temporal-spring-ai: per-model ActivityOptions registry#2855
temporal-spring-ai: per-model ActivityOptions registry#2855donald-pinckney wants to merge 20 commits intomasterfrom
Conversation
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Narrows the activity-summaries scope to cases where the plugin itself owns activity-stub creation. Activity-backed tool calls, Nexus tool calls, and @SideEffectTool calls are now explicitly out of scope; the first two would require per-call option overrides on user-owned stubs (no clean API), and the third writes MarkerRecorded events which have no Summary field. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Runs a workflow that drives a single chat call through ActivityChatModel, fetches the resulting history, and asserts the ActivityTaskScheduled event for callChatModel carries a userMetadata Summary that starts with "chat: default" and includes the user prompt. Intentionally fails against unmodified chat code — the implementation follows in a subsequent commit. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
ActivityChatModel.forModel() now stores the ActivityOptions it built and, on each chat call, rebuilds the stub with a per-call Summary of the form "chat: <model> · <first 60 chars of user prompt>". When a caller passes a pre-built stub directly via the public constructors, behavior is unchanged (no options known → no summary overlay). ActivityMcpClient.create() does the same and adds a callTool(clientName, request, summary) overload. McpToolCallback passes "mcp: <client>.<tool>". Also fixes the activity-type-name casing in ActivitySummaryTest — Temporal capitalizes the first character of method-name-derived activity types, so the event carries "CallChatModel", not "callChatModel". Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
44777c8 to
316d52b
Compare
Planning scratchpad — not part of the shipped artifact. Removed before merge. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
f967316 to
1310ba4
Compare
316d52b to
ba6a88a
Compare
The Summary now carries only the model label ("chat: <model>") instead
of "chat: <model> · <first 60 chars of user prompt>". Including even a
truncated prompt leaks whatever the prompt contains — PII, secrets,
internal identifiers — into workflow history, server logs, and the
Temporal UI, which is a surprising default for an observability label.
An opt-in API for callers who explicitly want the prompt in the
Summary can be added later if there's demand.
ActivitySummaryTest.chatActivity_carriesModelOnlySummary_neverLeaksUserPrompt
asserts the Summary equals "chat: default" exactly and defensively
checks that no part of the prompt leaked in.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
1310ba4 to
9bb3866
Compare
bad7fd6 to
e8405cc
Compare
f038d46 to
9528294
Compare
e8405cc to
1a2adb2
Compare
Addresses review feedback (thread on #2852) preferring typesafe Optional<T> over nullable fields and raw null delegation. Three sites changed: - ActivityChatModel: modelName and baseOptions fields + private ctor params are now Optional<String> / Optional<ActivityOptions>. getModelName() returns Optional<String>. Public ctors and factories still accept nullable String modelName at the API boundary (matches prior javadoc: "null for default"); they normalize via Optional.ofNullable before storing. Internal readers use .map / .orElse instead of null checks. - ActivityMcpClient: the 3-arg callTool's summary parameter is now Optional<String>. The 2-arg convenience overload passes Optional.empty(). The rebuild-with-summary branch uses .isPresent() + .get() instead of null checks. - McpToolCallback.call(...) wraps its generated summary string in Optional.of(...) before passing to callTool. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Reverses the previous Optional<> commit in favor of @nullable on the nullable fields/params. Matches the dominant convention in temporal-sdk (431+ @nullable usages on public API surface) and avoids a thicker runtime story for what is fundamentally a documentation / IDE-hint concern (no NullAway in the build either way). - ActivityChatModel: modelName and baseOptions fields + private ctor params are now @nullable String / @nullable ActivityOptions. Public ctors/factories accept nullable String modelName at the API boundary directly. getModelName() returns @nullable String. - ActivityMcpClient: baseOptions field and callTool(..., summary) param use @nullable instead of Optional<>. 