All notable changes to this project are documented in this file. Format follows Keep a Changelog.
- WRDS ISS Voting reference (
skills/wrds/references/iss-voting.md): vavoteresults and voteanalysis_npx tables, base-conditional turnout/forpct logic, CRSP CUSIP+ticker linking, director election agenda codes - WRDS TFN Ownership reference (
skills/wrds/references/tfn-ownership.md): 13-F S34 institutional ownership pipeline, S12 mutual fund holdings via MFLINKS, passive/index classification, as-of merge pattern - Voting + Ownership EDA notebook (
skills/wrds/examples/voting_ownership_eda.py): full Python/PostgreSQL translation of the SAS1-make.saspipeline — ISS votes, CRSP linking, 13-F IO, S12 MF holdings, merge_asof, summary stats and plots
- WRDS ISS Directors reference (
skills/wrds/references/iss-directors.md): two-table schema (risk.directors + risk.rmdirectors), type harmonization, 1996 gender backfill, S&P 1500 filter - WRDS ExecuComp reference (
skills/wrds/references/execucomp.md): CEO anncomp, legacy codirfin vs current directorcomp, firm-year aggregation, combining both tables - WRDS Compustat additions: business segments (compseg.seg_annfund), derived variables (tobins_q, roa, leverage, cusip6), SIC fallback, winsorization, 5 new gotchas
- WRDS CRSP additions: market index tables (msi/dsi), annual stock performance, 60-month rolling volatility, year-end market cap, 6 new gotchas
- Marimo
--watchflag: added to allmarimo editcommands in SKILL.md and marimo-pair finding-marimo.md
- New reference files consolidated in
skills/wrds/references/instead of duplicating in top-levelreferences/
- Two-track DS delegation: ds-delegate routes tasks by type (
engineeringvsanalysis)ds-engineeragent: pipeline/ETL tasks with determinism, schema validation, join audits, idempotency enforcementds-analystagent: analysis tasks with statistical validity, p-hacking prevention, robustness, SE specification- Keyword-based type detection heuristic when PLAN.md tasks lack explicit
typefield
references/ds-engineering-constraints.md(E1-E5): determinism, schema contracts, join audits, idempotency, error handlingreferences/ds-analysis-constraints.md(A1-A7): statistical validity, p-hacking prevention, robustness checks, sample selection, SE specification, visualization integrity, analysis-specific deviation rules- Type-aware methodology reviewer in ds-delegate: engineering checklist (schema, determinism) vs analysis checklist (stats, specification)
- DS workflow audit fixes (6.5 → 9.4 composite):
- Requirement IDs (
CAT-NNformat) in SPEC.md template with scope classification (v1/v2/out-of-scope) - Checkpoint type annotations on all 11 DS phase gates (human-verify/decision)
allowed-toolsrestrictions on all reviewer/verifier agents (ds-spec-reviewer, ds-plan-reviewer, ds-review parallel reviewers, ds-verify, ds-delegate methodology reviewer)- Context monitoring in ds-implement and ds-fix (Warning ≤35%, Critical ≤25%, auto-handoff)
- Structured task summaries with YAML frontmatter in ds-implement LEARNINGS.md
- Smart-discuss batching in ds brainstorm (batch 3+ independent questions)
- Requirement tracing in ds-review issue output and ds-verify success criteria
- Requirement IDs (
- Wired
test-gap-auditoragent in dev-test-gaps (was usinggeneral-purpose)
- GSD patterns in workflow-creator: checkpoint types (human-verify/decision/human-action), context monitoring (35%/25% thresholds), summary frontmatter (implements/requires/provides/affects), READ-ONLY verifier enforcement (allowed-tools), requirement traceability (CATEGORY-NN IDs), autonomous phase chaining (smart-discuss, blocker handling)
- Dev workflow audit fixes (8.1 → 9.5 composite):
- Requirement IDs (
CAT-NNformat) in SPEC.md, PLAN.mdimplementscolumn, VALIDATION.md mapping, review issue output, verification traceability allowed-toolsrestrictions on all reviewer/verifier agents (dev-spec-reviewer, dev-plan-reviewer, dev-review parallel reviewers, dev-verify, dev-delegate spec/quality reviewers)- Checkpoint type annotations on all 8 phase gates + handoff
- Context monitoring in dev-implement and dev-debug (Warning at ≤35%, Critical at ≤25%, auto-handoff)
- Structured task summaries with YAML frontmatter in dev-implement LEARNINGS.md
- Smart-discuss batching in dev-clarify (batch 3+ independent questions)
- Test-gap auditor tool restrictions (write tests only, never implementation)
- Requirement IDs (
- Fenced bang-backtick detection in validate-skill-paths.py hook — catches
!cat`` inside markdown fences (Claude Code parser ignores fences and executes them) - Fixed fenced bang-backtick examples in skill-creator and workflow-creator
- workflow-creator Mode 2 audit now scores 15 principles (was 9) — added checkpoint types, context monitoring, summary frontmatter, agent tool restrictions, requirement traceability, autonomous phase chaining
- workflow-creator Mode 3 fix patterns expanded with 7 new gap fixes
- workflow-creator Iron Laws: added NO VERIFIER WITH WRITE ACCESS, NO LONG WORKFLOW WITHOUT CONTEXT MONITORING
- Plugin validation hook (
plugin-validate.py): runsclaude plugin validateafter Write/Edit of plugin files (plugin.json, marketplace.json, SKILL.md, agent/command .md, hooks.json) - Scoped PostToolUse hooks in
skill-creatorandworkflow-creatorfrontmatter — validation runs automatically during plugin development skills only
- Writing workflow: full GSD adoption — deviation rules (R1 factual, R2 evidence, R3 structural, R4 argument restructuring STOP),
