@xiaozhiclaw/qlogicagent-core 0.6.5 → 0.6.7

package/README.md

@@ -200,8 +200,18 @@ with a different context projection requires explicit product migration; Core
 does not reinterpret its saved input or tool receipts. Checkpoints without an
 assistant context retain their existing binding protocol.
 Optional `goal` requires a product `verifyGoal` callback before semantic completion.
-Products may resolve a user-confirmed goal from receipts with `resolveGoal`, and
-return independent verification evidence to the same loop through `goalFeedback`.
+Products may resolve a user-confirmed goal from receipts with `resolveGoal`.
+Incomplete evidence returns to the same loop by default; `goalFeedback` can
+override its wording or return `[]` to stop. Unchanged incomplete evidence stops
+repeated completion claims. Additional read receipts or worker replacement do not
+prove business progress; this comparison reuses the persisted outcome.
+Two consecutive identical batches rejected before tool dispatch, with no new
+input, end with `TOOL_ARGUMENTS_STALLED`. The check uses checkpoint history;
+changed proposals, mixed valid work and new input keep their normal execution path.
+CloudTask terminal events report `taskOutcome: "satisfied" | "incomplete"` after
+independent verification; an incomplete outcome also has `turnStatus: "stopped"`.
+Without a confirmed goal, `taskOutcome` remains `"unverified"`; lifecycle completion
+does not prove a business result. Product budgets, authorization and waits still apply.
 The model can correct its work or use existing input/confirmation wait tools;
 the product continues to own the acceptance criteria and verification rules.
 Criteria may carry opaque product `parameters`; verification can return `expected`,
@@ -211,7 +221,7 @@ Context anchors carry explicit revision/supersession/source references through
 compaction, keeping unresolved conflicts visible. Product tools retrieve material.
 Optional `runtime.prepareContext(messages, context)` asynchronously retrieves and
 projects product task materials before inference. `context` carries the model,
-input token budget, session/turn IDs and cancellation signal.
+input token budget, resident tool schemas, session/turn IDs and cancellation signal.
 The budget is `Infinity` when context-capacity metadata is explicitly disabled;
 retrieval still runs and the provider remains responsible for its actual capacity.
 Return `{ messages, droppedCount, strategy }`; retrieval and fact updates are applied even
@@ -224,8 +234,22 @@ fills remaining capacity with recent history. These essential messages may excee
 capacity; Core does not silently drop them or pretend an estimate proves model fit.
 Optional `runtime.countMessageTokens(message, model)` supplies model-aware counts
 for the default window. Its fallback is an advisory serialized-message estimate.
-The input budget reserves configured maximum output tokens rather than a fixed
-percentage; provider context errors remain authoritative.
+The default projection reserves configured maximum output tokens and resident
+tool schema cost before selecting message history; provider context errors remain authoritative.
+Optional `runtime.summarizeContext({previousSummary, messages}, context)` summarizes
+only omitted complete protocol groups and merges the prior summary. Configure it
+through the product's existing metered model service. The callback receives a
+summary token allowance and cancellation signal; it owns model authorization,
+cost accounting, deduplication and unknown-outcome recovery for that work. Its
+usage is not automatically included in the main loop's model-operation ledger.
+The default selector reserves a quarter of message capacity for a summary, pins
+original intent, the latest amendment, facts and the latest complete protocol group,
+and passes remaining older groups to the callback. It saves one derived summary in
+the same checkpoint. Empty/oversized summaries or callback failures fail explicitly;
+they do not silently fall back to discarding history. A provider-acknowledged response
+replays its original projection without invoking the summarizer again.
+`prepareContext`, when supplied, owns the whole projection and takes precedence.
+Summaries are data, never instructions, authenticated facts or authorization.
 After retrieving current product facts, call public
 `prepareSlidingWindowContext(messages, context, countMessageTokens?)` to reuse
 the same canonical window implementation. Product retrieval must not accidentally
@@ -245,6 +269,15 @@ that contract and enforce the original ceiling atomically; it remains responsibl
 for supported currencies, precision and settlement. The model transport supplies
 normalized usage totals; source names in usage records stay opaque to Core.
 
+Products explicitly mark independent tools with `meta.executionMode: "parallel"`
+and set `config.maxConcurrentTools` to permit overlap. Omitted or `"sequential"`
+execution modes form barriers: all earlier calls settle before that call starts,
+and later calls wait for it. Consecutive parallel-safe calls share the same bounded
+executor. No safety is inferred from tool names, categories or model arguments.
+Results stream as each call finishes; model/checkpoint history retains original
+call order. Cancellation or a wait stops queued calls; already dispatched effects
+still need their original receipts and the existing reconciliation path.
+
 Cloud execution does not invent a round limit, per-response tool-call limit,
 retry-count limit, total presentation-block limit, compaction-count stop or default
 tool timeout. Products may supply per-tool `meta.executionTimeoutMs`; task deadlines