I wanted to upgrade my self-hosted AI gateway from Claude Opus 4.5 to Opus 4.6 and unlock the 1 million token context window.

It took me an entire day. Ten things broke. The final proof cost me $9.70 in API fees — in thirty minutes.

Here’s everything that tried to stop me.

🚀 Just Want the Working Config?

Skip the war story. Jump to the complete step-by-step setup guide →

What Broke

  1. Default context hardcoded to 200K
  2. Config pinned to old model
  3. Stale pi-ai dependency
  4. No git hook for dep sync
  5. systemd can’t find node
  6. Model registry reports wrong context size
  7. Custom header kills OAuth auth
  8. Build wipes the UI
  9. 1M not available on flat-rate billing
  10. API key in wrong file = 17 crash restarts

The Setup

I run OpenClaw, an open-source AI gateway, on a local workstation. It connects Claude to my WhatsApp, manages context windows, handles tool calls — basically turns an LLM into a persistent assistant that lives on my machine.

My environment:

  • OS: Fedora Silverblue 43 (immutable, rpm-ostree based)
  • OpenClaw: Forked from upstream openclaw/openclaw, running as a systemd user service
  • Node: v22.22.0 via nvm
  • Package manager: pnpm
  • GPU: NVIDIA (driver 580.119.02)
  • Auth: Anthropic OAuth token (Max subscription, flat-rate billing)

Anthropic had just released Opus 4.6 with support for 1 million token context. My gateway was running Opus 4.5 at 200K. Simple upgrade, right?

No.

Wall 1: The Default Was Hardcoded

The first thing I checked was src/agents/defaults.ts:

export const DEFAULT_CONTEXT_TOKENS = 200_000;

Hardcoded. Doesn’t matter what the model supports — if no override exists, you get 200K. I changed it to 1_000_000, committed, and opened PR #10536.

Lesson: Always grep for hardcoded constants before assuming config drives behavior.

# Fix: edit src/agents/defaults.ts
sed -i 's/200_000/1_000_000/' src/agents/defaults.ts
git add src/agents/defaults.ts
git commit -m "feat: increase DEFAULT_CONTEXT_TOKENS to 1M for Opus 4.6"

Wall 2: Config Override Stuck on the Old Model

My runtime config at ~/.openclaw/openclaw.json still had:

"agents": {
  "defaults": {
    "model": {
      "primary": "anthropic/claude-opus-4-5"
    }
  }
}

The source code knew about Opus 4.6, but my local config was pinning the old model.

Lesson: When debugging model issues, check both source code defaults and runtime config files. They can disagree.

# Fix: update the model in openclaw.json
# Find this line:
#   "primary": "anthropic/claude-opus-4-5"
# Change to:
#   "primary": "anthropic/claude-opus-4-6"
nano ~/.openclaw/openclaw.json

# Restart the gateway
openclaw gateway restart

Wall 3: “Unknown model: anthropic/claude-opus-4-6”

OpenClaw uses pi-ai as its model abstraction layer. My node_modules had [email protected] — which didn’t know Opus 4.6 existed. The package.json specified 0.52.6, but the lockfile hadn’t been synced after a recent pull.

Lesson: Stale dependencies are silent killers. Your package.json can say one thing while node_modules runs another.

# Fix: sync dependencies and rebuild
cd ~/openclaw
pnpm install
pnpm build
pnpm ui:build  # important — see Wall 8

# Restart
openclaw gateway restart

Wall 4: Stale Deps Prevention (Git Hook)

To prevent Wall 3 from happening again after every git pull, I created a post-merge hook:

Lesson: Automate the thing that just bit you.

# Fix: create git-hooks/post-merge
cat > git-hooks/post-merge << 'EOF'
#!/bin/sh
changed_files=$(git diff-tree -r --name-only --no-commit-id ORIG_HEAD HEAD)

if echo "$changed_files" | grep -q "pnpm-lock.yaml"; then
    echo "📦 pnpm-lock.yaml changed — running pnpm install..."
    pnpm install --frozen-lockfile
fi
EOF

chmod +x git-hooks/post-merge

# If your git hooks dir is configured:
git config core.hooksPath git-hooks

Wall 5: systemd Can’t Find Node

After pnpm build, the control UI failed to build inside systemd:

env: 'node': No such file or directory

OpenClaw’s service environment script (src/daemon/service-env.ts, lines 69/78) assumes ~/.nvm/current/bin exists — that’s an fnm convention. I use nvm, which puts node at ~/.nvm/versions/node/v22.22.0/bin/. systemd doesn’t source .bashrc, so the nvm PATH setup never runs.

Lesson: Self-hosted software and systemd have a PATH disagreement that will bite you. Check your service environment with systemctl --user show openclaw-gateway.service --property=Environment or just pre-build manually.