2-arg callTool passes null; McpToolCallback passes a plain String. - ChatModelTypes.ChatModelActivityInput record: @nullable on modelName and modelOptions fields so deserialized readers see the nullability in the signature (per the reviewer's concern about the deserialization-side typesafety). Consistent with org.springframework.lang.Nullable already used elsewhere in this module (TemporalChatClient, SpringAiPlugin). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Switch the five files in this module that imported org.springframework.lang.Nullable over to javax.annotation.Nullable to match the dominant convention in sdk-java (197 usages vs. 7). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…e error classification Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…ies UX wart The activity-summaries branch only overlays Summaries when the factories (forModel/create) built the stub. Users who need custom timeouts today fall back to the public constructor, which silently drops UI Summaries. The ActivityOptions overloads planned here are the proper fix: they let users customize the stub and keep Summary labels. Plan now also covers deprecating the public constructors with javadoc pointing at the factories. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
… AI errors - ActivityChatModel.forDefault(ActivityOptions) and forModel(String, ActivityOptions) overloads added. New public defaultActivityOptions() returns the plugin's default bundle so callers can tweak one field without losing the other sensible defaults. - ActivityMcpClient.create(ActivityOptions) + defaultActivityOptions() added, mirroring the chat side. - Default RetryOptions for chat calls now mark org.springframework.ai.retry.NonTransientAiException and java.lang.IllegalArgumentException non-retryable. Default options for MCP calls mark IllegalArgumentException non-retryable. User-supplied ActivityOptions pass through verbatim — the plugin does not augment them. - new ActivityChatModel(...) and new ActivityMcpClient(activity) constructors are @deprecated with javadoc pointing at the factories — they still work at runtime but skip the UI Summary labels the plugin-owned stub path attaches, which is now called out explicitly. - README: new "Activity options and retry behavior" section documents the defaults, how to customize, and the Summary/factory connection. - Tests: two new suites — ActivityOptionsAndRetryTest covers the non-retryable classification (1 attempt for NonTransientAiException, 3 attempts for transient RuntimeException, custom task queue landing on the scheduled activity); ActivitySummaryTest gains a regression test asserting forDefault(customOptions) still emits UI Summaries. - build.gradle: spring-ai-retry added as a testImplementation so tests can reference NonTransientAiException directly. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Planning scratchpad — not part of the shipped artifact. Removed before merge. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Now that forDefault(ActivityOptions) / forModel(String, ActivityOptions) / create(ActivityOptions) exist, the (Duration, int) convenience overloads are asymmetric dead weight — they expose two of N ActivityOptions fields as positional parameters, and callers wanting anything else (heartbeats, task queue, custom retry backoff, ...) have to drop to the ActivityOptions path anyway. Removed pre-release so the API surface is consistent: no-arg → plugin defaults; ActivityOptions arg → caller options. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
9528294 to
006d8ad
Compare
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
SpringAiPlugin now accepts an optional Map<String, ActivityOptions>
keyed by chat-model bean name. On construction the plugin publishes
the map to a new package-public static registry SpringAiPluginOptions,
which ActivityChatModel.forModel(name) and forDefault() consult when
building the activity stub. Entries resolve by bean name; the reserved
key SpringAiPlugin.DEFAULT_MODEL_NAME ("default") covers forDefault().
Callers who pass explicit ActivityOptions via forModel(name, options)
or forDefault(options) bypass the registry entirely — explicit options
always win. The registry has no effect on the (timeout, maxAttempts)
convenience factory either; that still builds options from its args.
Auto-configuration picks up a user bean named
"chatModelActivityOptions" (constant
SpringAiTemporalAutoConfiguration.CHAT_MODEL_ACTIVITY_OPTIONS_BEAN) of
type Map<String, ActivityOptions>. The explicit bean-name qualifier
avoids Spring's collection-of-beans auto-wiring for Map<String, T>.
