[BUG] Thinking block content empty (thinking_len=0) since v2.1.69 — only encrypted signature returned

Correction: The platform is Claude Max subscription (not direct API). Claude Code is authenticated via Max plan, not API keys. This may be relevant if thinking summary behavior differs between Max and API access paths.

Found 2 possible duplicate issues:

1. https://github.com/anthropics/claude-code/issues/30958
2. https://github.com/anthropics/claude-code/issues/31143

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

This is a duplicate of #30958 which reports the exact same regression (v2.1.68 → v2.1.69, thinking content empty, only signature returned). That issue has been labeled as regression and bug by maintainers.

Our report adds additional data points:

• Confirmed on macOS + Warp (not just Linux)
• Confirmed on Max subscription (not just API)
• Confirmed across two separate machines with the same account
• Tested on both v2.1.69 and v2.1.70 — both affected

Happy to close this as duplicate if appropriate.

I'm starting to think this is on purpose, THEY ALWAYS MESS UP thinking

Found 2 possible duplicate issues: 1. [[BUG] v2.1.69: thinking summaries empty in transcript and TUI — undocumented behavior change #30958](https://github.com/anthropics/claude-code/issues/30958) 2. [[FEATURE] Persist summarized thinking text to session JSONL #31143](https://github.com/anthropics/claude-code/issues/31143) This issue will be automatically closed as a duplicate in 3 days. If your issue is a duplicate, please close it and 👍 the existing issue instead To prevent auto-closure, add a comment or 👎 this comment 🤖 Generated with Claude Code

not a duplicate

I also noticed the same, but since Anthropic hasn't documented this yet, I asked Claude to dig into this, below is the report:

"""

How redact-thinking works in Claude Code v2.1.69+

The mechanism

Starting in v2.1.69 (March 5, 2026), Claude Code sends a redact-thinking-2026-02-12 beta header with every API request. This header tells the Anthropic API to strip the text content from all thinking blocks before returning them to the client. The thinking blocks still appear in the response with valid cryptographic signatures (proving thinking occurred), but the thinking property is an empty string.

The code path (v2.1.71)

Step 1: Beta header injection

When building API request parameters, Claude Code constructs the betas array. The relevant condition (deminified):

if (
  hasThinking &&                                          // thinking is enabled
  modelSupportsInterleavedThinking(model) &&              // model supports it
  !isVerboseMode() &&                                     // not in verbose/transcript mode
  getSettings().showThinkingSummaries !== true &&          // setting is NOT explicitly true
  getFeatureFlag("tengu_quiet_hollow", false)              // feature flag (defaults to false)
) {
  betas.push("redact-thinking-2026-02-12");
}

The showThinkingSummaries setting defaults to undefined (not documented, not set by default). Since undefined !== true evaluates to true, the condition passes and the redact header is always included unless the user explicitly sets "showThinkingSummaries": true in ~/.claude/settings.json.

Note: The tengu_quiet_hollow feature flag check uses p8("tengu_quiet_hollow", false) where false is the default — meaning when the flag is not set, it defaults to false, and the condition false would block the beta. However, this flag appears to be server-controlled and is currently returning true for users, enabling the redaction.

Step 2: API response

The API receives the redact-thinking-2026-02-12 beta header and returns thinking blocks like:

{
  "type": "thinking",
  "thinking": "",
  "signature": "EuEECkYICxgCKkBrmuins..."
}

The thinking text is empty. The signature is valid (proving the model did think).

Step 3: Rendering

The thinking content renderer (PL1 component) extracts the thinking property from the content block:

let { thinking } = contentBlock;
if (!thinking) return null;  // empty string is falsy → renders nothing

Since thinking is "" (empty string, falsy in JavaScript), the component returns null. Nothing is rendered — not in the live view, not in the transcript view (ctrl+o), nowhere.

The setting

// ~/.claude/settings.json
{
  "showThinkingSummaries": true
}

The setting's internal schema description is:

"Show thinking summaries in the transcript view (ctrl+o). Default: false."

This is misleading. The setting doesn't control a UI display toggle — it controls whether the API request includes a beta header that causes the server to strip thinking text from the response entirely.

Timeline

| Version | Date | Status |
|---------|------|--------|
| ≤ v2.1.63 | Feb 28 | No redaction. Thinking text sent normally. |
| v2.1.64 | Mar 3 | redact-thinking introduced |
| v2.1.66 | Mar 3 | Reverted (same day) |
| v2.1.68 | Mar 4 | Still reverted |
| v2.1.69 | Mar 5 | Re-introduced permanently |
| v2.1.70–72 | Mar 6–10 | Present in all versions |

Impact

• The ctrl+o transcript view (the official way to view thinking) shows nothing because the thinking text was never received
• alwaysThinkingEnabled: true still causes the model to think (and bill for thinking tokens), but the output is discarded before reaching the client
• The setting is not documented in the official settings documentation
• Users have no way to discover this setting exists unless they read the minified source code

Workaround

Add to ~/.claude/settings.json:

{
  "showThinkingSummaries": true
}

Then restart Claude Code.
"""

