WikiSkill 0.2.1 — BriefLoop fixes upstreamed

Upstreams the WikiSkill fixes that BriefLoop 0.26.3 carried in its vendored copy, so the standalone package and BriefLoop share one source.

• Windows: workspace locks use msvcrt on Windows and fcntl elsewhere; engine journals are read and written as UTF-8.
• Feedback-first API: feedback with origin: human and learning_intent: explicit_requirement is a binding constraint. Every compared pair needs evidence-backed requirement checks; a candidate is adopted only when all are fulfilled with no worse cases or regressions, otherwise the round is REVISION_REQUIRED and returns to the Maintainer. begin is idempotent and refuses changed inputs; skip closes a batch without comparable cases and makes no adoption decision.
• Native subagents: any fresh-context host may run the handoffs under its own short runtime name (for example opencode). Codex and Claude Code remain the recommended hosts with packaged role files.
• CI runs the feedback-loop and MCP tests on Windows, macOS and Linux.
• No research result was remeasured; historical records are unchanged.

WikiSkill 0.2.0 — MCP server and paper role adapters

• MCP server: python -m pip install 'wikiskill-research[mcp] @ git+https://github.com/Stahl-G/wikiskill.git' and wikiskill mcp expose the host-agent workflow (start, next, record, learn, propose, feedback, retry, status, report, preflight, scorer inspection, export, capabilities) to any MCP client. Tools accept output, pattern and skill text directly and make no model calls. Scorer authorization, skill installation and restoration remain terminal-only.
• Feedback-first host API: wikiskill.feedback_loop.begin/work/finish starts Wiki maintenance from saved user revisions and comments, with host-supplied paired comparisons and an optional better-than-worse adoption policy. Numeric scoring is unchanged.
• Portable Python scorers: {python} in a scorer command resolves to WikiSkill's own interpreter.
• Research adapters: engine.evolve() accepts domain_loader, maintainer_factory and proposer_factory; engine.initialize() accepts a Wiki template and role prompts. paper_alignment.PaperAgents runs the paper's appendix Maintainer and Proposer prompts on TRAIN-only evidence and requires four actual trace reads before a proposal. The engine still owns validation selection, rollback and the journal.
• Documentation: paper-to-code conformance map and a README section on WikiSkill inside BriefLoop.
• Package version drops the local +briefloop suffix. No research result was remeasured; historical records are unchanged.

WikiSkill 0.1.1 — Native subagent workflows

WikiSkill 0.1.1 adds native subagent orchestration for the host-agent workflow, building on the WikiSkill paper.

The main agent coordinates. Separate task executors, a Wiki Maintainer and a Skill Proposer work from role-specific context packages. The host creates the subagents with its native tools; WikiSkill records outputs, scores them and applies the validation gate.

What's new

• Codex and Claude Code assets: install the entry skill and three native role definitions with wikiskill agents install.
• Fresh role handoffs: baseline, training and candidate-validation tasks receive separate contexts. Maintainer submission precedes Proposer dispatch.
• Recorded delegation: dispatch, bind-agent, collect and fail preserve host IDs and integrate native work with the existing controller. Different requests cannot reuse the same child ID. Scoring retries reuse saved execution instead of rerunning the model.
• First-use improvements: readiness checks, readable status, journal-derived reports, and local skill installation with backups and restoration.
• A workbook-delivery example: public inputs and a checker for live formulas, saved values and preserved inputs/number formats.
• Consistent scoring: shared finite-score/strict-improvement rules; missing research scores no longer become zero. Historical research records remain unchanged.

Install

In a Python 3.11+ project environment:

python -m pip install git+https://github.com/Stahl-G/wikiskill.git@v0.1.1

Choose your host, in the project where you want to use WikiSkill:

wikiskill agents install --runtime codex --project .
# Or:
wikiskill agents install --runtime claude-code --project .

Reload host discovery if needed, then ask your agent to use WikiSkill for one round with your examples and evaluator. The package does not select a model or change host permissions. Existing differing host files are preserved for review.

Verification

Local tests: 144 passed, 4 skipped. The installed wheel exercised host-asset installation, the offline demo, bundled result verification and native-workflow reporting.

A real Codex desktop check used five distinct fresh-context subagents for baseline, training, maintenance, proposal and validation. The gate correctly rejected a 1.0/1.0 tie. Claude Code definitions were installation/syntax-checked; Claude live inference has not been verified here. Fresh child conversations do not provide filesystem or host-memory isolation.

Native workflow and validation details · Product guide · Changelog

WikiSkill 0.1.0

WikiSkill turns agent experience into a persistent Wiki and reusable skills. This first release provides an independent implementation of the WikiSkill paper, with a practical workflow for your own agent and tasks.

What's included

• An agent-driven improvement loop: run tasks, score outputs, maintain a Wiki, propose a skill, and retain it only when validation improves. Rejected proposals leave their lessons in the Wiki.
• Bring your own agent and scorer: use your current model and tools, a command-based evaluator, or human scores. Import an existing skill, add human feedback, and export the retained skill with provenance.
• An installable wikiskill skill: guides your agent through the loop without requiring a particular model or research environment.
• Local scorer trust: inspect and approve an external scoring command once; changed commands and directly referenced files require renewed trust. Approval stays outside shared workspaces.
• Research tools and results: benchmark adapters, preserved experiment records, and a three-run fixed-skill Spreadsheet study. See the repeatability report for results and methodology.
• English and Chinese documentation, architecture diagrams, and a runnable text-cleanup example.

Install

In a Python 3.11+ virtual environment:

python -m pip install git+https://github.com/Stahl-G/wikiskill.git@v0.1.0

Add the agent skill:

npx skills add Stahl-G/wikiskill --skill wikiskill

Start with the product guide.

Validation

CI passed for this release commit, including Linux Python 3.11/3.12 and product portability checks on Linux, Windows, and macOS. Published research measurements identify their original harness; they are separate from the product workflow checks.