# Fix: pre-build the UI manually (workaround)
cd ~/openclaw
pnpm ui:build

# Verify it's there:
ls dist/control-ui/index.html

# Restart
openclaw gateway restart

Wall 6: The Model Registry Lies (On Purpose)

Even after all the fixes, my context window still showed 200K. I dug into pi-ai’s model catalog:

// node_modules/@mariozechner/pi-ai/dist/models.generated.js
{
    id: "claude-opus-4-6",
    contextWindow: 200000,  // ← Not 1M
    // ...
}

This is intentional. The 1M context window requires a beta header (anthropic-beta: context-1m-2025-08-07). Without it, the API hard-rejects anything over 200K. So the model registry reports the safe default, not the maximum.

I filed issue #1329 upstream.

Lesson: Model catalogs report default capabilities, not maximum capabilities. The difference matters.

# Fix: override the context window in openclaw.json
# (see "The Config That Works" section at the end for the full config)
# The key field is:
#   "contextWindow": 1000000
# inside models.providers.anthropic.models[]

Wall 7: The Header That Ate Authentication

Here’s where it got ugly.

To send the beta header, I added a custom models.providers.anthropic config to openclaw.json with the context-1m-2025-08-07 header. Restarted. Got:

401 authentication_error: "OAuth authentication is currently not supported."

Wait, what? I’m using OAuth. It was working 5 minutes ago.

After deploying Claude Code’s agent teams feature (two parallel investigators — one tracing the sync path, one tracing headers), I found the kill chain:

  1. Adding models.providers.anthropic triggers ensureOpenClawModelsJson() in src/agents/models-config.ts
  2. normalizeProviders() in src/agents/models-config.providers.ts (line 233-256) sees the provider has models but no apiKey, so it injects the OAuth token as apiKey
  3. pi-ai’s mergeHeaders() in providers/anthropic.js uses Object.assign() — last write wins
  4. My custom anthropic-beta header overwrote the default one
  5. The default header included oauth-2025-04-20 — required for OAuth to work
  6. Without it: 401.

Lesson: When you override HTTP headers, you’re not adding — you’re replacing. If a framework merges headers with last-write-wins, you need to include everything the original had plus your additions.

# Fix: include ALL required beta features in the custom header
# In your openclaw.json models.providers.anthropic.models[].headers:
#
#   "anthropic-beta": "claude-code-20250219,oauth-2025-04-20,context-1m-2025-08-07,fine-grained-tool-streaming-2025-05-14,interleaved-thinking-2025-05-14"
#
# Missing any one of these will break a different feature.
# The critical ones:
#   oauth-2025-04-20          → required for OAuth auth
#   context-1m-2025-08-07     → required for 1M context
#   claude-code-20250219      → required for tool use

Wall 8: Build Nukes the UI

Every time I ran pnpm build, it wiped dist/ — including the pre-built control UI from Wall 5. I got burned by this three times before I learned.

Lesson: Document your build order. pnpm build && pnpm ui:build should probably be a single script.

# Fix: always rebuild UI after building
pnpm build && pnpm ui:build

# Or add a script to package.json:
# "build:all": "pnpm build && pnpm ui:build"

Wall 9: “The long context beta is not yet available for this subscription”

Final boss of the OAuth path. After fixing the header merge, I got:

400 invalid_request_error: "The long context beta is not yet available for this subscription"

This one isn’t a bug. It’s a billing limitation.

I was authenticated via OAuth using Anthropic’s Max subscription (flat-rate plan). The 1M context beta is only available with API key billing (pay-per-token, Tier 4).

Auth MethodBillingContext Limit1M Support
OAuth token (sk-ant-oat-...)Flat rate (Max plan)200K❌
API key (sk-ant-api03-...)Per token1M✅

Lesson: “Supports 1M context” means “supports 1M context on certain billing tiers.” Read the fine print.

# Fix: Get an API key from Anthropic
# 1. Go to https://console.anthropic.com/settings/keys
# 2. Create a new API key
# 3. Copy it (starts with sk-ant-api03-)
# 4. You'll need Tier 4 access for 1M context
#
# Don't put it in openclaw.json — see Wall 10.

Wall 10: The Crash Loop

I put the API key directly into openclaw.json with mode: "api-key" and apiKey: "sk-ant-api03-...". Restarted the gateway.

It crashed. systemd restarted it. It crashed again. And again. Seventeen times.

journalctl --user -u openclaw-gateway.service | grep -c "Started"
17

Two mistakes in one line:

  1. api-key vs api_key: The Zod schema validation expects api_key (underscore). The hyphenated version silently fails validation and crashes the process.
  2. Wrong file: API credentials don’t go in openclaw.json. They go in a separate auth-profiles.json file at ~/.openclaw/agents/main/agent/auth-profiles.json, with type: "api_key" and key: "...".

The config file declares what kind of auth. The credentials file stores the actual secret. Mixing them up doesn’t give you a helpful error — it gives you a crash loop.

Lesson: Zod validation errors in a systemd service are invisible unless you’re watching the journal. And schema field names with underscores vs hyphens will ruin your afternoon.

