TobyOnFitnessTech
←All Episodes
OpenClaw 2026.5.28, MiniMax M3, Claude Code 2.1.159, and Code Graphs — Episode 61 cover art
Episode 61·June 2, 2026·39:24

OpenClaw 2026.5.28, MiniMax M3, Claude Code 2.1.159, and Code Graphs

OpenClaw `v2026.5.28` leads with steadier agent and Codex runtime recovery, safer channel delivery, stricter browser and automation inputs, provider and media expansion, externalized plugin surfaces, and bounded release proof. MiniMax M3 follows as the model-discovery headline: an open-weight-positioned coding and agent model with MSA sparse attention, up to 1M context, native multimodality, API availability, MiniMax Code integration, and open weights promised after the launch window. Claude Code `2.1.159` is the quiet CLI hygiene lane. Then EP061 evaluates Understand Anything, agentgateway, MCPJungle, and CodeAlmanac as practical tools for codebase graphs, agent traffic control, MCP management, and persistent repo context. Show notes: https://tobyonfitnesstech.com/podcasts/episode-61/

🎧 Listen to Episode

Listen & Subscribe

SpotifyAppleiHeartYouTube MusicAmazonRSS

AgentStack Daily EP061 - OpenClaw 2026.5.28, MiniMax M3, Claude Code 2.1.159, and Code Graphs

Release Coverage Check

  • OpenClaw - Latest stable verified: v2026.5.28, published 2026-05-30T20:06:10Z from GitHub releases. Recent episode version tags detected: v2026.5.27, v2026.5.26, v2026.5.22, v2026.5.20, and prerelease text for v2026.5.28-beta.4. Selected missing versions: v2026.5.28. Candidate verification: stable release list begins v2026.5.28, v2026.5.27, v2026.5.26; the first matching stable tag in the recent notes is v2026.5.27, so v2026.5.28 is the selected block.
  • Hermes Agent - Latest stable verified: v2026.5.29.2 / v0.15.2, published 2026-05-29T13:37:26Z from GitHub releases. Recent episode version tags detected: v2026.5.29.2, v2026.5.29, v2026.5.28, and v2026.5.16. Selected missing versions: none. Candidate verification: the latest stable tag is present in the recent notes.
  • OpenAI Codex app/CLI - Latest stable verified: rust-v0.135.0, published 2026-05-28T17:31:35Z from GitHub releases. Recent episode version tags detected: rust-v0.135.0, rust-v0.134.0, and rust-v0.133.0. Selected missing versions: none. Candidate verification: the latest stable tag is present in the recent notes.
  • Claude Code CLI - Latest npm latest verified: 2.1.159, published 2026-05-31T17:50:06Z in npm metadata and documented in Anthropic's Claude Code changelog at 2026-05-31T19:47:03Z. npm stable dist-tag verified as 2.1.150; this draft uses npm latest because the recent episode baseline used npm latest. Recent episode version tags detected: 2.1.158, 2.1.157, 2.1.156, 2.1.154, 2.1.153, and 2.1.152. Selected missing versions from npm latest: 2.1.159. Candidate verification: the latest sequence begins 2.1.159, then 2.1.158; 2.1.158 is present in the recent notes.

Model Discovery Check

  • MiniMax M3 - Primary sources: https://www.minimax.io/blog/minimax-m3 and https://www.minimax.io/models/text/m3. Release date verified as 2026-06-01 from MiniMax's official announcement. Availability: API is live with model id MiniMax-M3; MiniMax Code and Token Plan access are live; MiniMax says the technical report and open weights will follow within 10 days. Core capabilities: open-weight positioning, frontier coding and agentic work, MiniMax Sparse Attention (MSA), up to 1M context with a guaranteed minimum of 512K tokens, native multimodality from training step zero, image/video input, computer-use capability through MiniMax Code, and coding-agent benchmark claims including SWE-Bench Pro, Terminal-Bench, MCP Atlas, BrowseComp, and long-horizon CUDA/kernel and paper-reproduction tasks. Caveat: many benchmark numbers are MiniMax-run or use specified scaffolds such as Claude Code/Codex, so the episode should treat them as important claims to test, not settled independent truth. Subscription relevance: fixed Token Plan tiers are advertised at Plus $20, Max $50, and Ultra $120 with shared usage pools. Decision: selected for Story Slate as the biggest model drop in the scan.

