InvestSkill-test

The sandbox for InvestSkill: a real test environment where the skills maintained there are run end to end on live market data, by a real LLM, on a schedule. Tracks InvestSkill v1.12.0 (34 files in prompts/), pinned in investskill.lock.json.

Project site: https://yennanliu.github.io/InvestSkill-test/

Showcase — all 27 frameworks on one basket (繁體中文, as of v1.11.0)

A worked example in Traditional Chinese covering the whole framework catalogue and all seven cookbook workflows, on a four-stock basket that all gapped down 7.8–9.6% on the same day (2026-07-28):

Page	Ticker	Angle
展示櫃總覽	—	The hub: why they fell together, plus a 27-framework coverage map
四檔對決	all four	stock-screener with all 22 sub-factors and their raw inputs shown
MU	MRVL·MU	Flagship 15-module report: $820 prices peak margins sustained for ever at 15×
SKHY	SKHY	Data-integrity audit — the framework refuses to give a target (47/100 confidence)
MRVL	MRVL	GAAP vs non-GAAP gap; three valuation methods disagree by an order of magnitude
SNDL	SNDL	Value-trap anatomy: 0.30× book, but book is melting 7.1%/yr
工作流 A–G	all four	Seven chained workflows — two correctly stop at the screening gate
產業鏈地圖	—	industry-map: HBM chain as 11 layers, 3 chokepoints, 4 second-order ideas

Three data defects are left in deliberately (KRW filings against a USD ADR, 13 days of price history, a bookValue field 27% below the filed balance sheet) so result-validator has something real to catch. Every figure is derived from one shared snapshot and is recomputable.

What this repo does

GitHub Actions clones the InvestSkill plugin, takes one yfinance snapshot per run, sends each analysis framework to an LLM with that shared snapshot, and commits the resulting Markdown report back to output/. Three loops keep it useful as a sandbox:

• Test any skill at any revision. skill_sandbox.yml runs one skill against an InvestSkill branch, tag, SHA or open PR (pull/<n>/head), so a skill change can be tried on real data before it merges upstream. Every report's frontmatter records the upstream commit that produced it (skill_commit).
• Follow upstream automatically. upstream_sync.yml checks InvestSkill daily. New upstream commits become a PR that bumps the lock and lists what changed. Anything this repo's code must follow (a new or removed skill, a full-report depth-tier change, a new alias stub) becomes an upstream-drift issue.
• Keep only recent output. cleanup_reports.yml deletes reports older than 30 days every week, or on demand. Git history keeps them.

Frameworks	34 files in InvestSkill's prompts/, all runnable via run_skill.py
Depth tiers	--depth quick (5) · standard (10) · comprehensive (25, default)
Default model	gemini-3.6-flash (also --provider openai / claude)
Report language	Traditional Chinese (繁體中文) by default, --language to change
Output	output/<type>/<ticker>/<type>_<date>_<model>.md

Usage

git clone https://github.com/yennanliu/InvestSkill.git InvestSkill   # frameworks + system context

python scripts/stock_eval_gemini.py AAPL
python scripts/dcf_valuation_gemini.py TSLA --model gemini-2.5-pro
python scripts/full_report_gemini.py NVDA --depth quick
python scripts/full_report_gemini.py MSFT --skills technical-analysis,bear-case
python scripts/run_skill.py earnings-preview NVDA          # any prompts/<skill>.md, by slug

dcf-valuation, fundamental-analysis and research-bundle are now alias stubs upstream: redirect cards merged into stock-valuation, stock-eval and full-report. They still run, but the CLI warns, because a report from a stub doesn't exercise the real framework.

Set GEMINI_API_KEY (or OPENAI_API_KEY / ANTHROPIC_API_KEY) in the environment, or as a repository secret for the workflows in .github/workflows/.

