Skip to content

changelog

v0.7.1 — 2026-04-20

New env knob: DOER_LOAD_TOOLS_FROM_DIR

  • Controls whether Agent(load_tools_from_directory=...) is True (default, hot-reload ./tools/*.py) or False.
  • Fixes threading race in hf_jobs/gen_dataset.py — Strands' hot-reload watcher isn't thread-safe at construction.
  • Default preserves old behaviour — 1 (on) unless explicitly unset/false.
  • Accepted falsy: 0, false, no, off, "" (case-insensitive). Everything else is true.
  • Use case 1: bulk/concurrent training-script runs — DOER_LOAD_TOOLS_FROM_DIR=0 doer ...
  • Use case 2: tiny HF job containers where ./tools/ doesn't exist and file watching costs RAM.
  • Use case 3: ephemeral/sandboxed agents that should not auto-load user tools.

gen_dataset.py now sets DOER_LOAD_TOOLS_FROM_DIR=0 by default so dataset generation is thread-safe out of the box.

v0.7.0 — 2026-04-20

Cloud dataset generation (HF Jobs)

  • New: doer --hf-jobs gen — generate dense training records on HF infrastructure, auto-append to cagataydev/doer-training
  • New: doer/hf_jobs/gen_dataset.py — single UV script, ~290 lines, thread-safe, dedupe-by-sha256
  • New: doer/hf_jobs/prompts.example.txt — 59 seed prompts covering pipe/shell/coding/self-awareness/agentic/meta/style patterns
  • Input flexibility: local file, hf://user/ds[:column], or stdin (-)
  • Provider matrix: Bedrock (default), Ollama, Anthropic, OpenAI — launcher auto-wires the right secret
  • Provenance: every generated record carries generated_by = "doer --hf-jobs gen @ <job_id>" — filter synthetic vs. human records trivially
  • Idempotent: rerun same prompts, nothing happens (sha256 dedupe against existing dataset)
  • Cost: ~$0.60 per 500 records (Bedrock Opus 4.7 @ concurrency=8)
  • Full loop is now cloud-native: generate → train → deploy, zero laptop involvement

Packaging

  • doer.hf_jobs package-data now includes *.txt so prompts.example.txt ships in the wheel
  • Version bump 0.6.1 → 0.7.0 (minor — new feature, fully backward-compatible)

Migration

No breaking changes. The gen subcommand is additive; existing text / vlm / omni / ps / logs / hw workflows unchanged.

v0.6.1 — 2026-04-20

Packaging polish

  • hf_jobs/ now ships inside the wheel — no more git clone required to use HF Jobs training
  • New doer --hf-jobs CLI shortcut — dispatches to the bundled launch.sh:
  • doer --hf-jobs — prints the bundled directory path
  • doer --hf-jobs text|vlm|omni — launch cloud training (was ./hf_jobs/launch.sh ...)
  • doer --hf-jobs ps|logs|hw — manage running jobs / see hardware pricing
  • Wheel size: 16 KB → 50 KB (still tiny; adds 5 bundled files)
  • Fixes: pipx/conda install conflicts — clean install now works across shells

Migration

  • Old: ./hf_jobs/launch.sh text
  • New: doer --hf-jobs text (or still works from a git clone)

Changelog

Compressed history. Newest first.

v0.6.0 — cloud training (HuggingFace Jobs)

  • hf_jobs/ suite — burn HF credits instead of battery for scale-up training
  • hf_jobs/train_text_lora.py — any causal LM → LoRA → merged push (Qwen3-1.7B default)
  • hf_jobs/train_vlm.py — Qwen2.5-VL-3B image+text LoRA
  • hf_jobs/train_omni.py — Qwen2.5-Omni-7B text+audio+image LoRA
  • hf_jobs/launch.sh — one-shot dispatcher (text/vlm/omni/ps/logs/hw)
  • One file per trainer, inline UV deps — no repo setup, no Dockerfile, hf jobs uv run handles everything
  • Raw JSONL loading via hf_hub_download — bypasses Arrow schema churn on heterogeneous multimodal records
  • Merge + push by default — output is a drop-in for transformers.AutoModelForCausalLM.from_pretrained, no peft glue on the consumer side
  • Tool calls preserved — Strands toolUse/toolResult become native <tool_call>/<tool_result> tags so the chat template lays down real tool-call tokens
  • Validated end-to-end: T4-medium, 522 records → 468/53, 50 steps / 33 min, eval_loss 0.149, token accuracy 97.6%, 3.44 GB merged model auto-pushed
  • Local --train / --train-vlm unchanged — cloud is opt-in, laptop-first stays default
  • New docs: train.md#train-in-the-cloud-huggingface-jobs

