Bug: claude-opus-4.7 via GitHub Copilot fails with 'output_config: Extra inputs are not permitted'

[jimicze]

Bug Report

Description

When using claude-opus-4.7 model via the github-copilot provider, every LLM request fails immediately with:

AI_APICallError: output_config: Extra inputs are not permitted

The model is completely unusable — no response is ever generated.

Steps to Reproduce

1. Configure OpenCode to use github-copilot provider with claude-opus-4.7 model
2. Start a session and send any message
3. Request fails with the error above

Expected Behavior

The request should succeed and the model should respond normally.

Actual Behavior

The API call to https://api.githubcopilot.com/v1/messages is rejected with a Pydantic validation error. OpenCode includes an output_config field in the request body that claude-opus-4.7 does not accept.

Log Evidence

From ~/.local/share/opencode/log/:

ERROR service=llm providerID=github-copilot modelID=claude-opus-4.7
  error=AI_APICallError url=https://api.githubcopilot.com/v1/messages

ERROR service=session.processor
  error=output_config: Extra inputs are not permitted
  stack=AI_APICallError: output_config: Extra inputs are not permitted

The error occurs on every request attempt, twice per attempt (first try + retry).

Root Cause (Hypothesis)

OpenCode unconditionally includes an output_config field in the Anthropic Messages API request body. The claude-opus-4.7 model as exposed via GitHub Copilot proxy (api.githubcopilot.com/v1/messages) does not support this parameter and rejects it with Pydantic strict-mode validation error.

output_config appears to be a newer/experimental Anthropic API feature (possibly for extended thinking or output format configuration) not yet supported by all models or provider proxies.

Suggested Fix

Conditionally omit output_config from the request body when the model/provider does not support it, or implement a fallback retry without the parameter when this specific error is received.

Test Results

Version	Config	Result
v1.14.29	claude-opus-4.7 without reasoningEffort	✅ works
v1.14.18	claude-opus-4.7 + reasoningEffort: medium	❌ fails (original bug)
dev (0.0.0-dev-202604281959)	claude-opus-4.7 + reasoningEffort: medium	✅ fixed

Environment

• OpenCode version: v1.14.18 (latest at time of report; current latest is v1.14.29)
• Provider: github-copilot
• Model: claude-opus-4.7
• Platform: macOS (darwin)
• API endpoint: https://api.githubcopilot.com/v1/messages

[github-actions]

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

• claude-opus-4.7 via GitHub Copilot fails with thinking.type.enabled not supported #23500: Same model (claude-opus-4.7) and provider (GitHub Copilot) failing with unsupported request body fields — root cause is the same (opencode sending fields that GitHub Copilot's claude-opus-4.7 proxy rejects)
• 1.4.10 breaks Haiku through GitHub Copilot #23342: Same error pattern ("Extra inputs are not permitted") via GitHub Copilot, different field (eager_input_streaming)
• Copilot: output_config.effort "max" is not supported by model claude-opus-4.6 #22788: Related output_config issue with GitHub Copilot rejecting unsupported output_config values

If these don't fully cover your case, please add a note on what's different so maintainers can track it separately.

[jimicze]

@github-actions no, its different bug

[jimicze]

Response to Duplicate Check

After thoroughly investigating all three referenced issues, I can confirm this is not a duplicate of any of them. Here is a detailed breakdown:

---

#23500 — claude-opus-4.7 + thinking.type.enabled not supported

Superficially similar: Same model (claude-opus-4.7) and same provider (github-copilot).