Keeping up with InvestSkill

git clone https://github.com/yennanliu/InvestSkill.git InvestSkill
python scripts/upstream_sync.py                 # what changed since the lock, and what drifted
python scripts/upstream_sync.py --check         # exit 1 if this repo's config has drifted
python scripts/upstream_sync.py --write-lock    # adopt the clone's revision as the new baseline

upstream_sync.yml runs the same script daily and on workflow_dispatch. To get a run as soon as InvestSkill changes, add this step to a workflow there, using a token that can dispatch to this repo:

gh api repos/yennanliu/InvestSkill-test/dispatches -f event_type=investskill-updated

For the bot to open PRs, enable Settings → Actions → General → Allow GitHub Actions to create and approve pull requests. PRs opened with the default token don't trigger tests.yml. To run CI on them, add a fine-grained PAT (contents, pull requests and issues: read/write) as the SYNC_TOKEN secret. test_committed_lock_matches_config fails any lock bump that adds or drops a skill without the matching ANALYSIS_TYPES row.

Report retention

python scripts/cleanup_reports.py --dry-run     # list reports older than 30 days
python scripts/cleanup_reports.py --days 60     # delete reports older than 60 days

A report's age is the generation date in its filename, or failing that its frontmatter date:. File mtime is never used, and a report with no readable date is kept. cleanup_reports.yml runs every Sunday, and can also be started by hand with days / dry_run inputs. It commits the deletions to main.

Layout

scripts/analysis/     config · prompts · data · llm · pipeline · publish
scripts/*_gemini.py   thin per-skill entrypoints
scripts/run_skill.py  any skill by slug (the sandbox entrypoint)
scripts/upstream_sync.py   InvestSkill change + drift detector (investskill.lock.json)
scripts/cleanup_reports.py report retention
scripts/showcase/     zh-TW showcase generator + committed yfinance snapshot
scripts/validate_html.py  stdlib HTML/link/a11y checker for docs/
scripts/tests/        pytest suite (run: python -m pytest)
.github/workflows/    scheduled report jobs + CI
docs/                 project website (GitHub Pages)
output/               generated reports from the last 30 days, committed

Rebuilding the showcase

The showcase pages are generated, not hand-written. Every figure is recomputed from scripts/showcase/fixtures/snapshot.json on each build, so a given snapshot always produces byte-identical HTML — no network access and no API key needed.

python scripts/showcase/build.py            # regenerate docs/showcase/
python scripts/showcase/build.py --check    # verify committed HTML is current
python scripts/showcase/derive.py           # print the derived-metrics digest
python scripts/validate_html.py docs        # structure · links · a11y · artifacts

If you edit a generator module, run build.py and commit the regenerated HTML — CI fails if the two disagree. Editing docs/showcase/*.html by hand will be reverted by the next build, so don't.

CI

Workflow	Checks
tests.yml	pytest on Python 3.11 + 3.13. All LLM SDKs and yfinance are faked, so no keys or network are needed.
upstream_sync.yml	Daily: InvestSkill vs. the lock and analysis.config. Changes → lock-bump PR; drift → upstream-drift issue
skill_sandbox.yml	Manual: one skill × one ticker × one InvestSkill revision; report uploaded as an artifact, and committed only if asked
cleanup_reports.yml	Weekly + manual: delete reports older than the retention window (default 30 days)
site.yml	build — docs/showcase matches a fresh render, and two renders are byte-identical · validate — every page in docs/ passes tag balance, head metadata, heading order, img/svg labelling, dead internal links and anchors, unrendered template artifacts, and flush borders on the CJK signal blocks · lint — ruff (pyflakes rules) plus a Python 3.11 grammar check

Lint scope is deliberately narrow (see ruff.toml): pyflakes catches real defects — undefined names, unused imports, dead locals — while style rules stay off so CI fails on bugs rather than on formatting opinions.

Disclaimer

Not investment advice. Reports are LLM-generated output over free public data and can be wrong, stale, or internally inconsistent. Reports that later proved incorrect are not singled out for removal: every report is deleted by age alone, and stays in git history afterwards.