Prompt caching

Prompt caching lets a model provider reuse an unchanged prompt prefix (system/developer instructions, tool definitions, other stable context) across turns instead of reprocessing it every request. This cuts token cost and latency on long-running sessions with repeated context.

OpenClaw normalizes provider usage into cacheRead and cacheWrite wherever the upstream API exposes those counters. Usage summaries (/status and similar) fall back to the last transcript usage entry when the live session snapshot lacks cache counters; a nonzero live value always wins over the fallback.

OpenAI-compatible routes accept both nested prompt_tokens_details.cached_tokens and top-level cached_tokens counters, including the forms documented by Together and StepFun. When a provider omits cache counters, its cache hit rate is unknown; the absence of telemetry does not mean the provider processed every token again.

Provider references:

• Anthropic prompt caching
• OpenAI prompt caching

Keep model settings stable

Prompt-cache reuse depends on provider request configuration as well as prompt text. Changing the model always starts a different cache lineage. Changing the thinking or reasoning level can also invalidate reuse even when the prompt and model stay the same. Supported native OpenAI Responses requests preserve the original effort and append turn-scoped configuration controls, including after transport expiry or a Gateway restart when saved replay metadata and history still match. See OpenAI reasoning changes. Other OpenAI models or incompatible modes can still reprocess the full prefix. Anthropic likewise documents cache invalidation when its thinking budget, effort, or mode changes.

If cache continuity matters, choose the model and thinking level when creating the session and keep both stable. Start a new session for a planned change. Invalidating reuse means the next request misses that cached state; it does not necessarily delete the provider's older cache entry before its normal expiry.

Keep replayed tool results stable

OpenClaw applies the same tool-result text cap before the first model request and when persisting the result. Subsequent turns therefore replay the same bounded text instead of introducing a different truncation notice. Redaction still runs before model-visible output is recorded. Large batches can exceed the aggregate result budget when reducing them would rewrite already-sent history; session pruning and compaction remain the owners of intentional history changes.

OpenAI Chat Completions also preserves valid native tool-call IDs during replay and disambiguates repeated IDs deterministically. Changing result text or IDs between turns can invalidate an otherwise reusable provider prefix. Stable request inputs support cache reuse; actual hits still depend on the provider.

Internal browser conversations reuse their opaque session cache key for private subagent completions, so the completion does not switch to a different affinity key between chat turns. The key remains scoped to the agent, provider, model, and conversation and does not expose the conversation identifier.

Primary knobs

Worker turns

Gateway-proxied worker inference uses the same OpenAI cache-key derivation as local turns: an explicit Gateway key takes precedence; otherwise the key combines the session ID with the authoritative transcript's reset and compaction boundary count. The Gateway retains these facts with the admitted turn, rather than accepting a worker-provided cache key or deriving one from trimmed replay history.

Delivered worker skills use stable session-scoped, content-addressed paths and a deterministic catalog order. Unchanged skills therefore keep the system prompt prefix stable between turns. Skill refreshes still deliver current verified bytes; a content change intentionally changes that skill's path. Worker and local prompts remain different in scope: workers load bounded workspace AGENTS.md and the Gateway's supplied instructions with their restricted tool set. Moving a session between runtimes or workspaces can still invalidate its cached prefix.

cacheRetention

Values: "none" | "short" | "long". Configurable as a global default, per model, and per agent. "standard" is not an alias; use "short" for the provider's default cache window. Invalid values are ignored with a warning.

agents:  defaults:    params:      cacheRetention: "long" # none | short | long    models:      "anthropic/claude-opus-4-6":        params:          cacheRetention: "short" # overrides the global default for this model  list:    - id: "alerts"      params:        cacheRetention: "none" # overrides both defaults for this agent

Merge order (later wins):

1. agents.defaults.params - global default for all models
2. agents.defaults.models["provider/model"].params - per-model override
3. agents.entries.*.models["provider/model"].params - agent-specific per-model override
4. agents.entries.*.params - agent-wide override, matched by agent id

Source: src/agents/embedded-agent-runner/extra-params.ts (resolveExtraParams).

contextPruning.mode: "cache-ttl"

Prunes old tool-result context after the cache TTL window elapses, so a post-idle request does not re-cache oversized history.

agents:  defaults:    contextPruning:      mode: "cache-ttl"      ttl: "1h"

See Session pruning for full behavior.

