Changelog

aiflow follows Keep a Changelog and Semantic Versioning. The authoritative, always-current changelog lives in the repository: CHANGELOG.md.

0.8.0 — an orchestrated agent network, binding architecture rules, Windows/WSL

Highlights:

• The agent roster became a network, with an orchestrator. Agents knew the rules but not each other, so routing lived in your head. The new orchestrator agent and /orchestrate <goal|bead> decide which specialist handles each step, dispatch exactly one at a time, judge the result and route on — planner → architect (when boundaries are crossed) → implementer → tester (risky changes) → reviewer → close. It never writes code, and every handover is written into the bead, so it survives a /compact. Every agent file gained a “Net & handoffs” section, and the audit agents stay outside the delivery loop: they only file beads, which re-enter at the top. See Agents.
• Architecture rules are now binding (AGENTS.md §2). What was an [EDIT ME] placeholder is a MANDATORY section in every language: layered with inward dependencies, interfaces at every seam, DAO as the only thing that talks to a store, DTOs across process boundaries, domain objects that never leave the domain. §2c makes it stick — a task that does not fit is not implemented as is: the agent stops before writing code, asks in PO language with options and consequences, and the answer is recorded. The reviewer treats each rule as a BLOCKER.
• Model tiers per activity. Instead of “Haiku for five audit agents”, aiflow apply stamps a model per kind of work: reasoning (Opus, or Fable) for architecture, planning, review and security; Sonnet for implementation and tests; Haiku for mechanical scans. Override a whole tier or move a single agent, both surviving aiflow change-settings. See Models.
• Windows without MinGW. aiflow’s CLI runs in PowerShell + Git Bash, but anything that compiles native code belongs in WSL. docs/installation.md gained a Windows prerequisites section, aiflow doctor reports WSL, WSL2, the in-distro toolchain and firmware virtualisation — and warns when a MinGW/MSYS gcc sits on the Windows PATH. install-deps installs build-essential inside WSL and never falls back to MinGW.
• New projects land on main, not master. aiflow init never passed --initial-branch, so on older git the mainline was master — while the whole branching governance references main by name and therefore never applied. init forces main now, and aiflow apply renames an existing master, printing the remote-migration commands.

0.7.0 — queue mode, four CI guards, hardened git hooks

Highlights:

• Queue mode — a closed task no longer ends the session. Beads is treated as an authoritative work queue: after closing an issue the agent refreshes the queue and starts the next task instead of asking what to do. aiflow next ranks the candidates (priority → unblocks-most → continuation of the issue just closed) and exits 3 when nothing is actionable; a Claude Code Stop hook hands the next task back once if the agent forgets. Four documented reasons to stop, and the agent must name the one that applies. Opt out with beads.queueMode or AIFLOW_QUEUE_MODE=off. See Workflows.
• Four mechanical guards in CI, each with its own self-test. The twin guard now looks inside a .sh/.ps1 pair (a step added to one half only is reported), a new guard rejects [ -x ] gates in front of interpreted calls (the exec bit is invisible to bash and Windows drops it), the rendered-copy guard covers the agent/command/skill roster, and the Python CI helpers are finally compiled and linted themselves.
• Git hooks that hold up in practice. pre-commit now judges the staged content rather than your working tree, hooks warn about issue ids that do not exist, path lists are read with core.quotePath=false so non-ASCII filenames stop slipping through, merge and revert commits pass commit-msg, and scopes may be path-shaped (docs(.aiflow)).
• aiflow project-update reaches further and still deletes nothing. It now refreshes .githooks/* and router-config.example.json under the *.bak rule — so a shipped hook improvement finally arrives in existing projects — and reports helpers aiflow no longer ships instead of removing them.
• Runs on macOS’s system bash again. Empty arrays under set -u and mapfile (bash 4+) broke project-update and ollama pull on bash 3.2; a macOS CI job now proves the fix, after asserting the runner really is on 3.2.

