chore(frontend): bump A2A, AI SDK v6, streamdown v2; re-sync AI Elements

[malinskibeniamin]

Summary

Upgrade the AI Agents chat toolchain to current majors and pull in the
vercel/ai-elements features the UI needs (MCP tool-approval flow,
conversation markdown export), preserving all Redpanda customisations.

Versions

Package	Before	After	Kind
@a2a-js/sdk	^0.3.10	^0.3.13	patch
ai	^5.0.101	^6.0.168	major
streamdown	^1.4.0	^2.5.0	major

@modelcontextprotocol/sdk is intentionally not bumped here — it
ships in a separate PR.

Why this is safer than it looks

• ai v6. Our custom A2aChatLanguageModel implements LanguageModelV2.
  v6 keeps the V2 types exported and accepts them via the public
  LanguageModel alias (V3 | V2), so no adapter rewrite is needed.
  The one shape change we hit is LanguageModelUsage requiring
  inputTokenDetails / outputTokenDetails; useContextUsage was
  updated accordingly.
• streamdown v2. Purely internal remark/katex plumbing changes;
  our response.tsx consumer is source-compatible.
• ai-elements. Every upstream file was diffed against ours and
  deltas applied surgically. Redpanda-specific customisations
  (theming, token-count avatars, Badge variants, artifact metadata)
  are preserved verbatim.

What we adopted from upstream ai-elements

• ConversationDownload + messagesToMarkdown helper
• DynamicToolUIPart support and three new tool states:
  approval-requested, approval-responded, output-denied
• Context provider memoisation and zero-token guards
• Module-level cached motion components in shimmer

What we deferred (follow-up PRs)

• Full prompt-input.tsx merge (heavy divergence, needs its own focused pass)
• loader.tsx → shadcn/ui Spinner migration (upstream deprecation)
• Vendoring ChainOfThought, Reasoning, Attachments,
  PromptInputActionAddScreenshot (not wired into the agents UI yet)

Risk and verification

• Typecheck, unit, and integration suites all green; 19 new tests added
  for the adopted behavior (tool, conversation, use-context-usage).
• No new lint errors introduced on touched files.
• No call sites broken by v6 — adapter compiles unchanged; only the
  usage hook needed the new sub-object fields.

Screenshots

Captured via vitest browser mode (Chromium, headless, viewport 1440×900,
reduced-motion). Source tests live under
frontend/src/components/ai-elements/*.browser.test.tsx.

Tool states

State	Screenshot
input-streaming
input-available
output-available
output-error
approval-requested (new)
approval-responded (new)
output-denied (new)
DynamicToolUIPart with toolName override (new)

ConversationDownload

State	Screenshot
Idle
Hover

Context hover-card

State	Screenshot
Zero tokens (guards hide sub-rows)
Populated usage with sub-objects

Shimmer

State	Screenshot
Loading (frozen frame under reduced-motion)

Test plan

• bun install && bun start, open an AI Agents chat.
• Trigger a tool call; confirm Tool card renders with Redpanda badge
  variants, duration, and toolCallId.
• Trigger an MCP tool call requiring approval; confirm the three
  new states render (Awaiting Approval → Responded / Denied).
• Open the Context hover-card; percentages and per-section costs
  still display, zero-token rows hidden.
• Verify streaming text, artifact rendering, and task transitions.
• If wired in, confirm ConversationDownload produces a well-formed
  markdown file.

🤖 Generated with Claude Code

[malinskibeniamin]

@claude please review this PR. Specifically focus on:

• A2A adapter v6 compatibility — is the implementation correct given v6's LanguageModel = V3 | V2 alias, and are there any edge cases in doStream that might regress
• The vendored ai-elements re-sync — did I preserve all Redpanda customisations correctly, any upstream features I missed adopting that would benefit the agents UI
• useContextUsage shape — correctly populating both new inputTokenDetails/outputTokenDetails and deprecated top-level fields for the Context component that still reads legacy
• Test coverage — are there critical-path behaviors still untested that we should cover in this PR vs follow-up

Flag any concerns even if minor. Don't worry about pre-existing lint noise.

[claude]

Claude finished @malinskibeniamin's task in 3m 16s —— View job

---

PR Review

Overall this is a well-executed bump. The v6 compatibility path is sound and the ai-elements re-sync is surgical. A few findings below, ordered by severity.

