Harness spec 3.1.0: the prefix stays still, in the CLI

The system prompt is made once per session (start, compaction, model
switch), kept in session_meta and resent byte for byte; it holds no
date. What changes mid-session -- the date, a mode the person chose, a
plan or memory or instruction file they edited, a personality, MCP
instructions that arrived later -- is a <session-note> on the next user
message, stored with it. The tool list is fixed for the session:
plan_submit is always offered and answers unavailable outside plan
mode, the wrap-up keeps the tools and runs no call, a withdrawn tool
stays listed and answers its refusal; the first request waits up to
mcp_wait_ms for MCP servers. Compaction asks for its summary as the
next turn of the same conversation. Cached input tokens are parsed
(incl. llama.cpp timings), stored per request and shown in /context.

Spec: context.md §7/§10, loop.json context.note (+ plan_ended,
plan_own, personal, empty, mcp_wait_ms), modes.json plan.refused,
compact.md without the transcript heading. Conformance area prefix.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

src/session/context.ts

@@ -41,7 +41,7 @@ export function contextBreakdown(engine: Engine, system: string, instructions: {
   if (imageCount) rows.push({ label: `images (${imageCount}, rough)`, tokens: by.images })
   const estimated = rows.filter((r) => !r.indent).reduce((n, r) => n + r.tokens, 0)
   const reported = engine.usage ? engine.contextUsed() : undefined
-  return { rows, estimated, reported, window: contextWindow(engine.model), model: engine.model.ref }
+  return { rows, estimated, reported, window: contextWindow(engine.model), model: engine.model.ref, last: engine.lastRequest }
 }
 
 /** Plain text, for the transcript. */
@@ -58,5 +58,14 @@ export function renderBreakdown(b: ReturnType<typeof contextBreakdown>): string
   const used = b.reported ?? b.estimated
   lines.push(`  ${"in use".padEnd(33)} ${fmt(used).padStart(7)}  ${bar(used)}${b.window ? `  ${Math.round((used / b.window) * 100)}%` : ""}`)
   lines.push(b.reported !== undefined ? "  (the server's own count from the last reply, plus an estimate since)" : "  (estimated: ≈4 characters a token — no reply yet to count from)")
+  // What the server reused of the last request (harness context.md §10 rule 6): a slow reply with
+  // little of it cached was a re-read.
+  const last = b.last?.usage
+  if (last && !last.estimated)
+    lines.push(
+      last.cached !== undefined
+        ? `  last request${b.last!.kind === "compaction" ? " (compaction)" : ""}: ${fmt(last.input)} in, ${fmt(last.cached)} of it cached${last.input ? ` (${Math.round((last.cached / last.input) * 100)}%)` : ""}`
+        : `  last request: ${fmt(last.input)} in — the server did not say how much was cached`,
+    )
   return lines.join("\n")
 }