[BUG] No working mechanism to review previous context after compaction, plan-mode clear, or branch navigation — data preserved but UI inaccessible #27242

Activity

1. scapeshift-ojones commented on Feb 23, 2026

   

   scapeshift-ojones

   on Feb 23, 2026

   Author

   More actions

   Related: #26123 partially fixed but still broken (49 👍, closed prematurely)

   #26123 (/resume broken) was closed as "completed" after v2.1.47 fixed the picker limit and Windows case sensitivity. But two of the original bugs remain unfixed as of v2.1.50:

   Bug 1 — Index staleness: sessions-index.json is not updated for new sessions. @vbhavsar reports 68 .jsonl files on disk but only 25 indexed. The v2.1.47 workaround (scanning .jsonl directly up to 50) masks this, but users with 50+ sessions per project still can't find their work. Tracked separately in #25032.

   Bug 4 — /rename names invisible in picker: @ThatDragonOverThere reports that custom session names set via /rename display correctly in terminal tab titles but do not appear in the resume picker — it shows auto-generated summaries instead. With 6+ concurrent sessions, there's no way to tell which session is which. Confirmed on v2.1.50.

   These compound directly with the bugs in this issue:

   Layer	Problem	Issue
   Find sessions	Index staleness — sessions not indexed	#26123 Bug 1
   Identify sessions	/rename names missing from picker	#26123 Bug 4
   Navigate branches	Right-arrow does nothing in picker	This issue, Bug 3
   View history	No pre-compaction content in transcript mode	This issue, Bug 1
   Link plan sessions	No connection to predecessor session	This issue, Bug 2

   The resume picker is the single entry point to all historical context, and it's degraded at every level. Even if you manage to find and open a session, you can't browse its branches or view what happened before compaction.

   #26123 has 49 👍 and should be reopened — the partial fix in v2.1.47 didn't address the most user-visible symptoms. I attempted to reopen but the repo restricts this to maintainers.
2. scapeshift-ojones commented on Feb 25, 2026

   

   scapeshift-ojones

   on Feb 25, 2026

   Author

   More actions

   Hey anyone who is coming to this thread please comment how its affecting you so we can get Anthropic to pay attention to this and solve the entire suite of these regressions once and for all rather than inconsistent whac-a-mole style
3. chickensintrees commented on Feb 25, 2026

   

   chickensintrees

   on Feb 25, 2026

   More actions

   Just hit this tonight. 509-turn session died at context overflow during a save operation. The JSONL file is 24MB of perfectly preserved data — every tool call, every correction, every branch point. All sitting on disk. Zero way to access it through the UI.

   I've built custom recovery tooling (session-rescue.sh) that parses the JSONL to extract last state, files written, and corrections from crashed sessions. It works, but the fact that I had to build it at all is the bug. The data layer is flawless. The access layer is missing.

   My specific pain points, in order:

   1. Post-compaction Ctrl+O shows nothing useful. I rely heavily on compaction for long creative sessions. After compact, the pre-compaction reasoning is gone from the UI. I've resorted to writing manual state summaries to disk before every compact — essentially reimplementing what the transcript viewer should do.

   2. Crashed sessions are black boxes without custom tooling. When a session hits context overflow mid-operation, the only recovery path is grep/jq on raw JSONL. The data is all there. Branch trees intact. parentUuid linkage perfect. Just no way to see it.

   3. Plan-mode context loss. When accepting a plan with "clear context," the exploration session that informed the plan becomes a separate, unlinkable JSONL. The architectural decisions that led to the plan are orphaned.

   The poster is right — this is purely a UI/rendering gap. The data engineering is solid. Would love to see even a basic read-only transcript viewer that loads the full JSONL and renders it chronologically. That alone would eliminate hours of manual recovery work per week.
4. stevenpetryk commented on Feb 25, 2026

   

   stevenpetryk

   on Feb 25, 2026 · Pinned by stevenpetryk

   Last edited by stevenpetryk

   More actions

   Thanks for reporting all this, appreciate the feedback. I'm going to raise these to the team.

   Plan-mode context loss. When accepting a plan with "clear context," the exploration session that informed the plan becomes a separate, unlinkable JSONL. The architectural decisions that led to the plan are orphaned.

   Minor correction here: "Clear context" is actually more like "implement plan in new conversation"—you can /resume to go back to the session in which you were iterating on the plan. I think the action naming could be clearer (should probably say "Implement plan in new conversation" or similar to avoid this frequent source of confusion).

   👍React with 👍4Reacted by Martin Atrin, Jed Tabaczynski, Ben Weinberg and JCJinvesting
5. anthropics

   deleted a comment from mantrahq502 on Feb 25, 2026
6. vonscons commented on Feb 26, 2026

   

   vonscons

   on Feb 26, 2026

   More actions

   I’m consistently loosing last 15 minutes of work, especially when there are multiple subagents in background
7. tassa-yoniso-manasi-karoto commented on Feb 27, 2026

   

   tassa-yoniso-manasi-karoto

   on Feb 27, 2026

   Last edited by tassa-yoniso-manasi-karoto

   More actions

   every... time... every time I update there are majors bugs and/or regressions. Claude Code has been steadily like this for year... idk what to say. Idk how people still have the patience to dig into minified JS and draft well-crafted issues like these when the devs constantly prove they have no regard for stability. They simply don't deserve these efforts.

   reverting to 2.1.19 fixed it for me.

   👍React with 👍4Reacted by rublev, Mevolove, metrovoc and Guanwen Xie
8. syncletica commented on Feb 27, 2026

   

   syncletica

   on Feb 27, 2026

   More actions

   Adding another scenario to this: background agents doing web searches.

   When background agents are running, they often keep searching well beyond what I believe is needed. I can see they've already visited enough sources and I'm ready to move forward, but there's no way to say "stop and give me what you have." If I interrupt, the main agent reports it has no results from those agents. All work is lost.
9. ianbmacdonald commented on Mar 5, 2026

   

   ianbmacdonald

   on Mar 5, 2026

   More actions

   I just noticed this on 2.1.69 .. I was away from claude for a week; On return, I though ghostty lost my scrollback settings .. something changed as I could scrollback before compaction previously .. this creates an impossible situation to supervise claude. My claude-cli auto-updates in place .. what is the best way to roll-back avoiding the deprecated npx method?
10. ThatDragonOverThere commented on Mar 6, 2026

    

    ThatDragonOverThere

    on Mar 6, 2026

    More actions

    Confirming on v2.1.69 — Windows (MSYS2/Git Bash)

    Version: Claude Code v2.1.69, Windows 11 (MSYS_NT-10.0-26200), MSYS2/Bash shell, Node.js runtime

    Ctrl+E and Ctrl+O do not show session history from before compaction. The keybindings either do nothing or toggle minor UI elements (timing badges), but pre-compaction conversation content is inaccessible.

    Why this matters beyond convenience

    This compounds with compaction's context loss (#9796). For users running long-lived sessions with complex state (in my case, a trading system with multi-phase build plans, custom hooks, and plan files), the failure chain is:

    1. Compaction loses context — project instructions deprioritized, active plan state lost
    2. Post-compact hooks deprioritized — even with a SessionStart hook that re-injects rules (plan file path, sub-agent workflow, RAM safety), the model treats hook output as lower priority than its compressed summary
    3. Ctrl+E/O broken — user cannot manually review pre-compact state to verify what was lost or re-orient the session

    The net effect is that after compaction, the user has no recovery path at all. The automated recovery (hooks) is unreliable, and the manual recovery (Ctrl+E/O history browsing) doesn't work. The only option is parsing raw JSONL, which is not practical mid-session.

    Echoing @ianbmacdonald's comment above — this creates an impossible supervision situation for long-running sessions.

    👍React with 👍1Reacted by Aphillyated52
11. scapeshift-ojones commented on Mar 7, 2026

    

    scapeshift-ojones

    on Mar 7, 2026

    Author

    More actions

    Thanks for reporting all this, appreciate the feedback. I'm going to raise these to the team.

    Plan-mode context loss. When accepting a plan with "clear context," the exploration session that informed the plan becomes a separate, unlinkable JSONL. The architectural decisions that led to the plan are orphaned.