Real-World Reaction Check

Runtime Target

45-55 min. Release-heavy episode led by OpenClaw v2026.5.28, with MiniMax M3 as the major model-discovery story, Claude Code 2.1.159 folded into the release readout, and stack-adjacent code graph / MCP tooling after it.

Episode Title

OpenClaw 2026.5.28, MiniMax M3, Claude Code 2.1.159, and Code Graphs

Tagline

OpenClaw ships a stable recovery and hardening release, MiniMax M3 brings a new open-weight long-context model into the coding-agent race, Claude Code lands a quiet infrastructure update, and the project radar follows graph and MCP tools that make agents easier to aim.

Feed Description

OpenClaw v2026.5.28 leads with steadier agent and Codex runtime recovery, safer channel delivery, stricter browser and automation inputs, provider and media expansion, externalized plugin surfaces, and bounded release proof. MiniMax M3 follows as the model-discovery headline: an open-weight-positioned coding and agent model with MSA sparse attention, up to 1M context, native multimodality, API availability, MiniMax Code integration, and open weights promised after the launch window. Claude Code 2.1.159 is the quiet CLI hygiene lane. Then EP061 evaluates Understand Anything, agentgateway, MCPJungle, and CodeAlmanac as practical tools for codebase graphs, agent traffic control, MCP management, and persistent repo context.

Story Slate

1. OpenClaw v2026.5.28 hardens recovery, channels, providers, and release proof

OpenClaw v2026.5.28 is the lead release: agent and Codex runtime recovery, channel identity, mobile and realtime Talk updates, stricter browser and automation input validation, provider and media expansion, GitHub Copilot and Codex Supervisor surfaces, plugin hot-path cleanup, and bounded CI/release evidence. Claude Code 2.1.159 is the adjacent CLI lane: npm latest moved to 2.1.159, and Anthropic's changelog lists internal infrastructure improvements with no user-facing changes. Technical depth angle: explain session-lock survival, subagent cwd/workspace separation, prompt-local hook context, channel identity keys, malformed input rejection, provider timeout bounds, plugin metadata caching, Codex Supervisor boundaries, npm latest versus stable, and release-validation evidence. Actionability angle: explain who should upgrade carefully, who should hold, and why the Codex/plugin install path and channel/runtime surfaces matter more than a green "latest version" label. Listener hook: this is a recovery release that some operators will welcome and others may experience as another OpenClaw upgrade surprise. Primary links: https://github.com/openclaw/openclaw/releases/tag/v2026.5.28, https://code.claude.com/docs/en/changelog#2-1-159, https://www.npmjs.com/package/@anthropic-ai/claude-code

2. MiniMax M3 brings open-weight ambition to coding agents, 1M context, and native multimodality

MiniMax M3 is the model-discovery headline: MiniMax officially released it on June 1, 2026 as an open-weight-positioned model that combines frontier coding and agentic work, a 1M-context MSA sparse-attention architecture, and native multimodality. The API is live under MiniMax-M3, MiniMax Code has been updated around it, and MiniMax says the technical report and model weights will follow within 10 days. Technical depth angle: explain MSA sparse attention versus full attention, 1M context with a 512K guaranteed minimum, native multimodal training from step zero, coding-agent scaffolding, MiniMax Code's agent-team loop, benchmark methodology caveats, API thinking-mode/service-tier knobs, and why open-weight availability would matter for local/private agent stacks. Actionability angle: explain where M3 looks genuinely better or cheaper in early use, where people still prefer Opus-class planning, and why pending open weights/independent benchmarks matter. Listener hook: this is the rare model launch where the community is asking whether cheap million-token context changes real routing decisions, not just leaderboard screenshots. Primary links: https://www.minimax.io/blog/minimax-m3, https://www.minimax.io/models/text/m3

