Updates

Release notes for the @silverprotocol/* packages. Every cohort ships all packages in lockstep, npm-attested, with the verification evidence on each package's compatibility table.

1.0.0-rc.7

AgJSON 1.0.0-rc.7: Google ADK 2.4.0 streamed text, Interactions closes and function-response members; a message keeps its turn's thread

The seventh release candidate for 1.0.0. All six packages move to 1.0.0-rc.7 together. @silverprotocol/core, @silverprotocol/google-adk and @silverprotocol/claude-agent-sdk change; the other three have no code change. The spec doesn’t change: the packages still ship spec revision 1.0.0-rc.3 (AGJSON_VERSION).

Installing the release candidate

The release candidate is published under the rc tag. latest is on 0.12.11:

npm i @silverprotocol/core@rc

Install each framework package you use the same way, so every @silverprotocol/* package in your lockfile is on 1.0.0-rc.7. While a release candidate is the most recent publish, a package’s page on npmjs.com shows no README; each package still carries its README.

0.12.11 ships the same @silverprotocol/core fix and the same three Google ADK fixes on the 0.12 line, and 1.0.0-rc.7 carries everything 0.12.11 changes.

Who should upgrade

  • You run Google ADK in StreamingMode.SSE with a planner that writes state (@google/adk 2.4.0). Streamed text rendered twice on rc.6; it now renders once.
  • You run Google ADK on the Interactions route (useInteractionsApi) in StreamingMode.SSE. A response’s text now renders once, and the response closes one turn.
  • Your Google ADK tools return function responses with parts, scheduling or willContinue. Those members now reach the stream; rc.6 dropped them.
  • You resume Claude Agent SDK sessions (@anthropic-ai/claude-agent-sdk 0.3.317 or later). A session whose writes were lost on a full disk is now reported.
  • Anyone else gets the same output as from rc.6: all 448 replayed runs are byte-identical (see Wire version). Read “Upgrading from rc.6” first: the message-thread change applies to every framework package’s output.

Upgrading from rc.6

  • A message lands on its turn’s thread. When a message.start names a threadId other than the one its turn already has (from the turn’s turn.start, from its parent’s thread for a nested turn, or from an earlier message.start), the reducer now creates the message on the turn’s thread. rc.6 kept the message.start’s own threadId. The turn record’s thread doesn’t move, and the fold doesn’t resync. A message.start for a turn that has no thread yet still gives the turn its thread.
  • Google ADK: an unmatched function response that holds credential material is no longer carried. A functionResponse that carries no id (or an empty one) and answers no pending call under its name rides ext.google.unparsed whole. When it holds an ADK credential object or a response to adk_request_credential, at any depth, the package now emits one non-terminal error event in its place, with code "adk_orphan_credential_response" and a fixed message, and nothing of the response rides (spec §8.0, graceful degradation). rc.6 carried such a response whole, credential members included. The error doesn’t close the turn. An unmatched response without credential material still rides whole, a reply named adk_request_credential keeps its allowlisted carry, and a response that answers a call still maps to tool.done. A credential an application keeps in a shape of its own isn’t recognised, and rides unchanged.

@silverprotocol/core

  • The message-thread change above. No other fold rule changes.

Google ADK

  • A planner’s state record no longer doubles streamed text. From @google/adk 2.4.0, with a planner that writes state, ADK yields an event that only records the state change inside a model call, between its streamed chunks and its repeat of them in StreamingMode.SSE. rc.6 read that event as the end of the call, so the repeat rendered the text a second time. The record still maps to its state.delta, and it now holds the match open for its source’s next event: a chunk, a partial: false repeat or another such record keeps it, and any other event of the source ends it first. Events of other sources don’t touch it. The repeat renders nothing, as without the record.
  • A repeated close of an Interactions response maps once. On the Interactions route (useInteractionsApi) in StreamingMode.SSE, ADK closes a response at interaction.completed and again when the stream ends, and a completed status update can close it once more, each time with every part so far. rc.6 rendered the response again at the second close. A close that maps to exactly what the response’s previous close mapped to, its id and timestamp aside, with no other event of the response between them, now emits nothing: the response renders once and closes one turn, with or without the host-completion opt-in. A close with any mapped difference maps as before, and a partial: false event ADK yields mid-response (a media part, a finished function call) is not a close and always maps.
  • A function response’s parts, scheduling and willContinue are carried. A tool.done has no slot for these members of a genai FunctionResponse, and rc.6 dropped them. They now ride one provider-raw block, verbatim, right before that call’s tool.done, with the response’s own name and id. The carried value is reduced like every other provider-raw carry (spec §8.0 item 28), and a reply to adk_request_credential never takes it. A response without these members maps as before.
  • The unmatched-response change under “Upgrading from rc.6”.

