[BUG] Thinking blocks stripped from in-prompt context on every turn in active Opus 4.7 sessions (no idle-resume trigger)

Found 2 possible duplicate issues:

1. https://github.com/anthropics/claude-code/issues/10199
2. https://github.com/anthropics/claude-code/issues/32810

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

Auto-classifier flagged #10199 and #32810 as possible dupes — neither is a match. Different failure modes that share the word "thinking":

#10199 is about thinking blocks being immutable post-send: system tried to modify a previously-sent thinking block, API rejected with 400, session corrupted, loud error. The failure surfaces as an unrecoverable error loop. This issue's failure surfaces silently with no API error — the model just receives a prompt with prior thinking blocks already removed and has no signal that anything was stripped.

#32810 is about "thinking" field being empty "" in the JSONL since 2.1.72 — content not persisted to disk. This issue is the inverse: empirically, the JSONL on disk for the affected session contained 174 thinking blocks totaling ~219K chars (~55K tokens), all with full populated content (verified by grep -c '"type":"thinking"' + reading sample blocks). The persistence layer is working correctly; the prompt-construction layer is stripping them when assembling the next turn's context.

Diagnostic check that disambiguates from both:

F=~/.claude/projects/<project>/<session-id>.jsonl
python3 -c "
import json
total = 0; empty = 0
for line in open('$F'):
    d = json.loads(line)
    msg = d.get('message',{})
    c = msg.get('content')
    if not isinstance(c, list): continue
    for x in c:
        if isinstance(x, dict) and x.get('type')=='thinking':
            total += 1
            if not x.get('thinking',''): empty += 1
print(f'thinking blocks: {total}, empty: {empty}, populated: {total-empty}')
"

For my session: thinking blocks: 174, empty: 0, populated: 174. Persistence intact. Then ask the model to introspect its current context for prior thinking blocks — model reports zero visible. That's the gap this issue documents: persisted-on-disk → 0-in-prompt-context, with no idle-resume trigger to explain it under #53513's documented mechanism.

Not a dupe. Leaving open.

Followup empirical finding — the persistence layer is already showing post-API-summarization content, not raw reasoning. The bug this issue describes affects an already-reduced trace, and even fixing it would only restore summaries to context, not the actual model-side reasoning.

On-disk content shape on Opus 4.7

Same session JSONL inspected for content depth. 204 thinking blocks with non-empty content. Length distribution:

| percentile | chars |
|---|---|
| 10th | 168 |
| 25th | 300 |
| 50th | 772 |
| 75th | 1522 |
| 90th | 3210 |
| max | 10034 |

A median ~770-char "thinking block" is structurally a planning summary, not raw reasoning. The longest blocks (~10K chars / ~2500 tokens) read as polished response-plans ("the user wants X, the real work is Y, what I need to do is Z..."), not exploratory reasoning trace — no false starts, no weighing of rejected options, no mid-thought pivots.

Connection to #49268

#49268 confirms Opus 4.7 changed the API default for display from "summarized" to "omitted". With showThinkingSummaries: true the harness presumably opts back into summarized, which is what we see persisted. The model-side raw reasoning is discarded at the API layer regardless — summarized returns short post-hoc summaries, not the full thinking trace.

The full failure stack

1. Model produces extended thinking during turn generation (substantial, exploratory, includes false starts).
2. API summarizes it before returning. Raw reasoning is discarded at this layer.
3. JSONL persists the summary — what's inspected above.
4. Prompt-construction layer strips even the summaries from subsequent turns — the bug this issue documents.
5. Rendering layer compresses-with-stitch-markers what Claude Code displays to the operator.

So this issue's fix (restoring summaries to subsequent prompt context) is necessary but not sufficient for full reasoning continuity. The underlying gap — that Claude Code never persists or restores the actual model-side reasoning, only API-side summaries — would require display: "detailed" (or equivalent) support.

Could be split into a separate issue if the maintainers prefer; flagging here so the layered nature is on the record while triage of this report is in progress.

Oddly, I'm still seeing this bug in some sessions, but not others. Currently leaving a session open & idle for a few hours to see if I can reproduce the issue.

I don't think it's model-specific; in the bugged session, thinking continues to be lost when switching to Sonnet 4.6.

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

Still happening with Opus 4.8 too.

Exiting and resuming with claude --continue appears to (at least temporarily) fix the issue.

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.