[Copilot Chat Bug] HTTP 400 "thinking blocks cannot be modified" after conversation compaction — paid tokens wasted, session unrecoverable

[JUN-WEBLISS]

Summary

After conversation compaction occurs in a long agent session, the session crashes with HTTP 400:

thinking or redacted_thinking blocks cannot be modified
messages.N.content.N: input is invalid

This completely destroys the active session and all work-in-progress is lost. As a paid subscriber, this is unacceptable — I am paying for tokens and getting nothing in return when this error strikes.

---

Environment

• VS Code version: 1.120+ (GitHub Copilot Chat extension, latest)
• Model: Claude Sonnet 4.6 (via GitHub Copilot)
• OS: Windows 11
• Setting applied: "github.copilot.chat.agent.thinkingTool": false

---

Steps to Reproduce

1. Start a long agent session (multi-file edits, tool calls, etc.)
2. VS Code's automatic conversation compaction triggers
3. Compaction strips or modifies thinking / redacted_thinking blocks from prior assistant messages
4. On next tool call or agent turn, Anthropic's API rejects the modified message history with HTTP 400

---

Expected Behavior

One of the following should hold:

• Conversation compaction must preserve thinking / redacted_thinking blocks verbatim — they cannot be touched
• OR the compaction should gracefully exclude those blocks without altering them, in a way the API accepts
• OR if compaction is impossible without breaking the conversation, the agent should warn the user before proceeding

---

Actual Behavior

• No warning is given
• Compaction silently corrupts the message history
• The next agent turn crashes with HTTP 400
• All context, in-progress work, and paid token investment in that session is lost
• There is no retry mechanism — the session is permanently broken

---

Impact

• Paid tokens are burned with zero output — this is a billing integrity issue
• Long agent sessions (the primary use case for GitHub Copilot Pro/Max subscribers) are fundamentally unreliable
• Workarounds (e.g., thinkingTool: false, manual session breaks) are insufficient because:
  • thinkingTool: false only prevents new thinking blocks — it does not protect against blocks already in history
  • If the setting was not active from the very first turn, the session remains vulnerable

---

Requested Actions

1. Urgent fix: Ensure conversation compaction never modifies thinking / redacted_thinking blocks — preserve them byte-for-byte or omit them in a valid way per Anthropic API spec
2. Billing guarantee: When a session is crashed by a Copilot-internal bug (not user error), tokens consumed in that session should not count against the user's quota
3. User notification: If a session cannot be safely compacted, notify the user before compaction rather than crashing mid-session

---

Additional Notes

I am forced to manually implement a checkpoint escape protocol in my own copilot-instructions.md to proactively bail out of sessions before this error occurs. The fact that users need self-documented escape hatches for a first-party tool reliability bug is not acceptable for a paid product.

Please treat this as a P0 reliability bug. Long-running agent sessions are the primary selling point of Copilot's premium tiers. If they cannot be trusted, the product's core value proposition is broken.

[Diego-Mendes-Forter]

I recently upgraded to 1.121 and started getting the same issue, see the error message below:

Sorry, your request failed. Please try again.

Copilot Request id: 6182d2dc-da99-4e5d-a597-0260b24a72a8

GH Request Id: 64B1:33C011:37F720:3C401D:6A0F3823

Reason: Request Failed: 400 {"type":"error","error":{"type":"invalid_request_error","message":"messages.1.content.1: thinking or redacted_thinking blocks in the latest assistant message cannot be modified. These blocks must remain as they were in the original response."},"request_id":"req_011CbG4TV177SCavvKwCG1ou"}

Environment

Version: 1.121.0 (Universal)
Commit: f6cfa2e
Date: 2026-05-19T11:51:53+02:00
Electron: 39.8.8
ElectronBuildId: 13870025
Chromium: 142.0.7444.265
Node.js: 22.22.1
V8: 14.2.231.22-electron.0
OS: Darwin arm64 25.4.0

Metadata

requestType      : ChatMessages
model            : claude-sonnet-4.6
maxPromptTokens  : 167997
maxResponseTokens: 32000
location         : 7
otherOptions     : {"stream":true}
intent           : undefined
startTime        : 2026-05-21T16:51:48.960Z
endTime          : 2026-05-21T16:51:49.445Z
duration         : 485ms
ourRequestId     : 6182d2dc-da99-4e5d-a597-0260b24a72a8
requestId        : 6182d2dc-da99-4e5d-a597-0260b24a72a8
serverRequestId  : 64B1:33C011:37F720:3C401D:6A0F3823
tools (77)       : create_file, fetch_webpage, file_search, ....

Steps to Reproduce

• Start a new emtpy agent session and ask the agent to execute a tool call that will generate a large output
• VS Code's will receive the tool response and try to load into the context but it is too large for the context
• Copilot fail to continue the process and output the error messages.1.content.1: thinking or redacted_thinking blocks in the latest assistant message cannot be modified. These blocks must remain as they were in the original response."
• Follow up message will fail, session must be restarted from scratch

Additional Notes

I was previous on the version 1.116 and 1.118 and the same workflow was working as expected, including on large results.
If I am not mistaken, the previous versions, the tool result would be output to file and then parse it, but seems like the behaviour changed to ingest the output into the context directly.

Unfortunately I can't put an example attached because of sensitive data, but generating a dummy mcp that output a large json should likely help.

[martinzhang]

I've been hitting this exact bug repeatedly. My log shows the same 400 error — thinking or redacted_thinking blocks cannot be modified — triggered during conversation compaction in Agent mode with Claude Opus. The session crashes, work is lost, and the auto-retry just hammers the same failed request 3 times before giving up.

I'm a paying Copilot subscriber. This is not a minor edge case — it's a core reliability failure in the flagship Agent mode feature. Long-running agent sessions are precisely what the premium tier is supposed to deliver.

I want to be direct: the AI coding assistant market is intensely competitive right now. Cursor, Windsurf, and Cline are all gaining ground fast. The reason developers stick with Copilot is trust in Microsoft's execution — and bugs like this, left unresolved across multiple issues (#292394, #302859, #282240), erode that trust quickly.

Users should not have to resort to workarounds like "github.copilot.chat.agent.thinkingTool": false or writing escape protocols in copilot-instructions.md to paper over a first-party reliability bug.

Please treat this as P0. Fix the conversation compaction logic to properly preserve thinking blocks, or disable Extended Thinking for models where it cannot be safely managed. Don't let this be the reason developers walk away.

[martinzhang]

Update — same session, second bug hit minutes later.

After the thinking blocks crash I reported above, I started a new Agent session. This time compaction triggered and caused a different failure: complete context wipe.

The log shows Compacted conversation appearing twice in quick succession, and the Agent responded with:

"I don't see a specific question or task in your message — just the workspace context. What would you like me to do?"

There was an active task running. The Agent simply forgot everything mid-execution.

These are two different symptoms of the same broken compaction logic:

• Bug 1 (this issue): compaction corrupts thinking blocks → HTTP 400 → session crash
• Bug 2 (also Urgent: Copilot Agent Context loss after "Summarized conversation history" #251250): compaction discards task context → Agent amnesia → work lost silently

Both happen in normal-length Agent sessions. Both result in lost work. The compaction feature is supposed to extend sessions — right now it's killing them.

This needs to be fixed end-to-end, not just the thinking block handling in isolation.