vrr/openclaw
2
0
Fork 0
mirror of https://github.com/openclaw/openclaw.git synced 2026-10-03 09:39:25 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 30
openclaw/docs/reference/prompt-caching.md
Peter Steinberger cee0906f38
improve(agents): explain unknown tool outcomes and assert append-only request history (#163379) 
* fix(agents): describe unrecorded tool results as unknown outcomes

Use shared unknown-outcome guidance while preserving the Responses-family aborted convention. Retain synthetic provenance in model context so late real results can replace repair placeholders without changing existing detection or stored transcript rows.

* improve(agents): assert provider request history stays append-only

Track converted message prefixes and record declared compaction, pruning, runtime-context, and image rewrites. Notify every cache-affinity baseline for the same session identity. Memoize message/content and schema fingerprints, with strict assertions enabled only by an opt-in environment flag.

* fix(agents): pin the shared tool-result text for the request wrapper

Include packages/llm-core/src/types.ts in the PR wrapper's extracted source inventory. Main commit 93625aa3d1 pulled tool-result-pairing.ts into that graph, so its existing shared tool-result text import must be included for standalone wrapper execution.
2026-10-02 07:50:31 -05:00

413 lines
34 KiB
Markdown
Raw Blame History

---
summary: "Prompt caching knobs, merge order, provider behavior, and tuning patterns"
title: "Prompt caching"
read_when:
- You want to reduce prompt token costs with cache retention
- You need per-agent cache behavior in multi-agent setups
- You are tuning heartbeat and cache-ttl pruning together
---
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.
Provider references:
- [Anthropic prompt caching](https://platform.claude.com/docs/en/build-with-claude/prompt-caching)
- [OpenAI prompt caching](https://developers.openai.com/api/docs/guides/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](/providers/openai). 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.
## 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.
```yaml
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.
```yaml
agents:
defaults:
contextPruning:
mode: "cache-ttl"
ttl: "1h"
```
See [Session pruning](/concepts/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`).
```yaml
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.
- `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.
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](#chat-completions-cache-markers). `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](#chat-completions-cache-markers)
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](https://docs.modelstudio.console.alibabacloud.com/en/model-studio/explicit-cache-guide).
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](https://www.alibabacloud.com/help/en/model-studio/context-cache).
### OpenAI (direct API)
- Prompt caching is automatic on supported recent models; OpenClaw does not inject block-level cache markers.
- OpenClaw sends `prompt_cache_key` to keep cache routing stable across turns. 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](https://developers.openai.com/api/docs/guides/prompt-caching#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`. OpenAI's Chat Completions API does not document or emit a `cache_write_tokens` counter, but OpenClaw still reads `prompt_tokens_details.cache_write_tokens` there for OpenRouter-compatible and DeepSeek-style proxies that report a separate write count.
- 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](#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.
- 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](https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html)
and model cards for [Micro](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-amazon-nova-micro.html),
[Lite](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-amazon-nova-lite.html),
[Pro](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-amazon-nova-pro.html),
[Premier](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-amazon-nova-premier.html),
and [Nova 2 Lite](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-amazon-nova-2-lite.html)
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](#chat-completions-cache-markers), 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](https://openrouter.ai/docs/guides/best-practices/prompt-caching).
`contextPruning.mode: "cache-ttl"` is allowed for `openrouter/anthropic/*`, `openrouter/deepseek/*`, `openrouter/moonshot/*`, `openrouter/moonshotai/*`, and `openrouter/zai/*` model refs, because these routes handle provider-side prompt caching without needing OpenClaw's injected markers.
Source: `extensions/openrouter/index.ts` (`OPENROUTER_CACHE_TTL_MODEL_PREFIXES`).
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.
- 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 probe 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 probe; 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 probe 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.
- 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:
```yaml
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.
- `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:
```sh
OPENCLAW_LIVE_TEST=1 OPENCLAW_LIVE_CACHE_TEST=1 pnpm test:live:cache
```
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:
```sh
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:
| Scenario | `cacheRead` floor | Hit-rate floor |
| -------------------- | ----------------: | -------------: |
| Stable prefix | 4,608 | 0.90 |
| Tool transcript | 4,096 | 0.85 |
| Image transcript | 3,840 | 0.82 |
| MCP-style transcript | 4,096 | 0.85 |
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
```yaml
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)
| Variable | Effect |
| ------------------------------------ | ------------------------------------ |
| `OPENCLAW_CACHE_TRACE=1` | Enables cache tracing |
| `OPENCLAW_CACHE_TRACE_FILE=path` | Overrides output path |
| `OPENCLAW_CACHE_TRACE_MESSAGES=0\|1` | Toggles full message payload capture |
| `OPENCLAW_CACHE_TRACE_PROMPT=0\|1` | Toggles prompt text capture |
| `OPENCLAW_CACHE_TRACE_SYSTEM=0\|1` | Toggles system prompt capture |
### 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.
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. Set `OPENCLAW_PROMPT_CACHE_ASSERT=1` to throw at the first differing message during development or tests. Compaction, pruning, transient runtime-context removal, and image cleanup declare their rewrites as `compaction`, `pruning`, `runtimeContextCarrier`, and `imageCleanup`; model, transport, or retention changes start a new history series. 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](#openai-direct-api) 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.
- **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](/reference/token-use)
- [API usage and costs](/reference/api-usage-costs)
- [Usage tracking](/concepts/usage-tracking)
- [Anthropic](/providers/anthropic)
- [Session pruning](/concepts/session-pruning)
- [Gateway configuration reference](/gateway/configuration-reference)
Reference in a new issue View git blame Copy permalink