Repository navigation
Conversation
…anaflow-ai#1973) * test: cover queued restore-time terminal portal shift * fix: resync terminal portal after restore-time bind
Add keyword-based tab auto-naming to the `claude-hook stop` handler. When Claude Code finishes a turn, the transcript is analyzed to extract the most frequent meaningful keywords from user messages, which are used to generate a short (2-5 word, max 25 char) tab title. Key design decisions: - Uses keyword frequency extraction (not AI/LLM) for zero-latency naming - Supports CJK + Latin tokenization for multilingual conversations - Respects user /rename: if a custom-title is detected in the transcript, auto-naming is skipped entirely - Extracts first 3 + last 3 user messages to capture both the initial topic and recent focus shifts - Filters common stop words in English and Chinese This runs synchronously in the existing stop hook path with negligible overhead since the transcript is already being read for notification summaries.
|
@Horacehxw is attempting to deploy a commit to the Manaflow Team on Vercel. A member of the Team first needs to authorize it. |
📝 WalkthroughWalkthroughThis pull request introduces automatic tab renaming from transcript summaries, enhances UI test stability through retry logic with window/foreground activation, adds external geometry synchronization after terminal bind operations to handle layout shifts, and implements CI retry logic for a display resolution regression test with output capture. Changes
Estimated code review effort🎯 4 (Complex) | ⏱️ ~50 minutes Possibly related PRs
Poem
🚥 Pre-merge checks | ✅ 2 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (2 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches🧪 Generate unit tests (beta)
⚔️ Resolve merge conflicts
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
Greptile SummaryThis PR adds automatic tab titling for Claude Code sessions in cmux by extracting high-frequency keywords from user messages in the conversation transcript, and separately hardens the display-resolution UI regression CI job with a one-shot retry for a known launch-activation flake, plus a session-restore geometry fix in Key changes:
Confidence Score: 4/5
Important Files Changed
Sequence DiagramsequenceDiagram
participant CC as Claude Code
participant Hook as cmux CLI<br/>(claude-hook stop)
participant TS as readTranscriptSummary
participant GT as generateTabTitle
participant Server as cmux App<br/>(tab.action rename)
CC->>Hook: hook event: stop (transcript path)
Hook->>TS: read JSONL transcript
TS->>TS: collect firstUserMessages[0..2]<br/>lastUserMessages (ring buffer, max 3)<br/>detect custom-title entries
TS-->>Hook: TranscriptSummary { firstUserMessages, lastUserMessages, hasCustomTitle }
alt hasCustomTitle == true
Hook-->>CC: skip auto-rename (user /rename takes precedence)
else hasCustomTitle == false
Hook->>GT: generateTabTitle(transcript)
GT->>GT: combine first+last messages<br/>tokenize CJK runs + Latin words<br/>filter stop words (EN+ZH)<br/>rank by frequency → top 5<br/>build title ≤25 chars, 2–5 words
GT-->>Hook: tabTitle (or nil if no keywords)
Hook->>Server: tab.action rename { workspace_id, surface_id, title }
Server-->>Hook: OK
end
Hook->>Server: setClaudeStatus → Idle
Reviews (1): Last reviewed commit: "feat: auto-name tabs from Claude Code co..." | Re-trigger Greptile |
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
| /// Generate a short tab title from user messages using keyword extraction. | ||
| /// Preserves the conversation language (works with CJK, Latin, etc). | ||
| private func generateTabTitle(from transcript: TranscriptSummary) -> String? { | ||
| let allMessages = transcript.firstUserMessages + transcript.lastUserMessages |
Contributor
There was a problem hiding this comment.
firstUserMessages and lastUserMessages overlap for short conversations
When there are ≤5 user messages, some or all messages appear in both firstUserMessages and lastUserMessages. Concatenating them without deduplication inflates keyword frequencies for those overlapping messages:
- 3 user messages: every message is counted twice — ranking is effectively unchanged, but work is doubled.
- 4 user messages:
firstUserMessages = [m1, m2, m3],lastUserMessages = [m2, m3, m4], soallMessages = [m1, m2, m3, m2, m3, m4]. Tokens fromm2andm3are counted at 2× the weight ofm1andm4. - 5 user messages:
m3is in both windows and receives a 2× frequency boost overm1,m2,m4,m5.
This biases keyword scoring toward middle-of-conversation messages, potentially surfacing stale topics over the most recent focus (m4/m5).
