Changelog

All notable changes to this project will be documented in this file.

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

Unreleased

1.8.0 - 2026-09-29

Added

• A portable Agent Plugins 1.0 manifest, plugin.json at the repository root, which Copilot and Codex read before any host-specific manifest. The host manifests stay for Claude Code, Cursor and older Copilot and Codex, with the same name, version and description.
• Plugin manifests for three more hosts: .codex-plugin/plugin.json (Codex CLI), .cursor-plugin/plugin.json (Cursor) and .github/plugin/marketplace.json (GitHub Copilot CLI), with install instructions for each in the README. They shipped from the default branch after 1.7.0 while carrying 1.7.0.

1.7.0 - 2026-05-26

Changed

• Phase 0 "Start fresh" archives to .lab/archives/<slug>/ (using the existing research/<slug> branch slug; suffixed with -<timestamp> on collision) instead of the sibling .lab.bak.<timestamp>/. Keeps archived runs inside the sacred .lab/ tree so they survive git operations alongside live experiment state.

1.6.0 - 2026-05-03

Added

• Claude Code plugin scaffolding. New .claude-plugin/plugin.json (manifest: name, version, description, keywords) and .claude-plugin/marketplace.json (single-plugin marketplace listing). The repo is now installable as a Claude Code plugin via /plugin marketplace add krzysztofdudek/ResearcherSkill then /plugin install researcher@researcher-marketplace. Single-file drop-in still works for Codex, Cursor, and any other agent that reads markdown skills — copy skills/researcher/SKILL.md into the agent's skill directory.
• README "Install" section documenting both paths (plugin install + single-file drop-in for non-Claude-Code agents).
• CLAUDE.md "Plugin scaffolding" section describing the .claude-plugin/ layout and the rule that plugin.json version is bumped in lockstep with the CHANGELOG section header.

Changed

• Canonical skill body relocated from repo root (researcher.md) to skills/researcher/SKILL.md — the path Claude Code reads directly when the plugin is installed. The root-level researcher.md is removed; users on Codex / Cursor / other agents copy skills/researcher/SKILL.md instead. There is now exactly one canonical file.
• Skill description frontmatter rewritten to be trigger-focused (compliant with the AutoReview skill-description rule): starts with "Use when", third person, lists concrete user phrasings (make X faster, improve [metric], find best config, iterate overnight, run experiments until it hits N), names quantitative vs qualitative metric domains, and includes a "Skip for…" exclusion clause. Old description summarized the workflow ("agent interviews the user, sets up a lab, then explores freely"); the new one tells the agent WHEN to invoke. 471 chars (under the 500-char ceiling enforced by the rule).
• CLAUDE.md Versioning workflow updated to include the .claude-plugin/plugin.json version bump as an explicit step (was: implicit "in lockstep" mention only in the Plugin scaffolding section).

Removed

• Root-level researcher.md and the three-clone sync rule that kept it byte-identical with skills/researcher/SKILL.md and ~/.claude/skills/researcher/SKILL.md. The plugin install path makes the canonical-at-skills/researcher/SKILL.md model authoritative; clones are no longer maintained.

1.5.0 - 2026-04-06

Added

• Token Hygiene — new interview question (Phase 1) and setup step (Phase 2) for ecosystem-specific ignore/rules files and helper scripts (.lab/bin/) to reduce token usage
• Token Hygiene Standards reference section — hard rules for measure and data_head helpers, context safety rules preventing accidental agent blinding

Changed

• Configuration confirmation uses a compact table instead of full restatement; defaults shown as "default"

1.4.4 - 2026-03-29

Changed

• THINK phase now requires mandatory ## THINK — before Experiment N log entry with 4 named subsections (Convergence signals, Untested assumptions, Invalidation risk, Next hypothesis). Without the entry, THINK didn't happen.
• Repo-file vs lab-only boundary clarified: if you modify any file in scope, it's a repo-file experiment — no "quick test" exception
• 3-discard and 5-discard guardrails wrapped in <critical>, require mandatory log entries (## 3-Discard Guardrail, ## 5-Discard Fork) before proceeding

Added

• Discipline validation test framework (tests/discipline/): deceptive landscape scorer, automated 18-checkpoint verifier, test runner
• Lab 2 validation results (archive/lab2-skill-discipline-validation/): 5 synthetic tests + retest confirming fixes