0.6.0 — cheap-model audit routing, ponytail, memory-setup skill

Highlights:

• Model routing for audit-only subagents — modelRouting.enabled (on by default) stamps model: haiku into the frontmatter of the 5 subagents that only do mechanical/background checks (docs-sync, test-gap-advisor, dependency-auditor, performance-advisor, onboarder), so they run on Haiku 4.5 instead of the session’s main model — every other subagent keeps the session default since it needs real reasoning. Toggle it with aiflow change-settings.
• ponytail — a new YAGNI decision-ladder skill (off by default): before new code, dependencies, or abstractions, checks whether it needs to exist at all, is already in the codebase, is stdlib, a native platform feature, an installed dependency, or a one-liner — only then writes the minimum-viable new code. /ponytail-review audits a diff for over-engineering regardless of the toggle. Inspired by dietrichgebert/ponytail, reimplemented (not vendored) the same way as caveman.
• memory-setup skill — AGENTS.md’s Memory section shrank to a short toggle + essentials; the full context-routing stack, team preferences, and Ollama routing detail moved to a dedicated, auto-offered skill. REST/Database rules and the Ralph-loop decision (both needed reliably on every relevant task) stay inline in AGENTS.md rather than moving to a skill — skills are Claude-Code-only and pattern-matched, not a guaranteed load.

0.5.1 — reliability fixes for aiflow update and install-deps.ps1

Highlights:

• aiflow update no longer leaves stale files behind after the 0.5.0 OS-split — it now removes the old bin//lib/ before installing the new archive.
• install-deps.ps1’s DeepSeek/uv/bun auto-install path fixed — a PowerShell native-exe argument-passing quirk was silently mangling a jq filter and crashing the installer.

0.5.0 — native PowerShell, no more Git Bash on Windows

Highlights:

• Every core tool script now has a behavior-equivalent PowerShell twin (lib/*.ps1 for all 10 lib/*.sh scripts) — Windows needs no Git Bash or WSL at all anymore, removing the 0.4.1/0.4.2 launcher workarounds entirely.
• OS-scoped release archives — the Windows zip ships only PowerShell scripts, the Linux/macOS tarballs only bash; templates/ still ships both shells to every OS since a generated project may target either regardless of host.
• Fixed a jq default-value bug that could silently flip an explicitly-set false config value back to true, and a PowerShell 5.1 mis-parse of UTF-8 em-dashes in two tool scripts.

0.4.2 — Windows launcher fix (bash.exe resolution)

Highlights:

• aiflow init and every bash-backed subcommand fixed on Windows — plain bash on PATH often resolves to WSL’s stub, which can’t parse Windows-style paths at all. The launcher now resolves Git for Windows’ own bash.exe explicitly.

0.4.1 — Windows launcher fix (path normalization)

Highlights:

• aiflow init fixed on Windows — the PowerShell launcher built AIFLOW_HOME with backslashes, which mangled when interpolated into forward-slash bash paths. Normalized once, up front.

0.4.0 — real multi-agent tooling

Highlights:

• CLIs actually get installed now. aiflow install-deps installs Claude Code, GitHub Copilot CLI, and OpenAI Codex CLI per agents.* (previously only Claude Code was ever installed); aiflow doctor reports all three plus ralph/bun.
• Each agent gets a real token-saving mechanism, not just Claude Code: GitHub Copilot gets the token-optimization guide baked into .github/copilot-instructions.md; OpenAI Codex CLI gets an optional CodexSaver MCP router (codexsaver.enabled) that routes cheap/bounded work to a cheaper worker.
• Ralph loop rebuilt on open-ralph-wiggum — genuinely agent-agnostic (Claude Code/Codex CLI/Copilot CLI via --agent), replacing the old Claude-only hand-rolled loop. Known limitation: completion-promise auto-stop isn’t always reliable — --max-iterations remains the real safety bound.
• Docs swept throughout to reflect the real per-agent picture instead of Claude-only framing.