The simplest fix is to deduplicate in generateTabTitle before building allMessages:
// Remove messages already captured by firstUserMessages so they're not double-counted.
let uniqueLast = transcript.lastUserMessages.filter { !transcript.firstUserMessages.contains($0) }
let allMessages = transcript.firstUserMessages + uniqueLast(Or collect indices rather than values during parsing so deduplication is exact even when two messages happen to have identical truncated text.)
Comment on lines
+10978
to
+10987
| // Build title: 2-5 words, max 25 chars | ||
| var seen = Set<String>() | ||
| var deduped: [String] = [] | ||
| for word in top { | ||
| let low = word.lowercased() | ||
| if !seen.contains(low) { | ||
| seen.insert(low) | ||
| deduped.append(word) | ||
| } | ||
| } |
Contributor
There was a problem hiding this comment.
freq is a [String: Int] dictionary whose keys are already lowercased strings (built with let low = token.lowercased()). top is produced by sorting that dictionary and mapping each entry back through casing[$0.key] ?? $0.key — every entry has a unique lowercase key. Therefore seen.contains(low) is never true, and deduped always ends up identical to Array(top).
The entire seen/deduped block can be replaced with a direct assignment:
Suggested change
| // Build title: 2-5 words, max 25 chars | |
| var seen = Set<String>() | |
| var deduped: [String] = [] | |
| for word in top { | |
| let low = word.lowercased() | |
| if !seen.contains(low) { | |
| seen.insert(low) | |
| deduped.append(word) | |
| } | |
| } | |
| let deduped = top |
|
|
||
| // Tokenize: CJK runs (2+ chars) or Latin word tokens | ||
| var tokens: [String] = [] | ||
| let pattern = try? NSRegularExpression(pattern: "[\\u4e00-\\u9fff]{2,}|[a-zA-Z][a-zA-Z0-9]*", options: []) |
Contributor
There was a problem hiding this comment.
generateTabTitle call
NSRegularExpression(pattern:options:) performs full pattern compilation on each invocation. Since the pattern is a constant, it should be compiled once and reused — e.g., as a static let or a private computed property at struct scope:
private static let tokenPattern: NSRegularExpression? = try? NSRegularExpression(
pattern: "[\\u4e00-\\u9fff]{2,}|[a-zA-Z][a-zA-Z0-9]*",
options: []
)This is low-impact today (the hook runs only on session stop), but it is unnecessary per-call overhead.
![cubic-dev-ai[bot]](https://avatars.githubusercontent.com/in/1082092?s=60&v=4){kind=small}
There was a problem hiding this comment.
3 issues found across 5 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="Sources/AppDelegate.swift">
<violation number="1" location="Sources/AppDelegate.swift:2425">
P2: Nested UITest retry loops can overlap and enqueue redundant main-thread retries, increasing flakiness and timing churn.</violation>
<violation number="2" location="Sources/AppDelegate.swift:2448">
P2: macOS 14+ UI-test activation dropped `.activateIgnoringOtherApps`, so retrying on `!isActive` can fail to ever foreground the app.</violation>
</file>
<file name="cmuxTests/TerminalAndGhosttyTests.swift">
<violation number="1" location="cmuxTests/TerminalAndGhosttyTests.swift:2993">
P2: New test uses a fixed 50ms wait for async coordination, which can make geometry assertions flaky under CI/scheduler jitter.</violation>
</file>
Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.
| openNewMainWindow(nil) | ||
| } | ||
|
|
||
| moveUITestWindowToTargetDisplayIfNeeded() |
There was a problem hiding this comment.
P2: Nested UITest retry loops can overlap and enqueue redundant main-thread retries, increasing flakiness and timing churn.
Prompt for AI agents
Check if this issue is valid — if so, understand the root cause and fix it. At Sources/AppDelegate.swift, line 2425:
<comment>Nested UITest retry loops can overlap and enqueue redundant main-thread retries, increasing flakiness and timing churn.</comment>
<file context>
@@ -2416,6 +2411,46 @@ final class AppDelegate: NSObject, NSApplicationDelegate, UNUserNotificationCent
+ openNewMainWindow(nil)
+ }
+
+ moveUITestWindowToTargetDisplayIfNeeded()
+ activateUITestAppIfNeeded()
+
</file context>
| window.orderFrontRegardless() | ||
| } | ||
| if #available(macOS 14.0, *) { | ||
| NSRunningApplication.current.activate(options: [.activateAllWindows]) |
There was a problem hiding this comment.
P2: macOS 14+ UI-test activation dropped .activateIgnoringOtherApps, so retrying on !isActive can fail to ever foreground the app.
