Changelog

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

Format based on Keep a Changelog.

[0.3.1] - 2026-10-04

Changed

• Add a portable OpenAI plugin manifest alongside the Claude manifest.
• Keep skill instructions shared and make bundled resources reachable from the skill directory.
• Document installation scope and public ChatGPT listing status.

[0.3.0] - 2026-10-04

Changed

• Rename brainforge-autoresearch to brainharness-autoresearch and align repository links and installation names with BrainHarness.
• Update distribution metadata and documentation; preserve skill behavior and historical release notes.

[0.2.5] - 2026-04-23

Changed

• Renamed repo and skill: openclaw-autoresearch → brainforge-autoresearch
• SKILL.md name field: autoresearch → brainforge-autoresearch
• Updated README, plugin.json, and SKILL.md metadata to point at the new repo URL
• Added to the brainforge Claude Code plugin marketplace
• Published to ClawHub as brainforge-autoresearch; old slug openclaw-autoresearch merged in as redirect

Compatibility

• GitHub: old URL https://github.com/zning1994/openclaw-autoresearch auto-redirects
• ClawHub: openclaw skills install openclaw-autoresearch continues to work via ClawHub slug merge/redirect
• No behavior changes to the optimizer itself

0.2.4 - 2026-03-29

Fixed

• SKILL.md: pass/fail → pass_description/fail_description in LLM eval docs and examples (field names didn't match actual API)
• SKILL.md: contains/not_contains parameter documented as value (string) → values (list) to match actual usage
• SKILL.md: LLM eval table now includes required type and name fields

Changed

• SKILL.md self-optimized via autoresearch (2 rounds: MiniMax 48.6%→68.1%, Claude 80.6%→91.7%)
• Procedure rewritten: added 3-step summary upfront, renamed steps for clarity, added actionable result-review guidance
• Example section restructured: removed "confirm target" preamble, simplified to "create eval → run → review" flow
• Added self-eval.json for self-optimization (dogfooding eval definition)

Tested

• Self-optimization: Round 1 (MiniMax M2.7, 15 exp, 3 keep) and Round 2 (Claude, 15 exp, 1 keep) independently discovered same doc bugs

0.2.3 - 2026-03-25

Added

• --timeout CLI flag to control HTTP timeout per LLM call (default: 180s)
  • Large prompts (e.g., 14KB style guides) need longer timeouts for mutation calls
  • Previously hardcoded, now user-configurable

Changed

• Default HTTP timeout increased from 60s to 180s (was causing mutation failures on Anthropic API with large prompts)
• Timeout is now stored on the provider instance, passed through detect_provider()

0.2.2 - 2026-03-25

Fixed

• .gitignore: removed incorrect entries (.git, .github), added autoresearch-*/ and *.baseline
• README: added ClawHub badge and link, fixed install slug to openclaw-autoresearch

0.2.1 - 2026-03-25

Fixed

• ClawHub security flag: changed requires.env (all required) to requires.anyEnv (any one suffices)
• Declared optional env vars OPENAI_BASE_URL / OPENAI_API_BASE in metadata

0.2.0 - 2026-03-25

Fixed

• MiniMax extended thinking support — models with thinking blocks (M2.7, Claude) no longer crash the parser
• Fallback to thinking content when text block is missing (thinking exhausts max_tokens)
• LLM judge max_tokens increased from 16 to 256 (thinking models need headroom)
• Rule type field: accept both "rule" and "check" keys for backwards compatibility
• contains/not_contains: accept both "values" (list) and "value" (string)
• LLM eval fields: accept both "pass_description"/"fail_description" and "pass"/"fail"
• test_inputs: accept both plain strings and {"name": ..., "input": ...} objects
• contains rule now supports "match": "all" mode (require all values present)
• Convergence counter double-increment bug fixed
• regex rule now uses re.MULTILINE for correct ^/$ matching
• CJK-aware word counting in word_count rule (counts Chinese/Japanese/Korean characters individually)

Added

• --model CLI flag to override the default model per provider
• JSON fallback: runs_per_experiment and max_experiments from eval.json used when CLI defaults unchanged
• Shared _extract_text() helper for Anthropic-style response parsing across all providers

Tested

• Successfully optimized brain-search skill: 37.5% → 54.2% pass rate in 5 experiments with MiniMax M2.7

0.1.0 - 2026-03-24

Added

• Core autoresearch.py script — autonomous skill prompt optimizer
  • Zero external dependencies (Python 3.9+ stdlib only)
  • LLM providers: MiniMax, OpenAI, Anthropic (auto-detect from env vars)
  • Custom endpoint support via OPENAI_BASE_URL
  • --model flag to override default model per provider
• Hybrid eval system
  • Rule-based evals: regex, banned_phrases, word_count, contains, not_contains
  • contains supports match: "all" and match: "any" modes
  • LLM-as-judge evals with binary YES/NO scoring
  • CJK-aware word counting for Chinese/Japanese/Korean text
• Experiment loop
  • Baseline measurement before any changes
  • One mutation per experiment (targeted, not bulk rewrites)
  • Automatic keep/discard based on score comparison
  • Convergence exit: 95%+ pass rate for 3 consecutive experiments
  • Budget cap via --max-experiments
• Output artifacts
  • results.tsv — tab-separated score log
  • changelog.md — detailed mutation history with reasoning
  • results.json — structured data for tooling
  • dashboard.html — self-contained Chart.js dashboard (opt-in via --dashboard)
  • SKILL.md.baseline — original skill backup
• SKILL.md — agent instructions for OpenClaw / Claude Code / Cursor / Cline
• eval-guide.md — practical guide for writing binary eval criteria
• Example eval configs: weekly-report.json, search-skill.json
• Compatible with npx skills add (Vercel Labs skills ecosystem)