Claude Agent SDK

  • A session whose writes were lost is reported. From @anthropic-ai/claude-agent-sdk 0.3.317, the CLI sends a system message with subtype session_writes_lost as it exits when the disk was full or over quota and lines of a session transcript or changes to ~/.claude.json were not saved. rc.6 emitted nothing for it. It now rides ext.anthropic.frame with kind session_writes_lost, verbatim, after the turn’s close, and the turn’s outcome doesn’t change. A host that resumes one of the sessions it lists may find part of that conversation missing.

Vercel AI SDK

  • With @ai-sdk/openai 4.0.110 or later, an OpenAI Responses refusal reaches the stream. From 4.0.110, the provider streams a refusal as the assistant’s text and finishes the step with content-filter. This package maps that finish, as before, to turn.done with finishReason "safety_blocked", and sets no finishReasonRaw for it. With 4.0.109 or earlier, the refusal text never reached the stream: the host saw an empty assistant message, and the turn closed with finishReason "stop". The package’s code doesn’t change.

OpenAI Agents SDK

  • No code change. The package is now verified over the openai client 7.53.0, where rc.6 was verified over 7.22.0. Of the Responses fields the client typed in between, the usage detail cache_write_tokens_details (7.50.0) would ride the existing usage-details event, though no recorded stream carries it yet. tool_usage (typed in 7.46.0) and image.context_compression (7.51.0) are not carried.

Spec

Unchanged. The packages ship spec revision 1.0.0-rc.3, as rc.6 did, so no closed set, schema or fold rule changes.

Verified versions

Each framework package is verified end to end, tool calls included, against a live run at:

  • @anthropic-ai/claude-agent-sdk 0.3.319;
  • @openai/agents 0.21.0;
  • @google/adk 2.4.0, with @google/genai 2.32.0;
  • ai 7.0.166, with @ai-sdk/openai 4.0.111.

Wire version

1.0.0-rc.3, unchanged. Replaying every recorded stream and engine-built fixture in the release’s test corpus through the 1.0.0-rc.6 and 1.0.0-rc.7 packages, all 448 runs give byte-identical events. The replay compares the events each package emits, not a fold of them, and no recorded stream or fixture carries an unmatched credential response, a function response’s new members, a planner’s state record, a lost-writes message or a stream on the Interactions route; those changes rest on their own tests. A live capture on the Interactions route, recorded outside that corpus in StreamingMode.SSE at gemini-3.8-flash, shows the repeated-close fix: rc.6 renders the response twice, in 30 events, and rc.7 renders it once, in 19.

The 1.0.0-rc.7 packages are built from typescript-sdk commit e06c4523, the commit npm’s provenance for all six 1.0.0-rc.7 packages attests.

1.0.0-rc.6

AgJSON 1.0.0-rc.6: nested emphasis and continued list items in rich text, and a spec erratum

The sixth release candidate for 1.0.0. All six packages move to 1.0.0-rc.6 together. Only @silverprotocol/richtext changes; the other five change only their version. The spec revision stays 1.0.0-rc.3 (AGJSON_VERSION), and its text gains one erratum.

Installing the release candidate

The release candidate is published under the rc tag. latest is on 0.12.10:

npm i @silverprotocol/core@rc

Install each framework package you use the same way, so every @silverprotocol/* package in your lockfile is on 1.0.0-rc.6. While a release candidate is the most recent publish, a package’s page on npmjs.com shows no README; each package still carries its README.

0.12.10 ships the same rich text fix on the 0.12 line, and 1.0.0-rc.6 carries everything 0.12.10 changes.

Who should upgrade

  • You render chat text with parseRichText from @silverprotocol/richtext. Nested emphasis and indented continuation lines under a list item now parse as written. Read “Upgrading from rc.5” below first.
  • Anyone else gets the same events and fold results as from rc.5: @silverprotocol/core and the four framework packages change only their version.

Upgrading from rc.5

  • A table or heading line indented under a list item is now item text. rc.5 ended the list there: an indented pipe table became a top-level table, and an indented # heading line a paragraph. The line now joins the item’s text, as the README’s subset already states for block content inside a list item. The AST types don’t change: a RichTextListItem still holds inline content only.