Heartbeat keep-warm

Heartbeat can keep cache windows warm and reduce repeated cache writes after idle gaps. Configurable globally (agents.defaults.heartbeat) or per agent (agents.entries.*.heartbeat).

agents:  defaults:    heartbeat:      every: "55m"

Provider behavior

Anthropic (direct API and Vertex AI)

• When caching is enabled and the route supports tool cache control, the tool prefix is checkpointed separately from the system prompt.
• Up to two conversation checkpoints cover recent user messages or tool results. Keeping the previous checkpoint reachable avoids a cache miss when a new turn adds more than the provider's 20-block lookup window. Transient runtime context keeps an earlier stable checkpoint.
• cacheRetention is supported for anthropic and anthropic-vertex providers, and for Claude models on amazon-bedrock and custom anthropic-messages-compatible endpoints when cacheRetention is set explicitly.
• When unset, OpenClaw seeds cacheRetention: "short" for direct Anthropic (anthropic and anthropic-vertex providers only; other Anthropic-family routes require an explicit value).
• Native Anthropic Messages responses expose cache_read_input_tokens and cache_creation_input_tokens, mapped to cacheRead and cacheWrite.
• cacheRetention: "short" maps to the default 5-minute ephemeral cache. cacheRetention: "long" requests the 1-hour TTL (cache_control: { type: "ephemeral", ttl: "1h" }) when set explicitly. An implicit/env-driven long retention (OPENCLAW_CACHE_RETENTION=long with no explicit cacheRetention) only upgrades to the 1-hour TTL on api.anthropic.com or Vertex AI (aiplatform.googleapis.com / *-aiplatform.googleapis.com) hosts; other hosts keep the 5-minute cache.

Keep short for requests less than five minutes apart. Choose long for sessions with longer gaps: cache writes cost twice the ordinary input rate instead of 1.25 times for the five-minute cache, while reads normally cost 0.1 times the input rate. The TTL starts when the request begins, so generation time counts toward expiry. Neither TTL can reuse a prefix whose instructions, tools, images, or earlier messages changed, and each model has a minimum cacheable token count.

Source: packages/ai/src/transports/anthropic-payload-policy.ts (resolveAnthropicEphemeralCacheControl, isLongTtlEligibleEndpoint).

DeepInfra

For anthropic/* models, the managed and SDK Chat Completions paths use the shared marker layout. cacheRetention: "none" disables these markers. By default, both "short" and "long" use ephemeral markers without a TTL override; OpenClaw does not assume one-hour support for this route.

Model Studio / DashScope (Qwen)

Both Chat Completions builders enable the shared marker layout on native or default Model Studio / DashScope routes through compat.cacheControlFormat: "anthropic". Explicit model compat settings take precedence over detected defaults. Custom proxy endpoints receive no automatic format default; set compat.cacheControlFormat: "anthropic" explicitly only when the proxy supports these markers.

Model Studio includes tool definitions in the system cache and ignores markers on tools themselves, so OpenClaw omits that marker on detected native routes. Qwen3.5 and later support message-level checkpoints only: splitting the stable and volatile system content into blocks does not guarantee independent reuse of the stable block. See Model Studio explicit cache guidance.

Explicit cacheRetention values reach this transport without enabling compat.supportsPromptCacheKey; leave that flag unset because this route does not need OpenAI's prompt_cache_key or prompt_cache_retention fields. With no explicit retention, the transport keeps its "short" default. Model Studio uses a five-minute explicit cache window, so "long" keeps ephemeral markers without requesting a one-hour TTL.

To disable OpenClaw's explicit markers, set cacheRetention: "none". The current compat.cacheControlFormat schema accepts only "anthropic", not a disable value; omitting it uses the detected default. Alibaba's automatic implicit caching is separate and cannot be disabled. Supported models, minimum prompt lengths, and cache billing are described in Model Studio context caching.

