v0.14.3 — The context window is now real

The context window — how much conversation and code the model can see — is MandoCode's most important reliability setting, and until now it only took effect when MandoCode started the Ollama engine itself. Most users run the Ollama desktop app, whose already-running engine silently ignored the setting. This release makes it real in every setup: the window now travels with every chat request (num_ctx), overriding the desktop app's slider and Ollama's small ~4k default, applied from your next message with no restarts.

Live testing on small models drove two more fixes riding along:

• Conversations compact before they overflow. Ollama never errors on an oversized prompt — it silently discards the oldest content, including the model's own instructions, which surfaced as empty or confused replies that looked like the model breaking. MandoCode now measures each outgoing request first and, when it's near the limit, summarizes older history into a recap and tells you so in the reply — visible, graceful degradation instead of silent failure.
• The context floor is now 16k (under-7B models; 7B+ stays at 32k). At 8k, MandoCode's instructions and tool definitions consumed nearly the whole window — small models were living on ~1.5k tokens of working memory and inventing the rest. 16k costs a small model well under 1 GB and gives it ~6× the memory.

Upgrading from 0.13.0? You also get everything from the unpublished 0.14.0:

• MandoCode remembers — mandocode --continue / -c. Every conversation saves itself automatically at the end of each turn; close the terminal mid-task and pick up exactly where you left off, tool calls included. /clear deletes the stored session too.
• Date grounding. The system prompt now carries today's date, so genuine current events in search results are no longer dismissed as fabricated "future" content.

Also in this release:

• Music playlists actually advance. A track no longer plays once and goes silent forever while the player claims it's still playing (NAudio's WaveChannel32 never reports end-of-stream by default; it does now).
• Agent identity. Hosts can name an AI instance and the system prompt speaks as that name — built for MandoCode Desktop's agent tabs, where "the app" and "the voice in this tab" are different things. The CLI keeps the classic MandoCode identity, byte-for-byte.
• Guidance that described the old context mechanism — the in-chat "context window filled" notice, /learn, --config show, the README — now describes the new one: raise contextLength (live from your next message) or /clear; no restarts involved.

Update: dotnet tool update -g MandoCode

Full details: docs/CHANGELOG.md. 505/505 tests passing.