Which currently listed-Free Zen route can produce a usable, accessible one-file clock without repair—and what does it cost in time, privacy, and billing risk?

Create one self-contained `index.html` containing a polished analog and digital clock. It must show correct local time and date, animate seconds, provide 12/24-hour and light/dark controls, remain usable from 320px to desktop width, support keyboard navigation and visible focus, respect reduced-motion preferences, and use no libraries, external assets, network requests, build step, or additional files. Add a `[data-toybox-clock]` root that sets `data-ready="true"` only when initialized. Render the visible time in a `<time data-toybox-time datetime="…">` element whose `datetime` is updated with the clock. Expose the selected hour mode and theme through accessible controls and state attributes. Write the file and verify it. Do not return an explanation instead of the artifact.

Protocol v2: locked before any run

Toybox is a small, reproducible build protocol for a practical consumer question. Its prompt, seven-route manifest, round-robin order, timeout, evaluator, acceptance gate, and evidence requirements are locked before any official attempt. This tranche contains no model attempts, receipts, scores, generated clocks, or recommendations.

This is not a model leaderboard. It tests one prompt, one OpenCode harness, and one narrow artifact. If results are released later, every raw attempt remains visible. Operational failures and zero-success experiments remain valid evidence.

OpenCode Zen documentation still listed seven routes as Free when checked on July 29, 2026. Protocol v2 tests all seven three times in round-robin passes. There is no substitute. A route that disappears after the protocol locks receives unavailable records; a newly listed route waits for a later protocol version. “Listed Free” is a dated access observation, not a completed-run receipt.

Quick try

Download toybox-clock-v2.zip, verify it against SHA256SUMS, install the pinned dependencies, and evaluate one untouched index.html:

1
2

bun install --frozen-lockfile
bun run EVALUATOR.mjs --input ./index.html --output ./evaluation

The evaluator starts observing before navigation, blocks and records non-document requests, waits up to five seconds for the ready hook, samples time twice, captures screenshots, and writes machine-readable evidence. A quick try is a replication, not an official result.

Strict reproduction

Official reproduction runs 21 attempts in three round-robin passes. Use the exact PROTOCOL.md, ROUTES.json, RESULTS-TEMPLATE.csv, and privacy/billing checklist. Record the OpenCode version, exact route, clean environment, permissions, timestamps, terminal output, token reporting, billing state, actual charge, redacted receipt, raw artifact, screenshots, console/network logs, evaluator JSON, and checksums.

Each attempt is one-shot and no-repair. Iterative repair is outside Toybox #001.

Evidence boundary

In-progress publication may expose evidence-reviewed raw attempts and completion counts, but it cannot name a best route. A consumer decision table appears only when all 21 rows are terminal and reviewed. It may conclude there is no recommendation.

Zero charged cost is recorded as $0.00; missing cost is not recorded, never zero. Incomplete aggregate cost must disclose missing records. Acceptance and quality fields come from evidence rather than display defaults.

Community replications use the GitHub issue template. They remain unverified and are never ingested automatically.

Follow-along lane

Codex remains unscored. OpenAI currently says Codex is included with ChatGPT Free for a limited time and that usage limits vary by plan and task. A Codex replication must record its own access and charge state; it does not enter the Zen comparison.

Limitations

• OpenCode describes these routes as Free for a limited time. That is not proof of a $0 completed run.
• Zen requires billing details. Auto-reload and workspace limits require owner verification before running.
• Every listed-Free route carries a data-use exception. Do not submit private, personal, confidential, or proprietary material.
• Any future public billing evidence must be redacted before it is linked.
• Big Pickle is a stealth route. No underlying identity is inferred.
• The protocol does not create a blended model score and does not generalize beyond this prompt-harness combination.

Sources

• OpenCode Zen documentation — route list, setup, pricing, auto-reload, workspace limits, and route-specific privacy terms; checked July 29, 2026.
• OpenCode Zen models endpoint — live route IDs; checked July 29, 2026.
• Using Codex with your ChatGPT plan — current plan access and variable usage limits; checked July 29, 2026.

Related links

• Set up OpenCode and understand the Zen boundary
• Compare practical free-stack options
• Read benchmark claims without overgeneralizing
• Browse the AIHackers Lab