@tangle-network/agent-knowledge 17.1.10 → 19.0.0

package/CHANGELOG.md

@@ -1,5 +1,40 @@
+## 19.0.0
+
+Knowledge tools require the digest from knowledge_read when updating an existing page.
+Stale writes refuse the complete proposal under the shared mutation lock.
+An identical retry preserves the completed write.
+Write transactions retain the host actor and run identity, with unknown authors explicitly null.
+Knowledge tools retain prior and new page bytes by default; applications can explicitly disable history.
+Lower-level proposal writers can select the same conditional-write contract.
+
 # Changelog
 
+## 18.0.0 — 2026-09-29
+
+Require Eval 0.201 through 0.203, Interface 2.15, and TCloud 0.8.
+Require Node.js 22.12 to match TCloud's supported runtime.
+Use Zod 4.6.5 with the Core 0.10 and Trace Contract 2 dependency cohort.
+Knowledge 17 remains the release for older Eval contracts.
+Run packed-package and official optimizer checks against every supported Eval minor.
+Keep knowledge records, source formats, and public knowledge operations unchanged.
+
+## 17.1.12 — 2026-09-28
+
+Allow consumers to install public Eval 0.200 alongside Knowledge after packed-package checks. Keep the previously supported Eval 0.182 through 0.199 releases; the 0.200.1 development pin replaces 0.199.0.
+Eval 0.201 is not admitted yet: it requires Core 0.10 and Zod 4.6.5, while the Sandbox release TCloud installs still requires Core 0.9 and Zod 4.5.4, so a consumer would hold two copies of each.
+Build with tsdown 0.23 and pnpm 12. The exported symbols and their shapes are unchanged.
+
+## 17.1.11 — 2026-09-28
+
+Use published TCloud 0.6 for research chat and search.
+Require Interface 2.13 to match the installed Sandbox dependency.
+Keep the public `RouterError`, retry options, and injectable `RouterClient` contract.
+Forward cancellation into HTTP and attribute cost from each response.
+Pass run cancellation through default worker and verifier requests, including claim extraction.
+Stop research after cancellation before fallback queries, source writes, or round events.
+Set `usage().usd` to NaN if a successful response omits its cost receipt.
+Reject search results without URLs and retain the caller-controlled reasoning timeout.
+
 ## 17.1.10 — 2026-09-27
 
 Allow consumers to install public Eval 0.199 alongside Knowledge after packed-package checks. Keep the previously supported Eval 0.182 through 0.198 releases; the 0.199.0 development pin replaces 0.198.0.

package/README.md

@@ -9,13 +9,16 @@ Supply application callbacks for those decisions, or use `@tangle-network/agent-
 ## Install
 
 ```bash
-pnpm add @tangle-network/agent-knowledge@17.1.4 @tangle-network/agent-eval@0.189.0 @tangle-network/agent-interface@2.10.0
+pnpm add @tangle-network/agent-knowledge@18.0.0 @tangle-network/agent-eval@0.203.0 @tangle-network/agent-interface@2.15.0
 ```
 
-Requires Node.js 20.19 or later.
+Requires Node.js 22.12 or later.
 
-Eval 0.182 through 0.189 are supported.
-Package and official optimizer checks exercise every released minor in that window; development remains pinned to 0.182.0.
+Eval 0.201 through 0.203 are supported.
+Package and official optimizer checks exercise each supported minor; development uses 0.203.0.
+These releases use Core 0.10, Trace Contract 2, and Zod 4.6.5.
+Knowledge 18 requires TCloud 0.8 with Sandbox 0.58.4 or later in that minor.
+Use Knowledge 17 with older Eval releases.
 
 ## Choose an API
 
@@ -201,11 +204,21 @@ const tools = createKnowledgeTools({
 `knowledge_record` writes into this run's store through the intake gate.
 Each proposal must contain complete `---FILE: <page-path>---` / `---END FILE---` blocks, with delimiters on separate lines.
 The tool rejects malformed, unsafe, or empty proposals before writing any pages.
+Read an existing page before editing it.
+Pass expectedPageDigests with the page path and pageDigest returned by knowledge_read.
+The write checks each digest under the store lock.
+A stale edit refuses the whole proposal and names the current digest.
+New pages need no digest; an explicit null requires the path to be absent.
+Retrying an identical completed write is safe.
 It does not report a partial write as tool success.
 The lower-level `applyKnowledgeWriteBlocks` API retains its explicit `written` and `warnings` result for callers that inspect partial proposals.
 `knowledge_resolve` returns the resolution status of each reference.
 
-When a pursuit must preserve every edit for later refinement or branch reconciliation, pass `retainHistory: true`.
+Knowledge tools retain write history by default.
+Each transaction records the host actorId, runId, and exact prior and new bytes.
+An absent author remains explicitly null.
+An application can disable history with retainHistory: false.
+Lower-level writers opt into history with retainHistory: true and conditional edits with expectedPageDigests.
 Completed write transactions then retain their existing manifest and before/after snapshots under `.agent-knowledge/history/<transactionId>`.
 The move is atomic and a repeated finish after a lost acknowledgement is idempotent.
 The default remains cleanup after completion, so callers choose retention deliberately and account for its unbounded storage growth.