OpenAI (direct API)

• Prompt caching is automatic on supported recent models; OpenClaw does not inject block-level cache markers.
• OpenClaw sends a stable prompt_cache_key across turns. On models before GPT-5.6 this helps cache routing; on GPT-5.6 and later it separates cache accounting, while OpenAI handles routing automatically. Responses and Chat Completions requests to direct api.openai.com hosts get this automatically when a session or explicit cache key is available; compat.supportsPromptCacheKey: false disables it. OpenAI-compatible proxies (oMLX, llama.cpp, custom endpoints) need compat.supportsPromptCacheKey: true in model config to opt in - this is never auto-detected for a proxy.
• cacheRetention: "long" requests prompt_cache_options: { ttl: "30m" } for GPT-5.6 and later on both APIs. Earlier native models receive prompt_cache_retention: "24h" only for OpenAI's documented extended-retention models: GPT-5.5 / GPT-5.5 Pro, GPT-5.4, GPT-5.2, GPT-5.1 / Codex / Codex Max / Codex Mini / Chat Latest, GPT-5 / Codex, and GPT-4.1 (including dated snapshots). Other earlier native models receive no lifetime field. See OpenAI cache lifetime.
• Lifetime fields require both cache-key support and compat.supportsLongCacheRetention (true by default; Together AI and Cloudflare profiles disable it). Opted-in proxies use the same GPT-5.6+ TTL mapping and otherwise receive "24h"; disable long-retention support if the proxy rejects lifetime fields.
• cacheRetention: "short" sends the key without lifetime fields, leaving the provider's default lifetime in effect. cacheRetention: "none" suppresses the key and both lifetime fields; it does not disable OpenAI's automatic prompt caching.
• Native ChatGPT-backed Responses routes keep the session cache key, honor none, and omit both OpenAI lifetime fields.
• Cache hits surface via usage.prompt_tokens_details.cached_tokens (Chat Completions) or input_tokens_details.cached_tokens (Responses API), mapped to cacheRead.
• Responses API payloads can also expose input_tokens_details.cache_write_tokens, mapped to cacheWrite and priced at the model's cache-write rate; Responses payloads that omit the field keep cacheWrite at 0. Chat Completions payloads that include prompt_tokens_details.cache_write_tokens also populate cacheWrite; availability depends on the model and route.
• Captured consecutive Responses request bodies retained byte-identical history prefixes, so history rewriting did not explain the observed shortfall. Provider breakpoint placement can affect reported cacheRead: OpenAI documents message-end breakpoints for GPT-5.6 and later on the Platform API, while the ChatGPT-backed Responses route was observed to report hits in 1,024-token steps. See OpenAI live expectations below.
• On Responses routes the whole system prompt, including the volatile suffix below the cache boundary, is sent as instructions. A suffix change (date rollover, timezone, elevated level, watched sessions, model identity, Project Memory facts) re-caches from the changed point on the next request; the stable prefix and tools are not split into a separate cached block the way Anthropic checkpoints are. Cache observations report this as a systemPromptSuffix change.

Amazon Bedrock

• Anthropic Claude model refs (amazon-bedrock/*anthropic.claude*, plus AWS system inference profile prefixes us./eu./global.anthropic.claude*) support explicit cacheRetention pass-through.
• One-hour retention is requested only for Claude model generations documented by AWS as supporting it, including Haiku 5.5. Older cache-capable models keep five-minute checkpoints when cacheRetention: "long" is selected, without sending an unsupported TTL field. Application inference profiles use the supported window of their resolved backing models.
• The stable system prefix is checkpointed separately from dynamic runtime additions. Conversation checkpoints advance through retained history, including tool results; transient runtime-context carriers remain outside the cached prefix. Bedrock Mantle's Anthropic Messages transport also preserves the separate stable system boundary.
• Nova Micro, Lite, Pro, Premier (amazon.nova-{micro,lite,pro,premier}-v1:0), and Nova 2 Lite (amazon.nova-2-lite-v1:0) support explicit checkpoints in system and messages, including their AWS geographic inference profiles and foundation-model ARNs. Both short and long use Nova's five-minute TTL; none disables explicit checkpoints. OpenClaw does not add tool checkpoints for Nova.
• Other non-Claude Bedrock models remain at cacheRetention: "none".
• Nova explicit caching is opt-in: set cacheRetention explicitly to short or long. With retention unset, Nova requests keep their existing payload layout with no checkpoints; neither the default short window nor OPENCLAW_CACHE_RETENTION enables Nova checkpoints.
• Opaque Bedrock application inference profile ARNs (profile IDs that do not contain claude) also resolve to no cache retention unless cacheRetention is set explicitly, since the model family cannot be inferred from the ARN alone.

