[EPIC] Context window management

[ypldan]

Goal

Stop context overflow from being a dead end. Today a long conversation silently consumes the model's context window until the app stops working (#180), the user has no visibility into how close they are (#351), and the agent wastes tokens regenerating tool output it already produced (#184). The pieces are already being built as a unit — the history-compaction design PR (#385) is stacked on the context-usage branch — this epic makes that structure explicit.

Scope — child issues

• Display context window usage indicator during conversation #351
• History compaction #180
• Optimized tool response -> tool request flow #184
• Dynamic tool discovery #430

Out of scope

• Long-term memory across conversations — compaction may later offload into it, but persistence is a separate capability (design PRs feat: add memory architecture design and implementation documentation #134/feat: add design document for external memory provider integration #375 track it).

Dependencies and sequencing

• #351 first — usage accounting is both a user-facing feature and the trigger signal compaction needs; PR feat: history compaction #385 already stacks on its branch.
• #180 second, building on the accounting; the compaction strategy itself is still at the investigation stage.
• #184 is independent enough to run in parallel, but shares the token-economy motivation.
• #430 is likewise independent (design doc first); it attacks the pre-conversation share of the context budget that Display context window usage indicator during conversation #351/History compaction #180 don't touch.

Additional information

Related: PR #385 (compaction design, stacked on feat/context_window_usage), #421 (skills epic — skills add system-prompt weight, another consumer of the context budget).