Epic: server-side context management for preserved-thinking models #5410

Activity

1. jimbrub

   added

   enhancementNew feature or requestNew feature or request

   area-contextSession or context relatedSession or context related

   area-modelRelated to models or model providersRelated to models or model providers

   on Oct 22, 2026
2. github-actions

   added

   design

   on Oct 22, 2026
3. github-actions

   added the

   Design

   issue type on Oct 22, 2026
4. jimbrub

   self-assigned this

   on Oct 22, 2026
5. jimbrub

   changed the title [-]Epic: server-side context management for preserved-thinking models[/-] [+]EPIC: server-side context management for preserved-thinking models[/+] on Oct 22, 2026
6. jimbrub

   changed the title [-]EPIC: server-side context management for preserved-thinking models[/-] [+]Epic: server-side context management for preserved-thinking models[/+] on Oct 22, 2026
7. jimbrub

   added sub-issues on Oct 26, 2026

   • [FEATURE] Add an overridable Model.compact() and route summarization through it #5464
   • [FEATURE] Core: providerData on TextBlock for session restore and model switches #5402
   • [FEATURE] Summarize through Anthropic native compaction #5401
   • [FEATURE] Test native compaction on other providers that serve preserved-thinking Claude models #5465
   • [DOCS] Summarization on preserved-thinking models #5405
   • [FEATURE] Gate server-side compaction through model metadata capabilities #5466
   • [FEATURE] Clear tool results server-side on preserved-thinking models #5406
   • [FEATURE] Allow role "system" messages mid-conversation #5176
   • [FEATURE] Send tool list changes as mid-conversation system messages #5407
   • [FEATURE] Clear old thinking server-side #5408
   • [FEATURE] OpenAI Responses server-side compaction #5409
8. yonib05

   added this to Strands Agents Roadmapon Oct 28, 2026
9. github-project-automation

   moved this to Intake in Strands Agents Roadmapon Oct 28, 2026
10. yonib05

    moved this from Intake to We're Working On It in Strands Agents Roadmapon Oct 28, 2026
11. jimbrub

    added sub-issues on Nov 2, 2026

    • [FEATURE] Separate native compaction from base summarization, replacing the placeholder Model.summarize() #5733
    • [FEATURE] Evaluate base summarization strategies across models before fixing defaults #5734
    • [FEATURE] Refine the base summarizer system prompt #5735
    • [FEATURE] Summarize after message 0 when the cut lands on the last message #5736
12. jimbrub commented on Nov 2, 2026

    

    jimbrub

    on Nov 2, 2026

    ContributorAuthor

    More actions

    Status update on Part 1 after #5470. There are no changes to the checklist above; this comment adds new tickets and one reorder.

    #5464 (#5470). It shipped as an experimental Model.summarize(), not compact(). Treat it as a placeholder: the final API shape is tracked in #5733. The as-built differences are noted on #5464. One of them is a behavior change for base models: message 0 is kept, and some low-preserve_recent passes are skipped.

    Reorder: #5466 before #5401. Model.model_info capabilities landed in #5313, so #5466 doesn't need to wait for #5305. Gating on a capability before the first summarize() override lands avoids the override-detection problems found in #5470's review. Details on #5466.

    New tickets (added as sub-issues):

    • [FEATURE] Separate native compaction from base summarization, replacing the placeholder Model.summarize() #5733: separate native compaction from base summarization, and replace the placeholder Model.summarize(). A summary is stored as a TextBlock with metadata.offloaded, plus providerData from [FEATURE] Core: providerData on TextBlock for session restore and model switches #5402 for native results.
    • [FEATURE] Evaluate base summarization strategies across models before fixing defaults #5734: evaluate base summarization strategies across models (keep message 0 or not, request shape, summarizer model, prompt, placement), including downstream task success over repeated passes. This also covers Part 1's "comparison of native and client-side summaries".
    • [FEATURE] Refine the base summarizer system prompt #5735: refine the base summarizer system prompt, chosen from the evaluation's results (blocked by [FEATURE] Evaluate base summarization strategies across models before fixing defaults #5734).
    • [FEATURE] Summarize after message 0 when the cut lands on the last message #5736: summarize after message 0 when the cut lands on the last message. This closes the skipped-pass gap feat(models): add Model.summarize() with Anthropic on-demand compaction #5470 introduced for preserve_recent ≤ 1.

    Order for Part 1 from here:

    1. [FEATURE] Core: providerData on TextBlock for session restore and model switches #5402 and [FEATURE] Gate server-side compaction through model metadata capabilities #5466, in parallel.
    2. [FEATURE] Summarize through Anthropic native compaction #5401, then [FEATURE] Include compaction calls in token usage #5404.
    3. [FEATURE] Evaluate base summarization strategies across models before fixing defaults #5734, then [FEATURE] Refine the base summarizer system prompt #5735.
    4. [FEATURE] Separate native compaction from base summarization, replacing the placeholder Model.summarize() #5733, informed by the evaluation and by [FEATURE] Summarize through Anthropic native compaction #5401's override.
    5. [FEATURE] Summarize after message 0 when the cut lands on the last message #5736 at any point.