3. Understand Anything turns a repo into an explorable graph for agents

Understand Anything turns code into an interactive knowledge graph that humans and agents can explore, search, and question. Its GitHub metadata shows strong velocity and direct adjacency to Claude Code, Codex, Cursor, Copilot, Gemini CLI, and codebase-understanding workflows. Technical depth angle: explain repo graph construction, symbol and relationship browsing, natural-language graph queries, context compression, graph evidence versus raw grep, and how coding agents should use orientation before patch generation. Actionability angle: describe why developers are pairing graph orientation with coding agents and where graph evidence beats blind repo stuffing. Listener hook: the fastest coding agent is often the one that stops reading the wrong files. Primary link: https://github.com/Lum1104/Understand-Anything

4. agentgateway puts MCP and agent traffic behind a control boundary

agentgateway describes itself as a next-generation agentic proxy for AI agents and MCP servers. The timely angle is tool-call governance: agents now need routing, identity, observability, and failure isolation before they touch shared internal tools. Technical depth angle: explain MCP gatewaying, agent proxy boundaries, routing policy, tool-call admission, observability hooks, failure isolation, and blast-radius control when Codex, Claude Code, Hermes, or OpenClaw share private tools. Actionability angle: explain what changes when MCP/tool access moves from scattered client configs into a gateway/control-plane layer. Listener hook: a gateway turns tool access from a pile of client configs into an operator surface. Primary link: https://github.com/agentgateway/agentgateway

5. MCPJungle manages the server sprawl that agents now create

MCPJungle gives builders one place to manage and connect to MCP servers. That matters because MCP adoption is moving from one server per demo to many servers across many clients, where config drift and invisible failures become real operator problems. Technical depth angle: explain MCP server inventory, connection brokering, config drift, tool discovery, per-client setup friction, failure visibility, and why a central manager can reduce duplicated setup across local and hosted agents. Actionability angle: explain why MCP management starts mattering once a team has more than one agent client and more than one private tool server. Listener hook: MCP is useful, but unmanaged MCP becomes another thing agents can quietly misconfigure. Primary link: https://github.com/mcpjungle/MCPJungle

6. CodeAlmanac and Argyph keep repo context close to the source

CodeAlmanac is a codebase wiki for AI coding agents, focused on decisions, flows, invariants, and gotchas. Argyph is a local-first MCP server for structured semantic context over a codebase, making this a two-part context story: durable human knowledge plus local retrieval that still has to be checked against exact files. Technical depth angle: explain codebase-wiki structure, invariant capture, stale-context risk, local-first MCP retrieval, semantic context responses, privacy boundaries, latency tradeoffs, and why source/test verification remains the final proof. Actionability angle: explain the lived failure mode these tools address: agents miss project memory, then make confident edits that violate decisions the source code does not explain. Listener hook: agents need more than syntax; they need project memory and local context without forgetting that source remains the ground truth. Primary links: https://github.com/AlmanacCode/codealmanac, https://github.com/ezzy1630/Argyph