Prompt for AI agents
Check if this issue is valid — if so, understand the root cause and fix it. At Sources/AppDelegate.swift, line 2448:
<comment>macOS 14+ UI-test activation dropped `.activateIgnoringOtherApps`, so retrying on `!isActive` can fail to ever foreground the app.</comment>
<file context>
@@ -2416,6 +2411,46 @@ final class AppDelegate: NSObject, NSApplicationDelegate, UNUserNotificationCent
+ window.orderFrontRegardless()
+ }
+ if #available(macOS 14.0, *) {
+ NSRunningApplication.current.activate(options: [.activateAllWindows])
+ } else {
+ NSRunningApplication.current.activate(options: [.activateAllWindows, .activateIgnoringOtherApps])
</file context>
Suggested change
| NSRunningApplication.current.activate(options: [.activateAllWindows]) | |
| NSRunningApplication.current.activate(options: [.activateAllWindows, .activateIgnoringOtherApps]) |
| window.displayIfNeeded() | ||
| } | ||
|
|
||
| RunLoop.current.run(until: Date().addingTimeInterval(0.05)) |
There was a problem hiding this comment.
P2: New test uses a fixed 50ms wait for async coordination, which can make geometry assertions flaky under CI/scheduler jitter.
Prompt for AI agents
Check if this issue is valid — if so, understand the root cause and fix it. At cmuxTests/TerminalAndGhosttyTests.swift, line 2993:
<comment>New test uses a fixed 50ms wait for async coordination, which can make geometry assertions flaky under CI/scheduler jitter.</comment>
<file context>
@@ -2939,6 +2939,83 @@ final class TerminalWindowPortalLifecycleTests: XCTestCase {
+ window.displayIfNeeded()
+ }
+
+ RunLoop.current.run(until: Date().addingTimeInterval(0.05))
+
+ let shiftedAnchorFrameInWindow = anchor.convert(anchor.bounds, to: nil)
</file context>
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 5
🧹 Nitpick comments (2)
cmuxTests/TerminalAndGhosttyTests.swift (1)
2993-2994: Make the async settle step deterministic to reduce CI flakes.The fixed
RunLoopdelay can intermittently fail under load. Prefer the existingdrainMainQueue()helper so the test waits for queued main-thread work instead of wall-clock timing.♻️ Suggested test stabilization
- RunLoop.current.run(until: Date().addingTimeInterval(0.05)) + drainMainQueue()🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@cmuxTests/TerminalAndGhosttyTests.swift` around lines 2993 - 2994, The test currently uses a fixed delay via RunLoop.current.run(until: Date().addingTimeInterval(0.05)) which is flaky; replace that line with a call to the existing drainMainQueue() helper so the test deterministically waits for queued main-thread work to complete (i.e., remove the RunLoop.current.run(...) invocation and call drainMainQueue() in its place, ensuring drainMainQueue() is in scope for the test)..github/workflows/ci.yml (1)
510-520: Make the flake detector file-based and newline-safe.Line 513's predicate only matches if both fragments land on the same log line.
grepwon't let.*span newlines, so the retry can be skipped if XCTest wraps that activation error. Grepping the file directly also removes the unnecessaryOUTPUT=$(cat ...)copies.♻️ Proposed change
- OUTPUT=$(cat /tmp/display-resolution-ui-test-output.txt) set -e - if [ "$EXIT_CODE" -ne 0 ] && echo "$OUTPUT" | grep -q "Failed to activate application.*Running Background"; then + if [ "$EXIT_CODE" -ne 0 ] \ + && grep -q "Failed to activate application" /tmp/display-resolution-ui-test-output.txt \ + && grep -q "Running Background" /tmp/display-resolution-ui-test-output.txt; then echo "Display resolution UI regression hit launch activation flake, retrying once" pkill -x "cmux DEV" || true sleep 2 set +e run_display_resolution_ui_test | tee /tmp/display-resolution-ui-test-output.txt EXIT_CODE=${PIPESTATUS[0]} - OUTPUT=$(cat /tmp/display-resolution-ui-test-output.txt) set -e fi🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In @.github/workflows/ci.yml around lines 510 - 520, Change the flake detector to operate on the test output file itself and make the match newline-safe: instead of piping the shell variable OUTPUT into grep, run a newline-aware search like grep -qz "Failed to activate application.*Running Background" /tmp/display-resolution-ui-test-output.txt (or use perl -0777 -ne 'print if /Failed to activate application.*Running Background/' for portability), remove the redundant OUTPUT=$(cat /tmp/display-resolution-ui-test-output.txt) copies and use the file path /tmp/display-resolution-ui-test-output.txt and EXIT_CODE for the retry predicate so the retry triggers even when the two fragments are on different lines.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@CLI/cmux.swift`:
- Around line 10929-10969: The generateTabTitle(from:) function currently builds
the frequency map from raw messages, causing sensitive data (paths, emails,
UUIDs, tokens, ENV_VAR=..., long numbers) to appear in tab titles; before
tokenization, run the same redaction used by the Claude tag-extraction (apply
the unanchored sensitiveSpanPatterns across the combined subtitle+body string to
remove/replace sensitive spans), then proceed to tokenize the redacted combined
text, and additionally filter each resulting token against the anchored
sensitivePatterns (and skip any matches) before updating freq/casing; reference
generateTabTitle, combined, tokens, freq, and the existing
sensitiveSpanPatterns/sensitivePatterns utilities when implementing this change.