.planning/state folder,writing-validatephase (claim coverage),writing-handoffskill writing-validateskill: maps PRECIS claims to draft sections, 4-level validation (exists, substantive, supported, addresses claim), produces VALIDATION.mdwriting-handoffskill: structured session handoff with writing-specific frontmatter (section_in_progress, total_sections)- Deviation rules constraint in
writing-common-constraints.md - Handoff detection in
/writing,/writing-review, and/writing-reviseentry points
- All writing state files migrated from
.claude/to.planning/(PRECIS.md, OUTLINE.md, REVIEW.md, REVIEW_STATE.md, ACTIVE_WORKFLOW.md, completed-workflows/) writing-draftnow chains towriting-validate(not/writing-reviewdirectly)
- DS workflow: full GSD adoption — deviation rules (R1-R3 auto, R4a data assumptions, R4b methodology STOP),
.planning/state folder,ds-validatephase (DQ checks as test suite),ds-handoffskill for session pause/resume ds-validateskill: maps SPEC.md requirements to output artifacts, runs DQ1-DQ5 + M1 checks, produces VALIDATION.mdds-handoffskill: structured session handoff with YAML frontmatter and mandatory sections- C8 (Deviation Rules) constraint in
ds-common-constraints.md - Handoff detection in both
/dsand/ds-fixentry points
- All DS state files migrated from
.claude/to.planning/(SPEC.md, PLAN.md, LEARNINGS.md, REVIEW_STATE.md) ds-implementnow chains tods-validate(notds-reviewdirectly)ds-checks.mdcheck matrix updated withds-validatecolumn
dev-common-constraints.md: replaced relativeRead("real-test-enforcement.md")with cache discovery pattern for path portabilitydev-clarify,dev-design,dev-review: removed process hints from skill descriptions (trigger-only per enforcement pattern #10)
workflow-creator: GSD deviation rules, state folder convention, and session handoff patterns
- Dev workflow: GSD patterns — deviation rules (4-rule system),
.planning/state folder, test gap validation phase, session handoff (dev-handoffskill), goal-backward verification (dev-verifieragent) dev-test-gapsskill andtest-gap-auditoragent for requirement-to-test coverage mappingdev-handoffskill for structured session pause/resume
companionskill: launch and correctly frame companion (host Claude Code) sessions — pre-flight checklist, path mapping, prompt templates, rationalization prevention
assistantagent for personal productivity (email, calendar, tasks, notes, Google Workspace)- Wraps superhuman, morgen, obsidian CLI, and gws
- Delegates to
nlmskill for deep research via Skill tool
librarianagent: replaced inline NLM command table withSkill(skill="workflows:nlm")invocationnlmskill: extracted workflow recipes toreferences/workflows.md, added Readwise→NLM import docs
- Generic ETL strategy assessment in ds-plan phase
- SAS ETL performance enforcement in WRDS skill and ds workflow
- Google Scholar: BibTeX export, cite command, download command
- Anti-hallucination enforcement for Google Scholar results
- Bluebook-audit workflow skill for law review footnote auditing (7 phases)
- PATH-based CLI names with dependency checks
- Google Scholar skill integrated with librarian agent
- README documentation overhaul
- Moved all entry skills from
skills/to discoverableskills/root - Renamed midpoint skills for consistency (e.g.,
dev-edit->dev-debug)
- Redundant
/checkpointand/verifyskills
- Replaced slash commands with discoverable skills
- Removed
commands/directory entirely - Skills are now auto-discovered from
skills/directory
- session-start: substitute CLAUDE_PLUGIN_ROOT in using-skills content
- Promoted visual-verify from internal to discoverable skill
/visual-verifycommand for standalone render-vision-fix loops
- Extracted tinymist plugin to standalone repo
- session-start: load using-skills from correct lib/skills path
- Readwise: replaced opencode delegation with readwise-cli
- Agent team parallelization across all workflows (dev, ds, writing)
- continuous-learning: save skills as
name/SKILL.mddirectories
- Writing-legal: overhauled law review template formatting
- Extracted formatting reference for legal writing
- Writing-review: paragraph-level gates, quote verification, section mapping
- Updated plugin description
- Writing-review skill for hierarchical document diagnosis
- Agent team parallel implementation in dev workflow
- All writing source searches routed through librarian agent
- DS workflow: audit and harden, add linted subagents (ds-analyst)
- visual-verify skill for render-vision-fix loops (agentic mode)
- PHILOSOPHY.md: midpoint constraint loading pattern
- Writing-edit: load full domain skill before edit checks
- Two-entry-point architecture for all workflows (start fresh + midpoint re-entry)
- Readwise skill with server-side tag filtering
- Readwise CLI integration (replacing opencode delegation)
- look-at: agentic vision mode with code execution
- data-context skill for dataset knowledge extraction
- workflow-creator skill for designing LLM workflows
- nlm: research command, generation and content transformation commands
- Librarian: skill-based orchestrator with NLM-first hierarchy
- Writing-legal: user-facing skill with Volokh enforcement
- PreToolUse hook to block direct Readwise MCP calls in main chat
- gog skill (superseded by superhuman-cli)
- Writing workflow decomposed with writing-setup, writing-outline, writing-draft phases
- Librarian simplified to skill-based orchestrator
Highlights from earlier development:
- v2.34: Streamlined writing workflow with
/writingentry point - v2.33--2.30: LSEG Data Library references, WRDS documentation
- v1.x--v2.29: Consolidated to single workflows plugin, delegation patterns, DS workflow hardening
- v0.5.0: Major skill audit with TOCs and progressive disclosure
- v0.3.0: Initial commit with dev and ds plugins