    Minor correction here: "Clear context" is actually more like "implement plan in new conversation"—you can /resume to go back to the session in which you were iterating on the plan. I think the action naming could be clearer (should probably say "Implement plan in new conversation" or similar to avoid this frequent source of confusion).

    Sev2 issue existing for over a month now, clearly affecting a bunch of users, does anyone at Anthropic care? What is going on here this is an easy fix.

    👀React with 👀4Reacted by Susanne (Susi) Keller, Ollie from Wind, Ben Weinberg and Abhishu Oza
12. abhishuoza commented on Mar 11, 2026

    

    abhishuoza

    on Mar 11, 2026

    Last edited by abhishuoza

    More actions

    Thank you for detailing this bug, it's been the most frustrating one for me.

    Another temporary workaround: Use /fork every time you need a session up to a specific point saved.
    Assume by default that the branch of the session you see in front of you is the only branch that you can access again via the TUI. You won't have them all grouped in a tree like they should be, but you can still find it as a different session at the same location.

    Unless I'm mistaken, I think this works.

    Kinda stupid and annoying especially given how easy this seems to be to fix.
13. goldngreen commented on Mar 24, 2026

    

    goldngreen

    on Mar 24, 2026

    More actions

    This is the worst possible time to clear the screen. The consequence is that, if Claude has forgotten something following the compact, the user cannot manually scroll to remind themself. A --no-clear-on-compact option would be a big improvement.

    👍React with 👍1Reacted by Franek Antoniak
14. yurukusa commented on Mar 29, 2026

    

    yurukusa

    on Mar 29, 2026

    More actions

    Hook workaround: export conversation to markdown before compaction
    Until the TUI adds a way to browse pre-compaction history, a PreCompact hook can export the full conversation to a readable markdown file before compaction runs:
    pre-compact-transcript-export.sh:

    INPUT=$(cat)
    TRANSCRIPT=$(echo "$INPUT" | jq -r '.transcript_path // empty' 2>/dev/null)
    if [ -z "$TRANSCRIPT" ] || [ ! -f "$TRANSCRIPT" ]; then
        TRANSCRIPT=$(find ~/.claude/projects -name "transcript.jsonl" -mmin -60 2>/dev/null | tail -1)
    fi
    [ -z "$TRANSCRIPT" ] || [ ! -f "$TRANSCRIPT" ] && exit 0
    SNAPSHOT_DIR="$HOME/.claude/conversation-snapshots"
    mkdir -p "$SNAPSHOT_DIR"
    OUTPUT="$SNAPSHOT_DIR/$(date -u +%Y%m%d-%H%M%S).md"
    {
        echo "# Conversation Snapshot ($(date -u '+%Y-%m-%d %H:%M UTC'))"
        echo ""
        while IFS= read -r line; do
            ROLE=$(echo "$line" | jq -r '.message.role // empty' 2>/dev/null)
            case "$ROLE" in
                user)
                    echo "## User"
                    echo "$line" | jq -r '.message.content // empty' 2>/dev/null | head -20
                    echo "" ;;
                assistant)
                    echo "## Assistant"
                    echo "$line" | jq -r '[.message.content[]? | select(.type=="text") | .text] | join("\n")' 2>/dev/null | head -50
                    echo "" ;;
            esac
        done < "$TRANSCRIPT"
    } > "$OUTPUT"
    echo "Snapshot saved: $OUTPUT" >&2
    exit 0

    Install in .claude/settings.json:

    {
      "hooks": {
        "PreCompact": [
          {
            "hooks": [
              {
                "type": "command",
                "command": "bash /path/to/pre-compact-transcript-export.sh"
              }
            ]
          }
        ]
      }
    }

    Every time compaction fires, the full pre-compaction conversation gets saved as ~/.claude/conversation-snapshots/YYYYMMDD-HHMMSS.md. You can read these files anytime to review what was discussed before compaction.
    This doesn't fix the TUI limitation, but it gives you a persistent, human-readable record that survives compaction.
15. khaledh commented on Apr 3, 2026

    

    khaledh

    on Apr 3, 2026

    Last edited by khaledh

    More actions

    Another subtle case is when compaction is immediately followed by reaching the session limit. You have no recourse (unless you buy extra credits) to go back and check the conversation history, and have to wait until the next session window to ask Claude for a summary or previous details.
16. christophervankammen commented on Apr 6, 2026