- Around line 10263-10272: The auto-titling currently sends a regular tab.action
"rename" which gets persisted as a user/custom title and can be overwritten by
later transcript state; instead send a distinct, non-persistent signal so manual
renames aren't clobbered: in the block that calls client.sendV2 (around
generateTabTitle and transcript.hasCustomTitle), replace or augment the "rename"
action with an explicit auto-title marker (e.g., use action "auto_rename" or add
a parameter like "auto_generated": true) so the app can treat it as
non-persistent; keep the transcript.hasCustomTitle guard in place and ensure the
app-side handler for "auto_rename" does not write to the same custom-title
persistence path used by manual rename actions.
- Around line 10265-10266: The variable `transcript` is undefined where it's
being checked; update the flow so the stop/idle branch has access to a
TranscriptSummary instance: either change summarizeClaudeHookStop(...) to return
the TranscriptSummary (not just (subtitle: String, body: String)?) and propagate
that value through runClaudeHook(...) so you can read transcript.hasCustomTitle
where generateTabTitle is called, or bind/obtain the TranscriptSummary directly
in the stop/idle branch before the if-check; specifically modify
summarizeClaudeHookStop, runClaudeHook, and the call site so a TranscriptSummary
(with hasCustomTitle) is available to evaluate and pass into generateTabTitle.
- Around line 10950-10952: The tokenizer's regex (pattern) only matches Han
characters and Latin words, causing Japanese and Korean input to produce no
tokens and an empty frequency dictionary; update the NSRegularExpression used in
the tokenization (the pattern variable in CLI/cmux.swift) to include Hiragana
(\u3040-\u309f), Katakana (\u30a0-\u30ff) and Hangul (\uac00-\ud7af) ranges for
multi-character runs (e.g. use ranges like
[\u4e00-\u9fff\u3040-\u309f\u30a0-\u30ff\uac00-\ud7af]{2,}) or switch to Unicode
script properties (e.g. \p{Han}, \p{Hiragana}, \p{Katakana}, \p{Hangul}) so
tokens and the frequency dictionary are correctly populated instead of returning
nil.
In `@Sources/AppDelegate.swift`:
- Around line 2425-2438: The stabilizeUITestLaunchWindowAndForeground retry loop
is repeatedly calling moveUITestWindowToTargetDisplayIfNeeded() (which itself
schedules retries), causing overlapping timers; to fix, ensure
moveUITestWindowToTargetDisplayIfNeeded() (and optionally
activateUITestAppIfNeeded()) are invoked once before starting the retry loop in
stabilizeUITestLaunchWindowAndForeground (or add a guard inside
moveUITestWindowToTargetDisplayIfNeeded to no-op if a retry is already
scheduled), and keep writeUITestDiagnosticsIfNeeded(stage:) within the loop for
per-attempt logging so you avoid nested retry storms while preserving
stabilization retries.
---
Nitpick comments:
In @.github/workflows/ci.yml:
- Around line 510-520: Change the flake detector to operate on the test output
file itself and make the match newline-safe: instead of piping the shell
variable OUTPUT into grep, run a newline-aware search like grep -qz "Failed to
activate application.*Running Background"
/tmp/display-resolution-ui-test-output.txt (or use perl -0777 -ne 'print if
/Failed to activate application.*Running Background/' for portability), remove
the redundant OUTPUT=$(cat /tmp/display-resolution-ui-test-output.txt) copies
and use the file path /tmp/display-resolution-ui-test-output.txt and EXIT_CODE
for the retry predicate so the retry triggers even when the two fragments are on
different lines.
