Include Claude debug artifacts in Export Agent Host Debug Logs#327374
Draft
TylerLeonhardt wants to merge 7 commits into
Draft
Include Claude debug artifacts in Export Agent Host Debug Logs#327374TylerLeonhardt wants to merge 7 commits into
TylerLeonhardt wants to merge 7 commits into
Conversation
The export command previously surfaced only Copilot's logs. Advertise the Claude SDK `--debug` log and the session transcript JSONL through the AHP session `_meta` bag (well-known `debugArtifacts` key, mirroring `agentHost.git`), so the client reads concrete host-side paths instead of hard-coding a provider's on-disk layout — no protocol change. - Shared log-naming helpers (`agentHostLogNaming`) reused by the AHP JSONL logger and Claude debug-file naming (sanitize + file-safe timestamp). - `ClaudeAgentSession` points the SDK at a per-run debug file and publishes the discovered artifacts (log + transcript) into `_meta`. Discovery/publishing is extracted into a testable `ClaudeDebugArtifacts` class backed by an observable that is a pure projection of disk truth. - Export resolves the `AgentHostSessionAdapter` that carries `_meta` and streams each artifact via the resource proxy (`file://` local / `vscode-agent-host://` remote), with provider-agnostic remote authority resolution. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Contributor
There was a problem hiding this comment.
Pull request overview
Adds Claude SDK debug logs and session transcripts to Agent Host debug exports via session metadata.
Changes:
- Adds Claude artifact discovery, naming, and metadata publication.
- Exports advertised artifacts through local or remote resource proxies.
- Adds unit coverage for naming, discovery, metadata, and SDK options.
Show a summary per file
| File | Description |
|---|---|
agentHostLogSources.ts |
Generalizes remote connection lookup and log discovery. |
exportAgentHostDebugLogsAction.ts |
Collects and exports advertised artifacts. |
exportDebugLogsAction.ts |
Supplies Agents-window session artifacts. |
baseAgentHostSessionsProvider.ts |
Exposes adapter debug artifacts. |
claudeSdkOptions.test.ts |
Tests SDK debug-file projection. |
claudeDebugArtifacts.test.ts |
Tests artifact discovery and preparation. |
sessionDebugArtifactsMeta.test.ts |
Tests metadata serialization and validation. |
agentHostLogNaming.test.ts |
Tests naming and artifact projection. |
claudeSdkOptions.ts |
Configures the Claude SDK debug file. |
claudeDebugArtifacts.ts |
Discovers Claude logs and transcripts. |
claudeAgentSession.ts |
Publishes discovered artifacts into session metadata. |
sessionState.ts |
Defines the debug-artifact metadata contract. |
ahpJsonlLogger.ts |
Reuses shared naming helpers. |
agentHostLogNaming.ts |
Adds shared naming and projection utilities. |
Review details
Comments suppressed due to low confidence (1)
src/vs/workbench/contrib/chat/browser/actions/exportAgentHostDebugLogsAction.ts:386
- This still uses the Copilot-only parser, so
remote-<authority>-claude(and every other non-Copilot provider) returnsundefined, contrary to this function's updated contract. As a result, the workbench action treats a remote Claude chat as having no active Agent Host session and omits all session-specific logs. Use the provider-agnostic remote-session-type predicate here.
if (parseRemoteAuthorityFromScheme(resource.scheme)) {
return { resource, title, isLocal: false };
- Files reviewed: 14/14 changed files
- Comments generated: 4
- Review effort level: Medium
Comment on lines
+345
to
+346
| if (activeSession?.isLocal) { | ||
| const debugArtifacts = await resolveLocalDebugArtifacts(agentHostService, activeSession.resource); |
Comment on lines
+351
to
+353
| // The original `accessor` is invalid after the await above, so re-enter with | ||
| // a fresh one for the (accessor-threading) export helper. | ||
| await instantiationService.invokeFunction(accessor => exportAgentHostDebugLogs(accessor, activeSession)); |
| const token = claudeDebugLogSessionToken(this._sessionId); | ||
| try { | ||
| const stat = await this._fileService.resolve(joinPath(this._logsHome, CLAUDE_LOG_DIR)); | ||
| return (stat.children ?? []).filter(c => !c.isDirectory && c.name.endsWith('.log') && c.name.includes(token)).map(c => c.resource.fsPath); |
…ts, drop invokeFunction - claudeDebugArtifacts: match the exact `-<token>.log` suffix instead of `includes(token)`, so a sibling session whose id merely contains this token (e.g. `sess-abc` vs `sess-abc-extra`) can't leak its debug log into the export. - claudeAgentSession: let an empty projection clear the `_meta` debugArtifacts slot (withSessionDebugArtifacts already removes only this slot), keeping the advertised state a true projection of disk; skip only the initial empty write. - export action: resolve debug artifacts inside collectAgentHostDebugLogs for both local and remote sessions (remote via the matching IAgentConnection), and drop the invokeFunction re-entry so run() uses the accessor synchronously. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Member
Author
|
Addressed all four review comments in 668628a:
The Linux / Electron check failure was an unrelated flake — |
Contributor
|
Base:
|
Both IAgentHostService and a remote IAgentConnection structurally expose `listSessions()`, so resolveSessionDebugArtifacts can take the owning host directly rather than a `() => host.listSessions()` thunk. Pick the lister once (local service or remote connection) and pass it, collapsing the two branches. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
resolveSessionDebugArtifacts had one call site and ~3 functional lines; fold it into collectAgentHostDebugLogs. `sessions` now infers its type from the picked lister's listSessions(), so the IAgentSessionMetadata import drops too. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Three cleanups from PR review: - Compute the transcript path from the cwd instead of scanning every ~/.claude/projects directory. The CLI slug is `cwd.replace(/[^a-zA-Z0-9]/g, '-')` (verified against the on-disk project dirs), so resolving the transcript is a single `stat`, not O(projects). claudeProjectSlug / buildClaudeTranscriptPath are pure and unit-tested. - Drop the observable + autorun. ClaudeDebugArtifacts.refresh(cwd) now returns the projected artifacts and the session's _publishDebugArtifacts writes them to `_meta` directly, at the three change points (materialize / rematerialize / first turn). With one producer and one consumer the observable was pure indirection, and its autorun fired a wasted no-op run with [] on registration. - ClaudeDebugArtifacts is now stateless; tests await refresh() rather than reading an observable, plus a new test proves the transcript is addressed by cwd slug (a same-id file under a different project dir is ignored). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Applies two learnings recovered from the old copilot-chat computeFolderSlug (which was removed when Claude session handling moved to the SDK's listSessions): - Windows: VS Code lowercases the URI drive letter, but the CLI slugs its native uppercase process.cwd(), so claudeProjectSlug now re-uppercases a leading `<drive>:` (`c:\x` -> `C--x`). Without this the transcript was silently missed on Windows. - The slug is only a best-effort match (a symlinked cwd, an unusual path char the encoder gets wrong, or a CLI version change all make it miss), so _readTranscriptPath now computes the path first (one stat) and, only on a miss, falls back to a bounded scan of ~/.claude/projects/* for <sessionId>.jsonl (newest). O(1) in the common case; correct in the edge cases. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Contributor
There was a problem hiding this comment.
Review details
Comments suppressed due to low confidence (1)
src/vs/workbench/contrib/chat/browser/actions/exportAgentHostDebugLogsAction.ts:371
- Remote Claude sessions are still excluded here:
parseRemoteAuthorityFromSchemeonly acceptsremote-<authority>-copilotcli(seecopilotCliEventsUri.ts:85-92). In the workbench action this makesactiveSessionundefined, socollectAgentHostDebugLogsnever queries the remote Claude host's_meta. Use the provider-neutralisRemoteAgentHostSessionTypecheck; authority resolution is already handled later bygetRemoteConnectionForSession.
* `undefined` if the scheme does not belong to an agent-host session. Covers any
* local agent host (`agent-host-<provider>`, e.g. Copilot CLI or Claude) and
* remote agent hosts; the EH CLI extension's own `copilotcli:` sessions are
- Files reviewed: 14/14 changed files
- Comments generated: 2
- Review effort level: Medium
Comment on lines
+241
to
+243
| const resource = activeSession?.isLocal | ||
| ? URI.file(artifact.path) | ||
| : artifactAuthority ? toAgentHostUri(URI.from({ scheme: Schemas.file, path: artifact.path }), artifactAuthority) : undefined; |
Comment on lines
+248
to
+252
| // Capture the size so remote artifacts use the bounded read path (a large | ||
| // `--debug` log shouldn't be slurped whole over the proxy), and a missing | ||
| // file (e.g. one whose SDK subprocess hasn't written it yet) is skipped. | ||
| const { size } = await fileService.resolve(resource, { resolveMetadata: true }); | ||
| files.push(await createDebugLogFile(`${sanitizeFilePart(artifact.label)}${extname(resource)}`, resource, fileService, size)); |
Two remote-path review comments:
- A Windows agent host advertised raw fsPaths like `C:\Users\...`, which the
client cannot wrap (`URI.from`/`URI.file` mangle a foreign drive/UNC path). The
host now encodes each artifact as a `file://` URI string (it owns its own path
semantics); the client just parses it and, for remote, wraps it for the
connection authority. Renamed the IDebugArtifact field `path` -> `uri`.
- The AHP filesystem provider has no ranged read, so `readFileStream({length})`
still fetched the whole remote file into the renderer (then re-buffered it).
Cap by stat size in createDebugLogFile: above 50 MiB, export a short note
instead of slurping a runaway `--debug` log.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Member
Author
|
Addressed the two new review comments in e7a7618:
(The other four inline comments are the original 2026-07-24 review, already fixed in 668628a.) Verified: typecheck + eslint + layers clean, 12673 node tests passing. Both fixes are on the remote path, which I can't exercise in the dev sandbox — CI covers it. |
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.
What & why
The Export Agent Host Debug Logs command previously surfaced only Copilot's logs. This includes the Claude backend's debug artifacts in the export:
--debuglog (enabled per-run viaOptions.debugFile), andevents.jsonl).How it works
Rather than hard-coding a provider's on-disk layout in the client, the host advertises its artifacts through the AHP session
_metabag under a well-knowndebugArtifactskey — mirroring the existingagentHost.gitprecedent. No protocol change.agentHostLogNaming(new, common) — shared, pure log-naming helpers (sanitize + file-safe timestamp) reused by both the AHP JSONL logger and Claude debug-file naming, plus the purebuildClaudeDebugArtifactsprojection.ClaudeDebugArtifacts(new, node) — owns discovery + publishing: points the SDK at a per-run debug file, globs the log(s) (current + prior runs) and the transcript, and exposes them as an observable that is a pure projection of disk truth.ClaudeAgentSessionis a thin consumer that bridges the observable into_meta.AgentHostSessionAdaptercarrying_metaand streams each artifact via the resource proxy (file://local /vscode-agent-host://remote), with provider-agnostic remote authority resolution.Testing
npm run typecheck-client— cleannpm run valid-layers-check— cleanFileService+InMemoryFileSystemProvider):claudeDebugArtifacts.test.ts— glob discovery / filtering / numbering / transcript /prepareDebugFileagentHostLogNaming.test.ts,sessionDebugArtifactsMeta.test.ts,claudeSdkOptions.test.tsClaude-debug-log.log+Claude-transcript.jsonlwith a matching session id.Draft while the agent-host e2e replay suite runs in CI (the local Electron integration harness can't init in the dev sandbox).
🤖 Generated with Claude Code