Repository navigation
feat(chat-runtime): Claude harness adapter with recorded wire fixtures - #6193
Merged
Merged
Conversation
Contributor
|
Caution Review failedThe pull request is closed. ℹ️ Recent review info⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Pro Plus Run ID: ⛔ Files ignored due to path filters (1)
📒 Files selected for processing (20)
📝 WalkthroughWalkthroughAdds Claude Agent SDK integration for chat runtime sessions. The change introduces tool-use mapping, stream translation, prompt and approval handling, cancellation and disposal, fixture recording, JSONL fixtures, and adapter tests. ChangesClaude harness
Estimated code review effort: 4 (Complex) | ~45 minutes Sequence Diagram(s)sequenceDiagram
participant ClaudeAdapter
participant ClaudeAgentSDK
participant ClaudeTranslator
participant HarnessEventStream
ClaudeAdapter->>ClaudeAgentSDK: Start query with prompt stream and approval callback
ClaudeAgentSDK-->>ClaudeAdapter: Emit Claude session messages
ClaudeAdapter->>ClaudeTranslator: Translate each message
ClaudeTranslator-->>HarnessEventStream: Emit translated adapter events
HarnessEventStream-->>ClaudeAdapter: Deliver events to consumers
ClaudeAdapter->>ClaudeAgentSDK: Interrupt or abort the session
Possibly related PRs
✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
Contributor
🚀 Preview Deployment🔗 Preview Links
Preview updates automatically with new commits |
Contributor
Greptile SummaryThe PR introduces a Claude Agent SDK harness adapter, translating Claude wire messages and tool activity into chat-runtime protocol events using recorded fixtures.
Confidence Score: 4/5The PR should not merge until cancellation can preserve subsequent turns and attachment prompts reach Claude with usable content. The adapter permanently aborts its only query stream when one turn is canceled even though LiveSession can immediately reuse it for queued work, and its prompt conversion silently strips all usable attachment information. Files Needing Attention: packages/chat-runtime/src/harness/claude/claudeAdapter/claudeAdapter.ts
|
| Filename | Overview |
|---|---|
| packages/chat-runtime/src/harness/claude/claudeAdapter/claudeAdapter.ts | Adds the SDK lifecycle and approval bridge, but cancellation permanently disables later turns and prompt conversion drops attachment data. |
| packages/chat-runtime/src/harness/claude/translateStream/translateStream.ts | Adds stateful conversion of Claude stream envelopes into session, turn, item, delta, and tool-call events with broad recorded-fixture coverage. |
| packages/chat-runtime/src/harness/claude/mapToolUse/mapToolUse.ts | Adds tool categorization, titles, locations, terminal output, and edit/write diff rendering. |
| packages/chat-runtime/scripts/recordClaudeFixtures.ts | Adds an isolated recorder for representative Claude wire scenarios, including approval, denial, abort, and subagent activity. |
| packages/chat-runtime/package.json | Adds an exactly pinned Claude Agent SDK runtime dependency. |
Sequence Diagram
sequenceDiagram
participant Client
participant Session as LiveSession
participant Adapter as ClaudeAdapter
participant SDK as Claude Agent SDK
participant Translator as ClaudeTranslator
Client->>Session: prompt(UserContent)
Session->>Adapter: prompt(content)
Adapter->>SDK: streamed user message
SDK-->>Adapter: SDKMessage stream
Adapter->>Translator: translate(message)
Translator-->>Session: AdapterEvents
SDK->>Adapter: canUseTool(...)
Adapter-->>Session: approval_request
Client->>Session: approval decision
Session->>Adapter: respondToApproval(...)
Adapter-->>SDK: allow or deny
Client->>Session: cancelTurn()
Session->>Adapter: cancelTurn()
Adapter->>SDK: abort()
Prompt To Fix All With AI
### Issue 1
packages/chat-runtime/src/harness/claude/claudeAdapter/claudeAdapter.ts:186-188
**Cancellation permanently stops the adapter**
When a prompt is queued and the current turn is canceled, `cancelTurn()` permanently aborts the adapter's only SDK stream. `LiveSession` then delivers the queued prompt to the same adapter, but its pump has already ended, so the new turn never starts and the session stalls.
