v1.4.4

AI-KIT v1.4.4 — Optional Orchestration and Budget Tracking

AI-KIT adds optional task delegation through Codex helpers and an OpenRouter text executor, with explicit scope, configurable limits and recorded API costs.

Added

• Codex helper configurations for exploration, implementation and review, using AI-KIT’s model roles and inheriting parent permissions.
• OpenRouter text executor with an explicitly selected model, project input files and limits on requests, input size, output tokens and execution time.
• Budget controls: dedicated capped-key preflight, durable reservations before requests and project-wide locking. Uncertain billing or reservation overruns block further requests until resolved.
• Usage accounting: actual API cost, token counts, model identity and generation IDs. Native Codex costs remain separate.
• Review and recovery: source-bound result acceptance, reuse of matching completed results and billing reconciliation for known generations without repeating inference.
• CLI commands: aikit orchestrate run, status, review and reconcile.

Installation and Documentation

• Orchestration is an independent opt-in module through --with-orchestration; configuration starts disabled.
• Preview performs no network requests or writes.
• Existing project documents, settings and customized helper files retain normal preservation and conflict handling.
• Private orchestration state is excluded from Git, package archives and Docker build contexts.
• README includes detailed orchestration instructions and a prominent Docker Hub / PyPI installation

AI-KIT v1.4.3 — Owner-approved PyPI releases

What's changed

• Stable GitHub Releases now start checks -> owner approval -> PyPI publication automatically.
• Future stable vMAJOR.MINOR.PATCH tags work without updating the environment allowlist. Release tag, committed VERSION, runner ref and canonical repository must agree.
• Drafts and prereleases skip automatic production preparation. Manual dispatch remains available and defaults to build-only validation.
• Fifteen new regression tests cover release identity, future tags, invalid events, manual defaults and untrusted payloads.
• README includes the full publication/recovery procedure and a live PyPI version badge. CLI, package metadata and the local development image are versioned 1.4.3.

Publication

The release workflow builds and checks the package first. PyPI upload then requires the repository owner's confirmation: Actions -> Review deployments -> pypi -> Approve and deploy. Agents do not approve this step. Upload jobs use short-lived OIDC credentials and the same run's tested artifacts.

Docker Hub remains at the separately published tigusigalpa/ai-kit:1.4.1 image. The local 1.4.3 image is prepared and tested; no Docker Hub upload is part of this release.

See the release record and full changelog.

v1.4.1

AI-KIT v1.4.1

Release date: 2026-10-15

• Added Docker image freshness instructions: assess change impact, update build inputs, rebuild affected images and verify their behavior.
• Excluded local root maintainer documents from Docker build context, wheels and source distributions.
• Preserved application templates and documented public maintenance through CONTRIBUTING.md.
• Updated validation to support distributions without local maintainer files and reject public links depending on them.
• Verified upgrades from 1.4.0 in private/team modes, preserving project context, preferences, dialogue memory and local edits.
• Published tigusigalpa/ai-kit:1.4.1 for linux/amd64; independently verified the registry digest and runtime configuration.

Validation: 339 Linux tests — 337 passed, two platform skips; nine Windows CLI tests passed. Docker smoke, offline wheel installation, package exclusions and kit validation passed.

v1.4.0

AI-KIT v1.4.0 — Docker & Shared Dialogue Memory

Docker CLI

• Standalone image containing Python, Git, AI-KIT and optional integration resources.
• Multi-stage build, digest-pinned Python base and hash-verified build tools.
• Non-root execution, offline commands and support for a read-only root filesystem.
• Kubernetes preview Job and Docker smoke tests.
• Prepared image: tigusigalpa/ai-kit:1.4.0 for linux/amd64.

Shared dialogue context

• Import ChatGPT export JSON, OpenRouter-style messages JSON and text/Markdown.
• Preserve selected ChatGPT branches and source references.
• Review pending claims and resolve explicit keyed contradictions.
• Export immutable reference documents for ChatGPT and OpenRouter.
• Protect private memory through Git ignore policies and build/package exclusions.