I also noticed the same, but since Anthropic hasn't documented this yet, I asked Claude to dig into this, below is the report: """ ## How redact-thinking works in Claude Code v2.1.69+ ### The mechanism Starting in v2.1.69 (March 5, 2026), Claude Code sends a redact-thinking-2026-02-12 beta header with every API request. This header tells the Anthropic API to strip the text content from all thinking blocks before returning them to the client. The thinking blocks still appear in the response with valid cryptographic signatures (proving thinking occurred), but the thinking property is an empty string. ### The code path (v2.1.71) Step 1: Beta header injection When building API request parameters, Claude Code constructs the betas array. The relevant condition (deminified): if ( hasThinking && // thinking is enabled modelSupportsInterleavedThinking(model) && // model supports it !isVerboseMode() && // not in verbose/transcript mode getSettings().showThinkingSummaries !== true && // setting is NOT explicitly true getFeatureFlag("tengu_quiet_hollow", false) // feature flag (defaults to false) ) { betas.push("redact-thinking-2026-02-12"); } The showThinkingSummaries setting defaults to undefined (not documented, not set by default). Since undefined !== true evaluates to true, the condition passes and the redact header is always included unless the user explicitly sets "showThinkingSummaries": true in ~/.claude/settings.json. Note: The tengu_quiet_hollow feature flag check uses p8("tengu_quiet_hollow", false) where false is the default — meaning when the flag is not set, it defaults to false, and the condition false would block the beta. However, this flag appears to be server-controlled and is currently returning true for users, enabling the redaction. Step 2: API response The API receives the redact-thinking-2026-02-12 beta header and returns thinking blocks like: { "type": "thinking", "thinking": "", "signature": "EuEECkYICxgCKkBrmuins..." } The thinking text is empty. The signature is valid (proving the model did think). Step 3: Rendering The thinking content renderer (PL1 component) extracts the thinking property from the content block: let { thinking } = contentBlock; if (!thinking) return null; // empty string is falsy → renders nothing Since thinking is "" (empty string, falsy in JavaScript), the component returns null. Nothing is rendered — not in the live view, not in the transcript view (ctrl+o), nowhere. ### The setting // ~/.claude/settings.json { "showThinkingSummaries": true } The setting's internal schema description is: > "Show thinking summaries in the transcript view (ctrl+o). Default: false." This is misleading. The setting doesn't control a UI display toggle — it controls whether the API request includes a beta header that causes the server to strip thinking text from the response entirely. ### Timeline Version Date Status ≤ v2.1.63 Feb 28 No redaction. Thinking text sent normally. v2.1.64 Mar 3 redact-thinking introduced v2.1.66 Mar 3 Reverted (same day) v2.1.68 Mar 4 Still reverted v2.1.69 Mar 5 Re-introduced permanently v2.1.70–72 Mar 6–10 Present in all versions ### Impact The ctrl+o transcript view (the official way to view thinking) shows nothing because the thinking text was never received alwaysThinkingEnabled: true still causes the model to think (and bill for thinking tokens), but the output is discarded before reaching the client The setting is not documented in the official settings documentation Users have no way to discover this setting exists unless they read the minified source code ### Workaround Add to ~/.claude/settings.json: { "showThinkingSummaries": true } Then restart Claude Code. """

Two things:

1. is there a way to be in verbose mode by default (i believe verbose mode is just ctrl+o,)? the problem is that if i do ctrl+o, the tui doesnt show what claude is doing , from the point when i pushed ctrl+o.
2. How did you do this exploration? how did you learn this? (may be wrong place to ask, but didnt knwo how to dm you)

The thinking doesn't seem to persist and only shows the current thinking. Past thinkings are deleted. Is there a way to persist them?

The thinking doesn't seem to persist and only shows the current thinking. Past thinkings are deleted.

yet another bug (or 'feature'):
https://github.com/anthropics/claude-code/issues/16965

they are really trying hard to hide those thinking messages from us. I have Claude reguarly patch the minified JS to restore them inline as they used to appear before they started thinkering with them: https://github.com/aleks-apostle/claude-code-patches/pull/9

there are several other community patches that do the same. Dozens of similar issues related to thinking messages have been closed automatically after inactivity. They haven't bothered explaining if this is intentional design/policy decision, but it is a fact that they are consistently restricting visibility of thinking messages.

The setting above ("showThinkingSummaries": true) worked for me. Possibly because it's coupled with verbose mode, I'm now getting the thinking blocks inline again, as desired.

> The thinking doesn't seem to persist and only shows the current thinking. Past thinkings are deleted. yet another bug (or 'feature'): #16965 they are really trying hard to hide those thinking messages from us. I have Claude reguarly patch the minified JS to restore them inline as they used to appear before they started thinkering with them: aleks-apostle/claude-code-patches#9 there are several other community patches that do the same. Dozens of similar issues related to thinking messages have been closed automatically after inactivity. They haven't bothered explaining if this is intentional design/policy decision, but it is a fact that they are consistently restricting visibility of thinking messages.

The purpose of a system is what it does. Since early February, thinking mode has anecdotally been broken ~50% of the time for me.

I posit two collectively exhaustive but not mutually exclusive possiblities:

• Unstable/unreliable thinking display is a pragmatic defense against distillation.
• Nobody at Anthropic has thought to write an integration test which sends a query with "alwaysThinkingEnabled": true,, then checks to see if the TUI shows thinking output after ctrl-o.

Between malice and negligence, I hope it is negligence.

Closing for now — inactive for too long. Please open a new issue if this is still relevant.

This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.