AWS's prompt caching guide and model cards for Micro, Lite, Pro, Premier, and Nova 2 Lite list these limits: a 1K-token minimum, four checkpoints, and at most 20K cached tokens for Nova. Provider token limits still determine whether a checkpoint is cached.

Nova explicit caching has not been live-verified against AWS by OpenClaw maintainers yet. Live AWS acceptance proof remains a gap until a maintainer with Bedrock access runs it.

OpenRouter

For openrouter/anthropic/* model refs, both Chat Completions builders apply the shared marker layout, but only when the request still targets a verified OpenRouter route (openrouter on its default endpoint, or any provider/base URL that resolves to openrouter.ai). Repointing the model at an arbitrary OpenAI-compatible proxy URL stops automatic marker injection. cacheRetention: "long" requests ttl: "1h" on these verified routes; "none" disables markers. See OpenRouter prompt caching.

contextPruning.mode: "cache-ttl" is allowed for openrouter/anthropic/*, openrouter/deepseek/*, openrouter/moonshot/*, openrouter/moonshotai/*, openrouter/x-ai/*, openrouter/zai/*, openrouter/google/gemini-2.5*, and openrouter/google/gemini-3* model refs, because these routes handle provider-side prompt caching without needing OpenClaw's injected markers. Older Gemini models and other google/* models (for example Gemma) are excluded because OpenRouter caches them only with explicit cache_control markers, if at all.

Source: extensions/openrouter/index.ts (OPENROUTER_CACHE_TTL_MODEL_FAMILY).

DeepSeek cache construction on OpenRouter is best-effort and can take a few seconds; an immediate follow-up request may still show cached_tokens: 0. Verify with a repeated same-prefix request after a short delay, using usage.prompt_tokens_details.cached_tokens as the cache-hit signal.

Google Gemini (direct API)

• Direct Gemini transport (api: "google-generative-ai") reports cache hits through upstream cachedContentTokenCount, mapped to cacheRead.
• Eligible model families: gemini-2.5* and gemini-3* (excludes Live/preview variants outside that prefix match, for example gemini-live-2.5-flash-preview).
• When cacheRetention is set on an eligible model, OpenClaw automatically creates, reuses, and refreshes a cachedContents resource containing the stable system prefix above the cache boundary plus tools and tool configuration - no manual cached-content handle needed. TTL is 300s for cacheRetention: "short" and 3600s for "long".
• The volatile system suffix travels first inside the current turn's hidden runtime-context carrier, before other runtime facts. This carrier is transient, so suffix changes reuse the same resource without accumulating history. Stable-prefix or tool changes create a new resource. If creation fails or the prompt has no cache boundary, the complete system prompt stays inline.
• Automatic resources also belong to the effective request credentials and headers. Changing credentials, project headers, or other request-header overrides creates a new resource. Cached inference uses the same credentials as resource creation; OAuth token refreshes conservatively rebuild the resource. Reissued secret placeholders for the same credential preserve its identity.
• You can still pass a pre-existing Gemini cached-content handle through as params.cachedContent (or legacy params.cached_content); an explicit handle skips the automatic cache-management path entirely.
• This is separate from Anthropic/OpenAI prompt-prefix caching: OpenClaw manages a provider-native cachedContents resource for Gemini instead of injecting inline cache markers.

Source: src/agents/embedded-agent-runner/google-prompt-cache.ts.

CLI-harness providers (Claude Code, Gemini CLI)

CLI backends that emit JSONL usage events (jsonlDialect: "claude-stream-json" or "gemini-stream-json") go through a shared usage parser that recognizes several field-name variants, including a plain cached counter mapped to cacheRead. When the CLI's JSON payload omits a direct input-token field, OpenClaw derives it as input_tokens - cached. This is usage normalization only - it does not create Anthropic/OpenAI-style prompt-cache markers for these CLI-driven models.