Why it is NOT the same:

• Different error message: claude-opus-4.7 via GitHub Copilot fails with thinking.type.enabled not supported #23500 errors with "thinking.type.enabled" is not supported for this model. Use "thinking.type.adaptive" and "output_config.effort" to control thinking behavior. Our issue errors with output_config: Extra inputs are not permitted.
• Different trigger condition: claude-opus-4.7 via GitHub Copilot fails with thinking.type.enabled not supported #23500 only occurs when using a reasoning/thinking variant (effort level high or max). Our issue occurs on every single request, regardless of reasoning settings — no thinking/reasoning mode was configured at all.
• Different root cause: claude-opus-4.7 via GitHub Copilot fails with thinking.type.enabled not supported #23500 is caused by claude-opus-4.7 being missing from the isAnthropicAdaptive list in packages/opencode/src/provider/transform.ts. OpenCode sends the legacy thinking.type=enabled format, but opus-4.7 requires thinking.type=adaptive. The fix would be to add opus-4.7 to that list. Our issue is about output_config being sent unconditionally (even without any reasoning mode), and the model rejecting the entire field as unknown.
• Different API field rejected: claude-opus-4.7 via GitHub Copilot fails with thinking.type.enabled not supported #23500 rejects thinking.type=enabled; ours rejects output_config itself.

In short: #23500 is a reasoning-mode misconfiguration. Our issue is output_config being sent to a model that does not accept it at all, even without reasoning.

---

#23342 — Haiku 4.5 + eager_input_streaming: Extra inputs are not permitted

Superficially similar: Same error pattern (Extra inputs are not permitted).

Why it is NOT the same:

• Different model entirely: 1.4.10 breaks Haiku through GitHub Copilot #23342 affects Claude Haiku 4.5, not claude-opus-4.7.
• Different rejected field: 1.4.10 breaks Haiku through GitHub Copilot #23342 is about tools.0.custom.eager_input_streaming — a field injected into the tool definitions in the request body. Our issue is about output_config — a top-level field in the request body.
• Different OpenCode version / regression: 1.4.10 breaks Haiku through GitHub Copilot #23342 was introduced in OpenCode 1.4.10 as a regression (worked in 1.4.9). Our issue is not version-specific in that way.
• Different root cause: 1.4.10 breaks Haiku through GitHub Copilot #23342 is caused by OpenCode adding eager_input_streaming to tool schemas, which GitHub Copilot's proxy strips or rejects. Our issue is about a top-level output_config parameter being sent unconditionally.

In short: The only commonality is the generic Pydantic error text Extra inputs are not permitted. The field, model, and root cause are all different.

---

#22788 — claude-opus-4.6 + output_config.effort "max" not supported

Superficially similar: Both involve output_config and GitHub Copilot.

Why it is NOT the same — this is the most important distinction:

• Fundamentally different API rejection type:
  • Copilot: output_config.effort "max" is not supported by model claude-opus-4.6 #22788: The output_config field is accepted by the API. The error is about an invalid value inside it: output_config.effort "max" is not supported; supported values: [low medium high]. The field itself is fine, just the value max is disallowed.
  • Our issue: The output_config field is entirely rejected as an unknown/extra field (output_config: Extra inputs are not permitted). The model does not accept output_config at all, regardless of its value.
• Different model: Copilot: output_config.effort "max" is not supported by model claude-opus-4.6 #22788 is claude-opus-4.6; ours is claude-opus-4.7.
• Different implication for fix: Copilot: output_config.effort "max" is not supported by model claude-opus-4.6 #22788 can be fixed by capping effort to high instead of max. Our issue requires not sending output_config at all for claude-opus-4.7 in non-reasoning contexts.
• Different trigger: Copilot: output_config.effort "max" is not supported by model claude-opus-4.6 #22788 occurs specifically when the user selects the max reasoning variant. Our issue occurs on every request with no reasoning mode selected.

In short: Same field name, completely different error semantics. #22788 = wrong value in an accepted field. Ours = field not accepted at all.

---

Summary Table

	#23500	#23342	#22788	Our issue #24527
Model	claude-opus-4.7	claude-haiku-4.5	claude-opus-4.6	claude-opus-4.7
Provider	github-copilot	github-copilot	github-copilot	github-copilot
Rejected field	thinking.type	eager_input_streaming	output_config.effort (value)	output_config (field itself)
Error type	Wrong format	Unknown field in tools	Invalid value	Unknown top-level field
Requires reasoning mode	Yes (high/max)	No	Yes (max)	No — fails always
Root cause	Missing model in isAnthropicAdaptive list	eager_input_streaming added to tool schemas	max effort not supported by copilot	output_config sent unconditionally without reasoning
Fix	Add opus-4.7 to adaptive list	Remove eager_input_streaming from tools	Cap effort at high	Omit output_config when model rejects it / not in reasoning mode

---

Additional Technical Detail for Our Issue

From the OpenCode log (~/.local/share/opencode/log/2026-04-26T174704.log):

• The error occurs twice per message attempt (first try + immediate retry), both failing identically
• The request body includes max_tokens: 32000 — this is the orchestrator agent, not a simple query
• No thinking/reasoning was configured in the session
• The error is thrown synchronously before any tokens are streamed: +1699ms / +1869ms — immediate API rejection, not a mid-stream failure
• Stack trace points to chunk-3f8tjbdd.js:6:14583 (bundled OpenCode core)
• After two failures the session enters idle state and the user must retry manually

The fix should conditionally omit output_config from the request when:

1. No reasoning/thinking mode is active, OR
2. The model is claude-opus-4.7 (or similar) which does not accept this field without reasoning context, OR
3. Implement graceful retry without output_config upon receiving this specific error response

[elsammykins]

can confirm I'm seeing the exact same issue with 1.14.26

[rekram1-node]

How are you hitting it? Does github copilot not support setting reasoning for opus 4.7 at all?

[thesammykins]

How are you hitting it? Does github copilot not support setting reasoning for opus 4.7 at all?

I'm under the impression they've removed being able to set reasoning for opus 4.7, even via copilot-cli they've removed being able to change it.

[jimicze]

There was an option to set it only to medium.

But users are pissed off... By this limitation, they were unable to see this option, because it's the only one, and set 'medium' as the default, hiding the options to see it / change it.

[jimicze]

Clarification for @rekram1-node — How We Hit It + Config Details

To directly answer your question: No, GitHub Copilot no longer supports setting reasoning for claude-opus-4.7 at all.

How We Were Hitting It

At the time of the bug, our OpenCode config had:

"agent": {
  "plan": {
    "model": "github-copilot/claude-opus-4.7",
    "reasoningEffort": "medium"
  }
}

reasoningEffort: "medium" — not because we chose medium, but because it was the only option GitHub Copilot exposed for claude-opus-4.7. There was no high or max available for this model via Copilot.

This causes OpenCode to include output_config in the API request body. GitHub Copilot then silently removed all reasoning support for claude-opus-4.7, including medium — and now rejects output_config entirely with Extra inputs are not permitted.

The Timeline

1. GitHub Copilot exposed only medium reasoning effort for claude-opus-4.7
2. Users set reasoningEffort: "medium" as it was the only available option
3. GitHub Copilot removed reasoning support entirely for claude-opus-4.7 — including medium
4. Now any request with output_config in the body is rejected, breaking all sessions using this model with any reasoningEffort setting

Why This Is an OpenCode Issue (Not Just a User Config Error)

The failure mode is silent and confusing:

• OpenCode accepts reasoningEffort: "medium" in config without warning
• When GitHub Copilot removes support, the session fails with a cryptic API error
• There is no fallback, no warning, no graceful degradation
• The user has no indication from the error that removing reasoningEffort from config would fix it

Suggested Fix

1. Retry without output_config: On receiving Extra inputs are not permitted for output_config, retry without it and emit a user-visible warning
2. Capability detection: Track which models/providers support output_config and skip it when not supported
3. Config validation warning: When reasoningEffort is set for a model/provider known not to support it, warn at startup rather than failing silently at runtime

Workaround for Affected Users

Remove reasoningEffort from the claude-opus-4.7 agent config in opencode.json:

"agent": {
  "plan": {
    "model": "github-copilot/claude-opus-4.7"
  }
}

This stops OpenCode from sending output_config in the request and the model works normally.