### Issue 2
packages/chat-runtime/src/harness/claude/claudeAdapter/claudeAdapter.ts:127-131
**Attachment content is discarded**
When a prompt contains an attachment, `promptText()` reduces it to a name such as `[report.pdf]` and drops its ID and contents. Claude therefore cannot inspect the attachment or answer questions about it despite `HarnessAdapter.prompt()` accepting attachment content.
---
For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.Reviews (1): Last reviewed commit: "feat(chat-runtime): Claude adapter with ..." | Re-trigger Greptile
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
Comment on lines
+186
to
+188
| cancelTurn(): void { | ||
| this.abortController.abort(); | ||
| } |
Contributor
There was a problem hiding this comment.
When a prompt is queued and the current turn is canceled, cancelTurn() permanently aborts the adapter's only SDK stream. LiveSession then delivers the queued prompt to the same adapter, but its pump has already ended, so the new turn never starts and the session stalls.
Prompt To Fix With AI
This is a comment left during a code review.
Path: packages/chat-runtime/src/harness/claude/claudeAdapter/claudeAdapter.ts
Line: 186-188
Comment:
**Cancellation permanently stops the adapter**
When a prompt is queued and the current turn is canceled, `cancelTurn()` permanently aborts the adapter's only SDK stream. `LiveSession` then delivers the queued prompt to the same adapter, but its pump has already ended, so the new turn never starts and the session stalls.
---
For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.
Comment on lines
+127
to
+131
|
|
||
| function promptText(content: UserContent[]): string { | ||
| return content | ||
| .map((part) => (part.type === "text" ? part.text : `[${part.name}]`)) | ||
| .join("\n"); |
Contributor
There was a problem hiding this comment.
When a prompt contains an attachment, promptText() reduces it to a name such as [report.pdf] and drops its ID and contents. Claude therefore cannot inspect the attachment or answer questions about it despite HarnessAdapter.prompt() accepting attachment content.
Prompt To Fix With AI
This is a comment left during a code review.
Path: packages/chat-runtime/src/harness/claude/claudeAdapter/claudeAdapter.ts
Line: 127-131
Comment:
**Attachment content is discarded**
When a prompt contains an attachment, `promptText()` reduces it to a name such as `[report.pdf]` and drops its ID and contents. Claude therefore cannot inspect the attachment or answer questions about it despite `HarnessAdapter.prompt()` accepting attachment content.
---
For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.![cubic-dev-ai[bot]](https://avatars.githubusercontent.com/in/1082092?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
8 issues found across 21 files
Prompt for AI agents (unresolved issues)
Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.
<file name="packages/chat-runtime/src/harness/claude/mapToolUse/mapToolUse.ts">
<violation number="1" location="packages/chat-runtime/src/harness/claude/mapToolUse/mapToolUse.ts:44">
P3: Titles lose directory context for workspace paths whose first segment starts with `..`, because this containment check rejects `..cache/...` as if it were parent traversal. Check for `..` or a `..` path segment using the platform separator instead of testing the prefix alone.</violation>
<violation number="2" location="packages/chat-runtime/src/harness/claude/mapToolUse/mapToolUse.ts:94">
P2: NotebookRead, BashOutput, and KillShell produce raw SDK names as their UI titles instead of human-readable summaries because they fall through this default. Add explicit title cases for every recognized tool so the tool row does not expose internal names.</violation>
<violation number="3" location="packages/chat-runtime/src/harness/claude/mapToolUse/mapToolUse.ts:163">
P2: A failed Bash call can render with an empty terminal output even when Claude returned an error message; when the raw result object has no stdout/stderr, this branch discards `modelFacingText`. Fall back to the model-facing text when the stream output is empty so the settled tool row preserves the failure reason.</violation>
</file>
<file name="packages/chat-runtime/scripts/recordClaudeFixtures.ts">
<violation number="1" location="packages/chat-runtime/scripts/recordClaudeFixtures.ts:106">
P3: Each recording run leaks one temporary workspace per scenario; clean the workspace in a `finally` block after the query has stopped, including failure paths.</violation>
<violation number="2" location="packages/chat-runtime/scripts/recordClaudeFixtures.ts:132">
