Skip to content

CHANGELOG.md

Latest commit

 

History

History
186 lines (157 loc) · 10.2 KB

CHANGELOG.md

File metadata and controls

186 lines (157 loc) · 10.2 KB
Raw

Changelog

All notable changes to FinAI Research Workflow are documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

Fixed

  • TokenBucketRateLimiter per-server collisions: replaced hash()-slot buckets with exact per-server dict keys so rate_limit_per_server no longer flakes under PYTHONHASHSEED / xdist (CI test_per_server_rate_limit).
  • Writing pre-gate compat: stop inventing baseline_p=1.0 for negative_result_handler on the writing track (always blocked). Soft-skip unless real baseline stats are present in writing/refinement payloads; keep fail-closed for manuscript/reference/data_source import failures.
  • Deprecated CLI modes: research_framework/pipeline.py accepts data|analysis|draft|lit-review|novelty-check|regression again as aliases that print an actionable redirect and exit 2 (DeprecationWarning; no silent reimplementation). Makefile validate-novelty and lit-review docs updated.
  • Dual-track docs: clarified writing (agent_pipeline) vs empirical demo (research_framework/pipeline.py) vs production empirics (enhanced_pipeline / modern_did) in ARCHITECTURE.md §0, AGENTS.md, api_reference.md, README.md, CLAUDE.md; removed dead CLI modes (data/analysis/draft/ lit-review) from argparse and examples.
  • NoveltyGate real search: _check_novelty now queries Semantic Scholar + OpenAlex via literature_download, filters by lookback_years, soft-boosts top journals, scores token Jaccard; LLM heuristic only when search is down (documented in report search_status).
  • Writing pre-gate data_source_checker: fixed wrong import scripts.research_framework.data_source_checker → scripts.data_source_checker; added check_data_sources() adapter; gate reports are normalized to dicts; failures are fail-closed (passed=False) instead of silent greenlight.
  • HITL post-exec + resume semantics: AgentOrchestrator previously called HITLGate.hold() before agent.run(), so gated stages often never executed; resume_pipeline also clamped/skipped incorrectly. Gates now hold after real agent output (with stage_result in gate content). Approve continues from the next step; reject re-runs the same stage with prior_rejection_feedback; resume while still PENDING is a no-op (fail-closed). Prior stage results are carried via _resume_stage_results.
  • PDF / post-run bookkeeping dead path: end-to-end TeX/PDF generation and provenance/HITL bookkeeping lived under _auto_generate_arch_diagrams after return arch_paths (unreachable). Moved into AgentPipeline.run() (and mirrored on successful resume_pipeline), skipped while HITL-paused.
  • Agent-host entry robustness: isolation / non-interactive hosts previously had no fail-closed FinAI entry when LLM was missing and Mock was forbidden, so agents freestyled outside the official pipeline. Added scripts/agent_host_entry.py and scripts/core/agent_host_report.py to write canonical output/SKIPPED_CONFIG.md + output/FINAL.md, wired the same artifacts into agent_pipeline exit code 4, and documented the protocol in AGENTS.md / fin-full-pipeline skill.

Added

  • tests/test_orchestrator_hitl_resume.py: regression coverage for post-exec HITL, approve→continue, reject→rerun+feedback, pending no-op, final-stage approve.

[0.2.0a1] - 2026-08-07

Fixed

  • Corrected maintainer docs: audit_guard check count 17→25, MCP directory count 50→43, and primary entry documented as agent_pipeline.py (not agent.py).
  • Fixed Makefile pipeline-lit target (removed nonexistent --stage lit; points to literature_download.py).

Changed

  • Cleared stale PyPI publish workflow waits for mismatched GitHub tags (v1.0.0/v1.0.1/v0.2.0-alpha) that would have uploaded wrong/conflicting package versions.
  • Bumped package version to 0.2.0a1 for a clean Trusted Publishing release aligned with pyproject.toml.

Fixed (audit_fix_2026_07_12, carried in tree)

  • T001: Removed mechanism tests from scripts/us_esg_regression.py that constructed cds_proxy, rating_proxy, and analyst_cov_proxy as linear functions of treatment variables (endless tautology — would mechanically re-state the baseline DID coefficient). Table 5 now outputs an "Omitted" notice with reference to future work using genuine IBES / TRACE / S&P data.
  • T002: SyntheticControlEngine.sig no longer returns heuristic stars based on raw RMSPE ratio thresholds. New .sig reads permutation p-value from inference(). Legacy heuristic preserved as .rmspe_ratio_sig.
  • T003: Added T_post < 5 warning to process_data() in scripts/us_esg_regression.py per Roth & Sant'Anna (2023, Biometrika). Table 3 tablenotes updated with explicit "illustrative, not definitive" caveat.
  • T004: Fixed pyproject.toml force-include paths. Previous config referenced finai_research_workflow/<dir> (nonexistent source), causing hatchling to fail with Forced include not found. Now correctly points to actual source directories (config, templates, knowledge, mcp_servers).
  • T005: Added examples/_template/ skeleton (5 markdown files + README) for users to bootstrap new research projects. Updated .gitignore to allow template through.
  • T006: Added synthetic test fixtures (data/sample/): 250-obs ESG panel, 300-obs staggered DID panel, 5-entry BibTeX. README documents they are for offline testing only.
  • T007: Added 2 Jupyter notebooks (notebooks/00_quickstart.ipynb, notebooks/01_did_lab.ipynb) with Angrist-Pischke MHE Ch.4 walk-through.
  • T008: Added scripts/generate_fixtures.py for reproducible fixture generation. Defaults seed=42 for deterministic output.
  • T009: Fixed README_EN.md Zenodo DOI badge (PENDING → 21262689) and removed duplicate "Architecture overview" entry.
  • T010: Aligned README_EN.md structure with README.md — added "Why FinAI Research Workflow?", expanded Quality Gates, added comparison table with alternatives.

Added

  • 16 regression tests across 3 new files:
    • tests/test_us_esg_regression_t001_audit.py (7 tests, T001)
    • tests/test_synthetic_control_t002_audit.py (8 tests, T002)
    • tests/test_us_esg_t003_shortpanel.py (6 tests + 1 skipped, T003)

[0.2.0a0] - 2026-07-11

Added

  • mcp_servers/_shared/_version.py: single source of truth for APP_VERSION/APP_NAME, reads from top-level pyproject.toml (skips mcp_servers/pyproject.toml sub-package)
  • 3 GitHub Discussion seed posts (#135 release announcement, #136 ideas/arXiv auto-submit, #137 Q&A install) for community seeding
  • audit_guard.py Check 16: scans 260+ scripts/**/*.py files for hardcoded vX.Y.Z version drift from [project].version
  • tests/test_version_drift.py: 14 regression tests covering CLI banner, MCP servers (sec_edgar/cryptocompare/newsapi), gen_architecture_diagrams.py
  • GitHub Description updated with accurate numbers (43 MCP, 47 methods, 30 templates)
  • 3 GitHub Discussion templates + 4 kinds of category defaults enabled

Fixed

  • scripts/cli.py banner: was hardcoded v1.0.0, now dynamic from pyproject
  • scripts/cli.py version_cmd: fallback "1.0.0" → 5-tier resolution (pyproject → importlib.metadata → "0.0.0+unknown")
  • mcp_servers/user_sec_edgar/server.py: APP_VERSION = "1.0.0" → dynamic
  • mcp_servers/user_cryptocompare/server.py: same fix
  • mcp_servers/user_newsapi/server.py: same fix
  • scripts/gen_architecture_diagrams.py: header() default version="v1.0.0" → dynamic; inline "v1.0.0 标签" → dynamic
  • README.md: "~20 种计量方法" / "~30 Econometric Methods" → "47" (truth)
  • tests/test_research_directions_advanced.py: stale docstring "all 12" → 13
  • .gitignore: added MANUAL_TASKS.md (local-only runbook)

Changed

  • This is a pre-release (alpha) tag — used during the Star Audit follow-up phase. Next stable will be 0.3.0 once DemO GIF + topics + awesome-list promotion complete.

[1.0.1] - 2026-07-08

Fixed

  • Trigger Zenodo webhook for DOI minting. The v1.0.0 release did not auto-archive on Zenodo because the GitHub integration toggle was enabled after the release was created. This patch release forces a fresh GitHub → Zenodo webhook cycle and re-archives the same v1.0.0 source tree (no code changes).

[1.0.0] - 2026-07-08

Added

  • 8,048 pytest tests across 398 test files (32.2% coverage baseline)
  • 47 econometric methods (DID, IV, RDD, PSM, GMM, modern staggered DID variants, synthetic control, synthetic DID, panel quantile, interactive fixed effects, triple-diff, local projections, spatial regression, ...)
  • 43 MCP data-source servers (OpenAlex, ArXiv, Semantic Scholar, NBER, SEC EDGAR, Tushare, yfinance, FRED, World Bank, IMF, OECD, Eastmoney reports, ...)
  • 30 journal templates (经济研究 / 金融研究 / JF / JFE / RFS / JPE / Econometrica / ...)
  • 17 AI skills for Claude Code, Cursor, GitHub Copilot
  • GitHub Discussions enabled with 6 default categories
  • Smoke tests for scripts/start_research.py (0% → 70.5% coverage on entry point)
  • Dependabot configuration for Python ecosystem security updates

Fixed

  • scripts/research_framework/volatility_models.py::VolatilitySpillover.diebold_yilmaz: four silent bugs (statsmodels API drift, shape mismatch, pandas applymap deprecation, .loc slice) caused the main spillover path to silently fall back to a correlation-based approximation. All four fixed; main path now produces correct diebold-yilmaz spillover indices.
  • pyproject.toml: pytest/pytest-cov upper bounds widened to match the locked versions that dependabot already resolved.

Changed

  • Repository renamed from FinAI-Research-Workflow to finai-research. Old URL permanently 301-redirects to new URL. PyPI package name finai-research-workflow is intentionally unchanged to preserve pip install ergonomics.

Known Limitations

  • All LLM-generated outputs (papers, reviews, design docs) require human verification before submission. Hallucinated citations and inappropriate statistical claims are possible.
  • Windows is a secondary platform; macOS and Ubuntu are primary tested targets.