GitHub Project Radar

  • Understand Anything - https://github.com/Lum1104/Understand-Anything - Interactive knowledge graphs for codebases. GitHub API metadata verified 48,236 stars, 3,919 forks, MIT license, pushed 2026-06-01. Primary stack/tool adjacency: Claude Code, Codex, Cursor, Copilot, Gemini CLI, and codebase-understanding workflows. Stack improvement angle: give coding agents a navigable project graph before they spend context on broad file scans. Try now: run it on a repo with confusing entry points, ask for the request path, then compare the agent plan before and after graph orientation.
  • agentgateway - https://github.com/agentgateway/agentgateway - Agentic proxy for AI agents and MCP servers. GitHub API metadata verified 2,952 stars, 497 forks, Apache-2.0 license, pushed 2026-05-29. Primary stack/tool adjacency: MCP gateways, agent traffic control, self-hosted infrastructure, and shared tool surfaces. Stack improvement angle: put policy, routing, and observability in front of tool calls instead of trusting every agent client to wire private MCP servers correctly. Try now: place one read-only MCP server behind the gateway and inspect discovery, failures, and logs.
  • MCPJungle - https://github.com/mcpjungle/MCPJungle - One place to manage and connect to MCP servers. GitHub API metadata verified 1,072 stars, 140 forks, MPL-2.0 license, pushed 2026-05-20. Primary stack/tool adjacency: MCP inventory, local agent setup, tool discovery, and multi-client configuration. Stack improvement angle: reduce MCP config sprawl when several coding agents need the same private and public tools. Try now: inventory servers used by two clients, move one shared server into MCPJungle, and compare setup friction.
  • CodeAlmanac - https://github.com/AlmanacCode/codealmanac - Codebase wiki for AI coding agents. GitHub API metadata verified 114 stars, 6 forks, Apache-2.0 license, pushed 2026-05-31. Primary stack/tool adjacency: repo memory, decision records, invariant capture, and agent context. Stack improvement angle: preserve the why behind a system so agents do not infer design intent from source alone. Try now: document one invariant and one gotcha, then ask a fresh agent session to explain a risky change using that wiki context.
  • Argyph - https://github.com/ezzy1630/Argyph - Local-first MCP server for structured semantic codebase context. GitHub API metadata verified 16 stars, 3 forks, Apache-2.0 license, pushed 2026-06-01. Primary stack/tool adjacency: local MCP, semantic retrieval, and privacy-preserving repo understanding. Stack improvement angle: expose fast local semantic context to agents without sending a repository to a hosted indexing service. Try now: connect it to a disposable repo and compare retrieved context against exact rg and file-read evidence.

Extra Research Candidates

  • AWS Agent Toolkit for AWS - Primary source: https://aws.amazon.com/about-aws/whats-new/2026/05/agent-toolkit/. Technical depth angle: explain managed MCP, IAM guardrails, CloudWatch and CloudTrail observability, sandboxed multi-step operations, and validated skills for infrastructure, analytics, serverless, containers, and AI services.
  • Claude Opus 4.8 system card and Messages API update - Primary source: https://www.anthropic.com/news/claude-opus-4-8. Technical depth angle: compare effort controls, fast-mode pricing, tool efficiency, system entries inside message arrays, evaluation claims, and how harnesses should separate runtime policy from user intent.
  • OpenAI Codex remote access and tokens - Primary source: https://openai.com/index/work-with-codex-from-anywhere/. Technical depth angle: explain secure relay behavior, host-local project state, scoped access tokens, hooks, mobile supervision, and the boundary between remote steering and local execution.

Show Notes

OpenClaw `v2026.5.28` is the lead release for EP061, and Claude Code `2.1.159` is the adjacent CLI release lane. OpenClaw hardens agent and Codex runtime recovery, channel delivery, input validation, provider and media paths, plugin surfaces, and release validation proof. Claude Code `2.1.159` moves npm `latest` forward with internal infrastructure improvements and no user-facing change notes. The big model-discovery story is MiniMax M3: a June 1 launch that combines MSA sparse attention, up to 1M context, native multimodality, coding-agent benchmark claims, a live API, and MiniMax Code integration, with technical report and open weights promised after launch. After that, the episode moves into code graphs, agent gateways, MCP management, repo wikis, and local semantic context.