1.4.3 - 2026-03-29

Fixed

• Domain-neutral language: removed parameter-tuning examples (alpha/beta) from THINK, Strategy Diversification, convergence signals — replaced with domain-diverse examples
• Termination default explicitly says "do not self-impose experiment limits" — agent runs until user interrupts or target is reached
• Session continuity: documented that .lab/ persists across sessions, Phase 0 handles resume

1.4.2 - 2026-03-29

Fixed

• THINK checklist generalized — works for prompt optimization, code refactoring, not just parameter tuning
• Parking lot trigger added to 5+ discard guardrail — check for untested ideas before forking
• Phase 4 wrap-up clarified — explicit branch checkout, SHA handling for closed branches, commit message format

1.4.1 - 2026-03-29

Fixed

• Convergence signal plateau threshold aligned with guardrail (both 8+, was 5 vs 8)
• Reduced <critical> blocks from 3 to 1 (top) + 1 (execution discipline) to prevent desensitization
• Guardrails (step 5) placed directly after decide (step 4) — removed git log example that broke the flow
• THINK phase references critical rules and guardrails correctly (was pointing "above" to a section below)
• Qualitative baseline now explicitly requires Multi-Evaluator Protocol (3 subagents)
• Phase 4 re-validation explicitly mentions Multi-Evaluator Protocol for qualitative metrics
• Log entry format clarified as labeled fields, not ambiguous slash-separated line
• Orphaned SHA warning: "fork from discarded SHAs sooner rather than later" (git gc risk)

1.4.0 - 2026-03-29

Added

• Execution guardrails in Phase 3: 3+ discard review, 5+ discard mandatory fork, plateau detection (8+ experiments without global best improvement), mandatory re-validation every 10th experiment (<critical>)
• Strategy Diversification section — when forking due to stagnation: assumptions list, deliberate fork point selection (baseline/best/discard), mandatory assumption inversion (<critical>) on first experiment of new branch
• THINK phase structured checklist: convergence signal check, monotonicity verification (3+ data points per variable), interaction awareness
• Multi-Evaluator Protocol for qualitative metrics — main agent spawns 3 independent subagent evaluators per experiment, each receives only artifact + rubric, median aggregation, divergence flagging
• Orchestration model — skill is for the main agent; subagents receive scoped tasks via prompt, not the skill file
• Autonomous git operations — commits, resets, branch creation are systemic, no user permission needed
• New convergence signals: variable tested in one direction only, plateau detection, anchoring across branches, locally optimal with desperate variants
• Re-Validation moved from standalone section into guardrails (integrated into execution flow)

1.3.0 - 2026-03-28

Added

• Metric Revision procedure in Phase 3 — clean mid-series metric change with mandatory re-scoring of all keeps for comparability
• Convergence signal for unmeasurable dimensions (always-neutral scores suggesting metric revision)

1.2.0 - 2026-03-28

Added

• .lab/workspace/ — dedicated scratch directory for experiment files (scripts, test data, generated output, per-experiment subdirectories). Untracked and safe from git operations.

1.1.1 - 2026-03-28

Changed

• Experiment commit discipline now uses structured commit messages (experiment #{N}: description with branch, parent, hypothesis)
• Each branch holds only keeps — DISCARD/CRASH use git reset --hard HEAD~1 (SHA preserved in results.tsv)
• Forking from any experiment (including discards) via new branch from SHA, with branches.md tracking genealogy
• Distinguished repo-file experiments (tracked files) from lab-only experiments — commit rules only apply to the former
• git log --oneline on a research branch now reads as a clean line of progress (gaps in numbering = discards or other branches)

1.1.0 - 2026-03-24

Added

• CHANGELOG.md — changelog following Keep a Changelog format
• Before building a qualitative rubric, the skill now asks the user whether they already know their criteria or need help figuring them out
• GitHub Actions pipeline that automatically creates a GitHub Release on every version tag push, with release notes extracted from CHANGELOG.md

1.0.0 - 2026-03-23

Added

• Initial release of the Researcher Skill (researcher.md)
• Autonomous experimentation loop: interview, lab setup, think-test-reflect cycle
• .lab/ directory as local, untracked experiment log (git-safe)
• README.md with use cases, examples, and experiment results
• GUIDE.md — detailed usage guide
• FAQ