claude-code-cache-fix 3.7.1 → 3.9.0

package/proxy/extensions.json

@@ -1,20 +1,82 @@
 {
-  "bootstrap-defense": { "enabled": true, "order": 45 },
-  "ttl-tier-detect": { "enabled": true, "order": 75 },
-  "fingerprint-strip": { "enabled": true, "order": 100 },
-  "image-strip": { "enabled": true, "order": 150 },
-  "sort-stabilization": { "enabled": true, "order": 200 },
-  "fresh-session-sort": { "enabled": true, "order": 250 },
-  "identity-normalization": { "enabled": true, "order": 300 },
-  "smoosh-split": { "enabled": true, "order": 320 },
-  "content-strip": { "enabled": true, "order": 330 },
-  "tool-input-normalize": { "enabled": true, "order": 340 },
-  "microcompact-stability": { "enabled": true, "order": 350 },
-  "thinking-display": { "enabled": true, "order": 360 },
-  "cache-control-normalize": { "enabled": true, "order": 400 },
-  "messages-cache-breakpoint": { "enabled": true, "order": 410 },
-  "ttl-management": { "enabled": true, "order": 500 },
-  "cache-telemetry": { "enabled": true, "order": 600 },
-  "overage-warning": { "enabled": true, "order": 610 },
-  "request-log": { "enabled": false, "order": 700 }
+  "bootstrap-defense": {
+    "enabled": true,
+    "order": 45
+  },
+  "ttl-tier-detect": {
+    "enabled": true,
+    "order": 75
+  },
+  "fingerprint-strip": {
+    "enabled": true,
+    "order": 100
+  },
+  "image-strip": {
+    "enabled": true,
+    "order": 150
+  },
+  "sort-stabilization": {
+    "enabled": true,
+    "order": 200
+  },
+  "fresh-session-sort": {
+    "enabled": true,
+    "order": 250
+  },
+  "identity-normalization": {
+    "enabled": true,
+    "order": 300
+  },
+  "smoosh-split": {
+    "enabled": true,
+    "order": 320
+  },
+  "content-strip": {
+    "enabled": true,
+    "order": 330
+  },
+  "tool-input-normalize": {
+    "enabled": true,
+    "order": 340
+  },
+  "microcompact-stability": {
+    "enabled": true,
+    "order": 350
+  },
+  "thinking-display": {
+    "enabled": true,
+    "order": 360
+  },
+  "cache-control-normalize": {
+    "enabled": true,
+    "order": 400
+  },
+  "messages-cache-breakpoint": {
+    "enabled": true,
+    "order": 410
+  },
+  "ttl-management": {
+    "enabled": true,
+    "order": 500
+  },
+  "cache-telemetry": {
+    "enabled": true,
+    "order": 600
+  },
+  "overage-warning": {
+    "enabled": true,
+    "order": 610
+  },
+  "request-log": {
+    "enabled": false,
+    "order": 700
+  },
+  "usage-log": {
+    "enabled": true,
+    "order": 650
+  },
+  "rate-limit-log": {
+    "enabled": true,
+    "order": 660
+  }
 }

package/tools/MANUAL-COMPACT.md

@@ -10,10 +10,10 @@ When using the 1M context window hack (`DISABLE_COMPACT=1` + `CLAUDE_CODE_MAX_CO
 
 1. Extracts conversation turns from the session JSONL transcript
 2. Splits turns into three weighted segments:
-   - **Foundational** (first 20%) — truncated to 200 chars each
-   - **Working** (middle 40%) — truncated to 400 chars each
-   - **Active** (last 40%) — preserved up to 2000 chars each
-3. Sends the weighted extract to Claude Sonnet for summarization
+   - **Foundational** (first 20%) — truncated to 300 chars each
+   - **Working** (middle 40%) — truncated to 1500 chars each
+   - **Active** (last 40%) — preserved up to 8000 chars each
+3. Sends the weighted extract to Claude Opus for summarization
 4. Produces a structured summary optimized for agent handoff
 
 The weighting ensures recent active work (the part you're most likely to need) gets full detail, while earlier completed work is compressed.
@@ -142,7 +142,7 @@ Use the user context file to fill known gaps.
 
 Two costs to account for:
 
