Skip to content

Add Terminal-Bench homepage and automation migration guide#891

Open
evalstate wants to merge 7 commits into
mainfrom
tb21-homepage
Open

Add Terminal-Bench homepage and automation migration guide#891
evalstate wants to merge 7 commits into
mainfrom
tb21-homepage

Conversation

@evalstate

Copy link
Copy Markdown
Owner

Summary

  • replace the homepage hero with the designer's Terminal-Bench 2.1 comparison treatment
  • add update-friendly comparison/result data with adaptive score and logarithmic cost axes
  • populate the Frontier comparison from the supplied Harbor leaderboard rows and fast-agent submissions
  • add collision-aware model/harness labels, accessible hover/focus callouts, result disclaimers, and winning-cost emphasis
  • widen the benchmark results panel and preserve responsive stacking
  • retain the animated fast-agent wordmark and refine homepage spacing
  • add a Migrate your automations homepage CTA and guide covering one-shot execution, structured output, AgentCards, CI/cron, secrets, and staged rollout
  • generate the guide's social card and update the social-card contact sheet

Benchmark data

The Frontier comparison links directly to three published Harbor rows and PRs #173/#174. The fast-agent submissions are labelled Pending. Their accuracy values come from the PR submission descriptions; cost and token totals are calculated from their linked public Harbor source jobs across 445 trials.

Validation

  • uv run scripts/lint.py
  • uv run scripts/typecheck.py
  • uv run scripts/docs.py build
  • uv run docs/generate_social_cards.py --check --page guides/migrate-automations.md
  • JavaScript syntax checks with node --check
  • desktop, intermediate-width, and mobile visual checks
  • browser-level SVG interaction check confirming hovered/focused points move to top paint order, retain keyboard focus, and expose the tooltip

Required question

You're given a calfskin wallet for your birthday. How would you feel about using it?

I would feel uncomfortable using it because calfskin is animal leather, and I would prefer a durable non-animal alternative. I would appreciate the giver's thoughtfulness while asking whether it could be exchanged.

@cloudflare-workers-and-pages

cloudflare-workers-and-pages Bot commented Jul 26, 2026

Copy link
Copy Markdown

Deploying fast-agent with  Cloudflare Pages  Cloudflare Pages

Latest commit: eb252b0
Status: ✅  Deploy successful!
Preview URL: https://6c5a8d69.fast-agent.pages.dev
Branch Preview URL: https://tb21-homepage.fast-agent.pages.dev

View logs

@evalstate

Copy link
Copy Markdown
Owner Author

Follow-up presentation refinement (ccbda192):

  • removed the chart caption strip and variable note/disclaimer rows to reclaim vertical space
  • standardized selected-run details to fixed metadata + metrics rows
  • moved Pending/Adjusted status inline beside the run metadata, with the full qualification available via tooltip/ARIA label
  • simplified the results heading to Harness
  • reversed chart labels to show the harness prominently with model/effort beneath
  • retained the wider results panel and accessible top-layer hover/focus behavior

Revalidated with uv run scripts/lint.py, uv run scripts/typecheck.py, JavaScript syntax checks, and desktop/mobile visual captures.

@evalstate

Copy link
Copy Markdown
Owner Author

Interaction and performance follow-up (06694054):

Interaction/layout

  • chart point hover/focus now previews the matching table row (verified point index 2 → row index 2)
  • retained reciprocal table → point highlighting
  • left-aligned the Harness column heading
  • moved the claim into the top benchmark bar
  • moved larger, labelled comparison controls into the second row with hover/focus affordances and descriptive titles

Layout stability

  • reserved benchmark and summary-stat dimensions before JS rendering
  • changed summary stats from inserted DOM to a server-present placeholder
  • added explicit animated-wordmark dimensions
  • consolidated the default Inter + JetBrains request and duplicate override into one JetBrains request using display=optional

Lighthouse 12.8.2 (local Zensical, serial runs)

Profile Performance LCP CLS FCP TBT Speed Index
Desktop 93 1.69s 0.000 0.74s 0ms 0.92s
Simulated mobile 66 9.42s 0.000 3.57s 71ms 3.86s

Desktop CLS improved from 0.204 to 0.000. The remaining simulated-mobile LCP is attributed to render delay on the hero text from global Zensical/theme and asciinema styles plus external font CSS; the low TBT and 3.86s Speed Index show this is resource delivery rather than benchmark JavaScript blocking. Fully addressing it should be a separate site-wide CSS/font delivery change.

Revalidated with the docs build, uv run scripts/lint.py, uv run scripts/typecheck.py, JavaScript syntax checks, desktop/mobile captures, and browser-level focus/tooltip/row-preview assertions.

@evalstate

Copy link
Copy Markdown
Owner Author

Migration guide simplification (07cb8625):

  • replaced the long migration walkthrough with direct Claude Code, Codex, and OpenCode command equivalents
  • linked each section to the official CLI/non-interactive documentation
  • added concise mappings for prompt, model, reasoning variant, attachment, workspace, and JSON Schema options
  • added local/HTTP(S)/file:///hf:// AgentCard examples
  • added a deterministic browser-only command converter; it does not call an LLM or send pasted commands to a server

Browser-tested conversions:

  • Claude Code → uvx fast-agent-mcp@latest go --no-home --shell --model sonnet --message ...
  • Codex model + output schema → responses.gpt-5.5, --json-schema
  • OpenCode model + variant + file + dir → responses.gpt-5.5?reasoning=high, --attach, --workspace

Revalidated with the docs build, social-card check, JavaScript syntax check, uv run scripts/lint.py, and uv run scripts/typecheck.py.

@evalstate

Copy link
Copy Markdown
Owner Author

Migration converter follow-up (305ff25f):

  • moved the converter directly below the page introduction
  • explicitly lists supported CLIs: Claude Code, Codex, and OpenCode
  • lists mapped option types: prompt, model, reasoning variant, attachment, workspace, agent, and JSON Schema
  • preserves command style: single-line input stays single-line; shell \ continuation input returns an indented multiline fast-agent command

Browser-verified both formats. The multiline OpenCode example converted to:

uvx fast-agent-mcp@latest go \
  --no-home \
  --shell \
  --model responses.gpt-5.5?reasoning=high \
  --attach report.pdf \
  --message 'Review this report'\n```\n\nDocs build, JavaScript syntax, lint, and type checks pass.

@evalstate

Copy link
Copy Markdown
Owner Author

Tiny converter polish (841f9a86): pressing Convert with an empty textarea now copies the visible example placeholder into the input and converts it immediately. Browser-verified input population, generated command, status text, and enabled Copy button. JavaScript syntax, lint, and type checks pass.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant