Context Summarization Plugin

When a conversation is about to exceed the model's context window, this plugin
shrinks the prompt so the turn can still run. It keeps system prompts and the
most recent messages, and — by default — asks a model to summarize the middle
instead of dropping it silently. If summarization is off or fails, it falls
back to truncation.

It also retries after a context-window error from the model, using the same
shrink-and-retry path so a long conversation is more likely to continue than
to fail outright.

Configuration

plugins:
  - module: codumentor.plugins.context_summarization
    class: ContextSummarizationPlugin
    args:
      enabled: true
      max_context_tokens: 128000
      safety_margin: 1000
      min_recent_messages: 5
      split_ratio: 0.7
      enable_summarization: true

Parameters

• enabled (optional): Whether the plugin is active (default: true)
• max_context_tokens (optional): Context limit used when the model has no entry in model_context_limits (default: 128000)
• safety_margin (optional): Tokens reserved for the model's reply (default: 1000)
• min_recent_messages (optional): Minimum number of recent conversation messages to keep uncompressed (default: 5)
• split_ratio (optional): Fraction of the older conversation eligible for summarization / removal, from 0.0 to 1.0 (default: 0.7)
• enable_summarization (optional): Summarize the dropped middle with an LLM instead of only deleting it (default: true)
• summarization_prompt_template (optional): Prompt template used for that summary (default: "context_summarization.jinja2")
• min_messages_for_summary (optional): Do not bother summarizing unless at least this many messages would be condensed (default: 3)
• model_context_limits (optional): Per-model context limits. Omitted by default (empty). Example:

args:
  model_context_limits:
    "Qwen3-235B-A22B-Instruct-2507": 262144

Aggressive truncation for small models

plugins:
  - module: codumentor.plugins.context_summarization
    class: ContextSummarizationPlugin
    args:
      enabled: true
      max_context_tokens: 16385
      safety_margin: 500
      min_recent_messages: 3
      split_ratio: 0.8

What you see

When a turn is summarized or truncated, a transcript card reports how many
messages were condensed, how many were kept, and roughly how many tokens were
saved. A pending card appears while summarization is running.

Each user can hide those cards in Settings → Extensions → Context
Summarization (Show Summarization Cards, on by default). This is a web-UI
setting; the TUI does not read it yet
(doc/core-dev/improvements/plugin-settings.md).

/compact

Type /compact at the start of the composer (or pick it from the / list)
to create or refresh the compacted context without sending a model turn.
The transcript stays the source of truth; compaction is derived state used by
later prompt assembly.

The command reuses this plugin's summarization, cache, token estimates, and
event card. If the conversation is too short, or summarization is disabled,
it returns a notice and does nothing. Disabled plugins do not advertise the
command.

Notes

• System messages are always kept. Recent messages are kept up to min_recent_messages; older content is summarized (or removed).
• Summaries from earlier turns are reused when the older history has not changed, so you are not charged a fresh summarization call every turn.
• For a proactive, every-turn compression scheme with topic-shift detection, see Context Management.