[00:00] Opening hook: releases and model drops first
OpenClaw `v2026.5.28` changes how resilient long agent runs should feel: fewer stale continuations, cleaner timeout recovery, safer channel identity, stricter browser and automation inputs, and more bounded provider failures. Claude Code `2.1.159` is a quiet npm `latest` point release for infrastructure hygiene rather than a feature launch. MiniMax M3 is the model drop that should not be missed: an open-weight-positioned coding and agent model with 1M context, native multimodality, and a sparse-attention architecture aimed directly at long-running coding agents.

[03:00] OpenClaw v2026.5.28 release deep dive
OpenClaw `v2026.5.28` is a concrete operator release. The agent runtime changes focus on failure recovery: subagents keep cwd and workspace separation, hook context stays prompt-local, session locks release on timeout abort while live OpenClaw-owned locks survive cleanup, stale restart continuations are avoided, and Codex app-server or helper failures should no longer tear down shared runtime state. That is the difference between a long task that collapses into ambiguous state and one that can be resumed, debugged, or retried with a cleaner boundary.

The channel layer also moves. The release tightens delivery and identity across outbound hooks, Matrix room IDs, iMessage reactions and approvals, Slack final replies, Discord recovered tool warnings, runtime-config message actions, WhatsApp profile auth roots, Telegram polling, and Microsoft Teams service URL trust checks. The practical outcome is that approval, progress, and final messages have a better chance of landing in the intended session with the intended runtime config. That matters for agents because the wrong channel identity can turn a safe approval path into a confusing or risky one.

Input validation is another major surface. Browser tab indexes, viewport resizes, Gateway ports, cron retry handling, Discord component IDs, schema array refs, Telegram callback pages, geolocation options, screenshot timeouts, cookie expiries, and response-body limits are rejected earlier. Builders should expect sharper errors instead of silent drift. Provider and media paths are also more bounded: OAuth and token lifetimes, local service probes, generated media downloads, model requests, and provider auth checks should fail with proof instead of hanging a run.

The release also expands usable agent surfaces: Claude Opus 4.8 support, Fal Krea image schemas, NVIDIA featured model catalogs, MiniMax streaming music responses, encrypted PDF extraction, voice model catalogs, GitHub Copilot agent runtime support, and a Codex Supervisor plugin path. Migration risk is mostly around integrations that were accidentally relying on loose input parsing, stale channel state, or unbounded provider waits. The upgrade test should include one subagent, one channel approval, one browser action, and one provider/media call before production rollout.

[16:00] Claude Code 2.1.159 as fleet hygiene
Claude Code `2.1.159` is verified from npm `latest` and Anthropic's official changelog. The changelog entry says internal infrastructure improvements and no user-facing changes. That makes it a small release, but not a meaningless one. If a team follows npm `latest`, `2.1.159` is the current package. If a team pins Anthropic's `stable` dist-tag, the baseline is different; npm metadata during the scan showed `stable` at `2.1.150`.

The practical move is simple: update low-risk local installs normally, but do not promise a new CLI workflow from `2.1.159` alone. Use it to keep fleet state clean and reserve feature expectations for the nearby releases that changed visible behavior, such as managed-cloud auto mode and the plugin and background-agent improvements.

[20:00] MiniMax M3: sparse attention, 1M context, and coding-agent ambition
MiniMax M3 is the most important model discovery in this episode. MiniMax describes it as the first open-weight model to bring together frontier coding, million-token context, and native multimodality. The key technical idea is MSA, MiniMax Sparse Attention. Instead of paying full quadratic attention cost across very long contexts, MSA partitions the key-value cache into blocks and routes attention more selectively, with MiniMax claiming large prefill and decoding speedups at 1M context compared with its previous generation.