P1: The abort scenario hides unrelated SDK failures and can leave a corrupt fixture while reporting success; only accept the exception when this recorder's `AbortController` is actually aborted.</violation>
<violation number="3" location="packages/chat-runtime/scripts/recordClaudeFixtures.ts:140">
P2: The abort fixture can be recorded without an abort when Claude finishes before the first Bash tool use; validate that the controller was aborted before writing this scenario's fixture.</violation>
<violation number="4" location="packages/chat-runtime/scripts/recordClaudeFixtures.ts:157">
P3: A misspelled scenario name silently does nothing and leaves existing fixtures unchanged; reject unknown names before starting the recording loop.</violation>
</file>
<file name="packages/chat-runtime/src/harness/claude/claudeAdapter/claudeAdapter.ts">
<violation number="1" location="packages/chat-runtime/src/harness/claude/claudeAdapter/claudeAdapter.ts:130">
P2: Non-text `UserContent` entries are converted to `[name]` here, which silently strips attachment payloads before sending the prompt. That prevents Claude from using attachment contents; either pass attachment data through or fail fast for unsupported attachment types instead of degrading silently.</violation>
</file>
Architecture diagram
sequenceDiagram
participant Host as Host Application
participant Adapter as ClaudeAdapter
participant Queue as PromptQueue
participant SDK as @anthropic-ai/claude-agent-sdk
participant Translator as ClaudeTranslator
participant Mapper as mapToolUse
participant EventQ as EventQueue
participant Test as Test Fixture (JSONL)
Note over Host,EventQ: NEW: Claude Harness Adapter Flow
Host->>Adapter: start({ cwd, modelId })
Adapter->>Translator: new ClaudeTranslator({ cwd, now, mintId })
Adapter->>SDK: query({ prompt: Queue.stream(), options })
Note over Adapter,SDK: NEW: Injected query function—no process spawn
SDK-->>Adapter: AsyncIterable<SDKMessage>
par Host sends prompt
Host->>Adapter: prompt([UserContent])
Adapter->>Queue: push(user message)
Queue-->>SDK: yield SDK user message
and SDK streams messages
loop for each SDK message
SDK-->>Adapter: SDKMessage (system, assistant, stream_event, result)
Adapter->>Translator: translate(message)
Translator->>Translator: dispatch by type/subtype
alt system init
Translator-->>Adapter: { kind: session, session: { harness, modelId } }
else assistant content blocks
Translator->>Mapper: toolKindFor(name) / titleFor(use, cwd)
Mapper-->>Translator: ToolKind / title
alt text block
Translator-->>Adapter: { kind: item, item: agent_message }
else thinking block
Translator-->>Adapter: { kind: item, item: reasoning }
else tool_use block
Translator->>Translator: track tool (pending)
Translator-->>Adapter: { kind: item, item: tool_call (running) }
end
else stream_event content_block_delta
alt text/thinking delta
Translator-->>Adapter: { kind: delta, delta: { type: text, append } }
else input_json_delta
Translator-->>Adapter: { kind: delta, delta: { type: tool_input, append } }
end
else result
alt is_error
Translator-->>Adapter: { kind: turn, turn: { status: running } } (initial)
Note over Translator,EventQ: NEW: abort as thrown iterator error
Adapter->>Translator: interrupt(String(error))
Translator->>Mapper: contentFor(tool, null)
Translator-->>Adapter: { kind: item, item: tool_call (canceled) } for each pending
Translator-->>Adapter: { kind: turn, turn: { status: interrupted } }
else success
Translator->>Translator: build usage + cost
Translator-->>Adapter: { kind: turn, turn: { status: completed } }
end
Note over Translator,EventQ: NEW: updatedInput required at runtime on accept
end
Adapter->>EventQ: push(AdapterEvent)
end
and SDK asks for permission
SDK->>Adapter: canUseTool(toolName, input, { toolUseID })
Note over Adapter,SDK: NEW: allowedTools bypasses canUseTool (wire finding)
Adapter->>Adapter: requestApproval() – create pending approval
Adapter->>EventQ: push({ kind: item, item: approval_request (pending) })
Host->>Adapter: respondToApproval(id, decision)
alt accept / accept_for_session
Adapter->>Adapter: settle({ behavior: allow, updatedInput: input })
Adapter-->>SDK: PermissionResult (allow)
else decline
Adapter->>Translator: markDeclined(toolUseId)
Adapter->>Adapter: settle({ behavior: deny, ... })
Adapter-->>SDK: PermissionResult (deny)
else cancel
Adapter->>Adapter: cancelTurn() -> abortController.abort()
SDK-->>Adapter: throws – caught in run()
end
Adapter->>EventQ: push({ kind: item, item: approval_request (answered) })
end
Note over EventQ,Test: Test Mode (injected fixture replay)
Test->>Translator: translate(fixture lines)
Translator->>Mapper: contentFor, titleFor, etc.
