The AI Toolchain — issue -016, August 3, 2026

DSPy 3.3.0 adds Flex structure-optimizing programs, ReActV2 with native tool calling, and a typed provider-neutral LM boundary.

$ git clone --branch 3.3.0 https://github.com/stanfordnlp/dspy.git
# already have the repo? check out this version:
$ git checkout 3.3.0

import dspy

program = dspy.Flex("question -> answer")
optimized = dspy.GEPA(metric=metric, reflection_lm=reflection_lm).compile(
    program,
    trainset=trainset,
    valset=valset,
)

print(optimized.module_src)
optimized.save("flex_qa.json")

import dspy

try:
    result = lm(messages=[{"role": "user", "content": "Hello"}])
except dspy.LMError as e:
    print(f"LM call failed: {e}")

• ›Adds dspy.Flex, an experimental module that places program structure — predictors, control flow, DSPy primitives, and Python/LM balance — into the GEPA search space so the optimizer discovers decomposition instead of only tuning prompts; defaults to a single dspy.Predict baseline, or dspy.RLM when tools are supplied.
• ›Adds max_predictor_calls guard on Flex-generated programs to prevent runaway LM usage in optimizer-authored code, and supports a program_trace argument to metrics so programs can be scored on how a result was produced (e.g. penalizing excessive LM calls).
• ›Persists optimizer-discovered module_src as part of a Flex program's serialized state, so dump_state() / load_state() round-trips preserve the GEPA-authored implementation.
• ›Adds dspy.ReActV2, an experimental ReAct implementation built on native tool calling, using dspy.History, dspy.Tool, and dspy.ToolCalls (which can optionally store dspy.ToolCallResults) instead of custom next_tool_args / trajectory syntax.
• ›Adds parallel_tool_calls support to dspy.ReActV2, preserving each call/result pair by ID in both native and non-native mode.

The agent that grows with you

Hermes v2026.8.3 adds conversational voice with barge-in, A2A v1.0, signed webhooks, grounded citations, a plugin SDK, and a CLI power-user wave.

$ git clone --branch v2026.8.3 https://github.com/NousResearch/hermes-agent.git
# already have the repo? check out this version:
$ git checkout v2026.8.3

$ !git status

$ hermes /init

$ hermes import-agent

• ›Adds !command shell-escape mode to run a shell command from the CLI instantly without spending a model turn.
• ›Adds /init CLI command to scan a project and generate or update an AGENTS.md file.
• ›Adds /diff CLI command to show staged, all, or session-level changes from any surface.
• ›Adds /context CLI command to break down exactly what is filling the current context window.
• ›Adds /focus CLI command for a reduced-output view with hidden-line recovery.

GitHub Copilot CLI v1.0.78 adds /new-worktree and /permissions commands, ACP token usage exposure, and a new allowDevToolCaches sandbox setting.

$ git clone --branch v1.0.78 https://github.com/github/copilot-cli.git
# already have the repo? check out this version:
$ git checkout v1.0.78

$ /new-worktree

$ /permissions

• ›Adds /new-worktree (experimental) command to create a new git worktree and start a fresh conversation inside it.
• ›Adds /permissions command to switch between approval modes mid-session.
• ›Adds allowDevToolCaches sandbox setting (on by default) to grant sandboxed builds access to toolchain caches, registries, and installs; set to false to opt out.
• ›Adds forceRemoteSettingsRefresh managed setting to require a fresh managed-settings fetch on every startup.
• ›Adds /settings showToolDurations to control display of per-tool-call elapsed time in timeline headers (on by default, live-ticking for calls of at least 5 seconds).

OpenHands: AI-Driven Development

OpenHands v1.9.0 adds live agent activity in chat, a streamlined backend chooser, and an extension-manifest host.

$ git clone --branch v1.9.0 https://github.com/All-Hands-AI/OpenHands.git
# already have the repo? check out this version:
$ git checkout v1.9.0

• ›Adds a compact Cloud vs Agent-server add-backend chooser to streamline switching between backend types.
• ›Shows live agent activity directly in the chat UI so users can follow along in real time.
• ›Introduces a domain-neutral extension-manifest host to support broader extension compatibility.

$ git clone --branch v1.9.0 https://github.com/OpenHands/OpenHands.git
# already have the repo? check out this version:
$ git checkout v1.9.0

Qwen Code v0.21.4 adds new review/drive commands, configurable memory agents, session workflow toggles, and a desktop Web Shell.

$ git clone --branch v0.21.4 https://github.com/QwenLM/qwen-code.git
# already have the repo? check out this version:
$ git checkout v0.21.4

# In your Qwen Code config
memory:
  agentMaxTurns: 20

# In your Qwen Code config
experimental:
  sessionWorkflow: true

$ qwen review drive

• ›Adds experimental.sessionWorkflow setting to optionally enable Session Workflow features including Plan & Review mode and the Workflow DAG.
• ›Adds memory.agentMaxTurns setting to configure turn limits for all managed memory agents; set to 0 to disable the limit entirely.
• ›Adds qwen review drive command to poll for service readiness and verify completion facts instead of relying on fixed sleep delays.
• ›Adds qwen review mock-provider command to record OpenAI-compatible requests as JSONL for testing against a faithful outside-world simulation.
• ›Adds qwen review publish-assets command to host evidence images in a user-designated repository for embedding in PR review comments.

llama.cpp b10242 adds GPU-accelerated CUDA backend sampler for penalty handling including frequency, presence, and top-k penalties.

$ git clone --branch b10242 https://github.com/ggml-org/llama.cpp.git
# already have the repo? check out this version:
$ git checkout b10242

• ›Adds CUDA backend sampler for llama_sampler_penalties, offloading frequency, presence, and repeat penalty computation to GPU via the llama_sampler_backend infrastructure.
• ›Adds llama_n_ctx parameter to common_sampler_init so penalty_last_n defaults to the model context length when not explicitly set.
• ›Adds support for top-k penalties in backend sampling, with configurable positions in the sampler chain.
• ›Adds vocabulary-sized count tensor to replace the per-candidate penalty comparison matrix, improving scalability for large history windows.
• ›Adds validation that repeat penalty is finite and greater than 0, and preserves masked logits as -Inf to eliminate NaN generation.

$ git clone --branch b10238 https://github.com/ggml-org/llama.cpp.git
# already have the repo? check out this version:
$ git checkout b10238

$ git clone --branch b10237 https://github.com/ggml-org/llama.cpp.git
# already have the repo? check out this version:
$ git checkout b10237

Phoenix v19.14.0 adds a built-in hallucination evaluator to the evals library.

$ git clone --branch arize-phoenix-v19.14.0 https://github.com/Arize-ai/phoenix.git
# already have the repo? check out this version:
$ git checkout arize-phoenix-v19.14.0

• ›Adds a hallucination evaluator to the evals module for detecting hallucinated content in LLM outputs.

Langfuse v4.3.0 exposes semantic roots in the v2 API and adds media size display in trace previews.

$ git clone --branch v4.3.0 https://github.com/langfuse/langfuse.git
# already have the repo? check out this version:
$ git checkout v4.3.0

• ›Exposes semantic roots via the v2 API, enabling richer trace hierarchy data for downstream consumers.
• ›Shows media size in the trace/generation preview UI, giving practitioners visibility into payload footprint at a glance.