-1. **Summarization call** — the `claude --print` call through Sonnet. At ~50K extract tokens, expect ~1-2% Q5h.
+1. **Summarization call** — the `claude --print` call through Opus. With the relaxed recent-turn caps the extract is larger (and Opus costs more per token than Sonnet), so expect a few % Q5h rather than ~1-2%. The tradeoff buys markedly higher-fidelity summaries; override with `MANUAL_COMPACT_MODEL=claude-sonnet-4-6` if you need to minimize cost.
 2. **Cold start after /clear** — the first API call rebuilds the full cache from scratch. Real-world example from a 954K-token session:
 
 ```
@@ -153,11 +153,18 @@ Second call:    cache_read=957,253  cache_creation=5,569   (warm again)
 
 The cold rebuild consumed ~15% Q5h in one call on our Max 5x account. After that single rebuild, the session is warm again and cache hits resume at 99%+.
 
-**Total cost of a manual compact cycle:** ~17% Q5h (2% summarization + 15% cold rebuild). Compare to hitting the 1M wall and losing the session entirely.
+**Total cost of a manual compact cycle:** roughly ~15% cold rebuild plus a few % for the Opus summarization. Compare to hitting the 1M wall and losing the session entirely.
 
-### Requires Claude Sonnet access
+### Summarizer model
 
-The tool uses `claude --print --model claude-sonnet-4-6` for summarization. Sonnet is used instead of Opus to minimize Q5h impact. If Sonnet is unavailable, change the model in the script.
+The tool defaults to `claude --print --model claude-opus-4-7` for the highest-fidelity summary. Override with the `MANUAL_COMPACT_MODEL` env var — e.g. `MANUAL_COMPACT_MODEL=claude-sonnet-4-6` to minimize Q5h impact, or to point at a different model if Opus is rate-limited or retired.
+
+### Troubleshooting: empty summary output
+
+If `$OUTPUT` comes back empty, the most likely cause is that the extract exceeded the summarizer's context window — this tool runs near the 1M wall, and the relaxed recent-turn caps (active turns up to 8000 chars) make the extract large on exactly those big sessions. The summarizer call swallows stderr, so an oversized-input rejection surfaces as an empty file rather than a visible error. Fixes, in order of preference:
+
+- Use a 1M-window model for the summarization: `MANUAL_COMPACT_MODEL='claude-opus-4-7[1m]' manual-compact.sh ...`
+- Or lower the per-turn caps in the script's extraction block (the `text[:8000]` / `text[:1500]` / `text[:300]` slices).
 
 ## Why the 1M Hack Disables /compact
 

package/tools/manual-compact.sh

@@ -145,31 +145,33 @@ if total == 0:
     sys.exit(1)
 
 # Split into three segments with different detail levels:
-# - First 20%: truncate to 200 chars each (foundational context)
-# - Middle 40%: truncate to 400 chars each (working context)
-# - Last 40%: full text up to 2000 chars each (active work — most important)
+# - First 20%: truncate to 300 chars each (foundational context)
+# - Middle 40%: truncate to 1500 chars each (working context)
+# - Last 40%: full text up to 8000 chars each (active work — most important)
+# Recent-turn caps were relaxed (was 200/400/2000) so the summarizer sees the
+# active work in near-full detail; the stronger model (Opus, below) handles it.
 seg1_end = int(total * 0.2)
 seg2_end = int(total * 0.6)
 
 with open("$EXTRACT", 'w') as f:
     f.write("=== FOUNDATIONAL CONTEXT (early session) ===\n\n")
     for role, text in conversation[:seg1_end]:
-        f.write(f"[{role}]: {text[:200]}\n\n")
+        f.write(f"[{role}]: {text[:300]}\n\n")
 
     f.write("\n=== WORKING CONTEXT (mid session) ===\n\n")
     for role, text in conversation[seg1_end:seg2_end]:
-        f.write(f"[{role}]: {text[:400]}\n\n")
+        f.write(f"[{role}]: {text[:1500]}\n\n")
 
     f.write("\n=== ACTIVE WORK (recent — preserve in full detail) ===\n\n")
     for role, text in conversation[seg2_end:]:
-        f.write(f"[{role}]: {text[:2000]}\n\n")
+        f.write(f"[{role}]: {text[:8000]}\n\n")
 
 import os
 size = os.path.getsize("$EXTRACT")
 print(f"Extracted {total} turns ({size:,} bytes, ~{size//4:,} est. tokens)")
-print(f"  Foundational: {seg1_end} turns (truncated to 200 chars)")
-print(f"  Working: {seg2_end - seg1_end} turns (truncated to 400 chars)")
-print(f"  Active: {total - seg2_end} turns (up to 2000 chars)")
+print(f"  Foundational: {seg1_end} turns (truncated to 300 chars)")
+print(f"  Working: {seg2_end - seg1_end} turns (truncated to 1500 chars)")
+print(f"  Active: {total - seg2_end} turns (up to 8000 chars)")
 PYEOF
 
 # Build the summarization prompt
@@ -199,10 +201,14 @@ ADDITIONAL USER CONTEXT TO PRESERVE:
 $USER_CONTEXT"
 fi
 
+# Summarizer model. Defaults to Opus for highest-fidelity summaries; override
+# with MANUAL_COMPACT_MODEL (e.g. when Opus is rate-limited or retired).
+COMPACT_MODEL="${MANUAL_COMPACT_MODEL:-claude-opus-4-7}"
+
 echo ""
-echo "Sending to Claude for summarization..."
+echo "Sending to Claude ($COMPACT_MODEL) for summarization..."
 
-cat "$EXTRACT" | claude --print --model claude-sonnet-4-6 "$PROMPT" > "$OUTPUT" 2>/dev/null
+cat "$EXTRACT" | claude --print --model "$COMPACT_MODEL" "$PROMPT" > "$OUTPUT" 2>/dev/null
 
 SIZE=$(wc -c < "$OUTPUT")
 echo ""

package/tools/quota-statusline.sh

@@ -115,11 +115,13 @@ def draw_bar(consumed_pct, elapsed_pct, width=BAR_WIDTH):
     # Tick overlays a fill cell when consumed > elapsed, keeping bar width
     # constant — that's what makes the over-pace state legible (┃ inside the
     # filled run) rather than just pushing fill cells around.
-    fill = int(round(max(0, min(100, consumed_pct)) / 100 * width))
+    def to_cells(pct):
+        return int(round(max(0, min(100, pct)) / 100 * width))
+    fill = to_cells(consumed_pct)
     if elapsed_pct is None:
         tick = -1
     else:
-        tick = min(int(max(0, min(100, elapsed_pct)) / 100 * width), width - 1)
+        tick = min(to_cells(elapsed_pct), width - 1)
     cells = []
     remaining = fill
     for i in range(width):