Translator-->>Test: AdapterEvent[]
Note over Translator,Mapper: NEW: Agent/Task tool naming (wire finding)
Note over Translator,EventQ: NEW: 11 wire-truth findings logged for doc updates
Reply with feedback, questions, or to request a fix.
Re-trigger cubic
| } | ||
| } | ||
| } catch (error) { | ||
| if (!scenario.abortAfterToolUse) throw error; |
Contributor
There was a problem hiding this comment.
P1: The abort scenario hides unrelated SDK failures and can leave a corrupt fixture while reporting success; only accept the exception when this recorder's AbortController is actually aborted.
Prompt for AI agents
Check if this issue is valid — if so, understand the root cause and fix it. At packages/chat-runtime/scripts/recordClaudeFixtures.ts, line 132:
<comment>The abort scenario hides unrelated SDK failures and can leave a corrupt fixture while reporting success; only accept the exception when this recorder's `AbortController` is actually aborted.</comment>
<file context>
@@ -0,0 +1,166 @@
+ }
+ }
+ } catch (error) {
+ if (!scenario.abortAfterToolUse) throw error;
+ messages.push({
+ type: "system",
</file context>
| text(use.input.subject) ?? | ||
| "Subagent task" | ||
| ); | ||
| default: |
Contributor
There was a problem hiding this comment.
P2: NotebookRead, BashOutput, and KillShell produce raw SDK names as their UI titles instead of human-readable summaries because they fall through this default. Add explicit title cases for every recognized tool so the tool row does not expose internal names.
Prompt for AI agents
Check if this issue is valid — if so, understand the root cause and fix it. At packages/chat-runtime/src/harness/claude/mapToolUse/mapToolUse.ts, line 94:
<comment>NotebookRead, BashOutput, and KillShell produce raw SDK names as their UI titles instead of human-readable summaries because they fall through this default. Add explicit title cases for every recognized tool so the tool row does not expose internal names.</comment>
<file context>
@@ -0,0 +1,194 @@
+ text(use.input.subject) ??
+ "Subagent task"
+ );
+ default:
+ return use.name;
+ }
</file context>
|
|
||
| const stdout = text(output?.stdout) ?? ""; | ||
| const stderr = text(output?.stderr) ?? ""; | ||
| const combined = output |
Contributor
There was a problem hiding this comment.
P2: A failed Bash call can render with an empty terminal output even when Claude returned an error message; when the raw result object has no stdout/stderr, this branch discards modelFacingText. Fall back to the model-facing text when the stream output is empty so the settled tool row preserves the failure reason.
Prompt for AI agents
Check if this issue is valid — if so, understand the root cause and fix it. At packages/chat-runtime/src/harness/claude/mapToolUse/mapToolUse.ts, line 163:
<comment>A failed Bash call can render with an empty terminal output even when Claude returned an error message; when the raw result object has no stdout/stderr, this branch discards `modelFacingText`. Fall back to the model-facing text when the stream output is empty so the settled tool row preserves the failure reason.</comment>
<file context>
@@ -0,0 +1,194 @@
+
+ const stdout = text(output?.stdout) ?? "";
+ const stderr = text(output?.stderr) ?? "";
+ const combined = output
+ ? [stdout, stderr].filter((part) => part.length > 0).join("\n")
+ : (outcome?.modelFacingText ?? "");
</file context>
| if (!raw) return null; | ||
| if (!isAbsolute(raw)) return raw; | ||
| const inside = relative(cwd, raw); | ||
| return inside && !inside.startsWith("..") ? inside : basename(raw); |
Contributor
There was a problem hiding this comment.