---

1. Bug — CopyButton title uses toolName instead of displayName (minor)

frontend/src/components/ai-elements/tool.tsx, line ~193

<CopyButton
  content={textToCopy}           // ← correctly uses displayName
 
    ? `Copy: ${toolName} (${toolCallId})`   // ← uses toolName (undefined for static tools)
    : `Copy: ${toolName}`}
/>

For static-typed tools (type="tool-get-weather", etc.) toolName is typed never and is undefined at runtime, so the tooltip renders "Copy: undefined" or "Copy: undefined (id-123)". The adjacent textToCopy already uses displayName correctly. The fix is a one-liner:

```tsx ? Copy: ${displayName} (${toolCallId}) : `Copy: ${displayName}`}

[Fix this →](https://claude.ai/code?q=In%20frontend%2Fsrc%2Fcomponents%2Fai-elements%2Ftool.tsx%2C%20on%20the%20CopyButton%20title%20prop%20inside%20ToolHeader%2C%20replace%20%60toolName%60%20with%20%60displayName%60%20so%20static%20tools%20don%27t%20render%20%22Copy%3A%20undefined%22.%20The%20line%20reads%3A%20title%3D%7BtoolCallId%20%3F%20%60Copy%3A%20%24%7BtoolName%7D%20(%24%7BtoolCallId%7D)%60%20%3A%20%60Copy%3A%20%24%7BtoolName%7D%60%7D%20and%20should%20use%20displayName%20instead.%20Branch%3A%20chore%2Fbump-a2a-mcp-ai-elements&repo=redpanda-data/console)

---

### 2. Pre-existing bug surfaced by this bump — non-streaming fallback sends two requests

**`frontend/src/components/ai-elements/../a2a/a2a-chat-language-model.tsx`, `doStream` ~line 290**

```ts
if (!clientCard.capabilities.streaming) {
  const nonStreamingResponse = await client.sendMessage(streamParams);  // ← request 1
  if ("result" in nonStreamingResponse) { simulatedStream = ...; }
  if ("error" in nonStreamingResponse) { /* FIXME: error */ }            // ← silent swallow
}

const response = client.sendMessageStream(streamParams);                  // ← request 2, always
return {
  stream: (simulatedStream || convertAsyncIteratorToReadableStream(response)).pipeThrough(...)
};

Two issues:

• sendMessageStream is called unconditionally: even when simulatedStream is non-null and preferred, a live SSE request fires against the agent. The message is sent twice.
• The error branch is silently ignored; simulatedStream stays null and the code falls through to the streaming path — sending a second message after the first already failed.

Not introduced in this PR, but the v6 bump is a good moment to fix (or at minimum throw in the error branch). Fix this →

---

3. Minor — ToolOutput error container class is a dead ternary

tool.tsx ~line 277

className={cn("rounded-md", errorText ? "bg-muted/50" : "bg-muted/50")}

Both branches are identical — the ternary is dead. Either collapse to a constant or apply a different class for error state (e.g. bg-destructive/10).

---

4. A2A adapter v6 compatibility — assessment

The core claim holds. A2aChatLanguageModel implements LanguageModelV2 from @ai-sdk/provider, which is unchanged in the v6 public package bump. The removal of activeTextIds / text-end in the flush handler is correct — text-end is no longer a valid LanguageModelV2StreamPart member in v6. The specificationVersion = 'v2' stays valid because LanguageModel = V3 | V2.

One asymmetry worth noting: doGenerate sets X-Redpanda-Stream-Tokens: true but doStream does not. If this is intentional (e.g. the streaming path doesn't use token headers), a quick comment would help future readers.

The provider-level usage shape ({ inputTokens, outputTokens, totalTokens }) in doGenerate / doStream is from @ai-sdk/provider V2 types, which don't require the new inputTokenDetails/outputTokenDetails — so those two methods are fine without the v6 sub-objects.

---

5. useContextUsage shape — assessment

The dual-write strategy is correct: legacy fields (reasoningTokens, cachedInputTokens) and new v6 sub-objects are both populated. The Context component reads the legacy fields today so the display is stable.

One question worth verifying: totalTokens = cumulativeInputTokens + cumulativeOutputTokens. If cumulativeReasoningTokens is not already included in cumulativeOutputTokens (i.e. tracked separately by the streaming handler), the total would undercount when a reasoning model is used. If it is included, the current formula is correct. Worth a quick sanity-check against how latestUsage is assembled in use-message-streaming.ts.