Verification

• Post-label Linux suite: 336 tests, 334 passed and two Windows-only skips.
• Nine Windows CLI tests passed.
• Docker smoke, offline wheel installation and private-memory exclusion passed.
• Upgrades from the tracked 1.3.0 baseline preserved project data in private/team modes.
• Final kit validation: zero errors and warnings.

Conversation synchronization, account connections and API execution remain separate. Docker Hub publication is a separate step.

v1.3.0

AI-KIT v1.3.0 — Web Design

AI-KIT now includes an optional Web Design module for turning website briefs and annotated references into a coherent design, implementation and recorded UI acceptance.

Added

• --with-web-design installs four focused skills: web-design, laravel-blade-ui, ui-review and design-extract.
• Project design memory in docs/design/BRIEF.md, SYSTEM.md and ACCEPTANCE.md. Existing documents are preserved during installation and upgrades.
• Conditional guidance for Laravel/Blade, Tailwind, Bootstrap, Alpine, Livewire, daisyUI, Flowbite and Flux.
• Optional TypeUI and SkillUI workflows, with documented setup, provenance and execution boundaries.
• Explicit context selection for design, UI review and reference analysis.
• UI acceptance covering responsive layouts, keyboard/focus behavior, accessibility, component states and working application interactions.

Improved

• Removed the root README size quota and added a detailed Web Design guide with installation instructions, examples, starter requests and troubleshooting.
• Preserved existing project stacks and component systems.
• Added installation, conflict, upgrade, path-safety and client lifecycle coverage.
• Kept temporary design output private while ordinary project design documents remain shareable.

Validation

Before release labeling, 304 tests ran: 300 passed, four skipped because required Windows runtime capabilities were unavailable.

After labeling, eight CLI tests, clean kit validation, v1.3.0 wheel installation and a preservation upgrade passed. Installation and visibility were also verified for all 11 clients in private and team modes.

Enable

Preview installation with:

python scripts/install.py TARGET --with-web-design

Installation remains offline. External accounts, MCP connections, packages and deployment are separate steps. Live TypeUI/SkillUI execution and generated-site acceptance remain unverified.

v1.2.0

AI-KIT v1.2.0

This release adds an optional Context7 documentation workflow and expands the OpenRouter integration guide.

Context7 support

• Added --with-context7 and the short context7-docs skill, with shared source instructions and Claude Code native copies.
• Added guidance for checking installed dependency versions, resolving library IDs, retrieving focused documentation and citing sources.
• Added MCP setup examples for Codex, Claude Code and Cursor, using environment-based credentials or OAuth.
• Context7 remains optional and outside presets. Selections persist independently alongside security-audit and Strix.
• Installation stays offline and preserves existing MCP configuration and local skill modifications.

MCP connection and authentication are configured separately. Live documentation retrieval remains unverified.

OpenRouter documentation

Added a dedicated README section covering AI-KIT context assembly, API request structure, prompt caching, response caching, provider routing and observed usage accounting.

The guide describes compatibility and a proposed adapter design. A live OpenRouter API adapter remains future work.

Validation

• Full pre-label suite: 293 tests — 289 passed, four platform-dependent checks skipped.
• Eight CLI checks passed after version labeling.
• Kit validation completed without errors or warnings.
• The 1.2.0 wheel passed offline installation and stable rerun checks.
• A synthetic upgrade preserved project facts, settings values, local Context7 edits, MCP configuration and both security integration pins.

v1.1.0

AI-KIT v1.1.0 — Security workflows for coding agents

AI-KIT adds two optional security workflows: Cloudflare’s structured audit method and Strix’s pentesting skills. Enable either independently or combine them for security review, scoped testing and findings verification.

Cloudflare security-audit

• Complete upstream method covering reconnaissance, vulnerability hunting, evidence validation, coverage and reporting.
• Focused guidance and review, with full audits explicitly requested separately.
• Attack-class references, report schemas and Node.js validators.
• Pinned revision with the original MIT license.

Strix

• Nine upstream skills covering code, web applications, APIs, OWASP testing, findings and requested remediation.
• A short entry point for preparing scoped local scans and reviewing evidence.
• Explicit scope, model, budget, time limits and isolated writable target copies.
• Separate CI/cloud guidance, with a pinned revision and original Apache-2.0 license.