0.3.1 — hotfix: release-process correctness

Highlights:

• branching.sh no longer loses history when creating a missing local main/develop. Previously it always branched off current HEAD; now it creates the branch from origin/<branch> when that remote ref exists, so a diverged origin/develop is never silently ignored.
• Hotfix-branch detection now recognises GitHub PR-merge commits, not just local git merge ones — both aiflow release’s merge-into-develop step and the pre-push main-restriction guard previously only matched Merge branch 'X' subjects and silently let Merge pull request #N from owner/X commits through unchecked.
• aiflow release falls back to origin/<hotfix-branch> when the local branch was already deleted (e.g. GitHub’s “delete branch on merge” default), so the fix still lands on develop.
• release.yml never publishes a release for an in-progress -SNAPSHOT/-HOTFIX VERSION — it previously tagged and published whatever VERSION said verbatim, which could briefly be a literal -HOTFIX-suffixed string right after a hotfix PR merge and before aiflow release --yes stripped it.

0.3.0 — agent-agnostic core, gitflow automation, Skills, multi-host releases

Highlights:

• Agent-agnostic core — AGENTS.md is the shared source of truth for Claude Code, GitHub Copilot, and OpenAI Codex CLI; CLAUDE.md is now a one-line @AGENTS.md import. Sections only Claude Code can run automatically (subagents, hooks, slash-commands, the Ralph loop) are marked (Claude Code only).
• agents.{claude,copilot,codex} config renders per-agent MCP config from one server set: .mcp.json (Claude), .codex/config.toml (Codex CLI), .vscode/mcp.json (Copilot).
• Gitflow branching automation — bugfix/* alongside feature/*; main restricted to develop/hotfix/*/chore/* (enforced by pre-push); aiflow hotfix <name>; main may never carry a -SNAPSHOT/-HOTFIX version; aiflow release defaults to a dry run — only --yes actually cuts a release.
• Predefined release-publish workflows for GitHub, GitLab, Gitea, Forgejo, and Bitbucket, written once per project (never overwriting an existing one) from release-workflows/.
• GitKraken MCP toggle, independent of the chosen remote host.
• Skills (.claude/skills/<name>/SKILL.md) — a new mechanism distinct from slash-commands: Claude Code matches a skill’s description against the current task and offers to run it automatically. Ships with seo-optimization.
• aiflow update now works on archive installs, not just git checkouts — checks the GitHub Releases API, downloads the matching per-OS archive, and verifies it against SHA256SUMS.txt.
• aiflow project-update now also refreshes agent definitions — any customised file is kept as *.bak before being replaced, and reported so you can reapply your changes; .beads/, .claude/memory/*, and your project settings (aim, architecture) always survive untouched.
• SEO pass on the docs site (robots.txt, 404, favicon, JSON-LD structured data, Google Search Console verification) and refreshed llms.txt/llms-full.txt for context7.

0.2.0 — cross-platform scripts, self-update, no more broken nightly agent

Highlights:

• Removed the nightly aiflow-agent workflow — it ran an unattended Ralph loop on a cron in every scaffolded project and always failed without a configured token. Dropped from the templates; aiflow init no longer generates it.
• Every invoked project script now ships as a .sh + .ps1 pair — hooks (format/caveman/beads-sync), audits (security-check/quality-check/ requirements-check), ralph-headless, run-agent, release/version/protect, bd-close-sync, and docker/run. apply.sh writes the OS-correct interpreter into .claude/settings.json’s hook commands based on dev.os — Windows projects no longer need Git-Bash to run their own hooks/checks.
• aiflow update self-updates the CLI install; aiflow project-update refreshes a single project’s mechanical scripts from the installed templates. Projects now stamp the aiflow version they were generated with and get prompted to upgrade when it falls behind.
• “Built with aiflow” README badge inserted idempotently on apply — visible provenance without ever overwriting an existing README.

0.1.1 — quality gates & senior-engineer agents

Highlights:

• implementer as senior engineer — mandatory pre-analysis (architecture, change impact, effort, complexity) before code; architecture-fit check with targeted refactoring; SOLID/DRY/KISS/YAGNI, testable by design (DI, deterministic); proven open-source frameworks and design patterns over self-implementations; no duplication, reusable/generic solutions; PO-level clarification questions with recorded decisions.
• reviewer as architect + quality gate (one agent) — architecture/design/risk review (layers, module boundaries, tech debt, over-/under-engineering, vulnerabilities, concurrency, breaking changes) plus an objective release checklist; verdict PASS or CHANGES REQUIRED; suggestions persisted as [suggestion] beads for the next loop.
• tester as test/QA engineer — negative/edge/boundary/exception/invalid-input tests plus test-quality audit (assertions, determinism, independence); runs adaptively when the pre-analysis flags high risk/complexity.
• Objective metric targets — 0 % new duplication, no new smells, 0 architecture violations, 0 linter/compiler warnings, 0 high/critical security findings (CLAUDE.md §3a table).
• Production readiness & architecture hygiene — production-ready output only (low-maturity tech flagged by reviewer + tester), small classes/KISS with divide & conquer + interfaces, monolith avoidance, state-of-the-art check (SOAP/XML/legacy-MQ requests are questioned as PO decisions), deliberate Redis/SQLite/Elasticsearch consideration.
• New checkers (on demand, outside the loop) — accessibility-checker (aiflow a11y-check, strict WCAG 2.2 AA → [accessibility] beads + E2E-a11y-tool recommendation) and modernization-advisor (aiflow modernize-check, brownfield modernisation concepts as a report for the architect: microservices, REST/cloud-native, git over svn, supported stacks, missing test frameworks).
• Quality gates (CLAUDE.md §3a) — static analysis on every implementation (tool or the agent itself), > 80 % line coverage + all non-static methods tested, unit + BDD end-to-end always, integration/system where sensible, leveled logging required.
• REST versioned + secured (§3b) — /api/v1/… from day one; OAuth2/OIDC, JWT, or managed API keys — Basic Auth is insufficient; every new/changed endpoint ships an IDE-testable .http file (host/port/test credentials from .env).
• Database rules (§3c) — new models: ≥3NF, real FKs, constraints, precise types, junction tables, no needless surrogate keys (R1–R20). Brownfield: existing schemas handled with care (shared consumers, rollback to older app versions) — improvement potential becomes recommendation beads, never side-effect fixes; commissioned changes need consumer check + rollback plan (B1–B8).
• Ralph loop — the implementer decides automatically from its pre-analysis; manual triggers win (/implement <bead> ralph|no-ralph, or a directive written into the issue).
• --no-token-saving — aiflow init / aiflow change-settings flag that switches caveman + rtk off for full, unfiltered output.
• Project aim — onboarder proposes an aim from brownfield code and asks the user to confirm; aim-writing guidance (where + how) added to READMEs, docs, and the CLAUDE.md template.
• Positioning — aiflow = one strong, universal base config with deliberately generic, customisable agents; ~70–80 % less configuration effort than starting Claude blank.
• Terminal GIFs — install, init Q&A, and change-settings demos in the READMEs and docs.
• New docs — AI Basics (plain-language primer) and the example-project walk-through (all defaults + first feature end-to-end); honest token framing (quality rules spend tokens; first-pass-production-ready saves them).

0.1.0 — first public release

Highlights:

• Setup — aiflow init interactive Q&A → .aiflow/config.json → renders the whole project; change-settings, install-deps, doctor; installer offers git/svn/Ollama.
• Version control & remotes — git / svn / none; token-based GitHub, GitHub Enterprise, GitLab, self-managed GitLab, Bitbucket, Forgejo, Gitea, or custom — with the matching host MCP wired.
• Models — Claude API key or OAuth; Ollama local models (qwen3-coder recommended); model routing.
• Memory — graphify structural graph + cocoindex-code semantic RAG + context7 docs + a retrieval routing policy; aiflow index refreshes both indexes.
• Team — shared Dolt issue graph, session-start auto-pull, atomic claiming, pull-before-push.
• Agents & workflow — delivery + audit + brownfield agents, slash skills, the Ralph loop.
• Quality, git & releases — Google style, Conventional Commits, enforcement hooks, branching models, aiflow release.
• Token savings — caveman + rtk on by default; graph/RAG retrieval; aiflow cost.
• Containers & CI/CD — Podman/Docker headless runs; ci.yml, release.yml, pages.yml.
• Docs & project — extensive README (EN/DE), this documentation site, MIT license, no funding prompts.

See the full list in CHANGELOG.md.