P3: Titles lose directory context for workspace paths whose first segment starts with .., because this containment check rejects ..cache/... as if it were parent traversal. Check for .. or a .. path segment using the platform separator instead of testing the prefix alone.
Prompt for AI agents
Check if this issue is valid — if so, understand the root cause and fix it. At packages/chat-runtime/src/harness/claude/mapToolUse/mapToolUse.ts, line 44:
<comment>Titles lose directory context for workspace paths whose first segment starts with `..`, because this containment check rejects `..cache/...` as if it were parent traversal. Check for `..` or a `..` path segment using the platform separator instead of testing the prefix alone.</comment>
<file context>
@@ -0,0 +1,194 @@
+ if (!raw) return null;
+ if (!isAbsolute(raw)) return raw;
+ const inside = relative(cwd, raw);
+ return inside && !inside.startsWith("..") ? inside : basename(raw);
+}
+
</file context>
Comment on lines
+157
to
+159
| const selected = only.length | ||
| ? SCENARIOS.filter((scenario) => only.includes(scenario.name)) | ||
| : SCENARIOS; |
Contributor
There was a problem hiding this comment.
P3: A misspelled scenario name silently does nothing and leaves existing fixtures unchanged; reject unknown names before starting the recording loop.
Prompt for AI agents
Check if this issue is valid — if so, understand the root cause and fix it. At packages/chat-runtime/scripts/recordClaudeFixtures.ts, line 157:
<comment>A misspelled scenario name silently does nothing and leaves existing fixtures unchanged; reject unknown names before starting the recording loop.</comment>
<file context>
@@ -0,0 +1,166 @@
+async function main(): Promise<void> {
+ mkdirSync(FIXTURES_DIR, { recursive: true });
+ const only = process.argv.slice(2);
+ const selected = only.length
+ ? SCENARIOS.filter((scenario) => only.includes(scenario.name))
+ : SCENARIOS;
</file context>
Suggested change
| const selected = only.length | |
| ? SCENARIOS.filter((scenario) => only.includes(scenario.name)) | |
| : SCENARIOS; | |
| const unknown = only.filter( | |
| (name) => !SCENARIOS.some((scenario) => scenario.name === name), | |
| ); | |
| if (unknown.length > 0) { | |
| throw new Error(`Unknown scenario(s): ${unknown.join(", ")}`); | |
| } | |
| const selected = only.length | |
| ? SCENARIOS.filter((scenario) => only.includes(scenario.name)) | |
| : SCENARIOS; |
| } | ||
|
|
||
| async function record(scenario: Scenario): Promise<void> { | ||
| const cwd = mkdtempSync(join(tmpdir(), `claude-fixture-${scenario.name}-`)); |
Contributor
There was a problem hiding this comment.
P3: Each recording run leaks one temporary workspace per scenario; clean the workspace in a finally block after the query has stopped, including failure paths.
Prompt for AI agents
Check if this issue is valid — if so, understand the root cause and fix it. At packages/chat-runtime/scripts/recordClaudeFixtures.ts, line 106:
<comment>Each recording run leaks one temporary workspace per scenario; clean the workspace in a `finally` block after the query has stopped, including failure paths.</comment>
<file context>
@@ -0,0 +1,166 @@
+}
+
+async function record(scenario: Scenario): Promise<void> {
+ const cwd = mkdtempSync(join(tmpdir(), `claude-fixture-${scenario.name}-`));
+ scenario.seed?.(cwd);
+
</file context>
This was referenced Aug 5, 2026
Merged
Contributor
There was a problem hiding this comment.
3 issues found across 15 files (changes from recent commits).
Not reviewed (too large): packages/chat-runtime/src/harness/claude/fixtures/subagent-task.jsonl (~331 lines), packages/chat-runtime/src/harness/claude/fixtures/edit-declined.jsonl (~287 lines), packages/chat-runtime/src/harness/claude/fixtures/edit-approved.jsonl (~258 lines) - if these are generated or fixture files, add them to ignored paths to exclude them from future reviews.
Prompt for AI agents (unresolved issues)
Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.
