeded

[contextlevy]

🤖 ContextLevy

This PR adds ~20.1k estimated net-new AI-context tokens.

Risk level: 🔴 High

Added	Finding
+10.7k	examples/agent-debt-pr/pnpm-lock.yaml Lockfiles add bulk context; agents often need dependency names but not every resolved URL.
+7.9k	examples/agent-debt-pr/dist/app.js Build artifacts are rarely useful agent context.
+1.3k	examples/agent-debt-pr/.agents/skills/onboarding/SKILL.md Agent instruction files are high-signal but add persistent context overhead.

Estimated worst-case input cost if read by an agent
Based on configured input-token pricing. Estimates may vary ±50% depending on model tokenizer. Output tokens and caching are not included.

Pricing profile	Est. input cost (±50%)
GPT-5.5	~$0.05–$0.15/session
Kimi K2.6	~$0.01–$0.03/session

Suggestions

• Commit lockfiles when your team policy requires reproducible installs — just expect higher context cost.
• Keep build output out of version control.
• Keep agent instructions concise and scoped to what agents must know.
• Consider excluding these paths from agent indexing (tool-agnostic):
• examples/agent-debt-pr/dist/app.js

Many tools honor .gitignore; others support dedicated ignore files such as .cursorignore.

Different models tokenize differently, and agents may not read every changed file. ContextLevy estimates context risk, not exact billing.

ContextLevy runs locally in CI and does not send code to an external API.