[rekram1-node]

@jimicze undo whatever config change u made to get around bug and show me opencode debug config I think a plugin or soemthing is causing errors for you, output_config should work

[jimicze]

@rekram1-node Here is opencode debug config:

{
  "$schema": "https://opencode.ai/config.json",
  "agent": {
    "plan": {
      "model": "github-copilot/claude-opus-4.7",
      "reasoningEffort": "medium",
      "options": {
        "reasoningEffort": "medium"
      },
      "permission": {}
    },
    "build": {
      "model": "github-copilot/claude-sonnet-4.6",
      "reasoningEffort": "high",
      "options": {
        "reasoningEffort": "high"
      },
      "permission": {}
    },
    "general": {
      "model": "github-copilot/claude-sonnet-4.6",
      "reasoningEffort": "high",
      "options": {
        "reasoningEffort": "high"
      },
      "permission": {}
    },
    "explore": {
      "model": "github-copilot/claude-sonnet-4.6",
      "reasoningEffort": "medium",
      "options": {
        "reasoningEffort": "medium"
      },
      "permission": {}
    }
  },
  "provider": {
    "lmstudio": {
      "name": "LM Studio (local)",
      "npm": "@ai-sdk/openai-compatible",
      "options": {
        "baseURL": "http://127.0.0.1:1234/v1"
      },
      "models": { ... }
    }
  },
  "mcp": {
    "playwright": { "type": "local", "command": ["npx", "@playwright/mcp@latest"] },
    "context7": { "type": "local", "command": ["npx", "-y", "@upstash/context7-mcp@latest"] }
  },
  "compaction": { "auto": true, "prune": true, "reserved": 20000 },
  "plugin": [],
  "username": "xxxx"
}

No plugins installed ("plugin": []). Only MCP servers (Playwright and Context7).

The only non-default things are:

• Custom agents with reasoningEffort settings
• LM Studio as a local provider (not used for claude-opus-4.7)
• Two MCP servers

[rekram1-node]

This may fix your issue:
0acac21

I'm just not really sure how this is even possible?

Maybe you can send me a session you hit it in or soemthing?

[jimicze]

Session has some extra log?

[jimicze]

Here's the relevant session log excerpt (two attempts — first try + immediate retry, both fail identically):

Attempt 1 (+1699ms — immediate API rejection):

ERROR service=llm providerID=github-copilot modelID=claude-opus-4.7 agent=orchestrator mode=primary stream
ERROR service=llm providerID=github-copilot modelID=claude-opus-4.7 agent=orchestrator mode=primary
  url: https://api.githubcopilot.com/v1/messages
  requestBodyValues: { model: "claude-opus-4.7", max_tokens: 32000, output_config: { effort: "medium" }, ... }
  AI_APICallError: output_config: Extra inputs are not permitted
ERROR service=session.processor error=output_config: Extra inputs are not permitted
  stack: AI_APICallError at chunk-3f8tjbdd.js:6:14583

Retry (+1869ms — identical failure):

ERROR service=llm providerID=github-copilot modelID=claude-opus-4.7 agent=orchestrator mode=primary
  url: https://api.githubcopilot.com/v1/messages
  requestBodyValues: { model: "claude-opus-4.7", max_tokens: 32000, output_config: { effort: "medium" }, ... }
  AI_APICallError: output_config: Extra inputs are not permitted
ERROR service=session.processor error=output_config: Extra inputs are not permitted
  stack: AI_APICallError at chunk-3f8tjbdd.js:6:14583

Both attempts reject instantly before any tokens stream. Session goes idle after the retry.

Also — does the fix in 0acac21 address this? From reading the patch it looks like it now reads reasoning_effort from the Copilot API response dynamically, so if Copilot reports no reasoning support for opus-4.7, no variants are built and output_config won't be sent. That matches exactly what we're seeing.