<file name="packages/chat-runtime/src/harness/claude/claudeAdapter/claudeAdapter.test.ts">
<violation number="1" location="packages/chat-runtime/src/harness/claude/claudeAdapter/claudeAdapter.test.ts:131">
P2: The dispose test doesn't verify that dispose() actually resolves. Because abort fires synchronously inside dispose(), the `aborted()` assertion holds even if dispose hangs at `await this.pump`; the Promise.race with a 50ms timeout silently passes either way. Assert on the race outcome (e.g. both settled in Promise.all and that disposal resolved) so a regression that leaves dispose pending is caught rather than missed.</violation>
</file>
<file name="packages/chat-runtime/src/harness/claude/mapToolUse/mapToolUse.ts">
<violation number="1" location="packages/chat-runtime/src/harness/claude/mapToolUse/mapToolUse.ts:169">
P3: subagentContent re-implements the same `rawOutput` extraction block (ducked through `typeof ... object` and cast to `Record<string, unknown>`/`null`) that `editContent` and `terminalContent` already contain. That's now three copies of the identical idiom in one file, so a future change to how outcomes are parsed (or how rawOutput is structured) has to be made in three places and can drift. A small shared helper like `function rawOutputOf(outcome): Record<string, unknown> | null` that centralizes this extraction would keep the three call sites consistent without changing behavior.</violation>
</file>
<file name="packages/chat-runtime/scripts/recordClaudeFixtures.ts">
<violation number="1" location="packages/chat-runtime/scripts/recordClaudeFixtures.ts:107">
P2: The committed fixtures still expose recording-machine path data and are not reproducible: `system.init.cwd` remains `/private/workspace`, while `memory_paths.auto` retains random `mkdtempSync` names such as `...-3urBTa`. `redact` only replaces the raw temp `cwd`/home strings and a macOS-specific path pattern, so it misses SDK-derived path fields (and other platform temp paths). Normalizing the parsed message path fields—including the SDK's derived `memory_paths` and the process working directory—before serializing the fixture would keep reruns portable and stable.</violation>
</file>
Reply with feedback, questions, or to request a fix.
Re-trigger cubic
| const disposal = harness.adapter.dispose(); | ||
| expect(harness.aborted()).toBe(true); | ||
| harness.emit(result()); | ||
| await Promise.race([ |
Contributor
There was a problem hiding this comment.
P2: The dispose test doesn't verify that dispose() actually resolves. Because abort fires synchronously inside dispose(), the aborted() assertion holds even if dispose hangs at await this.pump; the Promise.race with a 50ms timeout silently passes either way. Assert on the race outcome (e.g. both settled in Promise.all and that disposal resolved) so a regression that leaves dispose pending is caught rather than missed.
Prompt for AI agents
Check if this issue is valid — if so, understand the root cause and fix it. At packages/chat-runtime/src/harness/claude/claudeAdapter/claudeAdapter.test.ts, line 131:
<comment>The dispose test doesn't verify that dispose() actually resolves. Because abort fires synchronously inside dispose(), the `aborted()` assertion holds even if dispose hangs at `await this.pump`; the Promise.race with a 50ms timeout silently passes either way. Assert on the race outcome (e.g. both settled in Promise.all and that disposal resolved) so a regression that leaves dispose pending is caught rather than missed.</comment>
<file context>
@@ -0,0 +1,136 @@
+ const disposal = harness.adapter.dispose();
+ expect(harness.aborted()).toBe(true);
+ harness.emit(result());
+ await Promise.race([
+ disposal,
+ new Promise((resolve) => setTimeout(resolve, 50)),
</file context>
| } | ||
|
|
||
| // Fixtures are committed and public: never ship the recording machine's paths. | ||
| function redact(line: string, cwd: string): string { |
Contributor
There was a problem hiding this comment.
P2: The committed fixtures still expose recording-machine path data and are not reproducible: system.init.cwd remains /private/workspace, while memory_paths.auto retains random mkdtempSync names such as ...-3urBTa. redact only replaces the raw temp cwd/home strings and a macOS-specific path pattern, so it misses SDK-derived path fields (and other platform temp paths). Normalizing the parsed message path fields—including the SDK's derived memory_paths and the process working directory—before serializing the fixture would keep reruns portable and stable.