Claude Code has no OpenClaw-controlled cache_control breakpoint on --append-system-prompt-file, so OpenClaw keeps its complete system prompt in that transport. When the bounded version check finds Claude Code 2.1.98 or newer, bundled claude-cli also passes --exclude-dynamic-system-prompt-sections. The first CLI execution or direct Anthropic OAuth request starts the shared check; concurrent executions reuse it, and API catalog discovery does not start it. That Claude Code flag moves only Claude's own per-machine cwd, environment, memory-path, and Git-status sections out of its native system prompt; an older, unknown, or failed check keeps the established argv. cacheRetention still has no effect on this path.

One-shot helper runs dispatched through a CLI backend, such as Active Memory recall on claude-cli, get a new session key on every run. Those runs carry the Runtime facts line (agent, session, model, channel) in their only user turn instead of the system prompt, so repeated recalls of one agent send a byte-identical system prompt and can reuse Claude's prompt cache. Normal CLI turns keep the Runtime line in the system prompt.

Source: src/agents/cli-output.ts (toCliUsage).

Other providers

If a provider does not support any of the above cache modes, cacheRetention has no effect.

Chat Completions cache markers

OpenAI-compatible routes with compat.cacheControlFormat: "anthropic" share one marker policy across the managed transport, SDK builder, and provider wrappers:

• The last tool definition is marked when the route supports tool markers; tools remain sorted by name.
• The stable system/developer block is marked, with the volatile suffix in a separate unmarked block.
• The latest eligible user text or tool result is marked, advancing through tool loops and new turns while skipping transient runtime-context carriers.

The layout normally uses three markers and stays within the four-breakpoint budget. cacheRetention: "none" emits none. A conversation checkpoint covers all preceding tools, system content, and messages, so changing the volatile system suffix still invalidates that later checkpoint; the earlier stable-system checkpoint remains reusable where the backend supports block-level caching. Backend token minimums and cache lifetimes still apply.

With compat.requiresStringContent: true, managed requests keep message content as strings and omit message-block markers, including through provider wrappers. The tool-definition marker remains where supported.

Detected defaults request ttl: "1h" only on verified OpenRouter routes. For a custom endpoint that supports one-hour Anthropic caching, explicitly set compat.cacheControlFormat: "anthropic", compat.supportsLongCacheRetention: true, and cacheRetention: "long" to send that TTL. Omitting the capability override keeps custom-endpoint markers without a TTL; setting it to false disables the one-hour TTL even on OpenRouter.

System-prompt cache boundary

OpenClaw splits the system prompt into a stable prefix and a volatile suffix at an internal cache-prefix boundary. Content above the boundary (tool definitions, skills metadata, workspace files) is ordered to stay byte-identical across turns. Content below the boundary (for example runtime timestamps and other per-turn metadata) can change without invalidating the cached prefix.

Key design choices:

• Stable workspace project-context files are ordered before volatile per-turn metadata so routine churn does not bust the stable prefix.
• The boundary applies across Anthropic-family, OpenAI-family, Google, and CLI transport shaping, so all supported providers benefit from the same prefix stability.
• Codex Responses and Anthropic Vertex requests are routed through boundary-aware cache shaping so cache reuse stays aligned with what providers actually receive.
• System-prompt fingerprints are normalized (whitespace, line endings, hook-added context, runtime capability ordering) so semantically unchanged prompts share cache across turns.

If you see unexpected cacheWrite spikes after a config or workspace change, check whether the change lands above or below the cache boundary. Moving volatile content below the boundary (or stabilizing it) usually resolves the issue.

Chat Completions routes without an explicit message-cache breakpoint move the bounded Runtime facts line to the first emitted user message. This keeps session identifiers behind the system-and-tools prefix on compatible local servers. The line stays on that first message during follow-ups. Behavioral instructions, including hook additions, permission notices, and Git coauthor guidance, retain their system/developer role. Routes with explicit message breakpoints keep their existing system layout. Current-turn Runtime Context snapshots still use their separate transient carrier; they do not become permanent first-message context.

