Vertex Anthropic routing: google-vertex sends claude-* to publishers/google/ (404), and built-in subagent model bindings ignore user config

[navidRashik]

Summary

Two related bugs in google-vertex provider handling of Claude (Anthropic) models on opencode 1.18.5:

1. Wrong Vertex publisher namespace: google-vertex provider routes claude-* model requests to publishers/google/models/<claude-model> on Vertex, but Anthropic models on Vertex live at publishers/anthropic/models/<claude-model>. All Claude calls under google-vertex return Publisher model ... was not found.
2. Built-in subagent model bindings are hardcoded and ignore user config: The built-in subagents explore, general, junior-software-engineer, ops-debugger, qa-engineer are pinned to google-vertex/claude-haiku-4-5 (a non-existent model id — the current catalog only has claude-haiku-4-5@20251001). No user-config path can override this binding:
   • agent.<name>.model in opencode.jsonc — silently ignored
   • model: frontmatter in ~/.config/opencode/agents/<name>.md — silently ignored (verified by setting it to FAKE/xyz with no effect on the resolved config)
   • Adding mode: subagent to the same block — mode field is accepted but model is still ignored
   • Removing google-vertex from enabled_providers — still stuck

The same config shape works perfectly for any custom agent name (verified by adding an agent mycustom with the same config — resolved correctly).

Environment

• opencode 1.18.5 (brew, /opt/homebrew/bin/opencode)
• macOS 14.5 (Darwin 25.5.0)
• Vertex project: uses ADC auth via gcloud auth application-default login
• providers enabled: google-vertex-anthropic, google-vertex, github-copilot

Reproduction

~/.config/opencode/opencode.jsonc:

{
  "$schema": "https://opencode.ai/config.json",
  "model": "google-vertex-anthropic/claude-sonnet-4-6",
  "enabled_providers": ["google-vertex-anthropic", "google-vertex"],
  "provider": {
    "google-vertex-anthropic": {
      "options": { "project": "<PROJECT>", "location": "global" }
    }
  },
  "agent": {
    "explore":   { "model": "google-vertex-anthropic/claude-haiku-4-5@20251001" },
    "mycustom":  { "model": "google-vertex-anthropic/claude-haiku-4-5@20251001" }
  }
}

Run:

GOOGLE_VERTEX_PROJECT=<PROJECT> GOOGLE_VERTEX_LOCATION=global opencode debug config

Actual

agent.explore.model  = google-vertex/claude-haiku-4-5     ← ignored user config
agent.mycustom.model = google-vertex-anthropic/claude-haiku-4-5@20251001   ← honored

Invoking explore as a subagent then fails at runtime:

AI_APICallError: Publisher model
`projects/<PROJECT>/locations/us-central1/publishers/google/models/claude-sonnet-4-6`
was not found or your project does not have access to it.

and separately:

ProviderModelNotFoundError: Model not found: google-vertex/claude-haiku-4-5.
Did you mean: claude-haiku-4-5@20251001?

Note the location us-central1 in the URL even though region: "global" is set in the provider options — the region field also seems to be ignored for google-vertex.

Expected

• google-vertex provider should either route claude-* models to publishers/anthropic/ OR reject claude-* model ids with a clear "use google-vertex-anthropic instead" error.
• agent.<name>.model config for built-in subagent names should override the built-in defaults.
• Built-in subagent default model ids should reference model versions that actually exist in the current catalog (e.g., claude-haiku-4-5@20251001 rather than the versionless claude-haiku-4-5).
• provider.<name>.options.region should be respected.

Impact

Any user with a Vertex-anchored Claude setup can't use the built-in Haiku subagents (explore, general, junior-software-engineer, ops-debugger, qa-engineer) at all. Delegated work through these agent names fails immediately. Custom-named "shadow" agents work as a workaround.

Workaround

Duplicate each built-in .md under a new name (e.g., explore-vx.md with name: explore-vx) with model: google-vertex-anthropic/claude-haiku-4-5@20251001, then always call them via subagent_type=<name>-vx from the task tool. Full config for the custom name is honored. Users must remember to prefer the shadow, and skills/scripts that hardcode built-in names still break.

Related

• Bundled @ai-sdk/google-vertex/anthropic provider (google-vertex-anthropic) works correctly out of the box for Claude on Vertex — this is the code path the built-in agents should be using.
• strings /opt/homebrew/bin/opencode | grep -c "publishers/anthropic/models" returns matches, so the correct code path is compiled in.

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Subagent model config ignored — all subagents use primary model #36289: Subagent model config ignored — all subagents use primary model (same symptom: agent.<name>.model in config is silently ignored for subagents)
• bug(session): agent.model and agent.variant config ignored in handleSubtask #36250: bug(session): agent.model and agent.variant config ignored in handleSubtask (same root area)
• [Bug] Model parameter ignored when launching subagent with agent name #18615: Model parameter ignored when launching subagent with agent name (same symptom for named agents)
• Desktop: Vertex Anthropic models resolve to wrong region + invalid @default version in new worktrees #32550: Desktop: Vertex Anthropic models resolve to wrong region + invalid @default version (overlaps with the wrong region / wrong publisher namespace issue you describe)

If your issue covers ground not addressed in those, feel free to clarify what's distinct — particularly the publishers/google/ vs publishers/anthropic/ routing bug for google-vertex with Claude models, which may be a separate root cause.