@silverprotocol/richtext

  • An emphasis run closes the innermost open emphasis whose marker it matches exactly. *a **b** c* now parses as an em holding a strong; rc.5 closed the em at the strong’s ** and left * c* as literal text. *a *b* c* and _a __b__ c_ parse the same way. **a *b** c and *** runs parse as before, and emphasis still never closes across a [.
  • An indented line directly under a list item continues that item. A line indented by spaces or a tab joins the item’s inline content after a break, and the item’s lines parse as one inline run, so emphasis can span them. - item one, continued line, - item two is now one list of two items, where rc.5 gave a list, a paragraph and a second list. An unindented line or a blank line still ends the list.

Spec

The revision stays 1.0.0-rc.3, so no closed set, schema or fold rule changes. Its text carries the spec’s first erratum, in §15:

Erratum (2026-10-27) — the host-reading gloss in §4. It corrects rc.3; the sentence entered at rc.1. §4’s second informative parenthetical of the costUsd note, which read “— a turn.abort, or a turn.error or turn.done whose usage is absent or carries no token counter —”, now reads “— a turn.abort, or a turn.error or turn.done whose usage is absent or carries no top-level token counter —”. The token counters that parenthetical and §10 item 67 name are the seven top-level members of AgUsage (inputTokens, outputTokens, cacheReadTokens, cacheWriteTokens, reasoningTokens, toolUseInputTokens and totalTokens), and a nested byModel entry carries its own cumulative flag; the word pair states the reading the enumeration already gives and changes no meaning. The sentence is informative; no field, event, fold or requirement moves, and no version moves.

Verified versions

Unchanged from rc.5. Each framework package is verified end to end, tool calls included, against a live run at:

  • @anthropic-ai/claude-agent-sdk 0.3.310;
  • @openai/agents 0.21.0;
  • @google/adk 2.3.0, with @google/genai 2.31.0;
  • ai 7.0.153, with @ai-sdk/openai 4.0.102.

Wire version

1.0.0-rc.3, unchanged.

The 1.0.0-rc.6 packages are built from typescript-sdk commit 2b49c89c, the commit npm’s provenance for all six 1.0.0-rc.6 packages attests.

1.0.0-rc.5

AgJSON 1.0.0-rc.5: a cut-off Vercel AI SDK stream closes as an interruption

The fifth release candidate for 1.0.0. All six packages move to 1.0.0-rc.5 together. The spec doesn’t change: they still ship spec revision 1.0.0-rc.3 (AGJSON_VERSION).

Installing the release candidate

The release candidate is published under the rc tag. latest is on 0.12.9:

npm i @silverprotocol/core@rc

Install each framework package you use the same way, so every @silverprotocol/* package in your lockfile is on 1.0.0-rc.5. While a release candidate is the most recent publish, a package’s page on npmjs.com shows no README; each package still carries its README.

0.12.9 ships the same Vercel AI SDK fix on the 0.12 line, and 1.0.0-rc.5 carries everything 0.12.9 changes.

Who should upgrade

  • You run the Vercel AI SDK with @ai-sdk/openai 4.0.100 or later. A stream cut off before its completion signal now closes as an interruption, where rc.4 closed it as a success.
  • Anyone else gets the same output as from rc.4: all 446 replayed runs are byte-identical (see Wire version).

Upgrading from rc.4

  • Vercel AI SDK: a cut-off stream now closes turn.abort with reason "interrupted", not a success turn.done. From @ai-sdk/openai 4.0.100, the provider starts every stream (Responses, Chat Completions and Completions) at finish reason interrupted and moves off it only when its completion signal arrives, so a stream cut off mid-answer ends with finish-step and finish both interrupted. From ai 7.0.149, interrupted is a member of FinishReason and stops the step loop. rc.4 closed that turn as a success turn.done, with finishReason "unknown" and finishReasonRaw "interrupted". rc.5 closes it turn.abort with reason: "interrupted", as the spec requires for a native interruption (§8.0 item 17), and each step’s usage still rides its finish-step’s message.end. A UI that renders a success turn as complete now shows a cut-off answer as stopped. With @ai-sdk/openai 4.0.99 or earlier, the provider reports a cut-off stream as other and gives no interruption signal, so the same stream closes turn.done with finishReason "other", on this version and every earlier one.

