[Sync] llm stuff

common/.config/claude/statusline.py

@@ -3,7 +3,8 @@
 
 One line: vim mode | dir (repo-relative, hidden at the root) |
 git(repo/branch +dirty +ahead/behind) | model (+effort) |
-context% (+compact warn) | 5h usage (+reset eta) | cost (+burn rate) | disk. If the
+context% (+compact warn) | 5h usage (+reset eta) | cost (+burn rate) |
+prompt-cache hit% (+ttl, cold marker) | disk. If the
 rendered line would overflow the terminal, trailing (lowest-priority) segments are
 dropped until it fits.
 (Dormant, re-addable in main(): ram, cpu%, temp -- retired 2026-07-26 because the
@@ -89,7 +90,24 @@ def term_cols():
         return 0
 
 
-# --- shared transcript usage ------------------------------------------------
+# --- context usage ------------------------------------------------------------
+_USAGE_KEYS = ("input_tokens", "cache_creation_input_tokens", "cache_read_input_tokens")
+
+
+def payload_usage(data):
+    """`context_window.current_usage` from the status payload, shaped like `message.usage`.
+
+    Claude Code 2.1.x hands the status line its own token accounting, so the transcript
+    scan below is only a fallback for payloads that predate it (or lack a transcript).
+    """
+    cw = data.get("context_window")
+    cu = cw.get("current_usage") if isinstance(cw, dict) else None
+    if isinstance(cu, dict) and any(isinstance(cu.get(k), (int, float)) for k in _USAGE_KEYS):
+        return cu
+    return {}
+
+
+# --- shared transcript usage (fallback) ------------------------------------------
 def _scan_usage(text):
     """Last `message.usage` block among these JSONL lines, or {}."""
     usage = {}
@@ -113,7 +131,7 @@ TRANSCRIPT_TAIL = 1 << 20  # 1 MiB
 def last_usage(data):
     """Most recent `message.usage` block from the transcript, or {}.
 
-    Reads only the tail: this runs on every status refresh and transcripts grow to
+    Fallback only -- see payload_usage(). Reads only the tail: this runs on every status refresh and transcripts grow to
     tens of MB, so re-parsing the whole file each time is unbounded work for a value
     that lives at the end. Falls back to a full scan if the tail holds no usage block
     (a long stretch of tool results can push it past the window).
@@ -292,6 +310,11 @@ def context_segment(data, usage):
         )
         window = context_window(data)
         pct = used / window * 100 if window else 0
+        # Prefer Claude Code's own percentage when the payload carries it (2.1.x) -- it
+        # already accounts for the auto-compact reserve and window quirks we can't see.
+        cw = data.get("context_window")
+        if isinstance(cw, dict) and isinstance(cw.get("used_percentage"), (int, float)):
+            pct = float(cw["used_percentage"])
         col = bucket(pct, 50, 80)
         out = "📝 " + color(f"{pct:.0f}%", col) + color(f" ({used / 1000:.0f}k)", "gray", dim=True)
         if pct >= 80:
@@ -353,20 +376,31 @@ def velocity_segment(data):
         return ""
 
 
-def cache_segment(usage):
+def cache_segment(usage, data=None):
+    """Prompt-cache hit % (higher is better) + the cache TTL, `·cold` when the cache has
+    expired. Hit % from the 2.1.260+ payload `prompt_cache.hit_ratio` (a 0..1 fraction, null
+    while all counts are zero), else computed from the usage block; ttl/cold always from the
+    payload when present. A 1h TTL is what keeps a 10-minute pr-loop cadence warm."""
     try:
-        if not usage:
+        pc = (data or {}).get("prompt_cache")
+        pc = pc if isinstance(pc, dict) else {}
+        pct = None
+        if isinstance(pc.get("hit_ratio"), (int, float)):
+            pct = float(pc["hit_ratio"]) * 100
+        elif usage:
+            read = usage.get("cache_read_input_tokens", 0)
+            total = usage.get("input_tokens", 0) + usage.get("cache_creation_input_tokens", 0) + read
+            if total > 0:
+                pct = read / total * 100
+        if pct is None:
             return ""
-        read = usage.get("cache_read_input_tokens", 0)
-        total = (
-            usage.get("input_tokens", 0)
-            + usage.get("cache_creation_input_tokens", 0)
-            + read
-        )
-        if total <= 0:
-            return ""
-        pct = read / total * 100
-        return "♻️ " + color(f"{pct:.0f}%", bucket(pct, 50, 80, invert=True))
+        out = "♻️ " + color(f"{pct:.0f}%", bucket(pct, 50, 80, invert=True))
+        ttl = pc.get("ttl")
+        if isinstance(ttl, str) and ttl:
+            out += color(f" {ttl}", "gray")
+        if pc.get("warm") is False:
+            out += color(" ·cold", "yellow")
+        return out
     except Exception:
         return ""
 
@@ -532,7 +566,7 @@ def main():
         or data.get("cwd")
         or os.getcwd()
     )
-    usage = last_usage(data)
+    usage = payload_usage(data) or last_usage(data)
     top = git_top(cwd)
 
     # One row: work (left) then system (right), joined. If it would overflow the
@@ -540,7 +574,7 @@ def main():
     # Dormant helpers kept above for easy re-add: ram_segment, cpu_segment,
     # temp_segment (retired -- they only refresh on message events, so they sat
     # stale; the tmux status bar owns system metrics now), velocity_segment,
-    # cache_segment, api_segment, version_segment, style_segment.
+    # api_segment, version_segment, style_segment.
     work = [s for s in (
         vim_segment(data),
         dir_segment(cwd, top),
@@ -549,6 +583,7 @@ def main():
         context_segment(data, usage),
         usage_segment(data),
         cost_segment(data),
+        cache_segment(usage, data),
     ) if s]
     system = [s for s in (
         disk_segment(cwd),