So far this course has talked about Claude Code as a standalone Node CLI. Subtopic 1.3 is the next layer: what happens when you wire Claude Code into an OpenClaw agent so the agent can hand coding tasks off to it, monitor the build, and come back with a diff. The channel framed this as a tag — claude-code — that shows up on any OpenClaw update where the integration actually moves. Three videos in the syllabus cover that track: the Perplexity Computer side-by-side (cross-listed from 1.1), OpenClaw 3.7, and OpenClaw 3.8. The Minimax 2.7 video is cross-listed from 1.2; it sits in the same track because the settings.json swap is exactly what an OpenClaw agent does when it routes a coding job to Claude Code.

What you'll learn

  • OpenClaw 3.7 shipped the context engine plugin (the Discord thread problem is fixed) and ACP persistent binding, which makes Claude Code actually usable inside an OpenClaw agent — your agent stops having to re-brief Claude on every session.
  • OpenClaw 3.8 added ACP provenance — every coding task gets a session ID plus origin tracing, so the morning-after audit pattern the channel keeps recommending now has something to audit.
  • GPT 5.4, Gemini 3.1 Flash, and Longcat (China-side) joined the model list in 3.7; native OpenAI stack support is the relevant cost lever for coding workflows that don't need Opus.
  • Update OpenClaw via openclaw update over Termius, not through the agent itself — agent-driven updates hit permission errors and stall.

The Perplexity Computer side-by-side (cross-listed from 1.1)

This video is in the 1.3 syllabus because the question it asks — "what should an agent-driven coding workflow look like?" — is the same question the claude-code track in OpenClaw is answering. The full breakdown is in subtopic 1.1, but the load-bearing point for 1.3 is: Perplexity Computer's $176/month Max plan ships Opus 4.6 plus parallel orchestration, and the side-by-side showed Claude Code only matching it after being hand-tuned with plan mode and agent teams. That hand-tuning is exactly what an OpenClaw agent is supposed to do for you. The other moat is process visibility — Perplexity Computer hides the orchestration; Claude Code + OpenClaw ACP provenance (see 3.8 below) does not.

OpenClaw 3.7: context engine + ACP persistent binding

The 3.7 release dropped on March 7th with a changelog that screams 40+ features and 30+ bug fixes. The channel's honest take: life after the update is "slightly" better, and almost all the value is concentrated in two changes plus a model-list expansion.

Context engine plugin — the Discord thread problem is fixed

The biggest user-facing change is the context engine plug-in system. Before 3.7, opening a Discord thread for a new task meant the main branch "doesn't update the main chat" — your agent had no idea what you'd just discussed. The new plugin connects to the main branch itself, so context "ties together. It's more coherent, more cohesive." For anyone running multiple OpenClaw instances (the channel uses an Avengers-themed setup internally), this is the actual "huge upgrade" and the prerequisite for the 1.3 workflow.

ACP persistent binding — Claude Code stops forgetting

The other load-bearing change is ACP persistent binding. Every time the agent spins up Claude Code, Claude Code has to be re-briefed on the repo, the goal, and the partial diff. ACP persistent binding makes OpenClaw remember what Claude Code has been building across sessions, which "substantially" improves the loop with tools like Windsurf and Claude Code. The channel's own boxminingai.com site was built this way — agent brainstorms, Claude Code executes. The creator says the new update "is much easier to create with the new open cloud."

Model support — GPT 5.4, Gemini 3.1 Flash, Longcat

The headline model additions in 3.7 are GPT 5.4 and Gemini 3.1 Flash via the OpenAI stack (no longer locked to one provider) plus Longcat integration for China-side users. The cost contrast: the host running Opus burns $30–$60/day, while the co-host on the Minimax developer plan pays a flat $20/month and reports it working "substantially better" after about two weeks of context optimization. That's the same overnight-build arithmetic from subtopic 1.2, restated as "switch backends before you switch frameworks."

Update via terminal, not your agent

Skip the auto-updater when things break. Run openclaw update directly over Termius — agent-driven updates hit permission errors and disconnects, manual runs don't. Small but load-bearing: if the update lands badly, the coding track is the first thing that breaks.

OpenClaw 3.8: ACP provenance makes the coding track auditable

OpenClaw 3.8 dropped on March 8 — one day after 3.7 — with a single workflow-changing feature: ACP provenance. The video is short because the changelog is short, but the one thing it ships is the missing audit trail.

With ACP provenance, every coding task gets a session ID plus origin tracing, so you can see what passed, what failed, and where it broke. The host pair it with is Cloud Code — "really good at programming" — used to vibe-code the new boxmoneyai.com site (Windows 95 UI included). For a track that hands work from OpenClaw to Claude Code to Cloud Code, provenance is the thread that ties the three logs together.

The overnight builder workflow, restated

The concrete use case: kick off a build before you sleep. "You can say, Hey, Stark. Just run a debug on my website. Do any refactoring you needed. They'll do it at night and you'll see in the morning what happened." Same pattern as 1.2, but in 1.3 the "review the diff" step has a real audit surface — provenance lets you replay the session instead of trusting the agent's summary.

Thread awareness is 3.7, not 3.8

The video muddles the two versions. Thread context shipped in 3.7 (the context engine plugin), not 3.8. The 3.8 contribution is provenance and Telegram streaming; everything thread-related is 3.7.