Google ADK

  • With @google/genai 2.31.0, an error the service sends mid-stream is reported. From genai 2.31.0, an error that arrives as an event of an SSE stream (a rate limit or an overloaded model after the stream began) reaches ADK as a thrown error, and ADK 2.3.0 reports it as an in-band error with error code UNKNOWN_ERROR and the HTTP status inside its message. The package emits an error event and closes the turn turn.error, as for any in-band error (§8.0 item 35). Under genai 2.30.0 the same stream lost the error: a stream with text before the error closed turn.abort "stream-truncated", or a success turn.done with the host-completion event, and an error-only stream opened no turn. The package’s code doesn’t change; its README documents this.

Spec

Unchanged. The packages ship spec revision 1.0.0-rc.3, as rc.4 did, so no closed set, schema or fold rule changes.

Verified versions

Each framework package is verified end to end, tool calls included, against a live run at:

  • @anthropic-ai/claude-agent-sdk 0.3.310;
  • @openai/agents 0.21.0;
  • @google/adk 2.3.0, with @google/genai 2.31.0;
  • ai 7.0.153, with @ai-sdk/openai 4.0.102.

Wire version

1.0.0-rc.3, unchanged. Replaying every recorded stream and engine-built fixture through the 1.0.0-rc.4 and 1.0.0-rc.5 packages, all 446 runs give byte-identical events: 85 recorded streams and 361 runs of the engine-built Google ADK and OpenAI fixtures. No recorded Vercel AI SDK stream carries the provider’s interrupted finish reason. A cut-off stream built from parts captured at ai 7.0.153 with @ai-sdk/openai 4.0.102 closes turn.done on rc.4 and turn.abort "interrupted" on rc.5, and the same stream run to completion closes turn.done on both.

The 1.0.0-rc.5 packages are built from typescript-sdk commit 3aa8e45b, the commit npm’s provenance for all six 1.0.0-rc.5 packages attests.

1.0.0-rc.4

AgJSON 1.0.0-rc.4: @openai/agents 0.21, and 0.12.7's clone fix on the release candidate

The fourth release candidate for 1.0.0. All six packages move to 1.0.0-rc.4 together. The spec doesn’t change: they still ship spec revision 1.0.0-rc.3 (AGJSON_VERSION).

Installing the release candidate

The release candidate is published under the rc tag. latest is on 0.12.8:

npm i @silverprotocol/core@rc

Install each framework package you use the same way, so every @silverprotocol/* package in your lockfile is on 1.0.0-rc.4. While a release candidate is the most recent publish, a package’s page on npmjs.com shows no README; each package still carries its README.

1.0.0-rc.4 includes 0.12.7’s change, which rc.3 didn’t: its @silverprotocol/core no longer calls the global structuredClone. It also carries everything 0.12.8 changes.

Who should upgrade

  • You run @silverprotocol/core on an engine without a global structuredClone. rc.3 fails there; rc.4 runs, as 0.12.7 does.
  • You run @openai/agents 0.21. rc.4 is the first release whose @silverprotocol/openai-agents accepts it: the peer range widens from <0.21 to <0.22. Read “Moving to @openai/agents 0.21” below first.
  • You run the Vercel AI SDK with a provider that cites sources (ai 7.0.138 or later). Its citations now land on the text block they cite.
  • You use an OpenAI-compatible provider through @openai/agents that reports usage detail counters beyond the ones the OpenAI client’s types define. Those counters are now carried, where rc.3 dropped them.
  • You read per-model usage from the Claude Agent SDK across a query with more than one result, as @anthropic-ai/claude-agent-sdk 0.3.310 reports it.
  • Anyone else gets the same output as from rc.3: of 446 replayed runs, only three differ, each by a member named below (see Wire version).

Moving to @openai/agents 0.21

These are changes in @openai/agents 0.21 itself. They reach a host that upgrades, whichever version of this package it runs:

  • A run paused and saved on 0.21 can’t be resumed on 0.20. 0.21 saves run state at schema version 1.23, and 0.20 rejects it on resume with a UserError: “Run state schema version 1.23 is not supported. Please use version 1.22.” A run saved on 0.20 resumes on 0.21. Upgrade every process that resumes saved runs before any process saves one on 0.21.
  • Usage token details can be nested. In the SDK’s own Usage object, inputTokensDetails and outputTokensDetails are now records whose values are numbers or further records, so narrow each value to a number before you do arithmetic with it. The AgJSON usage this package emits doesn’t change shape: it reads the response’s own usage, and a detail key it doesn’t map rides the usage-details event described below.
  • The Runloop sandbox provider is removed, with its public API. OpenAI’s migration guide lists the supported alternatives. This package never touches the sandbox.

@silverprotocol/core

These two changes shipped on the 0.12 line as 0.12.7 and are new to the release candidate.

