eeymoo/claw-code
1
0
Fork 0
mirror of https://github.com/instructkr/claw-code.git synced 2026-04-26 00:26:13 +08:00
Code Issues Packages Projects Releases Wiki Activity

roadmap: #219 filed — Anthropic prompt-caching opt-in is structurally impossible: cache_control marker has zero codebase footprint (rg returns 0 hits across rust/ src/ docs/ tests/) despite the wire-side beta header 'prompt-caching-scope-2026-01-05' being unconditionally enabled at every Anthropic request (telemetry/lib.rs:16,452,469 + anthropic.rs:1443); five cacheable surfaces are uniformly locked: pub system: Option<String> at types.rs:11 is a flat string with no array form so no system-block cache_control slot exists; InputContentBlock variants Text/ToolUse/ToolResult at types.rs:80-99 have no cache_control field; ToolResultContentBlock variants Text/Json at types.rs:100-103 have no cache_control field; ToolDefinition at types.rs:105-110 has no cache_control field; openai_compat path translate_message at openai_compat.rs:946 and build_chat_completion_request at openai_compat.rs:850 emit flat-string system+content with no cache_control or Bedrock cachePoint translation; ~600 LOC of response-side cache stats infrastructure (prompt_cache.rs PromptCacheStats / PromptCacheRecord / PromptCache trait) accumulates a zero stream because no payload was opted in, and four hardcoded zero-coercion sites (openai_compat.rs:477-478, 489-490, 597-598, 1211-1212) discard upstream cache stats from Bedrock/Vertex/kimi-anthropic-compat/MiniMax-relay even when emitted; integration test at client_integration.rs:88-89 asserts the beta header is sent but no companion test asserts payload contains a cache_control marker because the data structures cannot produce one — a uniquely paradoxical false-positive opt-in shape: wire signal advertises caching intent and data-model structurally precludes it (Jobdori cycle #371 / extends #168c emission-routing audit / sibling-shape cluster grows to eighteen: #201/#202/#203/#206/#207/#208/#209/#210/#211/#212/#213/#214/#215/#216/#217/#218/#219 / wire-format-parity cluster grows to nine: #211+#212+#213+#214+#215+#216+#217+#218+#219 / cost-parity cluster grows to seven: #204+#207+#209+#210+#213+#216+#219 — #219 is the dominant cost-parity miss, ~90% input-token-cost reduction unattainable / cache-parity request/response symmetry pair: #219 (request-side opt-in absent) + #213 (response-side stats absent on openai-compat lane) / five-surface uniform-structural-absence shape: system+tools+tool_choice+messages+tool_result_content all locked, with no extra_body escape hatch since cache_control is a per-block annotation not a top-level field / false-positive-opt-in shape: novel cluster member where wire signal says yes and structure says no / external validation: Anthropic prompt-caching reference at platform.claude.com/docs/en/build-with-claude/prompt-caching documenting cache_control: {type: ephemeral} on system/tools/messages/content blocks with 5-min default TTL and 1-hour optional TTL and 90% cost reduction on cache-read tokens, Anthropic Messages API reference documenting system: Vec<SystemBlock> array form as the cacheable shape, Bedrock prompt-caching docs documenting cachePoint: {} block form for Bedrock-anthropic relay, claudecodecamp.com analysis of how prompt caching actually works in Claude Code, xda-developers article documenting claude-code's cache-token-budget knob proving caching is actively engaged, anomalyco/opencode#5416 #14203 #16848 #17910 #20110 #20265 (cache-related issues and PR for system-prompt-split-for-cache-hit-rate optimization), opencode-anthropic-cache npm package as third-party plugin proving the ecosystem expectation, LangChain anthropicPromptCachingMiddleware as first-class JS wrapper, LiteLLM prompt-caching docs with single-line cache_control pass-through for Anthropic+Bedrock, Vercel AI SDK Anthropic provider providerOptions.anthropic.cacheControl, prompthub.us multi-provider comparison treating opt-in as documented baseline, portkey.ai gateway-level pass-through, mindstudio.ai cost-impact analysis, OpenTelemetry GenAI semconv gen_ai.usage.input_tokens.cached as documented attribute — claw is the sole client/agent/CLI in the surveyed coding-agent ecosystem with zero cache_control request-side opt-in capability despite shipping the eligibility beta header on every Anthropic request)

Browse Source
This commit is contained in:
YeonGyu-Kim
2026-04-26 00:40:20 +09:00
parent 116a95a253
commit 2858aeccff
1 changed files with 188 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

189
ROADMAP.md
View File

File diff suppressed because one or more lines are too long