v0.5.0 — multimodal + dataset publishing

  • Multimodal input — --img, --audio, --video flags route to mlx-vlm automatically
  • vision-only → Qwen2.5-VL-3B
  • audio-only → gemma-3n-E2B-it
  • mixed (image + audio) → Qwen3-Omni-30B-A3B
  • VLM LoRA training — do --train-vlm [iters] trains on image/audio/video records
  • HuggingFace upload — do --upload-hf / do --upload-hf-public publishes the corpus as an HF dataset (private by default). Idempotent, one atomic commit, reuses huggingface-cli login.
  • --train-status refreshed — shows sha256, modality breakdown (text/image/audio/video), HF sync state
  • Structural refactor — same CLI surface, clearer internals (PR #5). ~730 lines total.
  • New env knobs: DOER_MLX_VLM_MODEL, DOER_MLX_AUDIO_MODEL, DOER_MLX_OMNI_MODEL, DOER_VLM_ADAPTER, DOER_HF_REPO, DOER_CACHE_PROMPT, DOER_BEDROCK_GUARDRAIL_ID/VERSION, DOER_ADDITIONAL_REQUEST_FIELDS
  • New opt-in extras: [vlm] (mlx-vlm + datasets), [hf] (huggingface-hub), [all]

v0.4.0 — closed the loop

  • Self-training — every do "..." call appends a dense, self-contained record to ~/.doer_training.jsonl (full system prompt + messages + tool specs)
  • In-process LoRA via do --train [iters] — calls mlx_lm.tuner directly, no strands-mlx trainer indirection (~50 lines)
  • Native tool-call tokens — _strands_to_openai() preserves tool_calls as structured data so tokenizer chat templates emit real <tool_call> tokens (Qwen/Llama), not string mimicry
  • MLX provider — DOER_PROVIDER=mlx for Apple Silicon on-device inference with LoRA hot-swap via DOER_ADAPTER
  • Corpus inspector — do --train-status shows turn count, KB, path
  • Auto-detect extended — provider order now bedrock → mlx (Apple Silicon) → ollama
  • New env knobs: DOER_MLX_MODEL, DOER_ADAPTER, DOER_DEBUG
  • Opt-in extra: pip install 'doer-cli[mlx]' pulls strands-mlx + mlx-lm (~500MB) — default install stays lean
  • ~420 LOC (up from 221) at the time — one file, one default dep
  • New docs: Train on yourself

v0.3.0 — frontier by default

  • Default model: global.anthropic.claude-opus-4-7 on Bedrock (was Ollama-only)
  • Auto-detect provider — Bedrock if AWS creds exist, else Ollama fallback
  • 1M context window auto-enabled via context-1m-2025-08-07 beta header
  • 128k max output (Opus 4.7 native cap; raise via DOER_MAX_TOKENS)
  • Opt-in temperature / top_p — Opus 4.7+ rejects non-default sampling, so doer skips them unless explicitly set
  • New env knobs: DOER_PROVIDER, DOER_BEDROCK_MODEL, DOER_BEDROCK_REGION, DOER_ANTHROPIC_BETA, DOER_ADDITIONAL_REQUEST_FIELDS
  • 221 LOC (up from 164) — still one file, still one dep

v0.2.1 — curl or pipx

  • do shortcut alongside doer (less typing)
  • One-line installer (curl | sh) planned via GitHub Releases
  • Renamed to doer-cli on PyPI (doer was squatted)
  • Repo moved to github.com/cagataycali/doer-cli
  • Docs: migrated to mkdocs-material (mobile-first, proper nav, dark/light, cookbook)

v0.2.0 — new brand

  • Bold, solid, pipe-first identity (orange #FF3D00 + black + paper)
  • Custom SVG logo
  • Clean README, stripped marketing copy
  • Auto-inject SOUL.md + AGENTS.md into system prompt

v0.1.x — the primordial soup

  • 164 LOC — fits on one screen (barely)
  • Only dep: strands-agents
  • Ollama-only (local, private, no keys)
  • Injects own source, $HOME/.bash_history, $HOME/.zsh_history, ~/.doer_history
  • Hot-reload tools from ./tools/*.py
  • PyInstaller + Nuitka standalone binaries (linux/macos)
  • Rename: tiny → doer (better verb)

pre-history

  • Spawned from DevDuck — 60+ tools, every protocol
  • DevDuck asked itself at 4am: what if we deleted almost everything?
  • Two hours later: doer.

the cathedral teaches you which stones are load-bearing.