In `@cmuxTests/TerminalAndGhosttyTests.swift`:
- Around line 2993-2994: The test currently uses a fixed delay via
RunLoop.current.run(until: Date().addingTimeInterval(0.05)) which is flaky;
replace that line with a call to the existing drainMainQueue() helper so the
test deterministically waits for queued main-thread work to complete (i.e.,
remove the RunLoop.current.run(...) invocation and call drainMainQueue() in its
place, ensuring drainMainQueue() is in scope for the test).
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro
Run ID: c2d8771a-0169-464b-bd7d-669efae1c0f8
📒 Files selected for processing (5)
.github/workflows/ci.ymlCLI/cmux.swiftSources/AppDelegate.swiftSources/TerminalWindowPortal.swiftcmuxTests/TerminalAndGhosttyTests.swift
Comment on lines
+10263
to
+10272
| // Auto-name the tab based on conversation content. | ||
| // Skip if user has set a custom title via /rename. | ||
| if let transcript, !transcript.hasCustomTitle, | ||
| let tabTitle = generateTabTitle(from: transcript) { | ||
| _ = try? client.sendV2(method: "tab.action", params: [ | ||
| "workspace_id": workspaceId, | ||
| "surface_id": surfaceId, | ||
| "action": "rename", | ||
| "title": tabTitle, | ||
| ]) |
There was a problem hiding this comment.
Don't persist auto-titles through the manual rename action.
This only honors transcript-side custom-title events. A title set from cmux itself (rename-tab, tab context menu, etc.) is not in the Claude transcript, and tab.action "rename" stores through the same custom-title path, so the next stop will overwrite that manual title. This needs a separate auto-title path or an app-side “skip if user-customized” guard.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@CLI/cmux.swift` around lines 10263 - 10272, The auto-titling currently sends
a regular tab.action "rename" which gets persisted as a user/custom title and
can be overwritten by later transcript state; instead send a distinct,
non-persistent signal so manual renames aren't clobbered: in the block that
calls client.sendV2 (around generateTabTitle and transcript.hasCustomTitle),
replace or augment the "rename" action with an explicit auto-title marker (e.g.,
use action "auto_rename" or add a parameter like "auto_generated": true) so the
app can treat it as non-persistent; keep the transcript.hasCustomTitle guard in
place and ensure the app-side handler for "auto_rename" does not write to the
same custom-title persistence path used by manual rename actions.
Comment on lines
+10265
to
+10266
| if let transcript, !transcript.hasCustomTitle, | ||
| let tabTitle = generateTabTitle(from: transcript) { |
There was a problem hiding this comment.
🧩 Analysis chain
🏁 Script executed:
#!/bin/bash
set -euo pipefail
sed -n '10239,10273p' CLI/cmux.swift
printf '\nBindings/usages named transcript:\n'
rg -n '\b(let|var)\s+transcript\b|\bif let transcript\b' CLI/cmux.swift -C2Repository: manaflow-ai/cmux
Length of output: 2448
🏁 Script executed:
#!/bin/bash
set -euo pipefail
# Get the function definition boundaries for runClaudeHook
rg -n 'func runClaudeHook\(' CLI/cmux.swift | head -1
# Extract a wider range around line 10265 to see the full context
echo "=== Full context from 10200 to 10290 ==="
sed -n '10200,10290p' CLI/cmux.swift
# Search for where transcript might be bound in this function
echo -e "\n=== All transcript bindings in broader context ==="
sed -n '10200,10290p' CLI/cmux.swift | rg -n 'transcript'Repository: manaflow-ai/cmux
Length of output: 4522
🏁 Script executed:
#!/bin/bash
set -euo pipefail
# Check the type of 'completion' return from summarizeClaudeHookStop
echo "=== summarizeClaudeHookStop return type and implementation ==="
rg -n 'func summarizeClaudeHookStop' CLI/cmux.swift -A 50 | head -60Repository: manaflow-ai/cmux
Length of output: 2485
transcript is undefined in this scope and will cause a compile error.
runClaudeHook(...) receives only a (subtitle: String, body: String)? return from summarizeClaudeHookStop(...), not the TranscriptSummary with hasCustomTitle property. The transcript variable exists only within summarizeClaudeHookStop at line 10831 and is never returned. Either return TranscriptSummary from summarizeClaudeHookStop, or bind transcript directly in the stop/idle branch before using it at lines 10265–10266.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@CLI/cmux.swift` around lines 10265 - 10266, The variable `transcript` is
undefined where it's being checked; update the flow so the stop/idle branch has
access to a TranscriptSummary instance: either change
summarizeClaudeHookStop(...) to return the TranscriptSummary (not just
(subtitle: String, body: String)?) and propagate that value through
runClaudeHook(...) so you can read transcript.hasCustomTitle where
generateTabTitle is called, or bind/obtain the TranscriptSummary directly in the
stop/idle branch before the if-check; specifically modify
summarizeClaudeHookStop, runClaudeHook, and the call site so a TranscriptSummary
(with hasCustomTitle) is available to evaluate and pass into generateTabTitle.
