[yusiwen]

What changed

• agent/compaction.go (new): compactRunMessages(ctx, messages, step) compacts the loop's own traffic
  at a step boundary once it crosses CompressionThreshold. Head (system prompt, history, this run's
  prompt) and the last eight messages stay verbatim; the rest becomes
  [COMPACTED RUN — REFERENCE ONLY]. The cut snaps forward to an assistant message, so a tool_calls
  message never loses its results. A failed summary logs and leaves the messages alone.
• agent/compression.go: summarizeText is now the single summary call site and it goes through
  recordCall, so the summary's tokens and cost reach UsageTotal, the budgets and the status bar —
  they were invisible before. describeMessages is shared by both compaction paths.
• agent/agent.go: the loop calls the compaction hook before the provider call.
• Pin persistence: chatMessage.Pinned → persistSession → historyFromSession; the TUI mirrors the
  agent's pins into the transcript by role and content (newest first), because the transcript also holds
  system and banner messages. PinnedTurnTexts is the view it mirrors from.
• Docs: CODEBASE.md (the two new functions, the account path), roadmap B5 (S1 and S2 landed, S3 and S4
  open), test count 1270 → 1277.

Why

Compaction ran only at the start of a run, on a.History. The messages the loop actually sends — every
step's assistant message and every tool result — were never compressed, so a long round grew until the
provider refused. A continuation round (D7) is one Run, which makes that the routine case rather than
the edge one.

The accounting gap is the same class of defect: the summary was a real provider call whose spend no
budget could see. Fixing it once, at the single call site, is cheaper than remembering it at each caller.

Verification

• go test ./... -count=1 → 16 packages ok; go vet and gofmt -l clean.
• Mutation check: the loop hook disabled → TestRunCompactsItsOwnMessages fails with
  no provider request carried a compacted run across 12 calls; restored, it passes.
• New tests: 5. Compaction summarizes the middle and keeps head and tail (and the summarizer's usage
  lands in UsageTotal); the cut never separates a tool call from its results; a short run is untouched
  with no provider call; compressHistory now counts its own call; and a full run through Run carries
  the compaction marker into a later request.
• TUI: /pin marks the transcript, persistSession writes the flag, historyFromSession restores it.
• git log --format='%h %G? %s' → G.

Honest scope

• The compaction threshold is the history threshold (compression_threshold, 500k by default): a run
  has to grow a long way before it compacts. That is deliberate for now — a lower run-level bound is a
  tuning decision with no measurement behind it.
• Path-targeted unload is not here. /drop <path> cannot act on the run's messages: they live inside
  Run, and the TUI refuses commands while a run is active. S2's acceptance sentence assumed a command
  channel that does not exist; the issue is corrected and the work is S4 (a pre-run rule compaction
  honours when it walks tool calls by path, or a tool the model calls).
• Pin persistence matches by role and content, so two turns with byte-identical text pin the newest one.
  Acceptable for a command a person types; noted rather than hidden.

Follow-ups

• [agent] Give context management turn-level control: pin, unpin, drop #306 S3 (a preview that calls no model) and S4 (path-targeted unload).
• [memory] Align the Hindsight adapter with a real deployment #302 (the Hindsight adapter's contract) still needs a real deployment.

Refs #306, #168