Add compact plugin for auto context compression

[GoDiao]

Summary

Closes #249

Adds a compact plugin that automatically compresses conversation context when it exceeds a configurable token budget, inspired by Claude Code's /compact mechanism.

How it works

1. Token budget check - Estimates token count of the conversation; if below threshold, passes through unchanged (zero overhead)
2. Context window detection - Tries to get context window from provider's /models endpoint (e.g. Ollama, vLLM expose context_length), then falls back to config
3. Split regions - Older turns are compressed via LLM into a structured summary; recent N turns are preserved verbatim
4. Structured summary - The LLM produces a summary with Scope, Key decisions, User preferences, Pending work, Key files, and Context
5. Graceful fallback - If LLM compression fails, returns original query unchanged

Composability

Works with the & operator for pipeline composition:

compact&moa    -> compact first, then Mixture of Agents
compact&bon    -> compact first, then Best of N

Configuration

Config	Env Var	Default	Description
compact_context_window	COMPACT_CONTEXT_WINDOW	128000	Max context tokens
compact_threshold	COMPACT_THRESHOLD	0.75	Trigger ratio (0.0-1.0)
compact_keep_recent	COMPACT_KEEP_RECENT	4	Turns to preserve verbatim

Tests

• 32 unit/integration tests in tests/test_compact_plugin.py
• Plugin registration test added to tests/test_plugins.py

50 tests passed (18 plugin + 32 compact)

End-to-end verification

Tested with a live API (MiMo):

• 31-turn conversation (10,119 tokens) compressed to 580 tokens (94% reduction)
• 29 older turns summarized, 2 recent turns preserved verbatim
• Context window detection: provider /models returned 400, gracefully fell back to config default

Test plan

• All existing plugin tests pass (18/18)
• Unit tests for token estimation, conversation parsing, config priority
• Integration tests for passthrough, compression, LLM failure fallback
• Edge cases: summary tag extraction, malformed env vars, embedded tags
• Manual end-to-end test with live API - plugin loads and executes correctly

[CLAassistant]

All committers have signed the CLA.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[GoDiao]

Hi! The CI failures in conversation-logging-tests and integration-tests appear to be unrelated to this PR. They're failing due to HuggingFace authentication issues when trying to access the gated model google/gemma-3-270m-it:

openai.InternalServerError: Error code: 500 - You are trying to access a gated repo.
Make sure to have access to it at https://huggingface.co/google/gemma-3-270m-it.

The unit-tests and test_plugins checks passed successfully. These failures seem to be a CI environment issue (missing HuggingFace token) rather than anything introduced by this change. Happy to fix if there's anything on our side though!