Comment on lines
+10929
to
+10969
| private func generateTabTitle(from transcript: TranscriptSummary) -> String? { | ||
| let allMessages = transcript.firstUserMessages + transcript.lastUserMessages | ||
| guard !allMessages.isEmpty else { return nil } | ||
|
|
||
| let combined = allMessages.joined(separator: " ") | ||
|
|
||
| // Stop words (English + Chinese) to filter out | ||
| let stopWords: Set<String> = [ | ||
| "please", "can", "you", "the", "a", "an", "i", "want", "to", "me", | ||
| "need", "help", "with", "this", "that", "for", "in", "on", "it", | ||
| "is", "my", "do", "let", "make", "also", "and", "or", "but", | ||
| "just", "some", "all", "any", "be", "have", "will", "would", | ||
| "should", "could", "from", "how", "what", "when", "where", "which", | ||
| "not", "no", "yes", "use", "code", "file", "run", "test", | ||
| "请", "帮我", "帮", "我", "你", "能", "可以", "一下", "看看", | ||
| "这个", "那个", "的", "了", "吗", "呢", "是", "在", "有", "和", | ||
| "把", "给", "让", "对", "到", "从", "用", "也", "都", "就", | ||
| "还", "又", "很", "不", "没", "怎么", "什么", "如何", "已经", | ||
| "然后", "现在", "应该", "可能", "需要", "想", "要", "会", | ||
| ] | ||
|
|
||
| // Tokenize: CJK runs (2+ chars) or Latin word tokens | ||
| var tokens: [String] = [] | ||
| let pattern = try? NSRegularExpression(pattern: "[\\u4e00-\\u9fff]{2,}|[a-zA-Z][a-zA-Z0-9]*", options: []) | ||
| let nsRange = NSRange(combined.startIndex..., in: combined) | ||
| pattern?.enumerateMatches(in: combined, range: nsRange) { match, _, _ in | ||
| guard let match, let range = Range(match.range, in: combined) else { return } | ||
| tokens.append(String(combined[range])) | ||
| } | ||
|
|
||
| // Count frequencies, excluding stop words | ||
| var freq: [String: Int] = [:] | ||
| var casing: [String: String] = [:] | ||
| for token in tokens { | ||
| let low = token.lowercased() | ||
| guard !stopWords.contains(low), low.count >= 2 else { continue } | ||
| freq[low, default: 0] += 1 | ||
| if casing[low] == nil { | ||
| casing[low] = token | ||
| } | ||
| } |
There was a problem hiding this comment.
Redact sensitive spans before counting keywords.
This frequency map is built from raw user prompts, so paths, emails, UUIDs, access tokens, ENV_VAR=..., and long numerics can surface directly as visible tab titles. Reuse the existing Claude tag-extraction redaction/token filters before tokenization.
Based on learnings, pre-redact sensitive spans (UUIDs, emails, access tokens, filesystem paths, ENV_VAR=..., long numerics) across the combined body+subtitle using unanchored sensitiveSpanPatterns before tokenization. Then tokenize and still filter each token with anchored sensitivePatterns.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@CLI/cmux.swift` around lines 10929 - 10969, The generateTabTitle(from:)
function currently builds the frequency map from raw messages, causing sensitive
data (paths, emails, UUIDs, tokens, ENV_VAR=..., long numbers) to appear in tab
titles; before tokenization, run the same redaction used by the Claude
tag-extraction (apply the unanchored sensitiveSpanPatterns across the combined
subtitle+body string to remove/replace sensitive spans), then proceed to
tokenize the redacted combined text, and additionally filter each resulting
token against the anchored sensitivePatterns (and skip any matches) before
updating freq/casing; reference generateTabTitle, combined, tokens, freq, and
the existing sensitiveSpanPatterns/sensitivePatterns utilities when implementing
this change.