Prompt for AI agents
Check if this issue is valid — if so, understand the root cause and fix it. At packages/chat-runtime/scripts/recordClaudeFixtures.ts, line 107:
<comment>The committed fixtures still expose recording-machine path data and are not reproducible: `system.init.cwd` remains `/private/workspace`, while `memory_paths.auto` retains random `mkdtempSync` names such as `...-3urBTa`. `redact` only replaces the raw temp `cwd`/home strings and a macOS-specific path pattern, so it misses SDK-derived path fields (and other platform temp paths). Normalizing the parsed message path fields—including the SDK's derived `memory_paths` and the process working directory—before serializing the fixture would keep reruns portable and stable.</comment>
<file context>
@@ -102,6 +103,16 @@ function optionsFor(
}
+// Fixtures are committed and public: never ship the recording machine's paths.
+function redact(line: string, cwd: string): string {
+ return line
+ .split(JSON.stringify(cwd).slice(1, -1))
</file context>
| } | ||
|
|
||
| function subagentContent(outcome: ToolOutcome | null): ToolContent[] { | ||
| const output = |
Contributor
There was a problem hiding this comment.
P3: subagentContent re-implements the same rawOutput extraction block (ducked through typeof ... object and cast to Record<string, unknown>/null) that editContent and terminalContent already contain. That's now three copies of the identical idiom in one file, so a future change to how outcomes are parsed (or how rawOutput is structured) has to be made in three places and can drift. A small shared helper like function rawOutputOf(outcome): Record<string, unknown> | null that centralizes this extraction would keep the three call sites consistent without changing behavior.
Prompt for AI agents
Check if this issue is valid — if so, understand the root cause and fix it. At packages/chat-runtime/src/harness/claude/mapToolUse/mapToolUse.ts, line 169:
<comment>subagentContent re-implements the same `rawOutput` extraction block (ducked through `typeof ... object` and cast to `Record<string, unknown>`/`null`) that `editContent` and `terminalContent` already contain. That's now three copies of the identical idiom in one file, so a future change to how outcomes are parsed (or how rawOutput is structured) has to be made in three places and can drift. A small shared helper like `function rawOutputOf(outcome): Record<string, unknown> | null` that centralizes this extraction would keep the three call sites consistent without changing behavior.</comment>
<file context>
@@ -116,23 +153,27 @@ function editContent(use: ToolUse, outcome: ToolOutcome | null): ToolContent[] {
+}
+
+function subagentContent(outcome: ToolOutcome | null): ToolContent[] {
+ const output =
+ outcome && typeof outcome.rawOutput === "object" && outcome.rawOutput
+ ? (outcome.rawOutput as Record<string, unknown>)
</file context>
…d tool mapping Claude-Session: https://claude.ai/code/session_01XRFCEPsfmJHu2SenDKYJ2h
saddlepaddle
force-pushed
the
chat-claude-adapter
branch
from
August 5, 2026 18:41
768503f to
5256c33
Compare
This was referenced Aug 5, 2026
Merged
This branch was previously deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stacked on #6167. ClaudeTranslator (pure SDKMessage→AdapterEvent core) + ClaudeAdapter wiring @anthropic-ai/claude-agent-sdk 0.3.201 (pinned exact, injected query — tests never spawn a process). Six real recorded scenarios as fixtures (text, bash, edit approved/declined, abort mid-tool, subagent); 69 tests + a live smoke (passed 14.3s, kept skipped in CI). Includes 11 wire-truth findings vs the mapping doc — notably allowedTools silently bypassing canUseTool, updatedInput required-at-runtime on accept, abort as thrown iterator error, and the Agent/Task tool naming — doc amendments to chat-harness-adapters.md follow separately.
https://claude.ai/code/session_01XRFCEPsfmJHu2SenDKYJ2h
Summary by cubic
Adds a Claude harness adapter and translator for
chat-runtime, wired to@anthropic-ai/claude-agent-sdk@0.3.201and validated with recorded wire fixtures. Fixes cancel/interrupt behavior, delta targeting, and tool mapping to match the SDK stream.New Features
AdapterEvents viaClaudeTranslatorusing an injectedquery(no external process in tests/CI).mapToolUsewith kind, titles, locations, and concise outcome text.Bug Fixes
Write→edit,BashOutput→execute) to align with real SDK behavior.Written for commit 5256c33. Summary will update on new commits.
Summary by CodeRabbit
New Features
Tests