Skip to content

Changelog

v0.2.0 (2026-04-29)

Features

  • Bob Shell runner — Alternative CLI backend via --runner bob or FACTORY_RUNNER=bob. Protocol-based runner abstraction (factory/runners/) with dry-run mode (FACTORY_BOB_DRY_RUN=1), per-cycle and per-day invocation ceilings, auth persistence for nested subagents, and streaming output with role-prefixed lines
  • CEO completion guard — Auto-resumes when the CEO exits before all planned work is complete. Cycle state persists in .factory/state/cycle.json across respawns, with cross-cycle scoping (timestamps in results.tsv) to prevent stale experiment contamination. Configurable max respawns via FACTORY_CEO_MAX_RESPAWNS, 24-hour staleness threshold, and budget-aware re-spawn gating
  • Interactive ideation mode — factory ceo "idea" --mode interactive launches a research → brainstorm → refine loop before any code is written. The new Distiller agent synthesizes research and user feedback into a structured project spec (idea.md). Up to 5 refinement iterations with targeted follow-up research
  • Focused mode — --focus "target" pins a single backlog item and scopes the entire pipeline: one item, one hypothesis, one experiment, done. Requires improve mode, mutually exclusive with --loop
  • Unified backlog — Replaces the old deferred-items system with .factory/strategy/backlog.md. The Strategist clears backlog items each cycle with convergence tracking — new items are capped, backlog must shrink not grow. CLI: factory backlog-list, factory backlog-add, factory backlog-remove
  • Session summaries — End-of-cycle reports via factory summary: what was built (kept experiments with score deltas), what was deferred, what needs human input. Written to .factory/reviews/session-summary.md
  • Experiment checkpoint/resume — Per-experiment CEO state saved via factory checkpoint for crash-resilient recovery. Includes completed agents, pending agents, hypothesis state, and completed experiment IDs
  • Auto-discovery — Managed projects auto-detected from projects directory. factory study scans sibling projects for cross-project insights
  • Citation backfill — Research grounding scores now extract and backfill citations from experiment history for more accurate research_grounding dimension scoring
  • Model override — --model flag and FACTORY_MODEL env var for controlling which model agent subprocesses use

Fixes

  • CEO no longer auto-merges PRs — Leaves merge for human review, matching the no-merge policy (#125)
  • Guard check ignores auto-generated lock files — uv.lock, package-lock.json, etc. no longer trigger scope violations
  • Interactive mode path resolution — Idea strings are no longer treated as file paths (#125)
  • Reviewer prompt — Corrected from "merge" to "approve PR" to match the no-merge policy
  • Discover auto-chains — Discovery mode proceeds directly to improve mode instead of stopping
  • Deferred item persistence — Deferred items survive strategy rewrites across cycles
  • Calendar-time estimate stripping — CEO gate rejects agent outputs containing time estimates like "8-10 weeks"

Quality

  • 1144 tests — Up from 878 at v0.1.0 (30% increase)
  • Codecov — Coverage reporting integrated into CI

v0.1.1 (2026-04-24)

Fixes

  • Vault path: Removed all hardcoded vault path references — vault path is now resolved exclusively via $FACTORY_VAULT_PATH env var
  • Mypy: Fixed type errors in factory/eval/hygiene.py and factory/dashboard/app.py
  • CI-safe tests: test_rewards_vault_sources and related growth tests no longer depend on local filesystem state

Docs & CI

  • GitHub Actions CI — pytest (3.11/3.12/3.13 matrix), ruff, mypy; runs on PRs only
  • MkDocs Material — hosted docs at akashgit.github.io/remote-factory, auto-deployed on push
  • Mermaid diagrams — README uses native GitHub-rendered Mermaid instead of external SVGs
  • Self-evolving messaging — README title and intro emphasize the factory's learning loop
  • Obsidian recommendation — docs highlight vault setup for persistent cross-project learning

v0.1.0 (2026-04-24)

Initial public release.

Core

  • CEO Agent — dedicated orchestrator with 5-state machine (no_repo, incomplete, no_factory, evals_pending_review, has_factory), automatic mode routing, and mandatory archival
  • 7 Specialist Agents — Researcher, Strategist, Builder, Reviewer, Evaluator, Archivist, each running as independent Claude Code subprocesses
  • Experiment Loop — every change is a hypothesis: measured before/after, kept or reverted based on composite eval score
  • Universal Input — accepts directories, GitHub URLs, idea file paths, or raw text prompts

Eval System

  • Three-tier composite scoring — hygiene (6 dimensions), growth (5 dimensions), and user-defined project eval
  • Configurable weight distribution — default 50/50, shifts to 30/20/50 with project eval
  • Hard precheck gate — 4 non-overridable checks (score direction, scope, anti-pattern, smoke test)
  • Guard rules — scope enforcement and eval immutability

Strategy

  • FEEC priority — Fix > Exploit > Explore > Combine hypothesis ranking
  • Stuck detection — forces category rotation after 3+ consecutive same-category reverts
  • Structured hypothesis budget — reserved fix/growth/flex slots, configurable per-run

Self-Improvement

  • ACE (Autonomous Context Engineering) — Reflect/Curate/Inject loop that evolves agent playbooks from real experiment outcomes
  • Cross-project learning — patterns from one project inform behavior on others
  • Helpful/harmful counters — evidence-based rule reinforcement and pruning

Operations

  • Live dashboard — FastAPI server with SSE-powered real-time UI (port 8420)
  • Continuous mode — heartbeat loop with configurable interval and max cycles
  • tmux integration — detached sessions that survive SSH disconnects
  • Crash resilience — CEO checkpoint save/load for resume after failures
  • Structured PR reviews — score tables, guard results, and code notes posted on GitHub PRs

Integrations

  • Obsidian vault — optional archival of experiment history and cross-project knowledge
  • MCP servers — Playwright for UI testing, extensible per-project
  • Claude Code agent registration — factory install for seamless integration

CLI

30+ subcommands including: ceo, run, agent, eval, precheck, guard, begin, finalize, history, diff, explain, study, insights, ace, dashboard, detect, discover, export, checkpoint, resume, tmux, digest, archive, notify, review, install, vault-init, self-update.

Quality

  • 878 tests with pytest
  • Type checking with mypy
  • Linting with ruff
  • Strict Pydantic v2 models throughout