Implement configurable Karl context compression strategy (heuristic + LLM summarizer)

[YulYen]

Motivation

• Provide a configurable context-compression system to replace the ad-hoc karl_prepare_quick_and_dirty trimming and allow an LLM-based summarization strategy while keeping the original heuristic available.
• Make all behavior explicit in config.yaml with no hidden defaults or silent fallbacks.
• Keep the implementation KISS, modular, and testable with clear error paths when LLM summarization fails.

Description

• Added a new context_management section to config.yaml with strategy: "heuristic" | "karl" and a required karl sub-section containing model, summary_max_tokens, keep_last_messages, and log_dir.
• Implemented src/core/context_summarizer.py introducing KarlSummarizer (pure service) and KarlSummarizationError, which accept an injected llm_core and raw config and implement summarize(messages: list[dict]) -> list[dict] returning one system summary plus the tail without mutating the original list.
• Integrated strategy routing into WebUI and TerminalUI so that heuristic preserves existing trimming and karl invokes KarlSummarizer; LLM failures raise KarlSummarizationError and only fall back to heuristic when fallback_strategy: heuristic is explicitly configured.
• Karl writes its own daily log files logs/karl_YYYY-MM-DD.log containing timestamp, number of messages summarized, summary character length, and model used, without altering global logging configuration.

Testing

• Added tests/test_context_summarizer.py (validates reduction, tail preservation, no-mutation, model/options, and log creation) and updated tests/test_web_ui.py / tests/test_terminal_ui.py to include context_management and to assert strategy switching and non-instantiation of Karl under heuristic.
• Ran pytest -q tests/test_context_summarizer.py tests/test_web_ui.py tests/test_terminal_ui.py and all tests passed (22 passed).
• The new behavior enforces explicit config validation paths and raises clear exceptions when required config keys are missing or when summarization fails without an explicit fallback.

---

Codex Task

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 061d84d55f

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve persona context options in Karl summary calls

The Karl summarization request is sent with options={"num_predict": ...} only, so persona-specific options like num_ctx are dropped for this call even though context compression is triggered specifically in near-limit scenarios. In setups where personas run with a larger num_ctx than the backend default, summarization can fail or truncate before compression happens, causing KarlSummarizationError (and a hard failure when no fallback is configured). Pass through relevant persona options (at least num_ctx) when building the summary request options.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Guard Karl log writes from breaking successful summaries

After a summary is successfully produced, _append_log_entry performs filesystem writes without error handling; if log_dir cannot be created or appended to (e.g., read-only path, permission issues, disk full), the turn raises and fails despite having a valid summary. Because this logging is non-critical telemetry, these I/O failures should be caught and downgraded (e.g., warning) so context compression still completes.

Useful? React with 👍 / 👎.