  • The fold no longer calls the global structuredClone. Reducer.push(), result(), toPersistable(), toPersistableWithReport() and the JSON Patch applier now copy with core’s own deep copy. On a JavaScript engine that has no global structuredClone, every fold used to fail with “structuredClone is not defined”; it now runs. For every AgJSON value, the copy gives the same result as before, so no event, fold result or toPersistable() output changes.
  • A host-placed non-JSON object is now shared, not copied. If a host puts a Date, Map, Set, RegExp, Error, ArrayBuffer, typed array or boxed primitive inside an event it pushes, the fold no longer copies it. The fold, every result() and toPersistable() share that object with the host’s event, so a host that changes it after push() sees the change in the fold. Nothing throws that didn’t on rc.3. A function or symbol in the fold still throws where it did, now with an error named CloneError (a TypeError) where rc.3 threw a DOMException named DataCloneError. No AgJSON value holds any of these, and none arrives through ingestAgEvent() or from the framework packages.

The supported engines don’t change: the packages still declare node >=22.12.0.

OpenAI Agents SDK

  • The package accepts @openai/agents 0.21: its peer range widens to >=0.2.0 <0.22. 0.21 needed no mapping change of its own.
  • Usage detail counters the package doesn’t map are no longer dropped. The package maps cached_tokens, cache_write_tokens and reasoning_tokens from a response’s input and output token details, and dropped every other key. The OpenAI client’s types (openai 7.43.0) define no other key; an OpenAI-compatible provider can send flat counters or a nested breakdown, and 0.21 keeps them through a streamed run. They now ride one ext.openai.unparsed event per response, name: "usage-details", before the round’s terminal, holding the unmapped entries as received. Nothing is emitted when every key is mapped. A mapped key whose value isn’t a number (a provider’s null) now sets no usage counter, where rc.3 put that null in usage, and rides the same event.

Vercel AI SDK

  • A URL citation lands on the text block it cites. From ai 7.0.138, a text-end part can carry citations. A URL-source citation now maps to a citation of kind "url" on the text block’s text.end, with the provider’s cited text, else the block text between its offsets, as the OpenAI package maps OpenAI’s own URL citations. Any other entry, such as a document source, rides ext.vercel.citations with its text id.
  • A tool call’s input encoding is carried. From ai 7.0.142, tool-input-start can carry inputEncoding; it rides ext.vercel.input-encoding with the call id.
  • A provider-executed tool call keeps providerExecuted on its tool.start when the call arrives without streamed input, so a client knows not to execute it.
  • The package drops its one structuredClone call, as in 0.12.7. Its output is unchanged.

Claude Agent SDK

  • Per-result model usage rides the result. At @anthropic-ai/claude-agent-sdk 0.3.310, a result frame reports what each model spent since the query’s previous result. It now rides ext.anthropic.result-meta as modelUsageSinceLastResult, each model’s entry mapped like a usage.byModel entry with cumulative: false. usage.byModel still holds the running totals. A frame without it, or with an empty one, emits as before.

Google ADK

  • No code change. The package’s README now documents, at @google/adk 2.3.0, how an SSE prompt block closes, a failed Interactions call that arrives without its error code, and the Chrome built-in backend’s fixed reply.

Spec

Unchanged. The packages ship spec revision 1.0.0-rc.3, as rc.3 did, so no closed set, schema or fold rule changes.

Verified versions

Each framework package is verified end to end, tool calls included, against a live run at:

  • @anthropic-ai/claude-agent-sdk 0.3.310;
  • @openai/agents 0.21.0;
  • @google/adk 2.3.0, with @google/genai 2.30.0;
  • ai 7.0.146.

Wire version

1.0.0-rc.3, unchanged. Replaying every recorded stream and engine-built fixture through the 1.0.0-rc.3 and 1.0.0-rc.4 packages, 443 of 446 runs give byte-identical events. The 446 are 85 recorded streams, among them the Claude Agent SDK (33), OpenAI Agents SDK (16), Google ADK (25) and Vercel AI SDK (10) streams, and 361 runs of the engine-built Google ADK and OpenAI fixtures. The three that differ each add one member described above: two Claude streams gain modelUsageSinceLastResult on their result metadata, and one Vercel web-search stream gains citations on a text.end. No recorded stream or fixture carries, in a response’s usage, a detail key the package doesn’t map.

The 1.0.0-rc.4 packages are built from typescript-sdk commit 3e03011d, the commit npm’s provenance for all six 1.0.0-rc.4 packages attests.