    

    christophervankammen

    on Apr 6, 2026

    More actions

    Its silly this is an issue? Its compaction data should be sent to a simple backend folder, file or internal memory if its going to compact, chat data should just be held locally and what is seen in the editor/extension/cli is simply a copy of that in an interactive wrapper?
17. wowitsjack commented on Apr 8, 2026

    

    wowitsjack

    on Apr 8, 2026

    Last edited by wowitsjack

    More actions

    Fix for this is over at #44772
18. frackham commented on Jul 13, 2026

    

    frackham

    on Jul 13, 2026

    More actions

    will this also fix the multiple claude remote issues of continuing history after disconnect that keep being closed? If so great :)
19. bcherny

    added

    area:tui

    on Aug 20, 2026
20. bestiaya commented on Sep 9, 2026

    

    bestiaya

    on Sep 9, 2026

    More actions

    Not a fix, and not a TUI answer. For anyone hitting this on long-running jobs: I stopped depending on the transcript being reviewable. State for a job lives in one markdown file outside the context window, written as decisions land; when a session compacts or ends, the next one reads that file instead of the history. It doesn't recover your pre-compaction context — the bug is real and I'd still like it fixed — it removes the dependency on it. Six of my own cases: 93,073 characters of state, 53,339 actually read on takeover. Self-measured, n=6.
21. vshulcz commented on Oct 14, 2026

    

    vshulcz

    on Oct 14, 2026

    More actions

    compaction doesn't touch the file, the turns before it are still in ~/.claude/projects//.jsonl (the compact summary even prints that path), and this dumps them in order:

    jq -r 'select(.type=="user" or .type=="assistant") | .message.content | if type=="string" then . else map(select(.type=="text").text) | join("\n") end' <file>

    i built a small cli that indexes those files so you can search across sessions too: https://github.com/vshulcz/deja-vu. deja show <id> prints the whole session, pre-compaction turns included.
22. hjk-sys commented on Oct 28, 2026

    

    hjk-sys

    on Oct 28, 2026

    More actions

    Workaround until this lands: compacted messages are still in ~/.claude/projects. npx decompact shows them all, offline. (I built it.) https://github.com/hjk-sys/decompact
23. pearonjames commented on Nov 1, 2026

    

    pearonjames

    on Nov 1, 2026

    More actions

    Adding the VS Code extension to this report, which was filed against the CLI. I'm on extension 2.1.314 on Windows. Two VS Code reports, #53091 and #79042, were auto-closed for inactivity and are now locked, so I'm commenting here. #102361 describes the same problem in the Desktop Code tab.

    I reopened a conversation in the VS Code panel and could no longer scroll to its earlier messages. The panel stopped at the point where the session had compacted, and the compaction row expands to the summary and nothing else. I had no warning that compaction would do this.

    The messages still exist. They're all in the session file on my machine, and the sessions doc says of a compacted session: "The messages that the last compaction summarized aren't loaded again." I searched the settings reference, the environment variable reference, the VS Code docs, and the extension's own settings, and found no option that changes this.

    To read my own conversation, I had Claude write a script that rebuilds a second, read-only copy of any compacted session with the compaction removed. It runs at session start and adds the copy to the sessions list as a separate entry. I work in one thread and scroll back in the other. I can't send messages in the copy, because that would compact it again, and the script depends on the internal session file format, so an update could break it. The copy also shows the panel can display the whole conversation: I scrolled it from the first message onward.

    The Claude web and desktop apps handle this better. Anthropic's help page says the Claude app also summarizes long chats near the limit, and that "Your full chat history is preserved so Claude can reference it even after summarization." I can open any long conversation there and scroll to the top.

    I also can't plan around this. The conversation that compacted didn't feel long, because by character count my messages and Claude's replies made up about 2% of what filled the context. File reads and tool output made up 81%, and Focus view hides those. Turning auto-compact off doesn't help either: from what I've read, the thread then stops at the limit and can't write its own handoff.

    Please load the full conversation history into the panel when a compacted session reopens, or add a setting that does. Reading my own conversation shouldn't take a workaround.

    👍React with 👍1Reacted by Kat