[Bug]: Anthropic prompt caching shows 0% cache hit rate on Haiku 4.5

[krakenlgk]

Affected component

provider

Severity

S2 - degraded behavior

Current behavior

Summary

Running ZeroClaw v0.5.0 on Raspberry Pi 4 with Anthropic Haiku 4.5 (claude-haiku-4-5-20251001). Anthropic dashboard consistently shows 0% Cache Rate — zero cache reads across all API calls.

Per the wiki (05.1 Built In Providers), ZeroClaw's Anthropic provider should support:

Automatic prompt caching (system prompts >3KB, conversations >4 messages, tools)

Environment

• ZeroClaw version: v0.5.0 (stable, March 18, 2026)
• OS: Raspberry Pi OS Bookworm (aarch64)
• Hardware: Raspberry Pi 4 (8GB)
• Provider: anthropic
• Model: claude-haiku-4-5-20251001
• Channel: Telegram
• Workspace files total: ~3,279 bytes

Relevant config

default_provider = "anthropic"
default_model = "claude-haiku-4-5-20251001"
max_context_tokens = 8192
max_history_messages = 5
response_cache_enabled = true

Current behavior

• Cache Rate: 0% on all calls — no cache writes, no cache reads
• Two messages sent 50 seconds apart show no cache reuse
• Every call pays full input token price (~10,000 tokens per message)
• grep -ri "cache_control\|ephemeral" ~/.zeroclaw/ returns nothing

Cost impact

At 50 messages/day on Haiku 4.5:

• Without caching: ~₹1,275/month ($15/month) — 10K input tokens × $1/MTok × 1,500 calls
• With caching working (90% cache read): ~₹400-500/month ($5-6/month)
• Wasted spend: ~₹800/month due to cache not activating

Investigation done

1. Confirmed workspace files total 3,279 bytes (above the documented 3KB threshold)
2. Verified response_cache_enabled = true in config (this is ZeroClaw's internal cache, not Anthropic's prompt cache)
3. No cache_control or ephemeral strings found anywhere in ~/.zeroclaw/
4. Anthropic dashboard shows only "Uncached Tokens" in the chart — Cache Rate line flat at 0%

Possible root causes

1. System prompt below Haiku's minimum cacheable token count (2,048 tokens)? Workspace files are ~800-900 tokens. Even with ZeroClaw's internal system prompt + tool schemas, the total might be under the threshold.

2. Dynamic content invalidating cache prefix? PR feat: refresh stale datetime in cached system prompt #3223 ("refresh stale datetime in cached system prompt") merged in v0.1.9a addresses this for timestamps. Unclear if this fix is included in v0.5.0.

3. cache_control headers not being sent? The Anthropic API requires cache_control: {"type": "ephemeral"} on content blocks, or the top-level cache_control field. ZeroClaw may not be injecting these for Haiku model IDs.

Questions

1. Does v0.5.0 include the datetime cache fix from PR feat: refresh stale datetime in cached system prompt #3223?
2. Is there a minimum system prompt token count for Anthropic caching to activate? If so, is it documented?
3. Is there a config flag to explicitly enable/verify Anthropic prompt caching (similar to OpenClaw's cacheRetention parameter)?
4. Can ZeroClaw expose cache_write/cache_read token counts in logs or zeroclaw status output for debugging?

Related issues

• ZeroClaw PR feat: refresh stale datetime in cached system prompt #3223 — feat: refresh stale datetime in cached system prompt
• ZeroClaw feat: Cost tracking and budget enforcement #213 — feat: Cost tracking and budget enforcement
• OpenClaw #19534 — Anthropic prompt caching broken - Cache Read always 0 (same root cause: dynamic system prompt content)
• OpenClaw #18963 — Anthropic prompt cache misses on most turns (detailed investigation)
• Anthropic docs — Prompt caching

Expected behavior

Expected behavior

• First message: cache write for the system prompt prefix
• Second message (within 5-minute TTL): cache read showing >0% cache hit rate
• Subsequent messages within TTL: consistent cache reads, reducing effective input token cost by ~90%

Steps to reproduce

## Steps to Reproduce

1. Configure ZeroClaw v0.5.0 with Anthropic Haiku 4.5
2. Send a message via Telegram (e.g., "Hello")
3. Wait 30-60 seconds
4. Send another message (e.g., "Hey")
5. Check Anthropic console → Rate Limit Use + Caching → filter by `claude-haiku-4-5-20251001`

Impact

Cost impact

At 50 messages/day on Haiku 4.5:

• Without caching: ~₹1,275/month ($15/month) — 10K input tokens × $1/MTok × 1,500 calls
• With caching working (90% cache read): ~₹400-500/month ($5-6/month)
• Wasted spend: ~₹800/month due to cache not activating

Logs / stack traces

ZeroClaw version

v0.5.0

Rust version

No response

Operating system

Raspberry Pi OS Bookworm (aarch64)

Regression?

Unknown

Pre-flight checks

• I reproduced this on the latest master branch or latest release.
• I redacted secrets, tokens, and personal data from all submitted content.

[whtiehack]

#3271
this pr resolve the datetime problem.
and you need close canary_tokens

[security]
canary_tokens = false

both are indispensable