feat: compact conversation (/compact) — summarise history to save context

[VforVitorio]

Problem

Long sessions accumulate tokens in the chat history until the model's
context window is full and quality degrades or responses are cut off.

Proposed behaviour — /compact

/compact

  summarising conversation…
  ✓ compressed 42 messages → 1 summary (saved ~3.2k tokens)

The command:

1. Sends the current chat history to the model asking it to produce
   a compact summary of what has been discussed and decided
2. Replaces Agent._chat with a new Chat containing only the
   system prompt + one assistant message with the summary
3. Prints how many messages were compressed and estimated token savings

UI for context fullness (separate sub-feature)

Show context usage next to the stats line after each response:

↑ 1.2k  ↓ 384  ·  45 tok/s  ·  2.3s   [████░░░░░░ 38% ctx]

Implementation note

Token counting for the context bar uses the accumulated
_session_prompt_tok + _session_pred_tok from issue #29 divided
by the model's get_context_length() (from issue #2 implementation).

[VforVitorio]

Additional UX requirements (2026-03-17)

Compacting spinner

While the summarisation is running, replace the standard thinking spinner with a dedicated one:

⠋ compacting…

Use a different spinner style (e.g. "bouncingBar" or "aesthetic") so the user can tell at a glance that compaction — not a normal model turn — is in progress.

Preview box

After compacting, show a styled summary panel before returning to the prompt:

┌─ compacted ──────────────────────────────────────────────────────────────────┐
│  42 messages → 1 summary                                                     │
│                                                                               │
│  The user asked about X, Y, Z. Key decisions: A, B. Open questions: C.      │
│  [… truncated at 300 chars …]                                                 │
│                                                                               │
│  saved ~3.2k tokens  ·  context now at 18%                                   │
└──────────────────────────────────────────────────────────────────────────────┘

The panel uses ACCENT border and shows the generated summary text (truncated if long).

Context usage line

The footer of the panel should include the updated context window percentage (ties in with issue #41 — context usage indicator).

[VforVitorio]

Implemented in v0.3.0. \ summarises the full conversation via the model, resets the chat, injects the summary as context, and shows a panel with the summary preview and message count. Uses a \ spinner while compacting.