flowai 0.4: packs, multi-agent pipeline, and skill adaptation

https://github.com/korchasa/flowai

TL;DR: Over the past two weeks we moved away from the monolithic architecture — the framework is now split into 6 independent packs. We added a multi-agent SDLC pipeline where 6 sub-agents pass work along a chain via file artifacts. We solved the problem of local edits being overwritten during skill updates: three-way merge is now in place. We also parallelized checks and made whiteboards per-session. Totals: 142 commits, 10 releases (0.3.5 → 0.4.8), 41 skills.

Splitting into packs

The monolithic approach turned out to be inconvenient: installing the framework into a project pulled in unnecessary dependencies (e.g., Deno skills into a Go project). We split the framework into 6 packs. A pack is a directory with a pack.yaml, skills, and agents — similar to Claude Code plugins. You can now install only what you need.

• core — basic skills (commit, review, plan, etc., 14 total)
• engineering — deep-research, tests, QA, browser automation (14 total)
• devtools — skill/agent creation, benchmarks (6 total)
• pipeline — multi-agent pipeline (3 skills + 6 sub-agents)
• deno and typescript — stack-specific tooling

Multi-agent SDLC pipeline

The orchestrator drives a task through a conveyor of 6 sub-agents: PM → Architect → Developer → QA → Tech Lead → Tech Lead Review.

Agents share no memory and know nothing about each other. Coordination happens exclusively through reading and writing artifacts in the .flowai/ folder. Two modes are supported: fully local and via GitHub Issues.

On small tasks (add an endpoint, write a CLI command) the pipeline runs reliably. On larger tasks there are still rough edges: the architect sometimes produces an overly abstract plan, causing the developer to lose context. We will be iterating on this.

Skill adapter (no more lost local edits)

Previously, running flowai sync would overwrite local skill commands (e.g., if the base test invocation had been replaced with pytest -x or ruff check . for the project).

Now the flowai-update command launches a dedicated adapter for each changed skill. It performs a three-way merge: combining the fresh upstream version, the previous local version (from git HEAD), and the project context from AGENTS.md. Updates are applied while project-specific commands are preserved. A diff is shown before applying changes.

Parallel checks

Linters, formatters, and tests (deno task check and flowai-review-and-commit) now run in parallel. To prevent console output from turning into a mess, it is buffered and printed only after each step fully completes.

Per-session whiteboards

The task plan (whiteboard) no longer overwrites a single documents/whiteboard.md file when a new session starts. Each task now gets its own file in the format documents/whiteboards/YYYY-MM-DD-slug.md, making it possible to revisit the history of past plans.

Session history reflection

The new flowai-reflect-by-history skill analyzes transcripts of past IDE sessions (not the current one) and looks for recurring patterns in agent behavior. The skill automatically detects the IDE by project markers and selects analysis depth based on the request — from the single most recent session up to 20.

What it looks for: recurring errors (the same error in 2+ sessions), manual workflows that are candidates for automation, ad-hoc solutions without rules. Each finding is classified by artifact type: Rule, Skill, Hook, Code Change, or Project Docs — and comes with a concrete fix suggestion citing evidence from specific sessions.

Other additions and removals

• New: Browser automation via Playwright, training material generation, hooks for generating IDE configs and pipelines.
• Removed: flowai-skill-executor (skills are now called directly), flowai-skill-ai-skel-ts (extracted into a separate library).

Remaining changes

• CLI: Added --local flag for local development, -y for config generation without prompts, --version with update checks.
• Command logic: Plan critique in flowai-plan is now automatic (removed the extra prompt), added Risks and Trade-offs sections. flowai-commit suggests reflection after execution. Documentation Audit became a mandatory step.
• Infrastructure: The .flow/ directory was renamed to .flowai/. The flowai- prefix is now mandatory for naming.
• Benchmarks: Added 25 new scenarios, introduced static mocks, LLM-judge via Claude CLI, and a ban on test-fitting.
• CI: GitHub Actions are pinned to SHA for security. Semantic versioning for individual skills is being adopted.