---

6. Test coverage — what's critical-path vs follow-up

Cover in this PR (high confidence in the untested surface):

• doStream non-streaming fallback: the simulatedStream path + the error branch currently has a FIXME. Even a unit test that stubs client.getAgentCard + client.sendMessage and asserts the stream emits a finish part would catch the double-send regression.
• Static-tool CopyButton title: add an assertion that the title attribute on the copy button matches the tool name, not "undefined".

Acceptable as follow-up:

• doGenerate end-to-end (network-boundary, reasonable to defer)
• ConversationDownload click-to-download DOM behavior (URL.createObjectURL mocking is noise-heavy)
• ContextReasoningUsage / ContextCacheUsage zero-guard behavior (low-risk guard; simple visual regression)

---

Summary

Solid upgrade. The v6 story is correct, the ai-elements re-sync is careful, and the token-usage dual-write is a clean forward-compat pattern. The CopyButton title bug (#1) is the only production-visible defect; the double-send bug (#2) pre-existed but is worth addressing. Everything else is polish.

[malinskibeniamin]

Fix screenshot paths. They don't render properly

[malinskibeniamin]

@claude review

[claude]

Claude finished @malinskibeniamin's task in 2m 8s —— View job

---

PR Review

• Gather context and read key changed files
• Review A2A adapter v6 compatibility
• Review vendored ai-elements re-sync
• Review useContextUsage shape and dual-write strategy
• Review test coverage

---

A2A Adapter v6 Compatibility

The specificationVersion = 'v2' + LanguageModelV2 implementation is correct. v6's LanguageModel = V3 | V2 alias accepts this without an adapter rewrite. The a2a-stream-mapper.ts extraction is clean — pure reducer with no side effects, well-typed state. The removal of text-end / activeTextIds bookkeeping from the flush handler is correct for v6 (those part types don't exist in LanguageModelV2StreamPart).

Bug — non-streaming fallback sends two requests (a2a-chat-language-model.tsx line 312)

This is the same pre-existing bug called out in the previous review, still unfixed:

// line 293-309
if (!clientCard.capabilities.streaming) {
  const nonStreamingResponse = await client.sendMessage(streamParams); // request 1
  if ("result" in nonStreamingResponse) {
    simulatedStream = new ReadableStream(...)
  }
  if ("error" in nonStreamingResponse) {
    // FIXME: error  ← silent swallow, falls through
  }
}

// line 312 — fires unconditionally
const response = client.sendMessageStream(streamParams); // request 2, always

Two problems remain:

• sendMessageStream fires even when simulatedStream is set — message sent twice.
• Error branch is swallowed; simulatedStream stays null and falls through to the streaming path, sending a second request after the first already failed.

The fix is two lines: move sendMessageStream into an else branch, and throw in the error branch instead of silently ignoring. Fix this →

Minor — X-Redpanda-Stream-Tokens header asymmetry

doGenerate sets X-Redpanda-Stream-Tokens: true (line 201); doStream does not. If this is intentional (e.g. the streaming endpoint ignores it), a brief comment would prevent future readers from "fixing" it.

---

Vendored ai-elements Re-sync

CopyButton title bug from prior review: confirmed fixed. Line 193 correctly uses displayName for both the toolCallId present/absent branches — no "Copy: undefined" regression.

Dead ternary in ToolOutput (tool.tsx line 277)

className={cn("rounded-md", errorText ? "bg-muted/50" : "bg-muted/50")}

Both branches are identical — the ternary is dead. Likely intended bg-destructive/10 (or similar) for the error state. Fix this →

ConversationDownload positioning assumption (conversation.tsx line 153)

The button renders with absolute top-4 right-4. This implicitly requires its parent to be position: relative. There's nothing enforcing this on the consumer side — a caller who forgets could see the button escape to a distant ancestor. Worth either noting in JSDoc or switching to a relative-positioned wrapper. Minor.

messagesToMarkdown drops non-text parts silently (conversation.tsx lines 102–106)

getMessageText filters to type === "text" parts only. Tool inputs, images, and data parts are silently dropped from the export. For tool-heavy conversations (the primary use case here) the markdown export would be incomplete. This should at minimum be noted as a known limitation in the JSDoc, or tool parts could be serialized as fenced JSON blocks. Fix this →

Shimmer: Module-level motionComponentCache and motion.create caching is correct — avoids creating motion instances during render. memo wrapper is appropriate.