Comment on lines
+10950
to
+10952
| // Tokenize: CJK runs (2+ chars) or Latin word tokens | ||
| var tokens: [String] = [] | ||
| let pattern = try? NSRegularExpression(pattern: "[\\u4e00-\\u9fff]{2,}|[a-zA-Z][a-zA-Z0-9]*", options: []) |
There was a problem hiding this comment.
🧩 Analysis chain
🏁 Script executed:
#!/bin/bash
python - <<'PY'
import re
pat = re.compile(r'[\u4e00-\u9fff]{2,}|[a-zA-Z][a-zA-Z0-9]*')
samples = [
"请修复登录问题",
"このテストを直して",
"버그 수정해줘",
]
for s in samples:
print(f"{s!r} -> {pat.findall(s)}")
PYRepository: manaflow-ai/cmux
Length of output: 118
🏁 Script executed:
cd /tmp && git clone --depth 1 https://github.com/manaflow-ai/cmux.git 2>&1 | head -20Repository: manaflow-ai/cmux
Length of output: 910
🏁 Script executed:
wc -l CLI/cmux.swiftRepository: manaflow-ai/cmux
Length of output: 80
🏁 Script executed:
sed -n '10940,10970p' CLI/cmux.swiftRepository: manaflow-ai/cmux
Length of output: 1526
🏁 Script executed:
sed -n '10900,10950p' CLI/cmux.swiftRepository: manaflow-ai/cmux
Length of output: 2593
🏁 Script executed:
sed -n '10960,10990p' CLI/cmux.swiftRepository: manaflow-ai/cmux
Length of output: 1088
The tokenizer misses Hiragana, Katakana, and Hangul, causing silent failure for Japanese and Korean prompts.
The regex [\u4e00-\u9fff]{2,}|[a-zA-Z][a-zA-Z0-9]* only matches Han ideographs and Latin words. Japanese prompts like このテストを直して and Korean prompts like 버그 수정해줘 produce no tokens, resulting in an empty frequency dictionary that triggers a silent nil return at line 10968. Widen the regex to include Hiragana (\u3040-\u309f), Katakana (\u30a0-\u30ff), and Hangul (\uac00-\ud7af) ranges, or switch to Unicode script properties.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@CLI/cmux.swift` around lines 10950 - 10952, The tokenizer's regex (pattern)
only matches Han characters and Latin words, causing Japanese and Korean input
to produce no tokens and an empty frequency dictionary; update the
NSRegularExpression used in the tokenization (the pattern variable in
CLI/cmux.swift) to include Hiragana (\u3040-\u309f), Katakana (\u30a0-\u30ff)
and Hangul (\uac00-\ud7af) ranges for multi-character runs (e.g. use ranges like
[\u4e00-\u9fff\u3040-\u309f\u30a0-\u30ff\uac00-\ud7af]{2,}) or switch to Unicode
script properties (e.g. \p{Han}, \p{Hiragana}, \p{Katakana}, \p{Hangul}) so
tokens and the frequency dictionary are correctly populated instead of returning
nil.
Comment on lines
+2425
to
+2438
| moveUITestWindowToTargetDisplayIfNeeded() | ||
| activateUITestAppIfNeeded() | ||
|
|
||
| let hasWindow = !NSApp.windows.isEmpty | ||
| let hasVisibleWindow = NSApp.windows.contains { $0.isVisible } | ||
| let hasKeyWindow = NSApp.keyWindow != nil | ||
| let stage = attempt == 0 ? "afterForceWindow" : "afterForceWindow.retry\(attempt)" | ||
| writeUITestDiagnosticsIfNeeded(stage: stage) | ||
|
|
||
| guard attempt < 20 else { return } | ||
| if !hasWindow || !hasVisibleWindow || !hasKeyWindow || !NSRunningApplication.current.isActive { | ||
| DispatchQueue.main.asyncAfter(deadline: .now() + 0.25) { [weak self] in | ||
| self?.stabilizeUITestLaunchWindowAndForeground(attempt: attempt + 1) | ||
| } |
There was a problem hiding this comment.
Avoid overlapping retry loops during UI-test stabilization
This retry loop calls moveUITestWindowToTargetDisplayIfNeeded() on every attempt, but that helper also schedules its own retries. The nested retries can create timer storms and flaky timing/diagnostics under CI load.
