v0.4.2 — Safety gate, memory write-back, domain frameworks, evidence-gap search

What changed

• Halt-vs-guess gate: scope-agent and execute-agent now classify every action before proceeding — reversible actions continue, irreversible+out-of-scope actions halt and confirm.
• Memory write-back hook: Stop hook detects corrections and completions, writes to procedural.md / episodic.md automatically.
• Handoff contracts: All 4 agents now emit next_agent routing field. Review-agent runs a consistency check against episodic.md.
• Domain sharpening: All 10 domain packs now include ## Frameworks (authoritative sources) and ## Live Facts (facts that go stale).
• Evidence-gap search: SKILL.md rule — execute-agent searches before acting when a Live Facts gap is identified.
• Smart push script: scripts/push.py — preflight report, version consistency, honesty checks, explicit confirmation before push.

Honesty note

No live benchmark run for v0.4.2. Unit tests pass (46/46). v0.4.1 gpt-5.5 numbers (+3.9 pts) are from the v0.4.0 run — gpt-5.5 was not re-run in v0.4.1 or v0.4.2.

Unit tests

46/46 passing.

v0.4.1 — PLUGIN_DATA memory bug fixed, gpt-5.4-mini retest confirmed

What changed

• Bug fix: context_inject.py was reading PLUGIN_DATA as the memory dir in Codex environments. No memory files exist there, so the hook injected zero context on every prompt.
• Fix: Removed the PLUGIN_DATA branch. Memory dir now resolves correctly.
• Added: ## Safe Defaults section to SKILL.md.

Test results

model	v0.4.0 delta	v0.4.1 result
gpt-5.4-mini · low	-10.5 pts	+2.6 pts (retest confirmed, n=40)
gpt-5.5 · medium	+3.9 pts	not re-run in v0.4.1 — v0.4.0 number stands

gpt-5.4-mini recovered +13.1 pts after the fix. gpt-5.5 was not re-run.

Unit tests

35/35 passing.

v0.4.0 — 10 domain packs, 32 tests, context hook, memory bootstrap

Everything AI v0.4.0

Hands the agent expert scope across 10 domains.

Added

• 7 new domain packs (10 total): coding, writing, health, learning, finance, life, research — joining startup, data-analysis, personal-productivity.
• Memory bootstrap — creates semantic.md, episodic.md, procedural.md on first install.
• Plain-language UX — QUICKSTART.md + output rule (no jargon, short by default).
• Context hook — injects today's date and working directory on every prompt (Codex).
• NPX install setup screen — shows what gets installed before copying.
• Benchmark expanded to 20 scenarios.

Proof

• 32/32 unit tests passing.
• 20/20 benchmark scenarios, medical-safety regression passing.

Install

npx --yes github:mitunmanav/everything-ai

Use --agent claude for Claude.

Everything AI v0.3.0

Everything AI v0.3.0

v0.3.0 is the main-branch proof release for safer broad delegation.

What changed

• Tightened Everything AI ask-gate behavior for non-technical users.
• Added urgent medical and high-stakes safety boundary.
• Added contradiction and stale-status defaults.
• Added v0.3.0 benchmark contract and release proof docs.
• Added raw with-skill vs without-skill result file.
• Added visual graph showing behavior score and visible-output token estimate.
• Added npm lockfile so dependency audit can run cleanly.

Test model and method

• Model: gpt-5.4-mini
• Reasoning: medium
• Execution: fresh-subagent manual scorecard
• Scenarios: 10
• With skill: 20/20
• Without skill: 14/20
• Lift: +6

Token data

• Visible-output token estimate, without skill: 210
• Visible-output token estimate, with skill: 295
• Delta: +85
• Method: word/punctuation split over saved visible responses, not API billing usage.
• Plugin-eval static skill budget: trigger 39, invoke 407, deferred 510, total 956 tokens.

Proof files

• Raw result: tests/results/v0.3.0-with-vs-without-skill.json
• Graph: tests/results/v0.3.0-with-vs-without-skill.svg
• Benchmark contract: tests/evals/everything_ai_benchmark.json
• Report: TEST_RESULTS.md

Validation

• npm test: passed
• npm audit --omit=dev --audit-level=moderate: 0 vulnerabilities
• skill quick_validate: passed
• plugin-eval: 100/100, Grade A, low risk
• tracked-file personal data scan: clean
• release excludes AGENTS.md and development-only repo rules
• GitHub PR checks: test, analyze, CodeQL passed
• PR #7 merged into main

Known gap

• Scorecard is manual-scored, not automated model-runner output yet.
• No real API billing token log was available, so token data is clearly labeled as a visible-output estimate.

Everything AI v0.2.0

Everything AI v0.2.0 public release. Includes persistent local memory rules, AI observability trace guidance, non-expert safe defaults, blocker-only questions, local installer, tests, security policy, contribution guide, CodeQL, Dependabot, branch protection setup, and README star history graph.