New tool states (approval-requested, approval-responded, output-denied): All three have Redpanda badge variants (warning-inverted, info-inverted, destructive-inverted respectively). Badge and icon choices are sensible. ✓

---

useContextUsage Shape

Dual-write strategy is correct. Legacy top-level fields (reasoningTokens, cachedInputTokens) are populated for ContextReasoningUsage/ContextCacheUsage which still read them. v6 sub-objects (inputTokenDetails, outputTokenDetails) are also populated for forward compat. The useMemo dependency array is complete and the stable-reference test confirms it.

totalTokens formula (use-context-usage.ts line 42):

totalTokens: usage.cumulativeInputTokens + usage.cumulativeOutputTokens,

In the AI SDK v6, reasoning tokens are a sub-category of output tokens — outputTokenDetails.reasoningTokens is included in outputTokens, so this formula is correct assuming cumulativeOutputTokens already includes reasoning. The existing test asserts totalTokens === 150 for input=100, output=50 which confirms the expected interpretation. If cumulativeReasoningTokens were ever tracked as additional (not included in output), this would undercount — worth a comment to lock in the invariant. Low risk but worth the comment.

---

Test Coverage

Strong coverage in this PR:

• a2a-stream-mapper.test.ts — all TaskState branches, isFirstChunk guard, includeRawChunks, finalizeStream. Well structured.
• a2a-chat-language-model.test.ts — mapFinishReason and getResponseMetadata exhaustively covered.
• use-context-usage.test.tsx — legacy fields, v6 sub-objects, memoisation stability, and recompute-on-change all covered.

Gaps worth addressing in this PR:

1. doStream non-streaming fallback path — the simulatedStream path and the FIXME error branch have zero test coverage. At minimum, a test that stubs getAgentCard to return { capabilities: { streaming: false } } and sendMessage to return a result would catch the double-send regression.

2. messagesToMarkdown with non-text parts — a test asserting what happens with a message containing only tool parts would document the current behavior and prevent silent regressions.

Acceptable as follow-up: doGenerate end-to-end, ConversationDownload DOM behavior, ContextReasoningUsage / ContextCacheUsage zero-guard visual regression.

---

Summary

The v6 compatibility story is solid, the ai-elements re-sync is surgical and the useContextUsage dual-write is a clean forward-compat pattern. The CopyButton title fix from the prior review is confirmed applied. The main concerns remaining are the pre-existing double-send bug in the non-streaming fallback (still unfixed, worth a targeted fix before merge) and the silent tool-part drop in messagesToMarkdown. The dead ternary is cosmetic.

[malinskibeniamin]

@claude review again

[claude]

Claude finished @malinskibeniamin's task in 3m 9s —— View job

---

PR Review

• Read key changed files
• Address author's inline review comments
• Re-check previously flagged bugs (CopyButton title, dead ternary, double-send)
• Verify new items: .gitignore, copyright years, NaN check, image.tsx change, error handling
• Post final review

---

Response to Inline Review Comments

.gitignore line 27 — "Not needed. Remove this"
Agreed. The entry frontend/docs/pr-screenshots/ was added in this PR with a comment explaining screenshots are hosted on an orphan branch. But since they're never committed to the working tree, the gitignore is redundant and should go. Fix this →

---

Copyright years — "Use 2026 copyright everywhere"
All new files added in this PR carry Copyright 2025. The existing repo convention (see ai-agent-list-page.browser.test.tsx which uses Copyright 2026) is to use the current year. Every new file in this PR needs a year bump. Affected:

• ai-elements/context.browser.test.tsx, context.test.tsx
• ai-elements/conversation.browser.test.tsx, conversation.test.tsx
• ai-elements/shimmer.browser.test.tsx, shimmer.test.tsx
• ai-elements/tool.browser.test.tsx, tool.test.tsx
• ai-elements/response.test.tsx
• a2a/a2a-stream-mapper.ts, a2a-stream-mapper.test.ts
• a2a/a2a-chat-language-model.test.ts
• a2a/chat/utils/parse-a2a-error.ts, parse-a2a-error.test.ts
• a2a/chat/hooks/use-context-usage.test.tsx