Controlled integration

Both workflows are off by default, remain outside presets and persist when explicitly selected. Codex and Claude Code receive short entry points with implicit invocation disabled. Offline installation verifies file coverage, sizes, Git blobs and SHA-256 while preserving local adaptations through existing backup and conflict handling.

Enable with --with-security-audit, --with-strix, or both. Invoke $security-audit / $strix in Codex, or /security-audit / /strix in Claude Code.

Documentation and validation

README now includes capability comparisons, installation and invocation examples, provenance and execution prerequisites.

Before release labeling, 285 Windows tests ran: 281 passed and four were skipped because of environment limitations. Release checks passed eight CLI tests, version/wheel verification, clean kit validation, offline package installation and rerun, plus a preserved upgrade from 0.13.0.

Installation provides instructions and references. Actual scans require a configured runtime and authorized scope; Cloudflare’s pinned validator CLI refuses unsupported native Windows inputs.

v0.13.0

AI-KIT v0.13.0

AI-KIT is a complete working environment for AI coding agents: verified project memory, stack rules, guardrails, adaptive model routing, and an offline measurement harness. v0.13.0 widens stack coverage from seven profiles to eleven — Rust, .NET, Java, and Terraform/Helm join Go, PHP, Laravel, Filament 5, Moodle, Python, frontend, and library guidance.

New stack profiles

Each profile follows the established format — confirm your actual toolchain first, then concise rules and primary-documentation sources:

• Rust — ownership and borrowing discipline, unsafe behind reviewed boundaries with safety comments, Result<T, E> error design (thiserror/anyhow), cargo fmt/clippy/test, Cargo.lock hygiene, MSRV/edition awareness.
• .NET — nullable reference types, async end-to-end with CancellationToken (no sync-over-async), deterministic disposal, typed exceptions with preserved causes, dotnet build/test, global.json SDK pinning.
• Java (Maven/Gradle) — wrapper-first builds (mvnw/gradlew), checked-exception boundaries, try-with-resources, Optional as return type only, pinned plugins/dependencies, wrapper test/verify gates.
• Terraform/Helm — remote state and no committed secrets, fmt → validate → plan → reviewed apply, pinned providers/modules/charts, helm lint/helm template before install, destructive-change review in CI.

Evidence-based detection

The installer now drafts facts for the new stacks per module, the same way it already handles Go, Python, Node, and PHP:

• Cargo.toml → RUST, language/version from rust-version, commands cargo test/build/clippy
• *.csproj / *.fsproj / *.vbproj / global.json → DOTNET, version from <TargetFramework>, dotnet test/build
• pom.xml / mvnw / build.gradle(.kts) / gradlew → JAVA, language level from compiler properties, wrapper-aware commands (./mvnw test, ./gradlew build)
• .tf files → Terraform (terraform fmt -check, terraform validate); Chart.yaml → Helm (helm lint .)

Everything stays a clearly-marked suggestion until bootstrap confirms it — detection never asserts stack facts.

Ecosystem integration

Four new registry skills (rust-work, dotnet-work, java-work, infra-work) point agents at the right profile; Claude native copies install automatically with the claude agent. Benchmark packs accept the new stack names, and the README feature tour, SKILL.md, and WIKI profile lists are updated.

Verification

• 254 behavioral tests pass locally (Windows/Python 3.14.6; 2 new detection suites), 1 runtime skip; kit validation reports zero errors and zero warnings, including skill-registry parity.
• End-to-end smoke: a fixture with Cargo.toml and a Terraform module gets RUST/INFRA profiles, a drafted module map, and source + native skills installed.
• Python 3.10 grammar verified.

Known limits

Profile rules are established stack practices sourced from official documentation, but no real-world Rust/.NET/Java/Terraform applications have been trialed with the kit yet — as with all kit claims, fixtures verify wiring, not outcomes. Hosted CI and publication remain unverified.