[navidRashik]

Additional testing (2026-07-27)

Confirmed the "user config ignored" bug is more subtle than initially reported. The trigger is the existence of a matching agent definition file anywhere opencode scans — the resolver seems to bind the built-in model default to that name unconditionally, ignoring both opencode.jsonc agent.<name>.model AND the .md file's own model: frontmatter.

Reproduction matrix

Tested with opencode debug config for agent explore (opencode 1.18.7, macOS). opencode.jsonc and every agent .md file with model: google-vertex-anthropic/claude-haiku-4-5@20251001 explicitly set:

Setup	Resolved model
~/.config/opencode/opencode.jsonc has agent.explore.model = "google-vertex-anthropic/claude-haiku-4-5@20251001"	google-vertex/claude-haiku-4-5 (❌ ignored)
~/.config/opencode/agents/explore.md has model: google-vertex-anthropic/claude-haiku-4-5@20251001	same (❌)
Both of the above simultaneously	same (❌)
~/.claude/agents/explore.md has model: claude-haiku-4-5 (bare, no provider prefix)	same (❌ — bare model likely gets google-vertex/ prepended)
~/.claude/agents/explore.md has model: google-vertex-anthropic/claude-haiku-4-5@20251001 (fully qualified)	same (❌ — user model still ignored)
~/.claude/agents/explore.md has NO model: line, file exists	same (❌ — file existence alone triggers broken default)
~/.claude/agents/explore.md renamed/removed	google-vertex-anthropic/claude-haiku-4-5@20251001 (✅ falls back to small_model)
Custom name (explore-vx) with same config	google-vertex-anthropic/claude-haiku-4-5@20251001 (✅)

Interpretation

For the 5 built-in subagent names (explore, general, junior-software-engineer, ops-debugger, qa-engineer), the resolver appears to short-circuit to a hardcoded/derived google-vertex/claude-haiku-4-5 model binding whenever any file with that name is discovered — regardless of what model: the user config specifies. The only two escape hatches are:

1. Remove every file (from ~/.claude/agents/, ~/.config/opencode/agents/, .opencode/agents/, etc.) that matches the name. Then opencode falls back to inheriting from small_model. Painful — deletes cross-tool configuration.
2. Use a different agent name (e.g. explore-vx, code-explorer) that isn't in the built-in list. This works perfectly but breaks any workflow that references the canonical names.

Two independent bugs, one symptom

This confirms there are two bugs stacked on top of each other:

1. Wrong Vertex publisher path: google-vertex provider routes claude-* models to publishers/google/ instead of publishers/anthropic/ — 404 at runtime.
2. Hardcoded/invalid built-in default model: The 5 built-in subagent names have a resolver default of google-vertex/claude-haiku-4-5 — a model id that doesn't exist in the catalog (correct id is claude-haiku-4-5@20251001) and points at the broken publisher path. This default cannot be overridden by user config.

Suggested fixes

• Make agent.<name>.model in opencode.jsonc win over the built-in default (respect the documented precedence).
• Update the built-in default model id to a version that actually exists in models.dev (claude-haiku-4-5@20251001).
• Route the built-in subagents through google-vertex-anthropic (or whichever provider is appropriate) rather than google-vertex, so Claude models don't hit the Gemini publisher path.

[StayPirate]

Can confirm this affects a broader surface than just the built-in google-vertex-pinned agents (explore, general, etc.).

Additional repro data point

• Provider: github-copilot (not google-vertex)
• Agent type: a project-scoped custom subagent defined in .opencode/agents/<name>.md (not a built-in agent name) — so this isn't limited to overriding a hardcoded built-in binding
• Model ID validity: confirmed valid via opencode models output (model ID appears in the provider's model list)

What was tried (both silently ignored, consistent with the report)

1. Frontmatter model: field directly in .opencode/agents/<name>.md:

   ---
   description: ...
   mode: subagent
   model: github-copilot/claude-opus-5
   permission: { ... }
   ---

2. JSON override in opencode.json under agent.<name>.model:

   { "agent": { "<name>": { "model": "github-copilot/claude-opus-5" } } }

Neither changed the resolved model when the subagent was invoked via the Task tool (task tool call with subagent_type: "<name>") from a primary agent — the subagent consistently ran with the invoking primary agent's model instead.

Source trace (may help narrow down the root cause)

Traced the relevant path in the OpenCode source:

• packages/opencode/src/tool/task.ts (around where the subagent Info is resolved via agent.get(params.subagent_type)): the fallback logic const model = next.model ?? { modelID: msg.info.modelID, providerID: msg.info.providerID } looks correct in principle — it should prefer next.model when defined.
• packages/opencode/src/agent/agent.ts (around the config-to-runtime agent registration loop): if (value.model) item.model = Provider.parseModel(value.model) — this is where the frontmatter/JSON model string should get parsed into { providerID, modelID } and attached to the agent's Info.

Given that both a frontmatter-only agent and a JSON-only override fail identically, the issue looks like it could be in whichever step populates next.model for the resolved Info object at the point the Task tool calls agent.get() — possibly a session/model resolution step downstream that doesn't read from the same Info the Task tool builds its fallback from (worth checking whether there's a second, V2-style agent/model resolution path that takes precedence for some invocation routes but not others, since the same config shape was reported to work for one custom agent name in the original report but not here).

Happy to provide more repro detail if useful — this is currently blocking pinning specific subagents to a heavier model for higher-quality automated review work in an unrelated project, so a fix would help beyond the Vertex-specific case in the original report.