Model cost reality check

The creator is burning ~$30/day on Opus running his bot, while Ron runs his on MiniMax's coding plan for under $1/day, with a $10/month tier as the entry point. Native OpenAI GPT model support is in OpenClaw and being tested against Opus and MiniMax — no verdict yet, but the floor is "Minimax coding plan," not "Minimax token plan." 3.8 also adds Telegram streaming — responses stream incrementally instead of all at once; useful if you watch long agent outputs live, skip it if you only read the final message.

Claude Code + Minimax 2.7: the config side of the same track

This video is in the 1.3 syllabus because the settings.json swap it covers is exactly the swap an OpenClaw agent makes when it routes a coding task to Claude Code. The full breakdown is in subtopic 1.2, but the points that matter for 1.3 are: the three env vars in settings.json (token-plan API key, ANTHROPIC_AUTH_TOKEN, base URL https://api.minimax.io/anthropic) are what an OpenClaw agent injects when it hands work to Claude Code; the "Minimax token plan resets" detail is the contract that makes overnight coding tracks safe; the trailing-comma JSON gotcha applies whether you're editing settings.json by hand or whether an OpenClaw skill is editing it (an unguarded edit will break Claude Code at startup); and the same swap works for Kilo Code, Open Claude, Grok CLI, and other Anthropic-compatible clients — so the agent can route to whichever client is already authenticated for the project.

Try it yourself

The hands-on goal: get an OpenClaw agent driving a Claude Code session overnight, and have ACP provenance waiting for you in the morning.

  1. Confirm you're on OpenClaw 3.8 or later. Run openclaw --version. The claude-code track is only coherent with both 3.7 (context engine + ACP persistent binding) and 3.8 (provenance).
  2. Update via Termius, not the agent. openclaw update over SSH/Termius. Agent-driven updates stall on permission changes.
  3. Install the context engine plugin on Discord or Telegram, whichever channel your OpenClaw agent listens on. Without it, opening a thread for a new task means the main branch loses context.
  4. Enable ACP persistent binding in settings.json before starting any Claude Code project. With it off, you pay re-discovery on every loop.
  5. Pick a backend. Default to Minimax ($20/month developer plan, or $10/month coding plan) for orchestration, with Opus reserved for short bursts ($30/day ceiling). Pin GPT 5.4 or Gemini 3.1 Flash for cost-sensitive tasks now that the OpenAI stack is native in 3.7.
  6. Schedule a low-stakes overnight build — a content-inventory script, a new settings page, an integration test. Send it through the agent, not Claude Code directly. The point is to exercise the agent → Claude Code handoff.
  7. Verify ACP provenance in the morning. The coding task should have a session ID and an origin trace. If it doesn't, provenance isn't actually on.
  8. Diff before merging. Even with provenance, keep Sonnet or Opus in reserve for the final review pass on security-sensitive or money-handling code. Minimax 2.7 is "near Opus, not Opus" — trust the trace, but verify the diff.
  9. Budget two weeks of context tuning on a new model before judging it. The channel saw "substantially better" results only after that window on Minimax.

Common pitfalls

  • Skipping the 3.7 → 3.8 upgrade order. The 3.7 context engine plugin is the prerequisite for the 3.8 provenance to be useful. Jump straight to 3.8 and the trace will be technically complete but practically incoherent.
  • Updating through the agent. Agent-driven updates hit permission errors and disconnects. Always openclaw update over Termius.
  • Re-briefing Claude Code on every session. If your agent isn't using ACP persistent binding, every loop pays the re-briefing cost. Turn it on.
  • Skipping provenance checks in the morning. The overnight pattern only works if you actually review the trace. Trust the agent's summary without checking the session ID and you're back to the pre-3.8 "no idea what's happening" state.
  • Confusing 3.7 and 3.8. Thread awareness is 3.7, provenance and Telegram streaming are 3.8. Tutorials that mix them are mixing versions.
  • Routing everything through Opus. The host burns $30–$60/day on Opus; the co-host on Minimax pays $20/month. Default is cheap, Opus is the exception.
  • Reading the cost comparison as a one-shot. The channel explicitly said they hadn't finished testing GPT vs Opus vs Minimax. Run your own benchmark on your own tasks before committing budget.
  • Letting the agent update settings.json unsupervised. The trailing-comma gotcha from 1.2 still applies.
  • Driving Claude Code directly when an OpenClaw agent could route it. If the task is multi-step or context-heavy, the agent handoff is the point of the 1.3 track.
  • Trusting the agent's morning-after summary. ACP provenance is a session ID and a trace. If the trace is empty, provenance isn't actually wired up.

Sources

  • Perplexity Computer Just KILLED Claude Code (Side-by-Side Test) — 18,483 views · video_id: E4gc4-9O1JE
  • NEW Openclaw 3.7 Update is INSANE! — 7,974 views · video_id: p3aB9Qd-PIs
  • NEW OpenClaw 3.8 Update is HUGE for Coding Workflows! — 3,744 views · video_id: S-h3FaBFXSs
  • Claude Code + Minimax 2.7: Unlimited AI Coding on a Budget — 6,532 views · video_id: dURSH_Fwu6s (cross-listed from subtopic 1.2)
  • Supabase query — SELECT video_id, title, views, summary_content, summary_key_takeaways FROM public.videos WHERE video_id = ANY(ARRAY['E4gc4-9O1JE','p3aB9Qd-PIs','S-h3FaBFXSs','dURSH_Fwu6s']); against project ttxdssgydwyogq.