Adds the minimum scaffolding to publish demo/app.py as a public
Hugging Face Space without polluting the GitHub README with YAML
frontmatter (which would render as an ugly table for recruiters).
requirements.txt — HF reads this at build time. Mirrors the [demo]
extra in pyproject.toml, with `-e .` at the bottom so the agent
package itself installs and demo/app.py can `import agent.*`.
docs/HF_DEPLOY.md — step-by-step deploy guide:
- Create the Space (Streamlit SDK, CPU basic, public)
- Add GROQ_API_KEY as a Space secret
- Add `hf` git remote
- Generate a Space-specific README with HF YAML frontmatter at
deploy time (NOT committed to GitHub) — this is the key trick
that lets one repo serve both GitHub (clean README) and HF
(frontmatter for Space metadata)
- Push to HF
- Embed back in the portfolio via iframe
No code changes — demo/app.py is already HF-safe:
- DATASETS_DIR uses Path(__file__).parent (works in any cwd)
- GROQ_API_KEY from os.environ (HF Spac…
Subagent caught at scaffold-time that Astro v6 removed the legacy
src/content/config.ts + type:'content' pattern. Config now lives at
src/content.config.ts and uses loader: glob(...). getCollection() API
is unchanged.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add implementation plan for the portfolio website
Discovered during Task 5 (figure capture test): the Agg backend in
ipykernel suppresses display_data messages that the sandbox captures
figures from. The default inline backend is what makes plt.show() emit
PNG. Update prompt to tell the LLM NOT to call matplotlib.use().
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The Phase 0 subagent discovered the plan's DABench assumptions were
wrong in several material ways. Patching the plan so downstream tasks
build on the actual schema:
- Task 11: system prompt now templates `constraints` and `format`
fields and instructs `@name[value]` answer tags (the regex the
official scorer extracts)
- Task 12: `run()` accepts `constraints` and `format_constraint`
kwargs; defaults empty for CLI use
- Task 16: loader now joins two JSONL files (questions + labels) on
`id`, uses `file_name` (not `data_file`), preserves
`common_answers` list-of-pairs
- Task 16 Step 1: explicit instructions to copy and re-license
`eval_closed_form.py` + `utils.py` from the InfiAgent repo (Path A2)
- Task 16: replace substring `score()` with `score_run()` that calls
`evaluate_responses` and aggregates ABQ; per-task results now store
full predicted response + common_answers so scoring is decoupled
from running
- Task 16: add `--score-only` flag for re-scoring without re-r…
docs(spec): record Phase 0 scorer-integration decision (A2)
After Phase 0 discovery: the InfiAgent scorer is two stdlib-only files
(eval_closed_form.py + utils/utils.py). Copying them with attribution
avoids the heavy unrelated deps of the full infiagent package
(streamlit, openai==0.27.7, fastapi, ipykernel).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- gitignore eval/results in Phase 0
- spell out all three wiring updates needed when adding --provider flag
Plan now APPROVED.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
23 tasks across 10 phases. MVP gate at Task 17 (CLI works, demo
works, 80-task benchmark number). Portfolio plan blocks on the
MVP gate per the sequencing in both specs.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Critical fixes:
- Agent §3, §8, §13: rewrite resume sentence and success criteria to
match what the ablation actually measures (two-model comparison +
retry-on/off), not a fabricated "beat baseline by 5 points" claim
- Agent §4a (new): threat model + isolation plan for the public
Streamlit demo, with a pre-vetted-dataset fallback when Docker
isn't available
- Agent §7, §8: nail down exact retry semantics so the ablation is
interpretable; specify primary metric; document rate-limit fallback
- Agent §10, §11: move Streamlit out of `agent/` into `demo/` so
CLI/eval users don't need it installed
- Portfolio §9: add SEO scaffolding required for Lighthouse ≥ 95 SEO
- Portfolio §13: add pre-launch checklist (404, OG previews, favicon,
mobile, etc.) and acknowledge legacy-project diagram effort
- Portfolio §14: make "agent first, then portfolio" sequencing
mandatory rather than recommended
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add design specs for data-analysis agent and portfolio site
Two design docs from the initial brainstorm:
- 2026-05-26-data-analysis-agent-design.md: code-as-action DA agent,
evaluated on InfiAgent-DABench
- 2026-05-26-portfolio-website-design.md: minimal-academic Astro site
with project case studies and blog
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