Fix this →

---

context.tsx line 156 — "Number.isNaN() or similar for completeness"
Current guard: maxTokens > 0 ? usedTokens / maxTokens : 0. This handles zero but not NaN — if either value is NaN (e.g. the usage hook returns undefined after a partial stream), the ratio would still be NaN and slip past the guard. Safer:

const usedPercent = Number.isFinite(maxTokens) && maxTokens > 0
  ? usedTokens / maxTokens
  : 0;

Fix this →

---

conversation.browser.test.tsx line 77 — "Use ADP UI equivalent naming convention"
The new tests define a local const shot = (name: string) => page.getByTestId(...).screenshot(...). The established pattern in this repo is toMatchScreenshot (see remote-mcp-inspector-tab.browser.test.tsx line 166). These serve different purposes — toMatchScreenshot is visual regression; the shot helper here intentionally writes to a path for PR documentation screenshots. The author is right that the local helper name shot is non-standard. Options: rename to captureScreenshot for clarity, or extract a shared utility in browser-test-utils.tsx. Given the helper is consistent across all four new browser test files, a shared utility would be the right fix.

---

image.tsx line 11 — "Why no longer using uint8 array?"
The change is correct and safe. The before (spreading all props) would have passed uint8Array as an attribute to the DOM <img> element, causing a React "Unknown prop" warning. The after destructures it as uint8Array: _uint8Array (underscore-prefixed to silence lint) to remove it from ...props. The image still renders via src={data:${mediaType};base64,${base64}} — uint8Array was never used for rendering. This is an upstream ai-elements fix picked up correctly. No functional regression.

---

a2a-chat-language-model.test.ts line 26 — "Backend compatibility / defensive checks"
The @a2a-js/sdk bump is 0.3.10 → 0.3.13 (patch). The test file covers mapFinishReason and getResponseMetadata which depend on the A2A wire event shapes (kind, status.state, taskId, etc.). The patch bump should not break these. One gap: the test file doesn't cover the contextId injection at line 281 (message.contextId = options.providerOptions?.a2a?.contextId). If a caller passes a non-string contextId (or the field name changes in 0.3.13), this cast would silently corrupt the message. A defensive typeof ... === 'string' check and a test asserting contextId is only set when present would harden this.

---

use-context-usage.test.tsx line 49 — "tokenlens package improvements"
The current hook manually maps raw cumulative counts to LanguageModelUsage fields. If the project uses or plans to use tokenlens (a token-budget visualization utility), the integration point would be at the useContextUsage return shape — not here. The current dual-write strategy is already the right approach; tokenlens would consume the output of this hook, not replace it. No immediate change needed here, but wiring tokenlens in a follow-up would be cleaner if there's an existing package in the workspace.

---

parse-a2a-error.test.ts line 109 — "MCP error code mapping"
This is a valuable follow-up item. parseA2AError currently handles A2A/JSON-RPC error codes but not MCP protocol errors. MCP uses its own error codes (e.g. -32601 Method Not Found, -32602 Invalid Params, JSON-RPC 2.0 standard codes). A parseMCPError function that maps these to user-readable messages (similar to how ConnectRPC codes are mapped) would complete the error surface. The current parseA2AError would need to be extended or a parallel parseMCPError added. This is appropriately a follow-up given MCP SDK changes are deferred to a separate PR.

---

Still-Unfixed Issues from Previous Reviews

1. Double-send bug — non-streaming fallback (a2a-chat-language-model.tsx line 312)

This was flagged in both prior reviews and remains unaddressed:

// line 312 — fires unconditionally even when simulatedStream is set
const response = client.sendMessageStream(streamParams);

When clientCard.capabilities.streaming === false, sendMessage is called (request 1) to build simulatedStream, then sendMessageStream fires anyway (request 2), sending the message twice. The "error" branch at line 306 is also still a silent swallow — simulatedStream stays null and falls through to the streaming path after the first request already errored.

Fix: move const response = client.sendMessageStream(streamParams) into an else branch; throw in the error case. Fix this →

2. Dead ternary in ToolOutput (tool.tsx line 277)

className={cn("rounded-md", errorText ? "bg-muted/50" : "bg-muted/50")}

Both branches are identical — error and success output containers look the same. Likely intended bg-destructive/10 for the error branch. Fix this →

---

New Observation

messagesToMarkdown silently drops tool parts (conversation.tsx line 102)

getMessageText filters to type === "text" only. In a tool-heavy agents conversation the exported markdown would contain only the prose, stripping every tool invocation and result. Given that this is the primary use-case (agents conversations), the export would be misleading. Minimum fix: add a JSDoc noting the limitation. Better: serialize tool parts as fenced JSON blocks.