That context size matters for agent work because coding sessions are no longer one-shot prompts. A real coding agent may carry repository files, terminal output, test failures, screenshots, design notes, and multi-round user corrections. MiniMax is explicitly training and evaluating for that world: long-horizon coding, terminal tasks, MCP-style agent benchmarks, paper reproduction, CUDA kernel optimization, and interactive developer collaboration. The strongest claims need independent testing because many reported numbers are MiniMax-run and use particular scaffolds, including Claude Code or Codex in some evaluations. But the claim itself is exactly in the zone this show should track.

The launch is also practical, not just a teaser. The API is live as `MiniMax-M3`, MiniMax Code has been updated around M3, and Token Plan users can use the new model through fixed subscription tiers. MiniMax says the technical report and open weights will follow within 10 days. The immediate recommendation is to test it on three workloads: one long-context repo-understanding task, one coding-agent task with tools and tests, and one multimodal or computer-use task. If the weights land as promised and the long-context behavior holds up outside MiniMax's harness, this becomes a serious candidate for private and self-hosted agent stacks.

[28:00] Understand Anything: graph orientation before edits
Understand Anything turns a codebase into an interactive knowledge graph. For a coding agent, the benefit is not a prettier diagram; it is a better starting point. A graph can show entry points, relationships, module boundaries, and likely paths through the system before the model burns context on broad text search.

The right use is graph first, exact files second. Ask the graph where a request or behavior flows, then make the agent open the concrete files, tests, and configs before patching. That pattern is especially useful in repos with same-name functions, generated files, hidden route tables, or service boundaries that grep alone can blur.

[34:00] agentgateway and MCPJungle: controlling tool sprawl
agentgateway and MCPJungle both respond to the same pressure: MCP and agent tools are becoming operational infrastructure. agentgateway puts agent and MCP traffic behind a proxy boundary where routing, policy, observability, and failure isolation can live. MCPJungle focuses on managing and connecting MCP servers from one place.

The failure modes are practical. One client has an old server path. Another has the wrong token. A third exposes a tool that should be read-only. A fourth has no useful logs when a tool call fails. Start with a read-only MCP server behind a gateway or manager, then check what the agent sees, what gets logged, how errors are reported, and how quickly the tool can be disabled.

[42:00] CodeAlmanac and Argyph: context that stays close to the repo
CodeAlmanac is a codebase wiki for AI coding agents. It captures decisions, flows, invariants, and gotchas: the context maintainers know but the code may not express. That can prevent an agent from inferring design intent from source alone.

Argyph is a local-first MCP server for structured semantic context over a codebase. It is useful when privacy and latency matter, but the safety rule remains: semantic retrieval is an orientation layer, not the final proof. Let it point the agent toward the right files, then verify with exact file reads and tests before accepting edits.

[50:00] Practical queue
For OpenClaw, test `v2026.5.28` on the paths that used to be brittle: subagents, aborts, resumes, channel approvals, browser automation, and provider timeouts. For Claude Code, treat `2.1.159` as npm `latest` hygiene. For MiniMax M3, run a careful bakeoff before believing the benchmark deck: long-context repo understanding, coding-agent tool use, and multimodal computer-use tasks. For the project radar, choose one control-plane tool and one context tool: agentgateway or MCPJungle for MCP sprawl, Understand Anything for graph orientation, CodeAlmanac for durable repo knowledge, and Argyph for local semantic retrieval.

Chapters

  • 00:00 Opening hook: releases and model drops first
  • 03:00 OpenClaw v2026.5.28 release deep dive
  • 16:00 Claude Code 2.1.159 as fleet hygiene
  • 20:00 MiniMax M3: sparse attention, 1M context, and coding-agent ambition
  • 28:00 Understand Anything: graph orientation before edits
  • 34:00 agentgateway and MCPJungle: controlling tool sprawl
  • 42:00 CodeAlmanac and Argyph: context that stays close to the repo
  • 50:00 Practical queue

Verified Links

🎙 Never miss an episode — subscribe now

SpotifyAppleiHeartYouTube MusicAmazonRSS
🎙 Subscribe to AgentStack Daily