# Fix: put the API key in the CORRECT file with the CORRECT field names

# 1. Edit the auth profiles file (NOT openclaw.json):
cat > ~/.openclaw/agents/main/agent/auth-profiles.json << 'EOF'
{
  "anthropic": {
    "default": {
      "type": "api_key",
      "key": "sk-ant-api03-YOUR-KEY-HERE"
    }
  }
}
EOF

# 2. Restart:
openclaw gateway restart

# 3. Verify it's running (not crash-looping):
openclaw gateway status

# If something's wrong, check the journal:
journalctl --user -u openclaw-gateway.service -f

The Proof

It worked. I pushed the conversation to 210,000 tokens — past the 200K wall that would have stopped me on OAuth. No compaction. No errors. Still coherent.

📚 Context: 210k/1.0m (21%) · 🧹 Compactions: 0

I also proved it via curl — sending a 250K token prompt directly to the API:

# This succeeds with the beta header:
curl -X POST https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "anthropic-beta: context-1m-2025-08-07" \
  -H "content-type: application/json" \
  -d '{"model":"claude-opus-4-6","max_tokens":100,"messages":[{"role":"user","content":"...250K tokens..."}]}'
# → 200 OK

# This fails WITHOUT the beta header:
# → 400: "prompt is too long: 250033 tokens > 200000 maximum"

Then I checked my Anthropic dashboard.

$9.70. In about 30 minutes.

The Economics

This is the part most people won’t tell you about 1M context:

At 210K context on Opus 4.6, every message costs ~$1.05 just for input tokens — because the entire conversation history gets re-sent with each exchange. The cost compounds with every turn.

MetricOAuth (Flat Rate)API Key (Per Token)
Context limit200K1M
Monthly costFixed (Max plan)Variable
Cost at 100K context$0~$0.50/message in
Cost at 500K contextN/A~$2.50/message in
Cost at 1M contextN/A~$5.00/message in

1M context is real and it works. But you probably don’t want it running all day. Use OAuth for daily chat, switch to API key when you genuinely need to ingest an entire codebase, a full book, or months of system logs in a single session.

The Complete Config That Works

For anyone running an OpenClaw fork on Fedora (or similar) and wanting 1M context, here’s the full process end to end.

Step 1: Get an Anthropic API Key

Go to console.anthropic.com/settings/keys and create a new key. You need Tier 4 access for 1M context. Copy the key — it starts with sk-ant-api03-.

Step 2: Store the Key

Create or edit ~/.openclaw/agents/main/agent/auth-profiles.json:

{
  "anthropic": {
    "default": {
      "type": "api_key",
      "key": "sk-ant-api03-YOUR-KEY-HERE"
    }
  }
}

Do not put credentials in openclaw.json. That file is for config, not secrets.

Step 3: Add the Model Config

Add this to your ~/.openclaw/openclaw.json:

{
  "models": {
    "providers": {
      "anthropic": {
        "baseUrl": "https://api.anthropic.com",
        "api": "anthropic-messages",
        "models": [
          {
            "id": "claude-opus-4-6",
            "name": "Claude Opus 4.6",
            "reasoning": true,
            "input": ["text", "image"],
            "cost": {
              "input": 5,
              "output": 25,
              "cacheRead": 0.5,
              "cacheWrite": 6.25
            },
            "contextWindow": 1000000,
            "maxTokens": 128000,
            "headers": {
              "anthropic-beta": "claude-code-20250219,oauth-2025-04-20,context-1m-2025-08-07,fine-grained-tool-streaming-2025-05-14,interleaved-thinking-2025-05-14"
            }
          }
        ]
      }
    }
  }
}

Step 4: Rebuild and Restart

cd ~/openclaw  # or wherever your fork lives
pnpm install
pnpm build
pnpm ui:build
openclaw gateway restart

Step 5: Verify

Check /status in your chat or look at the logs:

openclaw gateway status

You should see Context: Xk/1.0m in your status output.

What I Learned

  1. Self-hosted AI is real infrastructure. It breaks like infrastructure. Config drift, stale deps, header conflicts, billing quirks — this is ops work.

  2. The “1M context” marketing is true but incomplete. Yes, the model supports it. No, your auth method might not. No, you probably can’t afford to run it continuously.

  3. Debugging is the product. Every wall I hit is a wall someone else will hit. Writing it down is the most useful thing I can do.

  4. Agent teams are underrated. Using Claude Code’s parallel agent feature to investigate the header merge bug cut my debugging time in half. Two agents, two hypotheses, converging on the answer.

  5. The unsexy fix is usually right. After all that work, my daily driver is still OAuth at 200K. The 1M capability is there when I need it — for codebase analysis, ingesting months of logs, or proving a point for $9.70.

This is the first real post on bold.casa. I’m building in public — AI tooling, self-hosted infrastructure, and whatever else needs taking apart. Follow along via RSS or Atom.