Tests: PerModelActivityOptionsTest covers the three cases called out
in the plan — registry hit, registry miss (falls back to default
2-minute timeout), and explicit options bypass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Planning scratchpad — not part of the shipped artifact. Removed before merge. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
1a2adb2 to
d9baf99
Compare
|
How would you feel about having two ways to initialize? (1) default activity options, or (2) explicit options. Combining the default options with explicit options makes it confusing what the final options are. |
…-timeouts # Conflicts: # temporal-spring-ai/build.gradle # temporal-spring-ai/src/main/java/io/temporal/springai/model/ActivityChatModel.java
…-timeouts # Conflicts: # temporal-spring-ai/src/main/java/io/temporal/springai/plugin/SpringAiPlugin.java
|
@brianstrauch Just so we're clear, default options and explicit options don't get merged together ever, but what I think is confusing at the moment is that there is a fallthrough path of: no matching entry in I suggest a very simple tweak to this PR, where Specifically:
|
…' catch-all Address reviewer feedback on #2855: the three-way implicit fallback (specific entry → library defaults) was ambiguous to readers of forDefault() / forModel(name) — nothing locally tells the caller which rung they land on. But a strict either-or (global OR exhaustive per-model) breaks compositionality when a third-party starter contributes a ChatModel bean your config didn't enumerate. Resolution: keep per-model overrides, and reserve ChatModelTypes.DEFAULT_MODEL_NAME as a user-declared catch-all key in perModelOptions. Lookup is now map[name] ?? map["default"] ?? library defaults. SpringAiPlugin validates that every perModelOptions key either matches a registered ChatModel bean or equals the catch-all name; typos fail at plugin construction, not silently at call time. Updates javadoc on SpringAiPlugin, ActivityChatModel.forModel(String), and ChatModelActivityOptions. Collapses the duplicate "Activity options" README section into one with an example of the catch-all pattern. Adds tests for the catch-all lookup, specific-wins-over- catch-all, and typo-key rejection. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Based on
spring-ai/activity-summaries— adds the chat/MCP Summary plumbingspring-ai/retry-and-options— adds theActivityOptionsfactory overloads anddefaultActivityOptions()this PR builds onMerge those in order first; this PR will retarget each time an ancestor lands.
What was changed
SpringAiPluginOptionsstatic registry (packageio.temporal.springai.plugin).SpringAiPluginpublishes aMap<String, ActivityOptions>to it on construction;ActivityChatModel.forModel(String)/forDefault()consult it before falling back to the plugin defaults.SpringAiPlugingains a third constructor that accepts the per-model map; the two existing constructors delegate with an empty map. No constructor changes to existingSpringAiPlugin(ChatModel)/SpringAiPlugin(Map, ChatModel)call sites.ChatModelActivityOptionsrecord (packageio.temporal.springai.autoconfigure) — a thin wrapper aroundMap<String, ActivityOptions>— exists soSpringAiTemporalAutoConfigurationcan inject user options by type rather than by bean name. RawMap<String, ActivityOptions>injection would trigger Spring's collection-of-beans autowiring and sweep in any unrelatedActivityOptionsbean in the context. Using a dedicated wrapper type avoids that collision and keeps the design consistent with temporal-spring-ai: discover MCP clients by type, not by bean name #2859 (MCP bean lookup by type, not by name).SpringAiTemporalAutoConfigurationinjectsObjectProvider<ChatModelActivityOptions>and forwards the wrapped map to the plugin. No magic bean name, no@Qualifier.ActivityOptionsviaforModel(name, options)/forDefault(options)bypass the registry entirely. The(timeout, maxAttempts)convenience factory is unaffected — it still builds options from its arguments.PerModelActivityOptionsTestcovers the three cases in the plan — registry hit uses the registeredstartToCloseTimeout, registry miss falls back to the 2-minute default, explicit options bypass a populated registry entry.Why?
A single default (2 min start-to-close, 3 attempts) doesn't fit every model. Reasoning and thinking-mode models need more time; fast models want shorter timeouts so retries recover quickly. The Temporal AI integration guide specifically calls this out as a capability partners should expose. With this change, users register one bean and never have to hand-build activity stubs again.
Example user config: