fix(claude): keep IDE-context envelopes out of first_message previews

The VS Code extension often prepends a <ide_opened_file> or
<ide_selection> wrapper directly onto a real prompt in the same user
entry. PR kenn-io#1252 promoted standalone envelopes to hidden system
metadata, but this mixed case (envelope + real prompt in one message)
falls outside that strict standalone match by design, so the raw
markup stayed in first_message and the visible transcript.

Split a leading IDE-context envelope off from the rest of the
message: the envelope becomes its own hidden system-metadata message
(same ide_opened_file/ide_selection subtype as the standalone case),
and the remaining real prompt becomes an ordinary user message.
first_message and user turn counts are computed from the same
firstMessageAndUserCount pass used before, so they now derive from
the real prompt only.

Bumps the parser data version to 75 so existing rows re-parse.

Author: adammw

internal/db/db.go

@@ -361,7 +361,12 @@ const projectIdentityRemoteScrubCompletedKey = "project_identity_remote_scrub_v1
 // last_activity_at, the Devin fingerprint hashes only raw epoch integers and
 // zero-time metadata, so it is byte-identical before and after the fix and
 // incremental sync would skip the correction.)
-const dataVersion = 78
+// (79: Claude Code IDE context wrappers prepended onto a real prompt in
+// the same entry are now split into a hidden system-metadata message plus
+// the real prompt, instead of leaving the raw wrapper in first_message and
+// the visible transcript. Existing rows need re-parsing so first_message
+// and message content drop the leading markup.)
+const dataVersion = 79
 
 const tokenCoverageRepairStatsKey = "token_coverage_repair_v1"
 

internal/db/db_test.go

@@ -1008,10 +1008,9 @@ func TestMigration_ToolResultEventsTable(t *testing.T) {
 		"expected tool_result_events table after reopen")
 }
 
-func TestCurrentDataVersionDevinEpochSecondsReparse(t *testing.T) {
-	assert.Equal(t, 78, CurrentDataVersion(),
-		"version 78 reparses Devin sessions whose timestamps were read as "+
-			"milliseconds and stored as 1970 dates")
+func TestCurrentDataVersionClaudeIDEEnvelopeSplit(t *testing.T) {
+	assert.Equal(t, 79, CurrentDataVersion(),
+		"version 79 splits Claude IDE envelopes off mixed prompts")
 }
 
 func TestInsertMessages_PreservesToolResultEvents(t *testing.T) {

internal/parser/claude.go

@@ -1093,6 +1093,29 @@ func extractMessagesFrom(
 				ordinal++
 				continue
 			}
+			// The VS Code extension sometimes prepends an IDE-context
+			// wrapper directly onto a real prompt in the same entry.
+			// Split it into a hidden system-metadata message plus the
+			// real prompt, so first_message and the visible transcript
+			// show only the prompt.
+			if subtype, envelope, remainder, ok :=
+				splitLeadingClaudeIDEEnvelope(text); ok {
+				messages = append(messages, ParsedMessage{
+					Ordinal:          ordinal,
+					Role:             RoleUser,
+					Content:          envelope,
+					Timestamp:        e.timestamp,
+					IsSystem:         true,
+					ContentLength:    len(envelope),
+					SourceType:       "system",
+					SourceSubtype:    subtype,
+					SourceUUID:       e.uuid,
+					SourceParentUUID: e.parentUuid,
+					IsSidechain:      gjson.Get(e.line, "isSidechain").Bool(),
+				})
+				ordinal++
+				text = remainder
+			}
 			// Skip unclassified noise (e.g. non-caveat
 			// <local-command-*> envelopes).
 			if isClaudeSystemMessage(text) {
@@ -2125,6 +2148,29 @@ func extractMessages(entries []dagEntry) (
 				ordinal++
 				continue
 			}
+			// The VS Code extension sometimes prepends an IDE-context
+			// wrapper directly onto a real prompt in the same entry.
+			// Split it into a hidden system-metadata message plus the
+			// real prompt, so first_message and the visible transcript
+			// show only the prompt.
+			if subtype, envelope, remainder, ok :=
+				splitLeadingClaudeIDEEnvelope(text); ok {
+				messages = append(messages, ParsedMessage{
+					Ordinal:          ordinal,
+					Role:             RoleUser,
+					Content:          envelope,
+					Timestamp:        e.timestamp,
+					IsSystem:         true,
+					ContentLength:    len(envelope),
+					SourceType:       "system",
+					SourceSubtype:    subtype,
+					SourceUUID:       e.uuid,
+					SourceParentUUID: e.parentUuid,
+					IsSidechain:      gjson.Get(e.line, "isSidechain").Bool(),
+				})
+				ordinal++
+				text = remainder
+			}
 			if isClaudeSystemMessage(text) {
 				continue
 			}
@@ -2611,6 +2657,50 @@ func isStandaloneClaudeTaggedMessage(content, tag string) bool {
 		len(afterOpen)-len(closeTag)
 }
 
+// claudeIDEEnvelopeTags are the VS Code extension's IDE-context
+// wrapper tags: standalone messages using these are already
+// promoted to hidden system metadata by classifyClaudeSystemMessage.
+// splitLeadingClaudeIDEEnvelope handles the remaining case where the
+// extension prepends one of these wrappers directly onto a real
+// prompt in the same user entry.
+var claudeIDEEnvelopeTags = [...]string{"ide_opened_file", "ide_selection"}
+
+// splitLeadingClaudeIDEEnvelope detects a well-formed IDE-context
+// envelope at the very start of content that is followed by
+// additional real prompt text, and separates the two. The standalone
+// case (envelope with nothing else) is left alone here; that is
+// handled by classifyClaudeSystemMessage so the whole message
+// promotes to system metadata.
+//
+// Splitting keeps the envelope recorded as hidden system metadata
+// (same subtype as the standalone case) while letting first_message
+// and the visible transcript show only the real prompt that follows,
+// instead of raw IDE-context markup.
+func splitLeadingClaudeIDEEnvelope(
+	content string,
+) (subtype, envelope, remainder string, ok bool) {
+	trimmed := trimClaudeSystemMessagePrefix(content)
+	for _, tag := range claudeIDEEnvelopeTags {
+		openTag := "<" + tag + ">"
+		closeTag := "</" + tag + ">"
+		if !strings.HasPrefix(trimmed, openTag) {
+			continue
+		}
+		closeIdx := strings.Index(trimmed, closeTag)
+		if closeIdx < 0 {
+			continue
+		}
+		envelopeEnd := closeIdx + len(closeTag)
+		rest := strings.TrimSpace(trimmed[envelopeEnd:])
+		if rest == "" {
+			// Standalone: classifyClaudeSystemMessage handles this.
+			continue
+		}
+		return tag, trimmed[:envelopeEnd], rest, true
+	}
+	return "", "", "", false
+}
+
 func stripLeadingClaudeSystemReminderContent(content string) string {
 	trimmed := trimClaudeSystemMessagePrefix(content)
 	remainder, stripped := stripLeadingClaudeSystemReminderBlocks(trimmed)

internal/parser/claude_parser_test.go

@@ -324,6 +324,48 @@ func TestParseClaudeSession_SkippedMessages(t *testing.T) {
 		assert.Equal(t, "real user message", msgs[7].Content)
 	})
 
+	t.Run("splits IDE envelope prepended onto a real prompt", func(t *testing.T) {
+		content := testjsonl.JoinJSONL(
+			testjsonl.ClaudeUserJSON(
+				"<ide_opened_file>The user opened /workspace/app/README.md.</ide_opened_file> Explain this file.",
+				tsZero,
+			),
+			testjsonl.ClaudeUserJSON(
+				"<ide_selection>The user selected package main.</ide_selection>\n\nWhat does this do?",
+				tsZeroS1,
+			),
+		)
+		sess, msgs := runClaudeParserTest(t, "test.jsonl", content)
+		// Each entry splits into a hidden system-metadata message
+		// plus the real prompt that followed it.
+		require.Len(t, msgs, 4)
+		assert.Equal(t, 4, sess.MessageCount)
+		assert.Equal(t, 2, sess.UserMessageCount)
+		assert.Equal(t, "Explain this file.", sess.FirstMessage,
+			"first_message should show the real prompt, not the IDE envelope")
+
+		assert.True(t, msgs[0].IsSystem)
+		assert.Equal(t, RoleUser, msgs[0].Role)
+		assert.Equal(t, "system", msgs[0].SourceType)
+		assert.Equal(t, "ide_opened_file", msgs[0].SourceSubtype)
+		assert.Equal(t,
+			"<ide_opened_file>The user opened /workspace/app/README.md.</ide_opened_file>",
+			msgs[0].Content)
+
+		assert.False(t, msgs[1].IsSystem)
+		assert.Equal(t, RoleUser, msgs[1].Role)
+		assert.Equal(t, "Explain this file.", msgs[1].Content)
+
+		assert.True(t, msgs[2].IsSystem)
+		assert.Equal(t, "ide_selection", msgs[2].SourceSubtype)
+		assert.Equal(t,
+			"<ide_selection>The user selected package main.</ide_selection>",
+			msgs[2].Content)
+
+		assert.False(t, msgs[3].IsSystem)
+		assert.Equal(t, "What does this do?", msgs[3].Content)
+	})
+
 	t.Run("skill invocation shown as user message", func(t *testing.T) {
 		content := testjsonl.JoinJSONL(
 			testjsonl.ClaudeUserJSON(
@@ -869,6 +911,46 @@ func TestParseClaudeSessionFrom_IDEContext(t *testing.T) {
 	}
 }
 
+func TestParseClaudeSessionFrom_IDEContextPrependedToPrompt(t *testing.T) {
+	t.Parallel()
+
+	initial := testjsonl.JoinJSONL(
+		testjsonl.ClaudeUserJSON("hello", tsEarly),
+		testjsonl.ClaudeAssistantJSON("hi", tsEarlyS1),
+	)
+	path := createTestFile(t, "inc-ide-context-prompt.jsonl", initial)
+	info, err := os.Stat(path)
+	require.NoError(t, err)
+
+	appended := testjsonl.JoinJSONL(
+		testjsonl.ClaudeUserJSON(
+			"<ide_opened_file>The user opened /workspace/app/README.md.</ide_opened_file> Explain this file.",
+			tsLate,
+		),
+	)
+	f, err := os.OpenFile(path, os.O_APPEND|os.O_WRONLY, 0o644)
+	require.NoError(t, err)
+	_, err = f.WriteString(appended)
+	require.NoError(t, err)
+	require.NoError(t, f.Close())
+
+	newMsgs, _, _, err := callParseClaudeSessionFrom(path, info.Size(), 2, "")
+	require.NoError(t, err)
+	require.Len(t, newMsgs, 2,
+		"the entry splits into a hidden IDE-context message plus the real prompt")
+
+	assert.True(t, newMsgs[0].IsSystem)
+	assert.Equal(t, "system", newMsgs[0].SourceType)
+	assert.Equal(t, "ide_opened_file", newMsgs[0].SourceSubtype)
+	assert.Equal(t,
+		"<ide_opened_file>The user opened /workspace/app/README.md.</ide_opened_file>",
+		newMsgs[0].Content)
+
+	assert.False(t, newMsgs[1].IsSystem)
+	assert.Equal(t, RoleUser, newMsgs[1].Role)
+	assert.Equal(t, "Explain this file.", newMsgs[1].Content)
+}
+
 func TestParseClaudeSessionFrom_ReminderPrefixedCommand(t *testing.T) {
 	t.Parallel()