OpenClaw cache-stability guards

• Active exec sessions, subagent state, and media-generation progress travel in compact Runtime Context carriers after the current user message, so changes do not rewrite the system prompt ahead of conversation history. Project Memory facts, channel-specific ACP hints, delegation/orchestration mode, and the current elevated level stay below the system-prompt cache boundary; static recall, safety, and capability guidance stay above it.
• Delivery instructions live after the system-prompt cache boundary. Native Codex carries the current delivery and target policy in late turn context, so alternating delivery modes does not rebuild its static prompt or message tool catalog when the available capabilities remain unchanged. Actual capability changes still update the catalog.
• Subagent completion turns reuse the session's generic conversation metadata and direct/group guidance, so native Ollama does not lose the tools-and-history prefix just because the turn arrived through an announcement. Channel formatting follows the resolved delivery account and is omitted for turns without channel delivery. Custom channel-supplied group instructions, changed delivery policy, and explicit lightweight bootstrap contexts can still produce different prompts.
• User-message metadata stays intact when an active turn becomes history, including conversation context and reply targets. Timestamping uses the message's recorded time so subsequent requests replay the same text.
• Prompt-hook prepend/append context and model-prompt replacements are captured on their original user transcript record before dispatch. Later requests and reopened sessions replay those bytes while user-visible history keeps the original text. Legacy rows without a recorded projection retain their previous replay behavior; see stored projection compatibility.
• Bundled MCP tool catalogs are sorted deterministically (by server name, then tool name) before tool registration, so listTools() order changes do not churn the tools block and bust prompt-cache prefixes.
• Message-tool action enums are sorted after policy filtering, keeping identical capabilities stable across channel discovery order changes.
• Native Ollama requests sort tools by name so discovery order changes do not churn the tools prefix.
• Legacy sessions with persisted image blocks keep the 3 most recent completed turns intact (counting all completed turns, not just image-bearing ones). Older already-processed image blocks are replaced with a text marker so image-heavy follow-ups do not keep re-sending large stale payloads.

Tuning patterns

Mixed traffic (recommended default)

Keep a long-lived baseline on your main agent, disable caching on bursty notifier agents:

agents:  defaults:    model:      primary: "anthropic/claude-opus-4-6"    models:      "anthropic/claude-opus-4-6":        params:          cacheRetention: "long"  list:    - id: "research"      default: true      heartbeat:        every: "55m"    - id: "alerts"      params:        cacheRetention: "none"

Cost-first baseline

• Set baseline cacheRetention: "short".
• Enable contextPruning.mode: "cache-ttl".
• Keep heartbeat below your TTL only for agents that benefit from warm caches.

Live regression tests

OpenClaw runs one combined live cache regression gate covering repeated prefixes, tool turns, image turns, MCP-style tool transcripts, and an Anthropic no-cache control.

Full Release Validation's live-cache suite also checks serialized transport prefixes and real agent turns across instruction refresh, prompt hooks, and persisted-session reopen. Offline coverage includes authenticated Gateway chat around a spawned child's private completion and an in-process server stop/start, without restarting the OS process. See Prompt-cache regression coverage for the exact boundaries, commands, and estimated cost.

• src/agents/live-cache-regression.live.test.ts
• src/agents/test-helpers/live-cache-regression-runner.ts
• src/agents/live-cache-regression-baseline.ts

Run it with:

OPENCLAW_LIVE_TEST=1 OPENCLAW_LIVE_CACHE_TEST=1 pnpm test:live:cache

This command selects the existing cache regression test through the shared live runner. Set OPENCLAW_VITEST_RUNTIME=bun to run its provider calls on Bun; the default is Node. Provider usage summaries and advisory warnings remain visible, and hard regression floors fail the test.

The baseline file stores the most recently observed live numbers plus the provider-specific regression floors the test checks against. Each run uses fresh per-run session IDs and prompt namespaces so previous cache state does not pollute the current sample. Anthropic and OpenAI use different enforcement: an Anthropic floor miss is a hard regression (test fails), while an OpenAI floor miss is watch-only (recorded as a warning, does not fail the run). They do not share a single cross-provider threshold.