---

v0.12.0

AI-KIT v0.12.0

AI-KIT is a complete working environment for AI coding agents: verified project memory, stack rules, guardrails, adaptive model routing — and, since v0.11, an offline measurement harness that turns routing from guesswork into evidence. v0.12.0 hardens that harness: benchmark runs survive mistakes, resume cleanly, and model your real client setup.

Benchmark hardening

Configurable kit arm. Benchmark packs accept optional kit_agents and kit_preset (default codex/minimal). Kit arms now install the adapter set and preset your project actually uses — previously hard-coded to Codex/minimal — so measurements reflect your real setup. Both fields are validated against the installer registry at pack time.

Resumable runs. Re-running benchmark run --apply with an identical pack no longer fails with "already exists": units with an existing result.json are never re-executed or rebuilt (no repeated paid calls), missing units continue, and the preview shows existing_run and remaining_units. A changed pack is refused against the existing run via a SHA-256 plan identity, so completed work can never be silently reinterpreted under a different experiment.

Per-unit failure isolation. An infrastructure failure inside one unit — kit install conflict, unevidenced task stack, workspace errors — is recorded as an explicit failed unit with an error field instead of aborting the entire run. Failed units honestly block savings, can never be reviewed as correct, and the worker is never invoked for them: no hidden paid retries, no silently lost hours.

Correct path semantics. Task paths in packs are now validated against the benchmark project instead of the process working directory.

CLI packaging

The unified aikit dispatcher prefers installed aikit.scripts.* package modules over bare script-path imports, eliminating duplicate module objects inside the wheel. Behavior is unchanged from a checkout.

Docs and budgets

README documents the new benchmark capabilities; its checker budget is now a named constant (README_BUDGET = 22000, other documents stay at 12000) so onboarding content has real headroom. The benchmark contract (docs/BENCHMARK.md) and continuity procedures are updated to match.

Verification

• 252 behavioral tests pass locally (Windows/Python 3.14.6; 1 runtime skip), including new regressions for pack-path probing, kit-arm validation and installation, resume identity, per-unit failure records, and dispatcher package preference.
• check_kit.py reports zero errors and zero warnings; Python 3.10 grammar verified.
• End-to-end benchmark smoke: mixed-fixture run completes with recorded unit failures, resumes without retries, and refuses a changed pack.

Known limits

Synthetic workers verify harness behavior only — no savings, model quality, or native client activation is claimed. Live adapters and orchestration remain behind the ADR 0012 boundaries until real application trials (owner-deferred) supply evidence. Hosted CI and PyPI publication remain unverified.

---

v0.11.0

AI-KIT v0.11.0

October 11, 2026

Reproducible benchmarks, routing informed by measured results, and context selection by module.

Added

• Isolated A/B/C benchmarks through aikit benchmark run, review, report, and export. Runs use immutable Git snapshots, fresh workspaces, fixed A/B model controls, semantic acceptance, and JSON/Markdown reports.
• Metric schema 3 with revision, case, repetition, stack, task class, capabilities, and review metadata.
• Cost per Correct Result, actual-spend comparisons, and stack statistics. Failed attempts and retries count toward cost; missing or estimated costs cannot establish savings.
• Routing modes and profiles: manual, fixed, recommended, adaptive; economy, balanced, quality. Includes project/task pins and conservative recommendations based on sufficient comparable evidence.
• Immutable proposal history and explicit configuration restore with backups. Saving a proposal does not change active routing.
• Module context selection with duplicate detection and byte budgets that retain critical rules and architectural decisions.
• Orchestration research with configurable planner, executor, reviewer, and escalation role mappings.

Fixed

• Provider/model overrides carrying incompatible inherited effort.
• Gemini policy inconsistencies with the reviewed registry.
• Adapter audits failing on file-blocked native paths.
• Subcommand help forwarding and wheel resource completeness.

Compatibility and validation

Existing settings, provider selections, and metric schemas 1/2 remain supported.

Validated on Windows/Python 3.12.14: 244 behavioral tests passed, four environment-dependent skips, seven release CLI checks passed, and kit validation returned zero errors and warnings. The v0.11.0 wheel passed installation and command checks in a clean environment outside the source checkout.

Real application/model trials, live client switching, and runtime orchestration remain separate stages. No real-world savings have been established.