💡 Proposed fix
- moveUITestWindowToTargetDisplayIfNeeded()
+ moveUITestWindowToTargetDisplayIfNeeded(attempt: attempt, allowRetry: false)
activateUITestAppIfNeeded()- private func moveUITestWindowToTargetDisplayIfNeeded(attempt: Int = 0) {
+ private func moveUITestWindowToTargetDisplayIfNeeded(
+ attempt: Int = 0,
+ allowRetry: Bool = true
+ ) {
let env = ProcessInfo.processInfo.environment
guard let rawDisplayID = env["CMUX_UI_TEST_TARGET_DISPLAY_ID"],
let targetDisplayID = UInt32(rawDisplayID) else {
return
}
guard let screen = NSScreen.screens.first(where: { $0.cmuxDisplayID == targetDisplayID }) else {
- if attempt < 20 {
+ if allowRetry, attempt < 20 {
DispatchQueue.main.asyncAfter(deadline: .now() + 0.25) { [weak self] in
- self?.moveUITestWindowToTargetDisplayIfNeeded(attempt: attempt + 1)
+ self?.moveUITestWindowToTargetDisplayIfNeeded(
+ attempt: attempt + 1,
+ allowRetry: allowRetry
+ )
}
}
self.writeUITestDiagnosticsIfNeeded(stage: "targetDisplayMissing")
return
}
guard let window = NSApp.windows.first else {
- if attempt < 20 {
+ if allowRetry, attempt < 20 {
DispatchQueue.main.asyncAfter(deadline: .now() + 0.25) { [weak self] in
- self?.moveUITestWindowToTargetDisplayIfNeeded(attempt: attempt + 1)
+ self?.moveUITestWindowToTargetDisplayIfNeeded(
+ attempt: attempt + 1,
+ allowRetry: allowRetry
+ )
}
}
self.writeUITestDiagnosticsIfNeeded(stage: "targetDisplayNoWindow")
return
}
// ...
- if window.screen?.cmuxDisplayID != targetDisplayID, attempt < 20 {
+ if window.screen?.cmuxDisplayID != targetDisplayID, allowRetry, attempt < 20 {
DispatchQueue.main.asyncAfter(deadline: .now() + 0.25) { [weak self] in
- self?.moveUITestWindowToTargetDisplayIfNeeded(attempt: attempt + 1)
+ self?.moveUITestWindowToTargetDisplayIfNeeded(
+ attempt: attempt + 1,
+ allowRetry: allowRetry
+ )
}
return
}🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@Sources/AppDelegate.swift` around lines 2425 - 2438, The
stabilizeUITestLaunchWindowAndForeground retry loop is repeatedly calling
moveUITestWindowToTargetDisplayIfNeeded() (which itself schedules retries),
causing overlapping timers; to fix, ensure
moveUITestWindowToTargetDisplayIfNeeded() (and optionally
activateUITestAppIfNeeded()) are invoked once before starting the retry loop in
stabilizeUITestLaunchWindowAndForeground (or add a guard inside
moveUITestWindowToTargetDisplayIfNeeded to no-op if a retry is already
scheduled), and keep writeUITestDiagnosticsIfNeeded(stage:) within the loop for
per-attempt logging so you avoid nested retry storms while preserving
stabilization retries.
Author
|
Superseded by #2043 which uses AI-powered summaries instead of keyword extraction. |
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
2 participants
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.
Summary
claude-hook stophandler with zero additional latency (transcript is already being read)/rename— if a custom title exists in the transcript, auto-naming is skippedMotivation
Currently, cmux tabs running Claude Code sessions show generic names (e.g., the shell command or directory). When working with multiple Claude Code sessions across tabs, it's hard to distinguish which tab is doing what without manually
/rename-ing each one.This change automatically extracts meaningful keywords from user messages in the Claude Code transcript and uses them as the tab title. For example, a conversation about "debugging cmux hooks" would auto-name the tab something like "cmux hooks debugging".
How it works
readTranscriptSummarynow also extracts the first 3 and last 3 user messages (in addition to the existinglastAssistantMessage)generateTabTitleuses regex-based tokenization to extract CJK runs and Latin words, filters stop words (English + Chinese), and picks the top keywords by frequency/renamed the session (detected viacustom-titleentries in the transcript JSONL), auto-naming is completely skippedDesign decisions
Test plan
/renamecustom title is preserved (auto-naming skipped)Summary by cubic
Automatically names Claude Code tabs from recent conversation keywords so you can tell sessions apart. Also stabilizes UI test launch and fixes portal geometry after window restore.
New Features
claude-hook stophandler with no added latency; supports CJK and Latin./renamecustom title is found in the transcript.Bug Fixes
Written for commit 275ee5c. Summary will update on new commits.
Summary by CodeRabbit
New Features
Tests