Claude CLI prompt reuse has a separate Docker lane because it exercises Claude Code's native session transport rather than the direct Anthropic API. After a fresh turn and tool-bearing warmup resume, it allows a no-tool settlement resume to run hot or cold, dirties the workspace, and requires at least 90% reuse on the following resume without rotating the settled live-session generation. It also verifies that a thinking-level change rotates the generation and that the next steady resume restores at least 90% reuse:

pnpm test:docker:live-cli-backend:claude:cache

Anthropic live expectations

• Expect explicit warmup writes via cacheWrite.
• Expect near-full history reuse on repeated turns, because Anthropic's cache control advances the cache breakpoint through the conversation.
• Baseline floors for stable, tool, image, and MCP-style lanes are hard regression gates.

OpenAI live expectations

• Expect cacheRead only; cacheWrite stays 0 on Chat Completions.
• cacheRead can lag the full prompt and advance in provider-sized steps. The observed interval below is not a guaranteed bound on uncached input; routing, cache availability, and changed prompt content can also reduce reuse.
• Floors are watch-only (a miss is logged as a warning, not a test failure), derived from live behavior observed on gpt-5.4-mini and unchanged since 2026.4.5:

The most recently observed baseline numbers (from live-cache-regression-baseline.ts, recorded 2026-04-04) landed at: stable prefix cacheRead=4864, hit rate 0.966; tool transcript cacheRead=4608, hit rate 0.896; image transcript cacheRead=4864, hit rate 0.954; MCP-style transcript cacheRead=4608, hit rate 0.891.

Why the assertions differ: Anthropic exposes explicit cache breakpoints, while OpenAI's reported reuse depends on available matching cached prefixes. The dated measurements below showed stepwise growth, not a universal hit-rate guarantee. Comparing the two providers against a single cross-provider percentage threshold produces false regressions.

Observed on 2026-09-16 with openai/gpt-5.6-luna through the ChatGPT-backed Responses route (transport: auto, default cacheRetention), using both openclaw agent --local and an isolated Gateway:

• Every reported cacheRead was congruent to 512 modulo 1,024 (32256, 33280, 34304, 35328), consistent with 1,024-token reuse increments in this sample, not proof of a fixed provider-wide breakpoint policy. A session whose prompt grew from 32658 to 33098 tokens across four turns kept reporting 32256 until the prompt crossed the next interval; a 60-second pause between turns did not change this.
• Consecutive request bodies were byte-identical up to the newly appended items, including replayed runtime-context carriers, ruling out history rewriting in those captures without establishing why every token was not reused.
• Changing the host timezone mid-session (which rewrites the ## Temporal Context lines inside instructions) dropped one turn from 32256 to 22016 cached tokens on a 32974-token prompt; the following turn was back at 32256.
• A new session sometimes reused the shared instructions-and-tools prefix on its first request and sometimes started cold, with or without a shared prompt_cache_key; routing looked session-sticky rather than key-driven, which matches OpenAI's note that the key is not needed to optimize caching on GPT-5.6 and later.

diagnostics.cacheTrace config

diagnostics:  cacheTrace:    enabled: true

enabled defaults to false. Cache traces otherwise write to $OPENCLAW_STATE_DIR/logs/cache-trace.jsonl and include messages, prompt text, and the system prompt by default. Output-path and payload-inclusion overrides are environment-only controls for one-off debugging.

Env toggles (one-off debugging)

What to inspect

Prompt-cache observations record input, cacheRead, and cacheWrite per completed foreground model request alongside its stable system-prefix, volatile-suffix, and tools fingerprints, and flag cache-read drops from the previous request, including reported zero reads; billing totals remain separate. A flagged drop lists the tracked changes since the last request (model, cacheRetention, transport, streamStrategy, systemPrompt, systemPromptSuffix, tools, aggregateToolResultTruncation). Trace results require cache tracing (diagnostics.cacheTrace.enabled or OPENCLAW_CACHE_TRACE=1) and identify each request within its attempt.

When an adapter marks cache telemetry unavailable, observations omit its cache read/write counts and preserve the last measured comparison baseline. A later reported zero still counts as a measured miss. This distinction applies to local engines with optional cache metrics as well as compatible cloud APIs.

The comparison baseline resets when the session ID changes, even if an isolated cron job reuses its provider cache key. A fresh transcript can legitimately reuse fewer tokens than the previous run's final request. Within a session, no tracked cache input change means the tracked fingerprints stayed stable; it does not prove identical final provider payloads or diagnose cache expiry. Compare request timing and final payloads before attributing a drop to provider caching.