---

Summary

CopyButton title fix is confirmed applied. The v6 adapter and useContextUsage dual-write remain sound. The main items to address before merge: copyright year sweep (all new files), .gitignore entry removal, Number.isFinite guard in context, and the still-present double-send + dead ternary bugs from prior reviews. The shot naming convention, messagesToMarkdown tool-part gap, and MCP error mapping are reasonable follow-ups.

[malinskibeniamin]

@claude review

[claude]

Claude finished @malinskibeniamin's task in 2m 39s —— View job

---

PR Review

• Read key changed files and new commits
• Re-check previously flagged bugs (double-send, dead ternary)
• Review new commits since last review (safeUsedPercent, isKnownA2AEvent, MCP error codes, captureScreenshotFrame)
• Assess remaining concerns and overall merge readiness

---

What's been resolved since the last review

All inline review comments are addressed:

• ✅ .gitignore entry removed (09624c5)
• ✅ Copyright year swept to 2026 across all 18 added files (3dfe7e0)
• ✅ safeUsedPercent helper guards against NaN, ±Infinity, and negatives in all three context.tsx divide sites (fc0e4e9)
• ✅ captureScreenshotFrame shared utility added to browser-test-utils.tsx, all four browser test files now use it (8b3498c)
• ✅ isKnownA2AEvent type guard + UnsupportedFunctionalityError thrown for unrecognised stream event kinds (3a84337) — good: silent-drop manifests as a stalled chat, this surfaces a diagnostic error instead
• ✅ parseA2AError extended with human-readable title + actionable hint for every JSON-RPC 2.0, A2A-extension, and MCP error code (fcae41e) — the code table is complete and the A2AErrorBlock wiring is clean

---

Still-unfixed — both flagged in all three prior reviews

1. Double-send in non-streaming fallback (a2a-chat-language-model.tsx lines 291–316)

if (!clientCard.capabilities.streaming) {
  const nonStreamingResponse = await client.sendMessage(streamParams); // request 1
  if ("result" in nonStreamingResponse) { simulatedStream = new ReadableStream(...); }
  if ("error" in nonStreamingResponse) {
    // FIXME: error  ← still silently swallowed
  }
}

// line 312 — fires unconditionally regardless of simulatedStream
const response = client.sendMessageStream(streamParams); // request 2, always

Both problems remain:

• A non-streaming-capable agent receives the message twice (once via sendMessage, once via sendMessageStream).
• A sendMessage error is swallowed and execution falls through to the streaming path, sending a second request after the first already errored.

Two-line fix: move sendMessageStream into an else branch; throw in the error case. Fix this →

2. Dead ternary in ToolOutput (tool.tsx line 277)

<div className={cn("rounded-md", errorText ? "bg-muted/50" : "bg-muted/50")}>

Both branches are identical — error and success containers look the same. Intended bg-destructive/10 (or similar) for the error branch. Fix this →

---

New minor observation

JSON_RPC_DATA_REGEX doesn't handle nested objects (parse-a2a-error.ts line 32)

export const JSON_RPC_DATA_REGEX = /Data:\s*(\{[^}]*\})/i;

[^}]* stops at the first }, so nested data like {"field":{"reason":"expired"}} would be truncated to {"field": and JSON.parse would throw — the catch silently returns undefined. The test suite only covers flat payloads. Non-blocking since data is informational, but worth noting.

---

Architecture note (resolved concern)

a2aEventToV2StreamParts emitting only metadata parts (no text-delta) is correct by design. use-message-streaming.ts passes includeRawChunks: true and routes all content through the raw chunk path — text and artifacts come from status-update.message.parts / artifact-update, not AI SDK text-delta parts. The comment on line 330 of use-message-streaming.ts documents this invariant.

---

Summary

Strong improvement round. Every inline review comment is addressed and the new additions (isKnownA2AEvent, parseA2AError enrichment, safeUsedPercent, captureScreenshotFrame) are well-executed. The only remaining blockers are the two issues above — the double-send bug has been flagged in every review and is the only production-visible correctness issue; the dead ternary is cosmetic. Both are small, targeted fixes.