[BUG] Assistant's own response text + tool-result widget missing from model context in later turns

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Environment

• Claude Code (desktop app), model Opus 5, switched to Fable 5 mid-session via /model
• MCP: atlassian connector (Jira/Confluence), plus project skills
• Long session, heavy tool use, several very large tool results

Symptom
An assistant turn that (a) invoked a skill, (b) made MCP tool calls including a
createJiraIssue, and (c) ended with a summary line and a rendered ticket widget,
was later only partially visible to the model. In subsequent turns the model's view
of that turn ended at three earlier context-gathering tool results; the
createJiraIssue call, its result, the final assistant message ("Created and
verified: BD-1755 — ..."), and the ticket widget were all absent.

Impact
The model asserted twice that the Jira ticket had never been created, contradicting
the user, who could see both the summary line and the widget in the UI. Resolved only
by re-querying Jira through the API. In a workflow where the model reasons about what
it has already done (created tickets, published pages, applied edits), silent loss of
its own prior output causes confident false statements and risks duplicate side effects
(e.g. creating the same ticket twice).

Timeline anchors (from external systems, all UTC)

• 12:30:25 Confluence page update (visible in model context)
• 12:40:47 Jira BD-1755 created <-- turn whose tail went missing
• 12:56:03 Confluence page update (visible in model context)

So the lost content sits between two turns that remained fully visible, rather than at
a context boundary edge.

Possibly related
Earlier in the same session, two tool results were auto-persisted to files for being
oversized (a ~132KB Confluence fetch, a ~55KB shell output), so context management was
demonstrably active. Speculation only — but the loss pattern (assistant message + its
trailing tool call/result dropped, while surrounding turns survive intact) may point at
truncation/summarization not preserving the tail of a turn.

What Should Happen?

Expected
Assistant messages and tool calls that the user can still see in the transcript should
remain represented in the model's context, or the model should be signalled that a
portion was elided rather than seeing a silently shortened turn.

Error Messages/Logs

Steps to Reproduce

1. Ask the agent to create a Jira ticket.
2. Make sure it was created and there is a clear response from the agent about it.
3. Proceed with the conversation.
4. Later in the converstaion, ask the agent if the ticket was created.

Claude Model

Not sure / Multiple models

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

Claude 1.24012.11 (09114b) 2026-08-03T15:24:21.000Z

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

iTerm2

Additional Information

_No response_