For encoded provider requests, providerPrefix compares the serialized instructions, tools, and up to 512 input items. It names the first changed item, with bounded protocol fields for the first 32 items, such as message:3.encrypted_content, plus changed request parameters such as parameters:prompt_cache_key. Values and unknown field names are never logged. order means object key order changed without a tracked field-value change; it is wire evidence, not proof that the provider rendered different tokens. Beyond the item limit, a tail digest can detect changes but cannot locate the first differing item.

A continuation: prefix means either request used previous_response_id. Its wire-input items are deltas, so differences between them do not establish rewritten conversation history. Compare the separate historyRewrite signal and the changed request parameters before attributing a cache miss. A matching full-history wire prefix also does not guarantee a hit: OpenAI must find an available matching cache boundary.

OpenClaw also checks that each converted request history extends the previous request in the same session. An undeclared edit, removal, or reorder records historyRewrite and warns once for the session. The diagnostic names bounded changed fields (such as timestamp, __openclaw, or content[0]) without logging their values; fields outside the diagnostic limit are grouped. Set OPENCLAW_PROMPT_CACHE_ASSERT=1 to throw at the first differing message during development or tests. Compaction, pruning, and transient runtime-context removal declare their rewrites as compaction, pruning, and runtimeContextCarrier; model, transport, or retention changes start a new history series. Image blocks and attachment references remain in retained history instead of being replaced by age-based cleanup markers. Content-block fingerprints reuse hashes only while every primitive property still matches; unchanged large text and image data are not hashed again. Nested block values and message envelopes are checked on every observation, so in-place edits are detected even when wrappers or content arrays are reused. String-content hashes live with the bounded history baseline. Provider-owned tool schema declarations are fingerprinted once per object identity.

• Cache trace events are JSONL with staged snapshots like session:loaded, prompt:before, stream:context, and session:after.
• Per-turn cache token impact is visible in normal usage surfaces: cacheRead and cacheWrite show up in /usage tokens, /status, session usage summaries, and custom messages.usageTemplate layouts.
• For Anthropic, expect both cacheRead and cacheWrite when caching is active.
• For OpenAI, expect cacheRead on cache hits; cacheWrite is populated only on Responses API payloads that include it (see OpenAI above).
• OpenAI also returns tracing and rate-limit headers such as x-request-id, openai-processing-ms, and x-ratelimit-*; use those for request tracing, but cache-hit accounting should still come from the usage payload, not from headers.

Quick troubleshooting

• High cacheWrite on most turns: check for volatile system-prompt inputs; verify the model/provider supports your cache settings.
• High cacheWrite on Anthropic: often means the cache breakpoint is landing on content that changes every request. A volatile system-suffix change can preserve the stable system-prefix checkpoint while invalidating later conversation checkpoints; cache observations report systemPromptSuffix for that change.
• Low OpenAI cacheRead: verify the stable prefix is at the front, the repeated prefix is at least 1024 tokens, and the same prompt_cache_key is reused for turns that should share a cache. A volatile system-suffix change can cause a one-turn drop that recovers on the next request; enable cache tracing and look for systemPromptSuffix in the [prompt-cache] warning. A small or stepwise shortfall can reflect provider granularity, but does not rule out prompt changes or cache availability. The warning's requests field pairs the last measured cache hit with the dropping request. For encoded HTTP requests, it includes cache-key presence and hash, continuation-ID presence, exact request bytes, a SHA-256 hash of the first 4096 bytes (or the whole request if smaller), model and route hashes, and the number of encoded attempts within that model call. Available response metadata includes HTTP status, numeric openai-processing-ms, and hashes of allowlisted request-ID, model, account, and region headers. Raw keys, IDs, headers, and prompt content are never included. Missing fields mean the transport did not supply that evidence; retries across model calls and unobserved transport retries are not counted.
• No effect from cacheRetention: confirm the model key matches agents.defaults.models["provider/model"].
• Bedrock Nova requests without cache hits: set cacheRetention explicitly to short or long, verify that the model is one of the supported variants above, and check that the prefix meets AWS's token limits; long still uses a five-minute TTL.

Related

• Token use and costs
• API usage and costs
• Usage tracking
• Anthropic
• Session pruning
• Gateway configuration reference